UC Irvine

UC Irvine Electronic Theses and Dissertations

Title
An Interface-Fitted Mesh Generator and Numerical Methods for Elliptic Interface Problems

Permalink
https://escholarship.org/uc/item/8106t9px

Author
Wen, Min

Publication Date
2018

Peer reviewed|Thesis/dissertation

eScholarship.org Powered by the California Digital Library

University of California
 UNIVERSITY OF CALIFORNIA,
IRVINE

An Interface-Fitted Mesh Generator and Numerical Methods for Elliptic Interface Problems

DISSERTATION

submitted in partial satisfaction of the requirements

for the degree of

DOCTOR OF PHILOSOPHY

in Mathematics

by

Min Wen

Dissertation Committee:
Professor Long Chen, Chair
Professor Qing Nie
Professor Hongkai Zhao

2018
Chapter 1,2,3 ©2017 Journal of Computational Physics
All other materials © 2018 Min Wen
 DEDICATION

To

my parents, my husband and my son

for your love, support and understanding

ii
 TABLE OF CONTENTS

Page

LIST OF FIGURES v

LIST OF TABLES vi

LIST OF ALGORITHMS vii

ACKNOWLEDGMENTS viii

CURRICULUM VITAE ix

ABSTRACT OF THE DISSERTATION xi

1 Introduction 1
1.1 Elliptic Interface Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Moving Boundary Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Mesh Generation 9
2.1 Interface-fitted Mesh Generator: Two Dimensions . . . . . . . . . . . . . . . . . . 9
2.1.1 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.3 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Interface-fitted Mesh Generator: Three Dimensions . . . . . . . . . . . . . . . . . 17
2.2.1 Main Difficulty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.2 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.3 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3 Finite Element Methods for Elliptic Interface Problems 27

3.1 Sobolev Spaces and Weak Formulation . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Finite Element Methods in 2D . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Virtual Element Methods in 3D . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.4 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4 Moving Interface Problems 48

4.1 Tissue Growth Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1.1 Tissue Growth Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1.2 Kinematic Boundary Condition . . . . . . . . . . . . . . . . . . . . . . . 51

iii
 4.1.3 Static Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.1.4 Moving Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2 Tumor Growth Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2.1 Tumor Growth Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2.2 Numerical Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2.3 Tumor Growth Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . 65

Bibliography 66

A Appendices 77
A.1 The Elliptic Interface Equation and Weak Formulation . . . . . . . . . . . . . . . 77
A.1.1 The Elliptic Interface Equation . . . . . . . . . . . . . . . . . . . . . . . . 77
A.1.2 The Weak Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
A.2 Efficient Implementation of Virtual Element Methods in Two Dimensions . . . . . 82
A.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
A.2.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
A.2.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

iv
 LIST OF FIGURES

Page

2.1 Example of interface elements: (1) − (5) with one level set function and (6) − (9)
with two level set functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Cartesian mesh Ω and a circle interface Γ. The grayed elements are interface ele-
ments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Three steps to generate an interface-fitted mesh. . . . . . . . . . . . . . . . . . . . 13
2.4 Interface points and interface-fitted meshes when the interface is unconnected. . . . 14
2.5 Add one auxiliary point when two intersection points are diagonal. . . . . . . . . 15
2.6 The angle ∠F EG at the intersection point E. . . . . . . . . . . . . . . . . . . . . 16
2.7 Sliver exists (left) and is removed when the element is divided into polyhedra
(right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.8 The surface of the interface is embedded in the hexahedron. . . . . . . . . . . . . . 19
2.9 An interface element is divided into two polyhedra. . . . . . . . . . . . . . . . . . 21
2.10 An cube is divided into three parts. . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.11 Different angle cases in the interface triangles. . . . . . . . . . . . . . . . . . . . . 25

3.1 An interface mesh with maximal angle 112.8104◦ . . . . . . . . . . . . . . . . . . . 38

3.2 Two balls are embedded in the unit cube. The maximal angle is 130.4665◦ . . . . . . 42
3.3 The interface is an orthocircle with maximal angle 132.4673◦ . . . . . . . . . . . . 43
3.4 Twelve balls are embedded in the cube. The maximal angle is 131.3925◦ . . . . . . 47
3.5 Part of the interface-fitted mesh with sharp edges. . . . . . . . . . . . . . . . . . . 47

4.1 The domain of tissue growth model. . . . . . . . . . . . . . . . . . . . . . . . . . 50

4.2 Sketch of the geometry for simple one-dimensional example. . . . . . . . . . . . . 52
4.3 Three steps to generate a semi-unstructured mesh. . . . . . . . . . . . . . . . . . . 55
4.4 All elements are triangles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.5 The mesh with periodic boundary condition . . . . . . . . . . . . . . . . . . . . . 56
4.6 The interface-fitted mesh in the domain with different time when ∆x is 0.05. . . . 61
4.7 The interface(red) and the domain Ω− (green) when the Cartesian mesh size h is
0.125. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

v
 LIST OF TABLES

Page

2.1 The face array (left) and face2elem (right) for two polyhedra in Fig. 2.9. . . . . 23

3.1 Errors for Example 3.1: β − = 1 and β + = 10. . . . . . . . . . . . . . . . . . . . . 39

3.2 Errors for Example 3.1: β − = 1 and β + = 100. . . . . . . . . . . . . . . . . . . . 39
3.3 CPU time (in seconds) for Example 3.1: β − = 1 and β + = 10. . . . . . . . . . . . 39
3.4 Example 3.1: Iteration steps of AMG with fixed β − = 1 and various β + . . . . . . . 39
3.5 Errors for Example 3.2: β − = 1 and β + = 1. . . . . . . . . . . . . . . . . . . . . 40
3.6 Errors for Example 3.2: β − = 1 and β + = 100. . . . . . . . . . . . . . . . . . . . 41
3.7 CPU time (in seconds) for Example 3.2: β − = 1 and β + = 1. . . . . . . . . . . . . 41
3.8 Example 3.2: Iteration steps of AMG with fixed β − = 1 and various β + . . . . . . . 41
3.9 Errors for Example 3.3: β − = 1 and β + = 1. . . . . . . . . . . . . . . . . . . . . 44
3.10 Errors for Example 3.3: β − = 1 and β + = 100. . . . . . . . . . . . . . . . . . . . 44
3.11 CPU time (in seconds) for Example 3.3: β − = 1 and β + = 1. . . . . . . . . . . . . 44
3.12 Example 3.3: Iteration steps of AMG with fixed β − = 1 and various β + . . . . . . . 44
3.13 Errors for Example 3.4: β − = 1 and β + = 1. . . . . . . . . . . . . . . . . . . . . 46
3.14 Errors for Example 3.4: β − = 1 and β + = 10. . . . . . . . . . . . . . . . . . . . . 46
3.15 CPU time (in seconds) for Example 3.4: β − = 1 and β + = 1. . . . . . . . . . . . . 46
3.16 Example 3.4: Iteration steps of AMG with fixed β − = 1 and various β + . . . . . . . 46

4.1 Errors for Example 1: linear finite element methods when B = 0.1. . . . . . . . . 57
4.2 Errors for Example 1: quadratic finite element methods when B = 0.1. . . . . . . . 57
4.3 Errors for Example 2: linear finite element methods when B = 0.15. . . . . . . . . 58
4.4 Errors for Example 2: quadratic finite element methods when B = 0.15. . . . . . . 58
4.5 Errors for Example 3: linear finite element methods when B = 0.25. . . . . . . . . 58
4.6 Errors for Example 3: quadratic finite element methods when B = 0.25. . . . . . . 59
4.7 Errors for Example 4: linear finite element methods when B = 0.5. . . . . . . . . 59
4.8 Errors for Example 4: quadratic finite element methods when B = 0.5. . . . . . . . 60
4.9 Errors for straight line:Errors for P and h. . . . . . . . . . . . . . . . . . . . . . . 62
4.10 Errors for straight line h = e4t : Order of accuracy with maximum norm for P and h. 62

vi
 LIST OF ALGORITHMS

Page
1 2D Interface-fitted Mesh Generation Algorithm . . . . . . . . . . . . . . . . . . . 12
2 3D Interface-fitted Mesh Generation Algorithm . . . . . . . . . . . . . . . . . . . 18

vii
 ACKNOWLEDGMENTS

I would like to express my deepest appreciation and thanks to my advisor Professor Long Chen for
all insightful guidance and continuous supports during my research. I am so lucky to be one of his
graduate students since he could always offer countless advice every time whenever I got lost in the
research. His scientific attitude, patience and enthusiasm as a advisor have made a profound impact
on me. My dissertation would not have been possible without his constant help and guidance in
me.

I am grateful to my collaborator Prof. Huayi Wei and he provided me so much help on under-
standing Virtual Element Methods. Thanks for all enlightening discussions and the efforts made to
finish our projects together.

I am also grateful to my dissertation committee members Professor Qing Nie and Professor Hongkai
Zhao. I appreciate all the help they have provided for me during the past years. I have learned a lot
from them after taking their classes. What’s more, my deepest appreciation goes to all the current
and past members in the team for all their help.

I am so appreciated to Department of Mathematics at University of California, Irvine for its gen-

erous offer of scholarship in the past five years. My dissertation would not exist without its full
support.

Last but not least, I am greatly indebted to my family for their selfless support and encouragement,
which make my graduate life colorful.

Finally, I would like to thank everyone else who supports me spiritually throughout my graduate
life, as well as express my apology that I could not mention personally one by one.

viii
 CURRICULUM VITAE

Min Wen

EDUCATION
Doctor of Philosophy in Mathematics 2018
University of California, Irvine Irvine, California
Master of Science in Mathematics 2011
Wuhan University Wuhan, China
Bachelor of Science in Mathematics 2008
Yangtze University Jingzhou, China

RESEARCH EXPERIENCE
Graduate Research Assistant 2013–2018
University of California, Irvine Irvine, California

TEACHING EXPERIENCE
Teaching Assistant 2016–2017
University California, Irvine Irvine, California

ix
REFEREED JOURNAL PUBLICATIONS

An Interface-Fitted Mesh Generator and Virtual Element Meth- 2017

ods for Elliptic Interface Problems
Journal of Computational Physics

x
 ABSTRACT OF THE DISSERTATION
An Interface-Fitted Mesh Generator and Numerical Methods for Elliptic Interface Problems

By

Min Wen

Doctor of Philosophy in Mathematics

University of California, Irvine, 2018

Professor Long Chen, Chair

In this thesis, we propose different numerical methods for solving elliptic interface problems in
three dimensions and moving boundary problems in two dimensions. In the first part, a simple and
efficient interface-fitted mesh algorithm which can produce a semi-unstructured interface-fitted
mesh in two and three dimensions quickly is developed in this thesis. Elements in such interface-
fitted meshes are not restricted to simplices but can be polygons or polyhedra. Especially in 3D, the
polyhedra instead of tetrahedra can avoid slivers, which are the major difficulty in finite element
methods. Virtual element methods are applied to solve elliptic interface problems with solutions
and flux jump conditions. Algebraic multigrid solvers are used to solve the resulting linear al-
gebraic system. In the second part, we impose the interface-fitted meshes to moving boundary
problems and present the applications in biology simulation. Static interface problems with peri-
odic conditions are considered at first, then the higher order accuracy algorithm is also developed.
Finally moving interface problems are discussed. We couple the interface Poisson equation and
the transport equation for the level set function together to simulate the tissue and tumor growth
phenomenon. Numerical results are presented to illustrate the effectiveness of our methods.

xi
Chapter 1

Introduction

We consider finite element methods for solving elliptic interface problems which have a variety of
applications in different research fields, including fluid dynamics, material science and biological
systems, etc. [23, 63, 73, 83]. The importance of the coupling of the complex geometry of the
interface with the numerical methods has been recognized and received rapidly increasing interest
in recent years. In this thesis, we discuss the elliptic interface problems at first, and then consider
moving boundary problems arising from applications in biology.

1.1 Elliptic Interface Problems

Let Ω be an open and bounded domain in Rd (d = 2, 3), and Γ be a continuous interface embedded
in Ω. The interface Γ separates the domain Ω into disjoint regions Ω+ and Ω− , where Ω+ denotes
the exterior domain and Ω− is the interior domain enclosed by Γ. We consider numerical methods
for solving the following elliptic interface problems:

−∇ · (β(x)∇u(x)) = f (x), x ∈ Ω\Γ (1.1)

1
with prescribed jump conditions across the interface Γ:

[u]Γ = u+ − u− = q0 , (1.2)
− −
[βun ]Γ = β + u+
n − β un = q1 , (1.3)

and boundary condition:

u=g on ∂Ω. (1.4)

Here un denotes the normal derivative (∇u)·n with n being the unit norm direction of the interface
Γ pointing outward (from Ω− to Ω+ ). The superscripts + and − stand for the restriction of a
function on Ω+ and Ω− , respectively. The diffusion coefficient β(x) is assumed to be uniformly
positive and smooth on each subdomain, but may be discontinuous across the interface. Because
of that, the solution u is piecewise smooth but the global regularity is low [33, 65, 66].

Numerical methods for elliptic interface problems can be roughly classified into two categories by
using either an interface-fitted (also known as body-fitted or interface conforming) mesh or an un-
fitted mesh (e.g. a uniform Cartesian mesh) in the discretization of the domain. In the unfitted mesh
approach, a popular way to enforce the jump conditions is to modify the finite difference stencils or
the finite element basis near the interface. A lot of numerical methods in this direction have been
proposed such as the immersed boundary method [108], the immersed interface method [78, 79],
immersed finite element methods [49, 57, 72, 84], ghost fluid methods [87], matched interface and
boundary (MIB) methods [134, 140, 147], multiscale finite element methods [33], extended finite
element methods (XFEM) [44, 93, 95], and many others [51, 62, 61, 67, 86, 133]. The jump con-
dition can be also imposed based on the Nitsche’s method [99] by introducing penalty terms across
interfaces, see, for example, the earlier work by Babuška [5], Barrett and Elliott [8], unfitted FEM
by Hansbo and Hansbo [53], hp-discontinuous Galerkin method [92], CutFEM [19, 55], and many
others [9, 19, 20, 21, 53, 54, 55, 71, 127, 128]. The most attractive feature of the unfitted mesh

2
approach is the easiness of the mesh generation. Indeed, if the background mesh is Cartesian, there
is no need of meshing which is very convenient, especially when the interface is moving in time.

On the other hand, using unfitted mesh approach, it is difficult to capture the complex geometry of
the interface and to enforce jump conditions across the interface accurately, and the resulting linear
system may not be always symmetric which could cause problems for fast solvers. Furthermore
a rigorous error analysis is difficult. Recent progress on immersed finite element methods can be
found in [52, 142].

In this work, we focus on the interface-fitted mesh approach. Provided a mesh fitted to the interface,
one can use conforming finite element methods and get a symmetric system which can be solved
efficiently by fast solvers such as algebraic multi-grid. Rigorous error analysis is possible. Optimal
a priori estimates of linear finite element is given in [137, 16, 29] and in [80] for high order finite
elements. Recent work using hybridized discontinuous Galerkin (HDG) [68] and weak Galerkin
(WG) [98] method is also based on a shape regular and body-fitted triangulation. The challenge of
this approach is quickly generating an interface-fitted mesh, especially in three dimensions (3D),
which is the topic of this study.

There is a lot of work on the unstructured interface-fitted mesh generation [88, 105, 143]. The
unstructured mesh generator is, however, time consuming as it needs to modify the mesh for the
whole domain, not just near the interface. For example, extensive and non-trivial computational
effort are needed to generate a high quality 3D finite element mesh from biomedical image data or
geological image data etc [4, 35].

We are interested in the semi-structured and body-fitted mesh generation methods [12, 14, 110]
and will develop a simple and effective mesh generation algorithm. As an illustrative example,
to generate an interface-fitted mesh in two dimensions (2D), we start from a uniform Cartesian
mesh with N -grid points, and apply three steps: 1) find all the intersection points, the grid points
near the interface, and add few auxiliary points; 2) generate a Delaunay triangulation of these

3
points; 3) remove the unnecessary triangles and merge the regular meshes away from the interface.
The resulting triangulations can preserve the interface and the maximal angle is bounded by 135◦ .
Since the Delaunay triangulations are only on a local region near the interface, the dominant cost is
reduced to O(N 1/2 log N ). Due to the semi-unstructured mesh and localization near the interface,
some nice properties of structured grids are still preserved such as superconvergence in the energy
norm and fast convergence of algebraic multigrid methods [129].

The main restriction of this approach is the quality of the generated mesh especially in 3D. Most
finite element methods require discretizing a domain into a set of shape regular tetrahedra in three
dimensions. The accuracy of the simulations and the efficiency of the solvers could deteriorate by
the presence of badly-shaped elements. The problematic tetrahedra are so-called slivers, which are
a type of flat tetrahedra without small edges, but with nearly zero volume. Namely, four vertices
of a sliver are almost coplanar. Due to the presence of slivers, three-dimensional mesh generation
is much harder than the two-dimensional case, and removing slivers from a 3D tetrahedral mesh is
one of the major tasks in the field of mesh generation [39, 81, 97].

We propose a new way to solve this difficulty. We choose polyhedral meshes rather than tetrahedral
meshes. Then silvers will be merged into a polyhedron. The shape of the polyhedron or other
tetrahedron could be still degenerate but the maximal angle is bounded uniformly away from π.
Notice that finite element approximation retains accurate if the maximal angle condition [6] is
satisfied. Namely tetrahedra with small volumes are allowed as long as the four vertices are non-
planar [1, 38]. Similar results can be established for polyhedral meshes and theoretical justification
will be reported somewhere else.

Another difficulty is encountered in the implementation. Due to the large number of possible
intersections between the fixed mesh and the interface, a variety of interface-cells are generated
leading to an equally large number of treatments, which could result in complex coding logistics;
see [109, 110].

4
We propose an all-in-one solution. The connectedness of intersection points is obtained by the
Delaunay algorithm which is a well developed algorithm in computational geometry and efficient
implementation is available in many software packages. Our mesh generation algorithm in 3D
is similar to the 2D case only different in the step 3: post-processing. The additional work is to
merge tetrahedra into polyhedra. To facilitate the merging, the polyhedra are stored in the form of
faces and the index of the elements to which the faces belong. The resulting mesh retains these
nice properties: the interface is approximately preserved, the maximal angle condition is satisfied,
and cost-efficient. The Delaunay algorithm is only called for points near the interface and thus
the dominated cost is reduced to O(N 2/3 ) which is considerably smaller comparing with O(N )
assembling and solving of the linear algebraic system. The quality and efficiency of our mesh
generation algorithm are balanced and suitable for the finite element simulation. No additional
mesh smoothing process is needed in our algorithm. Of course, adding such a mesh smoothing
process will furthermore improve the quality of the mesh and probably improve the accuracy of
the finite element approximation. However, it will destroy the structure of the grid. In our mesh
generator, the background mesh is fixed. The Delaunay algorithm can be called element by element
and thus local modification is possible if only part of the interface is changed. These features are
important for moving interface problems.

A similar approach was introduced in [50], where the authors introduced the Voronoi diagrams and
Delaunay triangulation of a point set of a surface and more focused on the surface mesh generation.
Our algorithm seems simpler and more suitable for finite element simulation as we shall discuss
below.

Since elements in such interface-fitted meshes are general polyhedra, we shall apply virtual el-
ement methods (VEM) [10, 11], which can be considered as an extension of conforming finite
element methods to polyhedral meshes. The resulting linear algebraic system is symmetric and
positive definite and thus can be solved efficiently using algebraic multigrid solvers. Furthermore,
according to our mesh generation algorithm, we will get the polyhedra with triangular and square

5
faces which will be much easier when assembling the matrices in VEM compared to the original
approach in [11]. Optimal second order of convergence in the L2 and L∞ norms and a super-
convergence of energy norm is observed in several numerical examples. Details are presented in
Chapter 2 and 3.

1.2 Moving Boundary Problems

Many applications in fluids, materials and biology involve multi-connected domains. In multi-
connected problems the boundary between different subdomains can be dealt with moving in-
terfaces such as hot metal forming, mould filling and open channel flows [124, 56, 60, 90]. In
numerical simulations, the major difficulties are the position of the moving interface is not known
a priori and strongly coupled to the solutions of the flow equations.

Numerical methods for moving interface problems based on the representation of the interface
have been developed and they can be roughly classified into two categories. The first category is
interface tracking, in which the interface is explicitly represented. The front tracking methods [132,
48, 124] and the marker and cell ( MAC ) methods [130, 56] belong to this class. Main advantages
of the tracking methods are efficient and accurate representation of the interface and simplicity in
tracking the motion of particles. However, it is difficult for interfaces with complicated geometry
and topological change and particular in three dimensions.

The second category is interface capturing, in which the interface is implicitly embedded in a
scalar field function defined on a fixed mesh, such as a Cartesian grid. Two widely used interface
capturing methods are the volume of fluid method (VOF) [60] and the level set method [101]. Main
advantages of capturing methods are complex interface structures and topological changes which
can be captured quite naturally in two and three dimensions; see, e.g., [22, 101, 120]. However,
there are also a few disadvantages for these Eulerian approaches. For examples, due to the implicit

6
representation, capturing methods are usually less accurate and less efficient than tracking method
in terms of both interface representation, which lead to loss of high order accuracy at the moving
front.

The level set method is a popular method in multi-connected domain problems as it is good for
computing curvature driven flow problems [120, 119, 125, 123, 91, 90, 126, 32]. There are different
approaches to solve the level set equation numerically. For instance, finite difference methods such
as forward difference, essentially nonoscillatory (ENO) and weighted essentially non-oscillatory
(WENO) schemes are often used [141, 120, 119, 100, 125]. Also discretization of the level set
equation by means of finite volume schemes [74, 45, 46] or finite element methods [123, 111, 116]
can be found in the literature.

A lot of work has been done for solving partial differential equations with moving interfaces using
finite difference methods. For example, immersed boundary methods(IBM) [94], immersed inter-
face method ( IIM ) [138, 64, 34, 82], a level set/ghost fluid method [89], an embedded boundary
formulation [139], and other methods [144, 117].

In Section 4.1, it introduces the tissue growth problems. We start with static interface problems
with periodic boundary conditions in one dimension using finite element methods. We introduce
another artificial mesh which satisfies the periodic conditions to assemble the matrix. It could
avoid for loop in the code to speed up the performance. Then we use Discontinuous Galerkin
Method Finite Element Method (DG-FEM) [59] to solve the moving interface problems. While
in [103], the approach is transforming the deformed geometry and governing equations to the unit
square and then using finite difference methods. We focuses on the spatial discretization of the
level set equation by DG-FEM combined with a 4th order Runge Kutta time stepping scheme. In
each time step the whole coupled system is solved iteratively. High order accuracy algorithm is
also presented.

In Section 4.2, another model tumor growth simulation is presented. Here we focus on solving the

7
transport equation using 5th order WENO and 3rd order Runge Kutta Methods [115] to simulate
tumor growth problems.. We could use a set of fifteen nearby points to fit a circle to compute the
curvature. And the sign of the curvature depends on the position of the points in the fitted circle. If
one point is still in the tumor interface, the curvature is positive. Otherwise, it is negative. For each
time step, we use harmonic extension for each component of the velocity, which is different from
the EPC methods in [143]. Reinitialization should be taken into consideration if we use signed
distance to compute curvature [143], since the level set function will become very flat or steep near
the zero level set after several time steps. Reinitialization [120, 104] could be proposed to reshape
it to be an signed distance function, which could result in accurate numerical solutions.

8
Chapter 2

Mesh Generation

We present the interface-fitted mesh generation algorithm in two and three dimensions. Both of
them only consist of three steps. Our method is simple and efficient since we just deal with the
points near the interface. We also show some properties of the mesh obtained by our algorithm.
The interface will be approximately recovered in the triangulation generated by the algorithm and
the maximum angle of the triangulation is bounded.

2.1 Interface-fitted Mesh Generator: Two Dimensions

In this section, we introduce our interface-fitted mesh generator in 2D. We first describe the algo-
rithm and then give two examples to illustrate the algorithm. In addition, we prove the generated
mesh will preserve the interface and satisfy the maximal angle condition.

9
2.1.1 Algorithm

Let Γ be an interface embedded in a rectangular domain Ω. Assume Γ can be represented by the

zero-level set of a function φ(x), i.e., Γ = {x ∈ Ω : φ(x) = 0}. The interface Γ separates Ω into
subdomains Ω+ := {x ∈ Ω : φ(x) > 0} and Ω− := {x ∈ Ω : φ(x) < 0}. Note that Ω− could have
multiple connected components when Γ consists of two or more closed curves.

One can easily generate a uniform Cartesian mesh Ωh of Ω with a given mesh size h. A vertex
p of Ωh is said to be inside if φ(p) < 0, outside if φ(p) > 0, or on Γ if φ(p) = 0; an edge
(p1 , p2 ) is called a cut edge if φ(p1 )φ(p2 ) < 0; the point which the cut edge intersects with Γ is
called an intersection point; a square element K of Ωh which intersects with the interface Γ, i.e.
|K̄ ∩ Γ| =
6 ∅, is called an interface element. We can find interface elements by using one of the
following two rules:

1. There exists at least two vertices p and q with opposite sign, i.e., φ(p)φ(q) < 0;

2. There exists at least two vertices on the interface, namely the value of φ on these vertices is
0.

These two rules could detect all the interface elements in Fig 2.1 except case (3), which could
be avoided by choosing the initial mesh size h small enough. For disconnected interfaces (cases
(6) − (9)), we assume it is described by two level set functions (c.f. Example 2.2), and the in-
tersection points can be found by treating each level set function one by one. See Fig 2.1 for
the illustration. We remark that it is much more difficult to modify stencils or the basis for such
cases. In general, the modified finite difference stencils or modified finite element basis near the
interface is to introduce additional but local degrees of freedom near the interface and then use the
jump conditions to eliminate these degree of freedom by solving a small linear system element-
wise [135, 2, 13, 131, 43, 70, 96]. In almost all of these work, it is assumed the intersection meets
the edges of an interface element at no more than two intersections and intersects at different edges

10
for one element, c.f. [142, 57, 86, 85]. If this condition is violated, such as those cases (6) − (9)
in Fig 2.1, the local system will be much more involved since it depends how the interface cuts the
elements.

We define the interface points as the collection of intersection points, vertices of interface elements,
and some auxiliary points explained below. When the intersection points are diagonal, we need to
add the midpoints of corresponding elements, which are called auxiliary points.

!!!!!!!(1)! !!!!!!!(2)! !!!!!!!!!(3)!

!!!!!!!(4)! !!!!!!!(5)! !!!!!!(6)!

!!!!!!!(7)! !!!!!!!(8)! !!!!!!(9)!

Figure 2.1: Example of interface elements: (1) − (5) with one level set function and (6) − (9) with
two level set functions

Recall that a Delaunay triangulation for a set points P in a plane is a triangulation of the convex set
of P such that no point in P is inside the circumcircle of any triangle in this triangulation [40, 77].

Our 2D interface-fitted mesh generation algorithm is described as follows:

11
Algorithm 1 2D Interface-fitted Mesh Generation Algorithm
Input: Grid size, h, level set function, φ(x) and square domain, Ω;
Output: An interface-fitted mesh of Ω;
1. Find all the interface points.
2. Construct a Delaunay triangulation of these points.
3. Remove triangles not in the interface elements and merge all uncut elements.

2.1.2 Examples

We give two examples to explain Algorithm 1 in detail. The first example shows the simple case
when the interface is a circle. The second example illustrates a more complex case when the
interface is unconnected and some interface elements are divided into three parts.

E XAMPLE 2.1 (A circle). Consider the domain Ω = (−1, 1)2 and a circle interface Γ represented
by the level set function φ(x, y) = x2 + y 2 − r2 , with r = 0.5. The interface elements are shown
in Fig 2.2.

Figure 2.2: Cartesian mesh Ω and a circle interface Γ. The grayed elements are interface elements.

First, we construct a point set P which includes the intersection points between cut edges and Γ,
the vertices of all interface elements, and some auxiliary points. See Fig. 2.3(a) for the illustration.
Here we use the bisection method to compute the intersection points within the machine precision
tolerance.

12
Then we construct a Delaunay triangulation based on the point set P. In MATLAB, we just call
DT = delaunay(x,y) (see Fig. 2.3(b)).

In the last step, we keep the triangles in interface elements and merge the uncut elements to get the
final interface-fitted semi-unstructured mesh in Fig. 2.3(c)-(d).

(a) Step 1: Find all the interface points. (b) Step 2: a Delaunay triangulation of inter-
face points.

(c) Step 3: Remove triangles not in the inter- (d) Step 3: Merge all uncut elements to get an
face elements. interface-fitted mesh.

Figure 2.3: Three steps to generate an interface-fitted mesh.

E XAMPLE 2.2 (Two circles). Consider the domain Ω = (−1, 1)2 and the unconnected interface Γ

13
represented by the level set function

φ(x, y) = min (x + r)2 + y 2 − (1.1r)2 , (x − r)2 + y 2 − (0.8r)2 ,



with r = 0.4. We can apply the same algorithm and obtain the mesh in Fig 2.4. The only difference
is when computing intersection points, we compute them for each level set function separately. We
use this example to show our algorithm can handle unconnected interfaces.

(a) Interface points. (b) The interface-fitted mesh when the inter-
face is two disjoint circles.

Figure 2.4: Interface points and interface-fitted meshes when the interface is unconnected.

2.1.3 Properties

We explore properties of the mesh obtained in Algorithm 1. A triangle is called an interior element
when the center point of the triangle is inside. The interface Γ could be approximated by the
boundary of those interior elements and can be extracted easily. The obtained discrete interface is
denoted by Γh .

Proposition 2.1. The interface will be approximately recovered in the triangulation generated by
Algorithm 1. More precisely, we have dist(Γh , Γ) . h2 provided Γ is smooth enough and h is

14
small enough.

Proof. We shall use another characterization of Delaunay triangulations: a Delaunay triangulation

is the projection of the lower convex hull of points lifted to the paraboloid f (~ xk2 [18, 28,
x) = k~
42].

The function values of f (~

x) on the four vertices of a square will be on a plane. As the function f is
strictly convex, the function value of any intersection points which are different from the vertices
of the square will be below this plane. Then the lower convex hull when lifted to R3 will always
connect the intersection points. Thus, the interface will be recovered under this circumstance.

If there are two diagonal vertices of a square on the interface element in P (see Fig. 2.5 (c)), then
the Delaunay triangulation on this square is not unique. Using either diagonal of the square is a
valid Delaunay triangulation (see Fig. 2.5(a) and (b)). Therefore, we introduce the center of this
square as an auxiliary point to make sure the interface is preserved (see Fig. 2.5 (d)).

In both cases, Γh contains a piecewise affine approximation of Γ with nodes on the interface and
thus the distance is in the order of Ch2 with constant C depends on the curvature of Γ.

(a)$ (b)$ (c)$ (d)$

Figure 2.5: Add one auxiliary point when two intersection points are diagonal.

Proposition 2.2. Assume the mesh size h is small enough such that the interior of each edge has at
most one intersection point. Then the maximal angle of the triangulation generated by Algorithm 1
is bounded by 135◦ .

Proof. Let C be a square with vertices A, B, C, D which intersects with the interface, S the points
set including the vertices of C and the intersection points, and DT the Delaunay triangulation of

15
S.

The vertex of every angle in DT can be a vertex of the square or an intersection point. The angle at
a square vertex must be bounded by 90◦ as the two rays of the angle is inside the square. Next, let
us prove that the angle at an intersection point must be bounded by 135◦ . Let E be an intersection
point on edge AB, F and G are the other two points of angle ∠F EG and G is on the right of F
(see Fig. 2.6). Here F or G can be an intersection point or a vertex of the square.

By our assumption, F or G cannot be in the interior of edge AB and F and G cannot be in the
interior of edge CD simultaneously. So either F is on the edge AD or G is on BC. Without
loss of generality, we assume G is on BC. Then the angle ∠F CG ≥ 45◦ since F is on the
left of the diagonal AC. Note that the triangle 4F CG may not be in the DT . Nevertheless, if
∠F EG > 135◦ , then ∠F EG + ∠F CG > 180◦ which means the circumcircle of ∆F EG must
include vertices C violating the Delaunay property. So ∠F EG must be bounded by 135◦ .

D C D F C

F
G G
E E
A B A B
Figure 2.6: The angle ∠F EG at the intersection point E.

Let N be the number of nodes. Since we restrict the Delaunay triangulation on a local region near
the interface, the complexity of generating a Delaunay triangulation will be O(N 1/2 log N ) in 2D
which can be ignored compared with the O(N ) complexity of assembling the matrix and solving
the matrix equation. Such localization will make it possible to track the moving interface, which
will be discussed in Chapter 4.

The overall complexity of our mesh generation algorithm is: c1 N + c2 N 1/2 log N since we need to
compute the sign of the level set function at N vertices. In practice, however, the constant c1  c2

16
and the time scales like O(N 1/2 ).

2.2 Interface-fitted Mesh Generator: Three Dimensions

In this section, we present a novel mesh generation algorithm to generate an interface-fitted mesh
for a given smooth interface in three dimensions. We begin with a brief review on the difficulty of
3D mesh generation and then introduce our algorithm to overcome this difficulty.

2.2.1 Main Difficulty

Tetrahedral meshes are frequently used in classical finite element methods. The size and shape of
the tetrahedra influence the accuracy of finite element solutions [118]. The quality of the tetrahe-
dron’s shape can be measured by using the aspect ratio or the radius-edge ratio. The aspect ratio
of a tetrahedron is usually defined as its circumradius divided by its inradius and the radius-edge
ratio is the circumradius divided by the shortest edge length of the tetrahedron. The aspect ratio or
radius-edge ratio of a mesh is the largest corresponding ratio of all of its tetrahedral elements. If
the aspect ratio or radius-edge ratio of a mesh are small, we called the mesh well-shaped [40, 81].
Ideally we expect each element in the mesh is shape regular. But violation is allowed as long as
the so-called maximal angle condition is satisfied [1, 6, 17, 38, 75].

The difficulty of mesh generation in three dimensions is due to the existence of slivers. Slivers
have small radius-edge ratio, but large aspect ratio, which are considered as bad-shaped tetrahedral
elements. The results of the accuracy and convergence of finite element methods may not hold
anymore in the existence of slivers which violates the maximal angle condition. A lot of methods
have been developed to remove slivers; see e.g. [30, 31, 41]. Sliver removal methods, however,
involve the addition and rearrangement of points and thus destroy the semi-structure of the mesh.

17
We shall introduce polyhedral meshes near the interface to remove slivers. When we get the
interface elements (which is defined similarly to 2D and will be made clear later) and intersection
points (the definition is the same as in 2D), we can divide interface elements into polyhedra instead
of tetrahedra. Slivers will be eliminated and part of their faces will become the faces of polyhedral
elements. For example, when the interface cuts across one element with four almost coplanar
intersection points, if we divide the element into tetrahedra, then the four intersection points could
form a sliver (see Fig. 2.7). If we use a polyhedral mesh, however, the two well-shaped triangles
will become the boundary of two polyhedra.

Figure 2.7: Sliver exists (left) and is removed when the element is divided into polyhedra (right).

2.2.2 Algorithm

We write down the algorithm and then explain the details step by step.
Algorithm 2 3D Interface-fitted Mesh Generation Algorithm
Input: Grid size h, level set function φ(x), and a cubic domain Ω;
Output: Interface-fitted mesh Ω;
1. Find all the interface points.
2. Construct the Delaunay mesh DT on these points.
3. Post processing: remove unnecessary tetrahedra in DT , merge tetrahedra into polyhedra, and
merge with uncut elements.

Given a cubic domain Ω which includes the interface Γ described as the zero level set of φ, and
a mesh size h, we first generate the uniform Cartesian mesh of Ω with size h. The cubes in the

18
Cartesian mesh in domain Ω could be classified into exterior, interior and interface elements by
checking the sign of the centers of the cubes. We label them by 1, −1 and 0 respectively.

We define interface elements as elements satisfying one of the following rules:

1. There exists at least two vertices p and q with opposite sign, namely φ(p)φ(q) < 0;

2. There exists at least three vertices on the interface.

All interface elements will form a hexahedral mesh of a layer of the interface (see Fig. 2.8). All
boundary faces of this hexahedral mesh are extracted and will be used as faces of the polyhedral
mesh for the interface. Note that these boundary faces are square faces.

Figure 2.8: The surface of the interface is embedded in the hexahedron.

In step 1, similarly to two dimensions, we find cut edges and intersection points, and add auxiliary
points if necessary. The criterion of adding auxiliary points is the same as 2D: if a square face
contains two opposite vertices on the interface, we will add the center point of this square face as
the auxiliary point.

19

1
In step 2, we generate a Delaunay mesh DT of P, the set of interface points, whose definition is
the same as that in two dimensions.

In step 3, we post-process DT to get a polyhedral mesh near the interface and merge all uncut
cubic elements away from the interface.

Similar to the 2D case, we only keep tetrahedra inside the interface elements, which might contain
slivers, and remove tetrahedra not in the interface elements which can be easily marked by checking
the center of tetrahedra in DT .

Now we have a tetrahedral mesh of all the interface elements, and we still call this tetrahedral
mesh as DT for convenience. We could split the tetrahedron in DT into two categories: exterior
tetrahedral set DTE and interior tetrahedral set DTI . For a tetrahedron in DT , if the minimum
of the sign function of the φ value of the four vertex nodes is −1, we put it into DTI , otherwise,
we add it into DTE . The interface Γ could be extracted using the boundary faces of DTI and the
normal direction of the extracted surface mesh points outside of the interface. Tetrahedra in each
category will be merged into polyhedra element by element.

Instead of storing all vertices of a polyhedron, we shall store the polyhedral mesh by the data
structure face and face2elem. The array face records indices of three (triangular face) or four
(square face) vertices of all faces. The direction of all faces follows the right-hand side rule, that
is, the normal direction of each face is outwards. The array face2elem records the index of the
polyhedron to which the faces belong.

Fig. 2.9 is a simple example. Given a unit cube with three intersection points, it is divided into two
polyhedra. Each polyhedron is stored by faces and elements to which they belong. The values of
face and face2elem in Fig. 2.9 are shown in Table 2.1.

Notice that some face, e.g., [3 7 8 4] is stored as a square instead of two triangles since this face
is shared by another cube which is not included as an interface element. Those square faces are

20
 8 8
6 6 4
4 9 9
10 9 10 10
2
2 11 7
7
5 11 3 5
11 3
1 1
Figure 2.9: An interface element is divided into two polyhedra.

boundary faces of the hexahedral mesh which consists of all interface elements. All such square
faces have been extracted when we collect all interface elements.

Every interface element is divided into several polyhedra according to our method. For each poly-
hedron, we need to assign a unique index. This index map from face to element, face2elem, can
be generated in two stages. In most cases, the interface element is just divided into two polyhedra.
In the first stage, for the interior part, we use the original interface element index j and for the
exterior part, we append a new index j + N , where N is the number of elements in the initial
Cartesian grid. In some cases, however, one cube could be divided into three or more polyhedra
(see the three cases in Fig. 2.10). In the second stage, we use Euler’s formula to check the connect-
edness of the obtained polyhedral mesh. If a disconnected polyhedron is found, we group faces
into different connected components which is equivalent to dividing the original polyhedron into
more polyhedra. Thanks to our data structure, we only need to change face2elem when adding
and storing the new polyhedra.

In a nutshell, we could get a polyhedral mesh near the interface by storing the triangular and
square faces. The final interface-fitted mesh consists of polyhedra near the interface and uncut
(cube) elements away from the interface.

21
 13 12 15 14 14
6 6 4 4
4
10 9 11 11 12 13 13
11
9 12
14 5 9 5
10
10

8
138 12 158 14 14
6
10 9 11 11 12 13 13
2 2 2
11
7 7 9 7 12
14 9
5 3 10 3
10 3
1 1 1

Figure 2.10: An cube is divided into three parts.

22
 1 2 9 10 1 1
2 2 11 9 2 1
3 2 10 11 3 1
4 11 10 9 4 1
5 11 1 3 5 2
6 5 1 11 6 2
7 3 4 9 7 2
8 3 9 11 8 2
9 9 4 8 9 2
10 5 10 6 10 2
11 5 11 10 11 2
12 10 8 6 12 2
13 10 9 8 13 2
14 11 9 10 14 2
15 1 5 7 3 15 2
16 5 6 8 7 16 2
17 3 7 8 4 17 2
1 2 3 4 1

Table 2.1: The face array (left) and face2elem (right) for two polyhedra in Fig. 2.9.

2.2.3 Properties

The generated Delaunay triangulation will recover the interface by the lifting method. Namely
Proposition 2.1 also holds for the 3D case since the characterization of a Delaunay triangulation
as the projection of the lower convex hull holds in general dimensions. We formally summarize
below.

Proposition 2.3. The interface will be approximately recovered in the triangulation generated by
Algorithm 2. More precisely, we have dist(Γh , Γ) . h2 provided Γ is smooth enough and h is
small enough.

Next we shall show the maximum angle of the surface mesh is uniformly bounded by 144◦ .
In [109], the author considers 12 types of subdivision of boundary cells (not necessarily satisfying
the Delaunay property) in three dimensions and shows the same bound.

Proposition 2.4. The maximal angle of the triangular faces of the polyhedral mesh is bounded by

23
144◦ .

Proof. For simplicity, let C be a unit cube which intersects with the interface, and S the set of
points including the eight cube vertices and the intersection points. Let DT be the 3D Delaunay
triangulation on S.

For a 3D Delaunay triangulation, it satisfies the Delaunay empty sphere property such that no
point in S is inside the circumsphere of any tetrahedron of DT . Given a tetrahedron T in DT
which has a triangular face τ on one (denoted as F) of the six square faces of C. Since, on the
plane spanned by F, the circumcircle of τ is also on the circumsphere of T , then by Delaunay
empty sphere property, there is no point of S on F which is inside the circumcircle of τ , namely,
the boundary triangulation of DT on F is also Delaunay, and thus the maximal angle of these
triangles is bounded by 135◦ by Proposition 2.2.

Next we only need to consider the interface triangles with three vertices on the interface. For these
interface triangles, their angles can be divided into 16 cases (see Fig. 2.11).

In case (1) to (15), one can find the upper bound of the angle by calculus analysis. Here we
take the case (1) as an example to show how to find the upper bound, see Fig. 2.11 (1). Let vLA
be the vector from point L to point A and |vLA | the length of vLA . Similarly, we have vectors
vLH , vLN , vN A , vN H , then

vLA · vLH (vLN + vN A ) · (vLN + vN H )

cos ∠ALH = =p p
|vLA ||vLH | |vLN |2 + |vN A |2 |vLN |2 + |vN H |2
|vLN |2
=p p ≥ 0.
|vLN |2 + |vN A |2 |vLN |2 + |vN H |2

When |vLN | = 0, cos ∠ALH reaches the minimum value zero, namely, the maximum of ∠ALH
is 90◦ . By the similar method, one can get the upper bounds for other cases except case (16) in
Fig. 2.11.

24
 (1) ALH ≤ 90 ◦ (2) ALK ≤ 135 ◦ (3) ALC ≤ 135 ◦ (4) ALF ≤ 135 ◦
C F
K
L L L L
A H A A A
N
(5) ALJ ≤ 135 ◦ (6) ALE ≤ 45 ◦ (7) ILK < 101. 5370 ◦ (8) ILC < 101. 5370 ◦
C
J
K
I I
L E L L L
A A

(9) ILJ < 63. 6780 ◦ (10) ILD < 101. 5370 ◦ (11) ILF < 63. 6780 ◦ (12) FLC ≤ 90 ◦
F F C
J
I I I
L L D L L

(13) JLC < 63. 6780 ◦ (14) ELC < 101. 5370 ◦ (15) ELF < 63. 6780 ◦ (16) ALG ≤ 144 ◦
C C F
J Q G

E E P L
L L L
A

Figure 2.11: Different angle cases in the interface triangles.

25
In case (16), provided ∆ALG is an interface triangle and ∠ALG is the angle bigger than 144◦ . By
Algorithm 2, there must exist a vertex of C, for example, vertex Q and ALGQ is a tetrahedron in
the Delaunay triangulation. Let (1 − h1 , 0, 0), (1, 0, h2 ) and (1, h3 , 1) be the coordinates of A, L
and H, respectively. Then one can get the circumcenter O and circumradius r of the circumsphere
of ALGQ, then construct function f (h1 , h2 , h3 ) := r − |P − O|. By the assumption ∠ALG >
144◦ and the 2D Delaunay empty circle property on the boundary face of C, one can show that
f (h1 , h2 , h3 ) > 0, namely P is inside of the circumsphere of tetrahedron ALGQ, which contradicts
with the Delaunay empty sphere property.

Remark 2.5. For the proof of the 3D angle case (16), we use the region plot function in SageMath
[36] to show f (h1 , h2 , h3 ) > 0 under the given assumptions.

Remark 2.6. We emphasize that the 16 cases plotted above are used to prove the maximal angle
condition. In the algorithm, we get the mesh by directly calling Delaunay algorithm with all
interface points as input.

Again, we restrict the Delaunay triangulations on a local region near the interface. The overall
complexity of our mesh generation algorithm is: c1 N + c2 N 2/3 log N since we need to compute
the sign of the level set function at N vertices with c1  c2 . The meshing time scales like O(N 2/3 ).
See Section 3.4 for numerical results.

In summary, our mesh generator is simple and fast. The generated mesh is semi-unstructured. The
interface is approximately recovered, and the maximum angle of the surface mesh is uniformly
bounded.

26
Chapter 3

Finite Element Methods for Elliptic

Interface Problems

We start with Sobolev spaces and the weak formulation of the elliptic interface problem (1.1)-
(1.4). We then introduce the linear virtual element methods and discuss the implementation detail.
Finally we present several numerical results within the 3D setting for elliptic interface problems
with jump conditions across the interface.

3.1 Sobolev Spaces and Weak Formulation

Let D denote a bounded and open set in Rd , d = 2, 3 and W m,p (D) be the usual Sobolev space
with standard norm k·km,p,D and semi-norm |·|m,p,D . In particular, for p = 2, we denote H m (D) =
W m,p (D) and the corresponding norm and semi-norm by k·km,D = k·km,p,D and |·|m,D = |·|m,p,D ,
respectively. The space H01 (D) = {v ∈ H 1 (D) : v|∂D = 0} is the subspace of H 1 (D) with
zero trace. Let (·, ·)D and h·, ·i∂D denote the standard L2 inner products of L2 (D) and L2 (∂D)
respectively.

27
Domains are considered as open sets. Define Ω̃ = Ω− ∪ Ω+ and notice that Ω = Ω− ∪ Γ ∪ Ω+ =
Ω̃ ∪ Γ. For v ∈ W m,p (Ω̃), that is, v|Ω− ∈ W m,p (Ω− ) and v|Ω+ ∈ W m,p (Ω+ ), v may not be in
W m,p (Ω) due to the jump across the interface Γ.

To derive the weak formulation of elliptic interface problems (1.1)-(1.4), we multiply (1.1) with a
test function v ∈ H01 (Ω) and apply integration by parts. To address the jump of function values,
we chose a w− ∈ H 1 (Ω− ) with w− = q0 on ∂Ω− . With a slight abuse of notation, the zero
extension of w− to H 1 (Ω̃) is still denoted by w− . The model (1.1)-(1.4) is equivalent to: find
p ∈ Hg1 (Ω) = {v ∈ H 1 (Ω) : v|∂Ω = g} such that

(β∇p, ∇v)Ω = (f, v)Ω − hq1 , viΓ + (β∇w− , ∇v)Ω− , ∀v ∈ H01 (Ω), (3.1)

and set u = p − w− . It is easy to show that u solves equations (1.1)-(1.4). Even though the choice
of w− is not unique, the solution u does not depend on the choice of w− by the maximal principle.
The flux jump [βun ]Γ = q1 is imposed in H −1/2 (Γ) and the jump of function value is imposed in
H 1/2 (Γ).

3.2 Finite Element Methods in 2D

In this section, we present the numerical analysis of the standard finite element methods on the
two-dimension interface-fitted mesh generated by the Algorithm 1.

For simplicity of exposition, we assume the function value jump condition [u]Γ = 0. Let Th be an
interface-fitted triangular mesh with maximal angles uniformly bounded away from π. For each
τ ∈ Th , let hτ denote its diameter and h = maxτ ∈Th hτ . The vertices on Γ forms a polygon Γh
−
approximation of Γ. The polygon also splits Ω into two subdomains, Ω+
h and Ωh , which are the
−
approximations of Ω+ and Ω− , respectively. Each triangle τ ∈ Th is in either Ω+
h or Ωh and has at

most two vertices on Γ.

28
Let Vh be the linear finite element space on Th . The linear finite element approximation of (3.1) is
as follows: find uh ∈ Vh ∩ H01 (Ω) such that:

(βh ∇uh , ∇vh )Ω = (f, vh )Ω − hq̄1 , viΓh , ∀vh ∈ Vh ∩ H01 (Ω), (3.2)

where q̄1 = q1 (P0 (x)) and P0 (x) is a well defined projection from Γh to Γ (c.f. [129]).

We can get the nearly optimal L2 -norm and H 1 -norm estimates as the results in [136, 29].

T HEOREM 3.1. Let u be the solution of (3.1) and uh be the linear finite element approximation in
(3.2) based on the two-dimension interface-fitted mesh generated by the Algorithm 1. We have

kβ 1/2 (∇u − ∇uh )k0,Ω . h| log h|1/2 (kf k0,Ω + kq1 k2,Γ ), (3.3)

ku − uh k0,Ω . h2 | log h|(kf k0,Ω + kq1 k2,Γ ). (3.4)

Proof. For finite element approximation, we have the Ceá lemma,

kβ 1/2 (∇u − ∇uh )k0,Ω ≤ kβ 1/2 (∇u − ∇uI )k0,Ω .

Then the energy error estimate is reduced to the interpolation error estimate. In [6], the authors
proved that the local interpolation error estimate k(∇u − ∇uI )k0,τ . hkuk2,τ provided the maxi-
mal angle condition is satisfied which has been verified for the interface-fitted mesh generated by
Algorithm 1; see Proposition 2.2. Another difficulty is the mis-match of the curved interface and
the discrete interface. Then follow the proof in [136, 29], and replace the mesh regular condition
there by the maximal condition, we obtain the desired results.

A mesh is O(h2σ ) irregular means the total area of all adjacent triangle pairs in Th which do not
form an O(h2 ) approximate parallelogram is O(h2σ ). For the interface-fitted mesh generated by

29
Algorithm 1, only the adjacent triangle pairs near the interface is not O(h2 ) approximate parallel-
ogram and other adjacent triangle pairs away from the interface can exactly form a parallelogram.
That is σ = 0.5 for the mesh generated by Algorithm 1.

Then follow the proof procedure in [129], we can also prove the following superconvergence
result.

T HEOREM 3.2. If u ∈ H 1 (Ω) ∩ H 3 (Ω̃) ∩ W 2,∞ and Γ is of class C 2 , then for all vh ∈ Vh ,

1/2
kβh (∇uh − ∇uI )k0,Ω . h3/2 kuk3,Ω̃ + kuk2,∞,Ω̃ + kuk2,∞,Ω̃ + kq1 k0,∞,Γ .


(3.5)

Let hmin be the minimum element size of Th , by the discrete embedding result,

kvh k0,∞,Ω . | log hmin |1/2 |vh |1,Ω , for all vh ∈ Vh ∩ H01 (Ω), (3.6)

we have the error estimate for the maximal norm estimate.

Corollary 3.3. Assume the same hypothesis in Theorem 3.2. Then

1/2
kβh (∇uh − ∇uI )k0,∞,Ω . | log hmin |1/2 h3/2 (kuk3,Ω̃ + kuk2,∞,Ω̃



+kuk2,∞,Ω̃ + kq1 k0,∞,Γ )

3.3 Virtual Element Methods in 3D

In this section, we focus on solving three-dimensional elliptic equations by the virtual element
methods (VEM) developed by Brezzi’s group [10, 11].

Let Th be the interface-fitted polyhedral mesh generated by the algorithm in Section 2.2. Recall
that elements near the interface Γ are polyhedra with triangular or square faces and a uniform cubic
mesh away from the interface. We could not use the classical finite element methods which are

30
not well-defined on polyhedra. Instead, we shall apply virtual element methods [10] which can be
thought of as conforming finite element spaces defined on polyhedral meshes.

A local finite-dimensional vector space Vh (E) for a polyhedron E ∈ Th is defined as

Vh (E) := {v ∈ H 1 (E) : ∆v|E = 0, v|∂E is continuous and

piecewise linear (on triangles) or bilinear (on squares)}.

As a piecewise linear or bilinear function will be uniquely determined by its value on vertices,
dim Vh (E) = nvE , where nvE is the number of vertices of E.

We define the global virtual element space

Vh = {vh ∈ H 1 (Ω) : vh |E ∈ Vh (E) for all E ∈ Ωh }.

Let N (Th ) be the set of vertices of mesh Th and N = |N (Th )| be the number of vertices. We define
the operator dof i from Vh to R as dof i (vh ) = vh (xi ), for a vertex xi ∈ N (Th ). The canonical basis
{φ1 , · · · , φN } ⊂ Vh is chosen as dof i (φj ) = δij , i, j = 1, · · · , N . And the nodal interpolation
Ih : C(Ω̄) → Vh is defined as Ih u = N
P
i=1 u(xi )φi and denoted by uI = Ih u. The basis does not

need to be written explicitly which is the main difference between classical finite element methods
and virtual element methods.

As mentioned before, we could extract an approximate surface Γh which splits Ω into two subdo-
mains: Ω− + − +
h and Ωh , which are the approximation of Ω and Ω , respectively. Similarly, βh |τ = β
+

−
for all τ ∈ Ω+ −
h and βh |τ = β for all τ ∈ Ωh .

Let wI− be the nodal interpolation of w− in Vh . A simple construction is one that: interpolates q0 on
Γh and sets other coefficients to zero. The linear virtual element approximation of (3.1) is: finding

31
ph ∈ Vh ∩ Hg1 (Ω) such that:

(βh ∇ph , ∇vh )Ω = (f, vh )Ω − hq1 , vh iΓ + (βh ∇wh− , ∇vh )Ω− , ∀vh ∈ Vh ∩ H01 (Ω)

PN PN
and taking uh = ph − wh− . Suppose ph = j=1 pj φj , wh− = j=1 wj φj , by linearity, we have for
i ∈ 1, · · · , N ,

N
X N
X
(βh ∇φj , ∇φi )Ω pj = (f, φi )Ω − hq1 , φi iΓ + (βh ∇φj , ∇φi )Ω− wj . (3.7)
j=1 j=1

We define the matrix (A− − + +

h )ij = (βh ∇φj , ∇φi )Ω− , (Ah )ij = (βh ∇φj , ∇φi )Ω+ and (Ah )ij =
h h

(βh ∇φj , ∇φi )Ωh in Ωh . Then Ah = A−

h + A+
h. Define the vector b = (b1 , · · · , bN )t by bi =
(f, φi )Ω − hq1 , φi iΓ . Equation (3.7) is written in the matrix form as

Ah ph = b + A −
h wh , (3.8)

where Ah and A− 1 N t t
h are N × N matrices, ph = (ph , · · · , ph ) and w h = (w1 , · · · , wN ) . Since the

coefficient β is a positive constant, the matrices Ah and A−

h are symmetric and positive definite.

The algebraic system (3.8) could be solved stably and efficiently by using algebraic multigrid
methods.

For finite element methods, it suffices to compute the local stiffness matrix in each element and
−
then, based on that, the matrices A+
h , Ah are assembled by summing the contribution from each

element. Therefore, the major task is to compute (∇φj , ∇φi )E .

To do so, we introduce some projection operators at first. For each polyhedron E, the operator
Π∇ : Vh (E) → P1 (E) is defined as the H 1 projection to P1 (E) space, i.e.,:

(∇pk , ∇Π∇ vh )E = (∇pk , ∇vh )E for all pk ∈ P1 (E),

32
where P1 (E) is the space of linear polynomials. As it can be easily seen that the above condition
defines Π∇ vh only up to a constant. This can be fixed by prescribing a projection operator onto
constants P0 : Vh (E) → P0 (E) and requiring

P0 (Π∇ vh − vh ) = 0.

PnvE PnvE
One such choice is P0 vh = i=1 vh (xi )/nvE = i=1 dof i (vh )/nvE .

Using the projection Π∇ , we write the basis function φi ∈ Vh (E) as Π∇ φi + (I − Π∇ )φi and split
the entry of the local stiffness matrix as

(∇Π∇ φi , ∇Π∇ φj )E + (∇(I − Π∇ )φi , ∇(I − Π∇ )φj )E .

Again the second term is not computable since the basis φi is not known point-wise. Instead we
replace by a so-called stabilization term S E (·, ·)

(∇Π∇ φi , ∇Π∇ φj )E + S E ((I − Π∇ )φi , (I − Π∇ )φj ).

Let hE be |E|1/3 , where |E| means the volume of E. We use a scaled l2 inner product in the
stabilization term

nEv
X
S E ((I − Π∇ )φi , (I − Π∇ )φj ) = hE dof r ((I − Π∇ )φi ) dof r ((I − Π∇ )φj )
r=1

in order to satisfy the assumption of S E

c1 (∇v, ∇v) ≤ S E (v, v) ≤ c2 (∇v, ∇v), ∀v ∈ Vh (E) and Π∇ v = 0

for some positive constants c1 and c2 independent of E and hE . The explicit expression of the local

33
stiffness matrix of the virtual element method is:

v
nE
X
∇ ∇
(K E
h )ij := (∇Π φi , ∇Π φj )E + hE dof r ((I − Π∇ )φi ) dof r ((I − Π∇ )φj ).
r=1

We now give concrete formulae on the computation of the matrix representation of the operator
PnvE
Π∇ . Let xE = (xE , yE , zE ) be the center of E, i.e. xE = 1/nvE i=1 xi . We choose a scaled
monomial basis of P1 (E) as m1 = 1, m2 = (x − xE )/hE , m3 = (y − yE )/hE , m4 = (z − zE )/hE .

Let G4×4 be defined as

 
P m
 0 1 P m
0 2 · · · P m
0 4
  
 
 0 (∇m2 , ∇m2 )0,E · · · (∇m4 , ∇m2 )0,E  1
 0 
G :=  . =

 .. .. .. ..   
 . . . 
 0 h I
E 3
 
0 (∇m2 , ∇m4 )0,E · · · (∇m4 , ∇m4 )0,E

where I 3 is a 3 × 3 identity matrix.

Let B 4×nvE be a matrix defined as:

 
 P0 φ1 ··· P0 φ nvE

 
(∇m2 , ∇φ1 )E · · · (∇m2 , ∇φnv )E 
E
B :=  .
 
.. .. ..

 . . . 

 
(∇m4 , ∇φ1 )E · · · (∇m4 , ∇φnE )E
v

The formulae for the first row of B is P0 φ1 = P0 φ2 = . . . = P0 φnvE = 1/nvE . For the other
R R ∂m
components (∇mj , ∇φi )E , j = 2, 3, 4, we have (∇mj , ∇φi )E = − E ∆mj φi + ∂E ∂nj φi by
integration by parts. The first term is zero as ∆mj = 0 for linear polynomials. We only need to
compute the second term. Due to our data structure, all the faces on the ∂E are either triangles or

34
squares. Then

njf |f | njf |f |
Z P P
∂mj i∈triangular facef i∈square facef
φi = + , (3.9)
∂E ∂n 3hE 4hE

where nf = (nxf , nyf , nzf ) = (n2f , n3f , n4f ) is an outward unit normal direction on each face f and
|f | is the area for each face f .

Remark 3.4. Using our mesh generation algorithm in Section 2.2, we store the polyhedron in
the form of either triangles or squares which leads to the simple formula (3.9). When the faces
are general polygons, additional projection operators are needed in order to compute the integral
R ∂mj
φ (see [10, 11]). 
f ∂n i

To compute the stabilization term, we need one more matrix D nvE ×4

 
hE x1 − xE y1 − yE z1 − zE 
 
  hE x2 − xE y 2 − y E y2 − zE

D := dof i (mj ) = h−1
 
E  
· · · ··· ··· ··· 
 
 
hE xnE − xE ynE − yE znE − zE ,
v v v

where (xi , yi , zi ), i = 1, · · · , nvE are vertices in each polyhedron E.

P4
By definition, Π∇ vh = α=1 sα mα and the coefficients (sα ) is determined by the following linear
systems

(∇mα , ∇(Π∇ vh − vh ))E = 0 α = 1, . . . , 4.

The matrix representation of Π∇ : Vh (E) → P1 (E) relative to the basis (mα ) is Π∇ = G−1 B.

We will also need the matrix representation of Π∇ in the canonical basis {φi }. Let Π∇ φi =
PnvE ∇ v ∇ ∇
j=1 dof j (Π φi )φj , i = 1, · · · , nE , then the matrix representation Π∗ of the operator Π :
Vh (E) → Vh (E) in the canonical basis is given by Π∇ −1 ∇
∗ = DG B = DΠ .

35
Finally the matrix formulation of K E
h could be written as

∇ T ∇ ∇ T ∇
KE
h = [Π ] G̃Π + hE [I − Π∗ ] [I − Π∗ ],

where G̃ is the same with G except that the elements in the first row are all zero.

For the first term of bi in (3.8), we approximate f by a piecewise constant and approximate

X X
(f, φi )Ω = (f, φi )E ≈ |E|f (xE , yE , zE )/nvE .
E∈Ωh E∈Ωh

The second term of bi could be computed by Gauss quadrature on surface mesh Γh .

Remark 3.5. An abstract error estimate of VEM has been given in [10]. With a type of Ceá lemma,
the convergence analysis is reduced to the interpolation error estimate |u − uI |1 and |u − uπ |1,E ,
where uI is the nodal interpolation and uπ is a local approximation of u. Notice that uI ∈ Vh
is continuous but uπ is most likely discontinuous. To obtain optimal order of the interpolation
and approximation error, the authors in [10] further assume the shape-regular condition: there
exists a γ > 0 such that each domain E is star-shaped with respect to a ball of radius ρ ≥ γhE ,
where hE = diam(E). This shape regularity assumption will rule out elements generated by our
algorithm.

As we mentioned before, for linear finite element space defined on triangles, a refined analysis
shows that the optimal first order interpolation error estimate still holds if the maximum angle
is uniformly bounded away from π as h → 0 [6]. Such angle condition is generalized to three
dimensions, and to high order elements in [76, 3, 37]. Generalization to polyhedron, however, is
unknown and under investigation.

36
3.4 Numerical Experiments

In this section, we present numerical results for the elliptic interface problems in three dimensions.
We implement mesh generation and VEM based on the MATLABr package iFEM [25]. We also
solve the algebraic system by an algebraic multigrid (AMG) solver implemented in iFEM [25]. We
start with a simple spherical interface and then consider more complex geometric shapes, including
two spheres, an orthocircle shape and 12 intersecting spheres. We shall report the following errors:

 1/2
1/2 1/2
kuI − uh kA = kβh ∇(u− I − u− 2
h )k Ω−
+ kβ h ∇(u +
I − u+ 2
h )k Ω+
,
h h
n o
− − + +
kuI − uh k∞ = max kuI − uh k∞,Ω , kuI − uh k∞,Ω ,
− +
h h
 1/2
X X
kuI − uh k0,h = h3/2  (u− − 2
I (xi ) − uh (xi )) + (u+ +
I (xi ) − uh (xi ))
2
,
xi ∈N (Ω−
h) xi ∈N (Ω+
h)

−
where uh is the numerical solution obtained by the linear virtual element methods; u+
I and uI are
−
the nodal interpolation of the exact solution u in Ω+
h and Ωh respectively. Note that the squared
− −
energy norm kuI −uh k2A can be computed by (u− − T − + + T + + +
I −uh ) Ah (uI −uh )+(uI −uh ) Ah (uI −uh )

and k · k0,h is a good approximation of L2 -norm. The rate is obtained by the least square fitting of
the errors in the log log scale.

E XAMPLE 3.1 (One sphere). The domain Ω is (−1, 1)3 and the interface is defined by φ(x, y, z) =
x2 +y 2 +z 2 −r2 with radius r = 0.75. The coefficient β is piecewise constant. The analytic solution
is given by u+ = 10(x + y + z) and u− = 5 exp(x2 + y 2 + z 2 ) + 20. In this case, the solution is
discontinuous and the flux jump across the interface is also non-homogenous.

Fig. 3.1 shows the surface mesh extracted from the volume mesh generated by our algorithm for the
spherical interface. The maximal interior angle of triangular faces on the surface mesh is bounded
by 112.8104◦ . Tables 3.1 and 3.2 show the error for β − = 1, β + = 10 and β − = 1, β + = 100,
respectively. As it can be seen that near second order accuracy is attained in both k · k0,h and

37
k · k∞ norms. The convergence rate in the energy norm is near 1.5, which is consistent with the
superconvergence result obtained in [24]. This superconvergence occurs due to the nice properties
of our semi-unstructured mesh. It seems that the convergence rate is robust to the variation of β.

From Table 3.3, we can conclude that the runtimes of the mesh generation part can be ignored
compared with the assembling and solving parts. In Table 3.4, we present the variation of iteration
steps of the algebraic multigrid method with respect to the number of degrees of freedom (#dof)
and to the variation of jump coefficients (fix β − = 1 and change β + ). It indicates that the algebraic
multigrid method is a robust and efficient solver: robust to the number of degrees of freedom and
to the variation of jump coefficients.

Figure 3.1: An interface mesh with maximal angle 112.8104◦ .

38
 Table 3.1: Errors for Example 3.1: β − = 1 and β + = 10.

#dof h kuI − uh kA kuI − uh k∞ kuI − uh k0

10,323 0.1 4.45798e-01 5.24227e-02 3.62347e-02
72,713 0.05 1.64215e-01 1.81762e-02 8.97415e-03
547,881 0.025 6.62120e-02 5.09725e-03 2.62269e-03
4,240,529 0.0125 2.43899e-02 1.37664e-03 7.05429e-04
Rate 1.4 1.8 1.9

Table 3.2: Errors for Example 3.1: β − = 1 and β + = 100.

#dof h ||uI − uh ||A ||uI − uh ||∞ ||uI − uh ||0

10,323 0.1 6.65319e-01 5.14276e-02 3.40871e-02
72,713 0.05 2.41564e-01 1.84007e-02 8.42552e-03
547,881 0.025 8.98117e-02 5.21504e-03 2.50271e-03
4,240,529 0.0125 3.21041e-02 1.42298e-03 6.80724e-04
Rate 1.5 1.7 1.9

Table 3.3: CPU time (in seconds) for Example 3.1: β − = 1 and β + = 10.

#dof Assemble Solve Mesh

10,323 0.228973 0.47 0.21165
72,713 0.970122 2.59 0.4854
547,881 6.99419 13.55 1.8981
4,240,529 62.1644 121.59 8.0172

Table 3.4: Example 3.1: Iteration steps of AMG with fixed β − = 1 and various β + .

β+
10−3 10−2 10−1 1 10 102 103 104
#dof
7,921 10 10 10 9 9 9 9 9
63,111 11 11 11 11 11 10 10 10
509,479 12 12 12 12 14 13 13 13
4,086,927 13 12 13 15 17 16 16 16

39
E XAMPLE 3.2 (Two spheres). The domain Ω is (−1, 1)3 and the interface is defined by

φ(x, y, z) = min {φ1 , φ2 }

where

φ1 = (x + 0.5r)2 + (y + 0.5r)2 + (z + 0.75r)2 − r2 ,

φ2 = (x − 0.25r)2 + (y − 0.75r)2 + (z − r)2 − r2 ,

with radius r = 0.4. The coefficient β is piecewise constant. The analytic solution is given by
u+ = 10(x2 + y 2 + z 2 ) and u− = 5 cos (x2 + y 2 + z 2 ).

Fig. 3.2 shows that two spheres are embedded in the unit cube. When h = 0.1, i.e., the background
Cartesian grid is not fine enough, there exists an interface element which is divided into three
parts by these two spheres. The maximal interior angle of triangular faces on the surface mesh
is bounded by 130.4665◦ . Tables 3.5 and 3.6 show the numerical results for β − = 1, β + = 1
and β − = 1, β + = 100, respectively. Table 3.7 shows how the computational time grows with
respect to the number of degrees of freedom. Table 3.8 shows the number of iterations taken by
the algebraic multigrid method for various values of input parameters.

All results are consistent with our conclusion. It indicates that our algorithm works in a case when
the interface is unconnected.

Table 3.5: Errors for Example 3.2: β − = 1 and β + = 1.

#dof h ||uI − uh ||A ||uI − uh ||∞ ||uI − uh ||0

9,789 0.1 3.76723e-01 1.08148e-01 6.52076e-02
71,225 0.05 1.32276e-01 2.82699e-02 1.74225e-02
540,945 0.025 4.52335e-02 7.30380e-03 4.28321e-03
4,211,729 0.0125 1.60165e-02 1.92081e-03 1.11271e-03
Rate 1.5 1.9 2

40
 Table 3.6: Errors for Example 3.2: β − = 1 and β + = 100.

#dof h ||uI − uh ||A ||uI − uh ||∞ ||uI − uh ||0

9,789 0.1 3.67145e+00 1.15596e-01 5.54995e-02
71,225 0.05 1.35081e+00 3.08549e-02 1.51136e-02
540,945 0.025 4.78640e-01 8.22331e-03 3.74778e-03
4,211,729 0.0125 1.72934e-01 2.25395e-03 9.77991e-04
Rate 1.5 1.9 2

Table 3.7: CPU time (in seconds) for Example 3.2: β − = 1 and β + = 1.

#dof Assemble Solve Mesh

9,789 0.160716 1.01 0.259938
71,225 0.706959 5.84 0.36018
540,945 6.32812 27.26 1.78691
4,211,729 52.7369 133.86 10.067

Table 3.8: Example 3.2: Iteration steps of AMG with fixed β − = 1 and various β + .

β+
10−3 10−2 10−1 1 10 102 103 104
#dof
7,387 10 11 10 9 9 9 9 9
61,623 11 11 11 10 10 10 10 10
502,543 12 12 12 12 12 11 12 11
4,058,127 13 13 13 12 13 13 13 13

41
 Figure 3.2: Two balls are embedded in the unit cube. The maximal angle is 130.4665◦ .

E XAMPLE 3.3 (An orthocircle). The domain Ω is (−1.2, 1.2)3 and the interface is defined by

φ(x, y, z) = (x2 + y 2 − 1)2 + z 2 (x2 + z 2 − 1)2 + y 2 (y 2 + z 2 − 1)2 + x2 − 0.0752

   

1 + 3(x2 + y 2 + z 2 ) .
 

The coefficient β is piecewise constant. The analytic solution is given by u+ = 1 − x2 − y 2 − z 2

and u− = sin(πx) sin(πy) sin(πz).

Fig. 3.3 shows the interface-fitted mesh extracted as the boundary of Ω−

h . The maximal angle of

triangular faces of the interface mesh is bounded by 132.4673◦ . Tables 3.9 and 3.10 show the
numerical results for β − = 1, β + = 1 and β − = 1, β + = 100, respectively. Table 3.11 shows
the computational time accordingly. The mesh part is still quick even the interface with complex
geometry. Table 3.12 is the number of iterations of the algebraic multigrid (AMG) solver.

42
These test results indicate that the proposed interface problem solver works well even for complex
surfaces. Note that the maximal angles of the surface mesh in our examples are uniformly bounded
by 144◦ , then there is no need to be smoothed as a post-processing step.

Figure 3.3: The interface is an orthocircle with maximal angle 132.4673◦ .

E XAMPLE 3.4 (12 intersecting spheres). In this example, we consider a more complicated in-
terface formed by 12 intersecting spheres, which has many one-dimension sharp features. The
domain Ω is (−3, 3)3 and the interface is defined by

φ(x, y, z) = min {φ1 , φ2 , · · · , φ12 }

43
 Table 3.9: Errors for Example 3.3: β − = 1 and β + = 1.

#dof h ||uI − uh ||A ||uI − uh ||∞ ||uI − uh ||0

11,145 0.12 2.93834e-01 5.50055e-02 5.14120e-02
76,469 0.06 7.95795e-02 1.05921e-02 4.18525e-03
561,957 0.03 2.35627e-02 2.22961e-03 9.17810e-04
4,295,165 0.015 7.80143e-03 6.05301e-04 2.15878e-04
Rate 1.7 2.1 2.1

Table 3.10: Errors for Example 3.3: β − = 1 and β + = 100.

#dof h ||uI − uh ||A ||uI − uh ||∞ ||uI − uh ||0

11,145 0.12 3.44042e+00 9.63266e-02 1.24556e-01
76,469 0.06 7.72426e-01 1.21513e-02 1.56542e-02
561,957 0.03 2.06488e-01 3.04186e-03 3.34684e-03
4,295,165 0.015 6.30787e-02 7.45803e-04 8.10506e-04
Rate 1.8 2 2.1

Table 3.11: CPU time (in seconds) for Example 3.3: β − = 1 and β + = 1.

#dof Assemble Solve Mesh

11,145 0.493225 0.75 0.383038
76,469 1.41325 1.72 0.857616
561,957 8.1635 12.05 3.48605
4,295,165 62.6127 97.92 13.8495

Table 3.12: Example 3.3: Iteration steps of AMG with fixed β − = 1 and various β + .

β+
10−3 10−2 10−1 1 10 102 103 104
#dof
8,743 15 14 15 13 11 10 10 10
66,867 13 12 11 11 11 11 11 11
523,555 13 12 12 12 12 12 12 12
4,141,563 13 13 13 13 13 13 13 13

44
where φi , i = 1, . . . , 12 are equal radius balls with centers

(1.0, 0, 0), (−1.0, 0, 0), (0.5, 0.866025403784439, 0), (−0.5, 0.866025403784439, 0),

(0.5, −0.866025403784439, 0), (−0.5, −0.866025403784439, 0), (2.0, 0, 0), (−2.0, 0, 0),

(1.0, 1.73205080756888, 0), (−1.0, 1.73205080756888, 0), (−1.0, −1.73205080756888, 0),

(1.0, −1.73205080756888, 0)

and radius r = 0.7. The coefficient β is a piecewise constant. The analytic solution is given by
u+ = 10(x2 + y 2 + z 2 ) and u− = 5 cos (x2 + y 2 + z 2 ).

Fig. 3.4 shows the interface-fitted mesh extracted as the boundary of Ω−

h , which did not capture the

sharp one-dimension features of the interface, and this will lead to a loss of the solution accuracy.
From part of the interface-fitted mesh, there are geometric singularities with sharp edges at the
intersection of balls. See Fig. 3.5 for an illustration. In order to capture the features of complicated
interface and improve the solution accuracy, one need to use adaptive mesh near the geometric
features, and this will be our future work. But the maximal angle condition is still satisfied as
the maximal angle of triangular faces on the surface mesh is bounded by 131.3925◦ . Tables 3.13
and 3.14 show the numerical results for β − = 1, β + = 1 and β − = 1, β + = 10, respectively. As it
can be seen that the convergence rate in the energy norm is still near 1.5. But the convergence rates
in the k·k0,h and k·k∞ norms are not second order due to geometric singularities. Table 3.15 shows
how the computational time grows with respect to the number of degrees of freedom. Table 3.16
shows the number of iterations taken by the algebraic multigrid method for various values of input
parameters.

45
 Table 3.13: Errors for Example 3.4: β − = 1 and β + = 1.

#dof h ||uI − uh ||A ||uI − uh ||∞ ||uI − uh ||0

2,527 0.48 2.27225e+01 3.60464e+00 6.97070e+00
18,960 0.24 1.18700e+01 2.24845e+00 4.20002e+00
138,051 0.12 3.99601e+00 8.79537e-01 9.22094e-01
1,051,665 0.06 1.57695e+00 2.91154e-01 3.92484e-01
Rate 1.4 1.2 1.6

Table 3.14: Errors for Example 3.4: β − = 1 and β + = 10.

#dof h ||uI − uh ||A ||uI − uh ||∞ ||uI − uh ||0

2,527 0.48 8.71247e+01 5.29705e+00 8.63843e+00
18,960 0.24 3.90973e+01 3.07134e+00 4.54532e+00
138,051 0.12 1.38105e+01 1.37945e+00 1.15474e+00
1,051,665 0.06 5.49027e+00 4.88589e-01 3.78441e-01
Rate 1.3 1.1 1.6

Table 3.15: CPU time (in seconds) for Example 3.4: β − = 1 and β + = 1.

#dof Assemble Solve Mesh

2,527 0.213214 0.29 0.3674
18,960 0.431979 1.81 0.61863
138,051 2.22761 4.88 2.4452
1,051,665 29.9028 42.76 6.8316

Table 3.16: Example 3.4: Iteration steps of AMG with fixed β − = 1 and various β + .

β+
10−2 10−1 1 10 102 103 104
#dof
1,611 11 10 9 8 9 8 8
15,208 12 11 10 10 10 10 10
123,049 12 12 11 11 11 11 11
991,663 14 12 12 12 12 12 12

46
Figure 3.4: Twelve balls are embedded in the cube. The maximal angle is 131.3925◦ .

Figure 3.5: Part of the interface-fitted mesh with sharp edges.

47
Chapter 4

Moving Interface Problems

There are two major sections in this chapter. The first section is to discuss the 2D model of the
tissue growth problems. The second section is about the 2D model of the tumor growth problems.
In this chapter we will discuss the applications of solving the Poisson equation with a moving
boundary domain. Free interface problems have been applied heavily in different fields such as bi-
ology, physics, and engineering recently. However, due to the discontinuities of physical quantities
across the interface designing accurate numerical methods for such problems is still challenging.
The fundamental problem in exploring the numerical scheme is how to treat the discontinuities.
Many schemes have tried to use the different strategies to handle the discontinuities across the
interface. For example, one of the most popular strategies is to convert the discontinuities to sin-
gular terms by adding the regularization term. The method is very straightforward and robust with
various spatial discretization methods such as finite difference, finite element, and the finite vol-
ume method, and so on. Level-set method [102], immersed boundary method [106, 107] and front
tracking methods [112] use this approach. Another popular strategy is to use local differencing
schemes or basis functions near the interface to approximate the discontinuities. The matched in-
terface and boundary method (MIB) employed this strategy [145, 146]. The last popular strategy
applied interface-fitted mesh with interfaces passing through mesh node points [113]. Here, we use

48
the interface-fitted mesh strategy to solve the moving interfaced problems since we could generate
good quality mesh efficiently. In every time step of the interface evolution, we use the zero level
set of the level set function to identify the interface position and then generate the interface-fitted
mesh. The velocity of the flow is solved using the linear finite element method and then used it to
solve the transport equation to update the level set function. We iterate this process to evolve the
interface. In order to show our schemes in details, we applied our schemes to solve the interface
problems in biology.

4.1 Tissue Growth Problem

This section is to introduce the tissue growth problem with static interface at first. The linear and
quadratic element methods are showed for this case. The moving interface problem is discussed
afterwards.

4.1.1 Tissue Growth Model

The following model describes the process of tissue growth [103]. The boundary is either the
interface between neighboring tissues or an open external environment and the pressure on the
boundary is proportional to the curvature of the boundary. The model has periodic boundary
condition along the x-direction throughout the tissue. The growth of tissue is quantified by the
dynamic boundary of the system. First, we discuss derivation of equations, which is on the two
dimensional rectangle domain Ω = (0, 2) × (0, h).

−K∆P = Ψ, x ∈ Ω (4.1)

49
with boundary conditions:

∂P
v(x, y = 0, t) = (x, y = 0, t) = 0, (4.2)
∂y
∂ 2 h/∂x2
P (x, h, t) = ξκ = ξ , (4.3)
[1 + (∂h/∂x)2 ]3/2

and the periodic boundary condition:

P (x = 0, y, t) = P (x = 2, y, t). (4.4)

The growth of the dynamic boundary of the system is determined by h, which is governed by the
kinematic condition,
∂h ∂h
+ u(x, h, t) = v(x, h, t).
∂t ∂x

Here u = −KPx and v = −KPy [103], P is the pressure, the gradient obtained from ( 4.1)
(Px , Py ) gives the velocity of the tissue, Ψ represents the net rate of production, ξ is related to the
magnitude of the surface tension along the boundary and K is the permeability. See Fig. 4.1 for
illustration.

y = h(x, t)

0 2
Figure 4.1: The domain of tissue growth model.

50
4.1.2 Kinematic Boundary Condition

In order to link the velocity of a boundary to the particle adjacent to the boundary we need to
impose a condition. It is the motivation of the kinematic boundary condition, which is modeled by
the transport equation. Firstly we consider the general 2D equations for a free-surface flow with
the compressible fluid and the flow driven by the potential.

u = ∇P and −∆P = f are used to describe the fluid, where u = (u, w) and P = P (x, y, t).
y = h(x, t) is the surface of the fluid in motion, f is the force term and y = 0 is the lower boundary
where the vertical velocity vanishes.

∂p
0 = n · ∇p = = v.
∂y

In other words, we get

∂p
= 0 on y = 0.
∂y

The velocity of the fluid normal to the boundary is required to be equal to the velocity of the
boundary normal to itself. Otherwise, the fluid would either be flowing through the boundary or
separating from it. Neither of them is acceptable. Therefore, the kinematic boundary condition on
DS
a moving surface is Dt
= 0 and the free surface is the zero set of S(x, t). We define a function
based on the height function h(x, t).

S(x, y, t) = y − h(x, t),

It is zero at the free surface. Thus

∂ ∂ ∂ ∂h ∂p ∂h ∂p
0=( +u + v )(y − h(x, t)) = − − + on y = h.
∂t ∂x ∂y ∂t ∂x ∂x ∂y

51
Therefore, the kinematic boundary condition holds

∂h ∂h
+u =v on y = h. (4.5)
∂t ∂x

We chose the discontinuous Galerkin finite element (DG-FEM) scheme to solve the above transport
equation 4.5. By subdividing the interval Ω = [L, R] into a union of non-overlapping elements
Dk = [xkl , xkr ]

Ω ≈ Ωh = ∪K k
k=1 D ,

as illustrated in Fig. 4.2 .

Dk 1 Dk Dk+1
x
L= x1l xkr 1 = xkl xkr = xk+1
l
xkr =R
hk+1

Figure 4.2: Sketch of the geometry for simple one-dimensional example.

Consider the one-dimensional scalar conservation law for the solution u(x, t)

∂u ∂f (u)
+ = g. (4.6)
∂t ∂x

subject to an appropriate set of initial conditions u(x, 0) = u0 (x) and boundary conditions on the
boundary, ∂Ω, where f (u) is the flux and g(x, t) is some prescribed forcing function. To solve
Eq. 4.6, we define a space of test functions, Vh , to which the residual is orthogonal to in this space
in the weak sense as

Z
∂uh ∂fh
( + − gh )φh (x)dx = 0, ∀φh ∈ Vh .
Ω ∂t ∂x

52
Next, the local residual is formed on the k = 1, · · · , N E elements

x ∈ Dk : Rhk (x, t) = ∂t ukh + ∂x fhk − ghk

and require this to vanish locally in a Galerkin sense

Z
Rhk (x, t)lik (x)dx = 0, i = 1, . . . , Np , k = 1, . . . , N E
Dk

It forms the basis of a nodal DG-FEM scheme. Gauss’s theorem is applied to obtain the local
statement.

∂ukh k dljk
Z
k+1
lj − fhk − gljk dx = −[fhk ljk ]xxk .
Dk ∂t dx

The right-hand side is used to connect the elements. However, the uniqueness of the solutions at
interfaces between adjacent elements cannot meet. To overcome it a numerical flux f ∗ is employed
to approximate the physical flux. We applied Gauss’s theorem and get

∂ukh k dlk
Z
k j k+1
lj − fh − gljk dx = −[f ∗ ljk ]xxk .
Dk ∂t dx

Then we applied Gauss’s theorem once again,

Z
k+1
Rh (x, t)ljk dx = [(fhk − f ∗ )ljk ]xxk .
Dk

The two schemes are the weak and strong forms for the scalar conservation law. The key of the
scheme is how to choose the numerical flux, f ∗ . The two local elementwise schemes are written

53
using the terminology of the finite element scheme.

dukh
Mk − (S k )T f kh − M k g kh = −f ∗ (xk+1 )lk (xk+1 ) + f ∗ (xk )lk (xk ),
dt

and

dukh
Mk + S k f kh − M k g kh = (fhk (xk+1 ) − f ∗ (xk+1 ))lk (xk+1 ) − (fhk (xk ) − f ∗ (xk ))lk (xk ).
dt

Here local unknowns,ukh ,of fluxes,f kh , and the forcing,g kh are given on the nodes in each element.
dlk (x)
lk (x) = [l1k (x), . . . , lN
k T k
lk (x)ljk (x)dx and S kij = Dk lik (x) jdx dx are the
R R
p (x)] , and M ij =
Dk i

local mass matrix and stiffness matrix, respectively [58].

We construct the matrices of M kij and S kij as the similar way of assembling the mass and the
stiffness matrices in the finite element methods.

4.1.3 Static Problems

We consider the static problem to test our scheme first and then solve the problem using linear and
quadratic finite element methods. K and ξ are set as 1. The top boundary of the tissue is given by
h(x) = 1 + B cos(2πx) with different B for numerical tests. The exact solution of the pressure is

−4π 2 B cos(2πx) cos(2πy)/(1 + B cos(2πx))

P (x, h) = .
[1 + 4π 2 B 2 sin2 (2πx)]3/2

We need a larger box containing the domain Ω and generate the mesh for the rectangle box. In the
first step, we could get the interface points based on given ∆x. In the second step, we generate
Delaunay triangulations on interface points. In the last step, we remove triangles not in the interface
elements and merge all uncut elements below h to get a semi-unstructured mesh. Finally, we could
get the approximated mesh of Ω. Figure 4.3 shows the mesh generation procedure. In order to

54
reduce the complexity, we could divide all squares into triangles (see Figure 4.4) and use finite
element methods to solve the Poisson equations.

(a) Step 1: Find all the interface points on the curve h. (b) Step 2: a Delaunay triangulation on interface points.

(c) Step 3: Remove triangles not in the interface elements. (d) Step 3: Merge all uncut elements to get a semi-
unstructured mesh.

Figure 4.3: Three steps to generate a semi-unstructured mesh.

Based on the assumption in the X direction of the domain, there is a straightforward way to impose
periodic boundary conditions in the matrix form. The details are shown in [47]. It is, however, not
easy to implement the method for higher order schemes. In this thesis, we use an alternative way
to impose the periodic boundary conditions, which could be used for higher order algorithm easily.
At first, we get the original mesh as above, which is only used for all geometric information such

55
 Figure 4.4: All elements are triangles

as length, area. See the right graph in Figure 4.5. Then we introduce another artificial mesh which
requires that the index of nodes on right hand side of the points is the same as that of the left hand
side of the points. See the left graph in Figure 4.5 for illustration. It is used for assembling the
finite element matrix except geometric information. For linear finite element methods, we use a
map between the original mesh and artificial mesh called nodeMap to impose periodic conditions.
For quadratic finite element methods, another map named dofMap should be introduced to get the
relationship between middle points. It could be extended to any higher order elements.

26 52 26 26
25 51 25 25
24 50 24 24
23 49 23 23
22 48 22 22
21 47 21 21
20 46 20 20
19 45 19 19
18 44 18 18
17 43 17 17
16 42 16 16
15 41 15 15
14 40 14 14
13 39 13 13
12 38 12 12
11 37 11 11
10 36 10 10
9 35 9 9
8 34 8 8
7 33 7 7
6 32 6 6
5 31 5 5
4 30 4 4
3 29 3 3
2 28 2 2
1 27 1 1

(a) The original mesh based on initial boundary con- (b) Construct the artificial mesh which imposes pe-
dition of h. riodic conditions

Figure 4.5: The mesh with periodic boundary condition

In this subsection, we test the convergence rates to demonstrate the effectiveness of our methods.

1 1
56
The energy (|| · ||A ) and maximum norms (|| · ||∞ ) are considered. Here is the definition:

p
||PI − P∆x ||A = (PI − P∆x )T A(PI − P∆x ), (4.7)

||PI − P∆x ||∞ = max |PI − P∆x |, (4.8)

where PI is the nodal interpolation of the exact solution of P , P∆x is the numerical solution ob-
tained by finite element methods, A is the corresponding stiffness matrix and ∆x is the mesh size
of the initial Cartesian mesh.

E XAMPLE 4.1 (B = 0.1). Table 4.1 and 4.2 show the numerical convergence results for linear
and quadratic finite element methods respectively when B = 0.1.

Table 4.1: Errors for Example 1: linear finite element methods when B = 0.1.

#Dof ∆x ||PI − P∆x ||A ||PI − P∆x ||∞

455 5.00e-02 2.09603e-01 7.44661e-02
1715 2.50e-02 5.58425e-02 2.00167e-02
6635 1.25e-02 1.50097e-02 4.96067e-03
26075 6.25e-03 4.41325e-03 1.21513e-03
Rate 1.9 2

Table 4.2: Errors for Example 1: quadratic finite element methods when B = 0.1.

#Dof ∆x ||PI − P∆x ||A ||PI − P∆x ||∞

887 5.00e-02 5.82090e-02 3.14154e-03
3385 2.50e-02 8.69113e-03 4.42525e-04
13181 1.25e-02 1.34846e-03 5.94591e-05
51973 6.25e-03 2.21733e-04 7.68399e-06
206357 3.13e-03 3.71963e-05 9.75976e-07
Rate 2.7 2.9

E XAMPLE 4.2 (B = 0.15). Table 4.3 and 4.4 show the numerical results for linear and quadratic
finite element methods respectively when B = 0.15.

57
 Table 4.3: Errors for Example 2: linear finite element methods when B = 0.15.

#Dof ∆x ||PI − P∆x ||A ||PI − P∆x ||∞

460 5.00e-02 5.86786e-01 1.43121e-01
1725 2.50e-02 1.45334e-01 4.35102e-02
6655 1.25e-02 3.80448e-02 1.11352e-02
26115 6.25e-03 1.00820e-02 2.72108e-03
Rate 2 2

Table 4.4: Errors for Example 2: quadratic finite element methods when B = 0.15.

#Dof ∆x ||PI − P∆x ||A ||PI − P∆x ||∞

896 5.00e-02 1.68363e-01 1.02863e-02
3403 2.50e-02 2.53524e-02 1.31378e-03
13217 1.25e-02 3.88118e-03 1.85960e-04
52045 6.25e-03 6.20807e-04 2.48831e-05
206501 3.13e-03 1.04426e-04 3.19088e-06
Rate 2.7 2.9

E XAMPLE 4.3 (B = 0.25). Table 4.5 and 4.6 show the numerical results for linear and quadratic
finite element methods respectively when B = 0.25.

Table 4.5: Errors for Example 3: linear finite element methods when B = 0.25.

#Dof ∆x ||PI − P∆x ||A ||PI − P∆x ||∞

470 5.00e-02 4.02367e+00 6.86283e-01
1745 2.50e-02 6.66825e-01 1.31735e-01
6695 1.25e-02 1.66809e-01 3.81307e-02
26195 6.25e-03 4.26586e-02 9.52049e-03
Rate 2.2 1.9

E XAMPLE 4.4 (B = 0.5). Table 4.7 and 4.8 shows the numerical results for linear and quadratic
finite element methods respectively when B = 0.5.

We could get the conclusion that near second order of accuracy of our method is in ||·||A and ||·||∞
norm for linear finite element methods. We also consider high order finite element methods such

58
 Table 4.6: Errors for Example 3: quadratic finite element methods when B = 0.25.

#Dof ∆x ||PI − P∆x ||A ||PI − P∆x ||∞

914 5.00e-02 8.47067e-01 6.98431e-02
3439 2.50e-02 1.30214e-01 6.88447e-03
13289 1.25e-02 1.94364e-02 9.88083e-04
52189 6.25e-03 3.01006e-03 1.34961e-04
206789 3.13e-03 4.91737e-04 1.77970e-05
Rate 2.7 2.9

Table 4.7: Errors for Example 4: linear finite element methods when B = 0.5.

#Dof ∆x ||PI − P∆x ||A ||PI − P∆x ||∞

495 5.00e-02 1.08631e+02 4.41181e+01
1795 2.50e-02 9.94629e+00 1.39314e+00
6795 1.25e-02 1.68044e+00 2.27722e-01
26395 6.25e-03 4.15898e-01 6.59658e-02
Rate 2.7 2.2

as quadratic finite element methods. The convergence rates for || · ||A and || · ||∞ norms are nearly
three. It seems that the convergence rate is robust to the variability of the coefficient B. But the
iterative pseudo spectral approach fails to converge for some B [103]. What’s more, our method
could be extended to any higher order algorithm easily. The convergence rate in the energy norm
is higher than the theoretical results due to the nice structure of our grids. And this is known as
superconvergence.

4.1.4 Moving Problems

In this subsection, we test the accuracy of our approach for the moving interface problems with
known or unknown solutions. We use linear finite element methods and second order of DG-FEM
to solve the coupled system of both P and h. In addition, a 4th order Runge Kutta (RK) is employed
for time discretization.

59
 Table 4.8: Errors for Example 4: quadratic finite element methods when B = 0.5.

#Dof ∆x ||PI − P∆x ||A ||PI − P∆x ||∞

959 5.00e-02 9.09381e+00 1.27759e+00
3529 2.50e-02 1.74621e+00 1.24500e-01
13469 1.25e-02 2.61604e-01 1.19905e-02
52549 6.25e-03 3.72530e-02 1.67308e-03
207509 3.13e-03 5.55039e-03 2.31907e-04
Rate 2.7 3

E XAMPLE 4.5 (Moving problems with exact solution). The initial state of the dynamic boundary
h is described by h(x, 0) = 1, and a uniform influx of cells is assumed by Ψ = −4. The exact
solution are given by P = 2(y 2 − h2 ) and h = e−4t . Parameters chosen are K = 1 and ξ = 10−5 .
We choose the final time T = 0.004. Recall that we consider the two dimensional domain Ω =
(0, 2) × (0, h).

−K∆P = Ψ, x ∈ Ω

with boundary conditions:

∂P ∂ 2 h/∂x2
v(x, y = 0, t) = (x, y = 0, t) = 0, P (x, h, t) = ξκ = ξ ,
∂y [1 + (∂h/∂x)2 ]3/2

and periodic condition :

P (x = 0, y, t) = P (x = 2, y, t).

The growth of the dynamic boundary of the system, h is governed by the kinematic condition,
∂h
∂t
+ u(x, h, t) ∂h
∂x
= v(x, h, t), where u = −KPx and v = −KPy . Since the velocity of the
transport equation is changing in different time, we choose the time step ∆t = O(∆x/ max (u, v)),
where ∆x is a uniform span in the x direction. Given the inital boundary condition h, we generate
the interface-fitted mesh. The procedure is the same as 4.1.3. Then we use linear finite element

60
methods to solve P . For each boundary node, choose 5 closet neighbour mesh nodes in x and y
direction. Fit those points with cubic polynomial using polyfit, which is a built-in function in
MATLAB. Finally the velocity u and v could be the first derivative of the fitted cubic polynomial.
Then we solve the transport equation using DG-FEM to update the new h. The couple system
P and h will be solved iteratively. In Figure 4.6, it shows the interface-fitted mesh generated in
different time step. From Table 4.9, it shows all the order of spatial accuracy is 2 for P using
|| · ||∞ and || · ||A norms and for h using || · ||∞ norm.
1.2

1.2

1
1

0.8
0.8

0.6
0.6

0.4
0.4

0.2
0.2

0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

(a) time = 0.00 (b) time = 0.01

1.2
1.2

1
1

0.8
0.8

0.6
0.6

0.4
0.4

0.2 0.2 1
1

0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

(c) time = 0.02 (d) time = 0.04

Figure 4.6: The interface-fitted mesh in the domain with different time when ∆x is 0.05.

E XAMPLE 4.6 (Moving problems without exact solution(a straight line)). The initial state of the
dynamic boundary h is described by h(x, 0) = 1 and a uniform influx of cells is assumed by
Ψ = 4. Parameters chosen are K = 1 and ξ = 10−5 . We choose the final time T = 0.001. Recall
that we also choose
1
∆t = O(∆x/ max (u, v)) for time discretization. But the error formulae is 1

different from the above definition because there is not exact solution. Here we use maximum

61
 Table 4.9: Errors for straight line:Errors for P and h.

#Dof ∆x ||PI − P∆x ||∞ ||PI − P∆x ||A ||hI − h∆x ||∞
231 1.00e-01 3.03654e-03 9.18840e-03 5.72428e-06
861 5.00e-02 7.03826e-04 2.34927e-03 1.06895e-06
3,321 2.50e-02 1.30030e-04 3.83277e-04 2.43114e-07
12,880 1.25e-02 4.78007e-05 3.28391e-04 5.49985e-08
Rate 1.9 1.7 2.1

norm difference between successive approximations to compute the order of accuracy. That is,
max |P∆x −P∆x/2 |
log2 max P is the order for P in maximum norm. Similarly, we choose h instead of
| ∆x/2 −P∆x/4 |
P to get the convergence rate. Here we need to calculate the approximate curvature since we
do not know the exact solution. The idea is the same as computing the velocity. We use cubic
polynomial to fit the curve and then take the first and second derivative in the x direction. Table
4.10 demonstrates that the order of spatial accuracy in maximum norm for both P and h is 2.

Table 4.10: Errors for straight line h = e4t : Order of accuracy with maximum norm for P and h.

#Dof ∆x ||PI − P∆x ||∞ ||hI − h∆x ||∞

544 0.25 2.0748 2.7105
2,112 0.125 2.0155 2.6449
8,320 0.0625 1.9553 2.3873
33,024 0.03125 1.9603 1.7226

4.2 Tumor Growth Problem

4.2.1 Tumor Growth Model

In this section, we apply our mesh algorithm to simulate tumor growth phenomenon. Based on the
biological reality, we assume that the tumor occupies the region Ω− and the healthy tissue stays
the outside Ω+ = Ω\(Ω− ∪ Γ) with a free boundary Γ. The pressure and velocity of Ω+ are simply

62
assumed to be zero. In Ω− , the following equations govern the tumor growth: Find (u, p) such that

u = ∇p in Ω− , (4.9)

−∇ · u = G(n − A) in Ω− , (4.10)

p |Γ = κ, (4.11)

here n(x, t) is the nutrient (e.g. oxygen) concentration which satisfies

∇2 n − n = 0 in Ω− ,

n |Γ = 1,

u(x, t) is the velocity of the tumor, p(x, t) is the pressure, G and A are respectively the growth
and death rates without unit, and κ is the mean curvature on the boundary Γ [143]. To simulate the
tumor growth, we need to solve the following elliptic equation:

−∆p = G(n − A) in Ω− ,

p |Γ = κ,

The interface Γ is represented by a zero level-set function φ, that is, Γ(t) = {x | φ(x, t) = 0}. If
φ < 0, it is inside Ω− . If φ > 0, it is outside Ω+ . Suppose the tumor boundary is moving with
velocity u = ẋ(t) = ∇p. Then we have φ(x(t), t) = 0 by definition. Taking the derivative with
respect to time t on both sides, we get the following equation:

∂φ
+ u · ∇φ = 0. (4.12)
∂t

63
4.2.2 Numerical Scheme

The mesh generation step is the same as that for the tissue growth problem described in Section 4.1.
Then we get the mesh in Ω− for the initial boundary condition of φ. We could directly use the
boundary points on Γ to caculate the curvature using the way of our interface-fitted mesh. For
each point on Γ, we choose 15 nearby mesh nodes to approximate a circle. The radius of the fitted
circle is R. Here the curvature is signed. If one point in the circle is in Ω− , then the curvature is
1/R. Otherwise, the curvature is −1/R, which is negative. n and p could be solved using linear
finite element methods. The gradient of p is obtained using gradu.m and recovery.m based
on the MATLABr package iFEM [25]. φ is defined over the whole domain Ω. But the velocity
u = (px , py ) is only computed in Ω− . Then we extend each component of u from Γ to Ω+ using
harmonic extension. After this, we use WENO5 [115] and 3rd order Runge Kutta methods to
solve the transport equation to update φ for each time step ∆t = O(∆x/ max u). Then we apply
reinitialization step near the interface [104, 141]. The numerical scheme could be summarized as
follows:

Step 1: Generate a Cartesian mesh with size h. Based on the initial boundary condition of φ, we
get the interface-fitted mesh generated by Algorithm 1 in Chapter 2.

Step 2: Compute p, px , py , κ in Ω− .

Step 3: Extend px , py from Γ to Ω+ using harmonic extension with Dirichlet boundary condition.

Step 4: Solve the transport equation (4.12) on the Cartesian mesh in Step 1 to update φ using
WENO5 methods.

Step 5: Reinitialize φ near the interface to make it the sign distance function to Γ.

Step 6: Do Step 1 to Step 5 iteratively.

64
4.2.3 Tumor Growth Simulations

We compute the growth when G = 10, A = 0.5, and the initial shape of the tumor boundary
is (2 cos(θ), 2 sin(θ), θ ∈ [0, 2π]. The computational domain is chosen as Ω = [−6, 6]2 . The
evolution of a tumor is presented in Figure 4.7.

(a) iteration step = 1 (b) iteration step = 100

(c) iteration step = 300 (d) iteration step = 500

Figure 4.7: The interface(red) and the domain Ω− (green) when the Cartesian mesh size h is 0.125.

65
Bibliography

[1] G. Acosta and R. G. Durán. The maximum angle condition for mixed and nonconform-
ing elements: Application to the stokes equations. SIAM Journal on Numerical Analysis,
37(1):18–36, 1999.

[2] L. Adams and Z. Li. The immersed interface/multigrid methods for interface problems.
SIAM Journal on Scientific Computing, 24(2):463–479, 2002.

[3] Al Shenk. Uniform error estimates for certain narrow Lagrangian finite elements. Math.
Comp., 63(207):105–119, 1994.

[4] L. Antiga, J. Peiró, and D. A. Steinman. From image data to computational domains. In
Cardiovascular Mathematics, pages 123–175. Springer, 2009.

[5] I. Babuška. The Finite Element Method for Elliptic Equations with Discontinuous Coeffi-
cients. Computing, 5(3):207–213, 1970.

[6] I. Babuška and A. K. Aziz. On the angle condition in the finite element method. SIAM
Journal on Numerical Analysis, 13(2):214–226, 1976.

[7] E. Bänsch, F. Haußer, O. Lakkis, B. Li, and A. Voigt. Finite element method for epi-
taxial growth with attachment–detachment kinetics. Journal of Computational Physics,
194(2):409–434, 2004.

[8] J. W. Barrett and C. M. Elliott. Fitted and Unfitted Finite-Element Methods for Elliptic
Equatiosn with Smooth Interfaces. IMA Journal of Numerical Analysis, 7:283–300, 1987.

[9] R. Becker, E. Burman, and P. Hansbo. A Nitsche extended finite element method for incom-
pressible elasticity with discontinuous modulus of elasticity. Computer Methods in Applied
Mechanics and Engineering, 198(41-44):3352–3360, 2009.

[10] L. Beirão da Veiga, F. Brezzi, A. Cangiani, G. Manzini, L. Marini, and A. Russo. Basic prin-
ciples of virtual element methods. Mathematical Models and Methods in Applied Sciences,
23(01):199–214, 2013.

[11] L. Beirao da Veiga, F. Brezzi, L. Marini, and A. Russo. The hitchhiker’s guide to the virtual
element method. Mathematical Models and Methods in Applied Sciences, 24(08):1541–
1573, 2014.

66
[12] B. Bejanov, J.-L. Guermond, and P. D. Minev. A grid-alignment finite element technique for
incompressible multicomponent flows. Journal of Computational Physics, 227(13):6473–
6489, 2008.
[13] P. A. Berthelsen. A decomposed immersed interface method for variable coefficient elliptic
equations with non-smooth and discontinuous solutions. Journal of Computational Physics,
197(1):364–386, 2004.
[14] C. Börgers. A triangulation algorithm for fast elliptic solvers based on domain imbedding.
SIAM journal on numerical analysis, 27(5):1187–1196, 1990.
[15] P. Bourke. Calculating the area and centroid of a polygon. 1988.
[16] J. H. Bramble and J. T. King. A finite element method for interface problems in domains
with smooth boundaries and interfaces. Advances in Computational Mathematics, 6(1):109–
138, 1996.
[17] J. Brandts, S. Korotov, M. Křı́žek, et al. A geometric toolbox for tetrahedral finite ele-
ment partitions. Efficient preconditioned solution methods for elliptic partial differential
equations, pages 103–122, 2011.
[18] K. Q. Brown. Voronoi diagrams from convex hulls. Information Processing Letters,
9(5):223–228, 1979.
[19] E. Burman, S. Claus, P. Hansbo, M. G. Larson, and A. Massing. CutFEM : Discretizing
geometry and partial differential equations. International Journal for Numerical Methods
in Engineering, 104(December 2014):472–501, 2015.
[20] E. Burman and P. Hansbo. Fictitious domain finite element methods using cut elements:
I. A stabilized Lagrange multiplier method. Computer Methods in Applied Mechanics and
Engineering, 199(41-44):2680–2686, 2010.
[21] E. Burman and P. Hansbo. Fictitious domain finite element methods using cut elements: II.
A stabilized Nitsche method. Applied Numerical Mathematics, 62(4):328–341, 2012.
[22] Y. C. Chang, T. Y. Hou, B. Merriman, and S. Osher. A Level Set Formulation of Eulerian
Interface Capturing Methods for Incompressible Fluid Flows. Journal of Computational
Physics, 124(2):449–464, Mar. 1996.
[23] D. Chen, Z. Chen, C. Chen, W. Geng, and G.-W. Wei. Mibpb: a software package for
electrostatic analysis. Journal of computational chemistry, 32(4):756–770, 2011.
[24] L. Chen. Superconvergence of tetrahedral linear finite elements. International Journal of
Numerical Analysis and Modeling, 3(3):273–282, 2006.
[25] L. Chen. ifem: an integrated finite element methods package in matlab. University of
California at Irvine, 2009.
[26] L. Chen. Programming of finite element methods in matlab. arXiv preprint
arXiv:1804.05156, 2018.

67
[27] L. Chen, H. Wei, and M. Wen. An interface-fitted mesh generator and virtual element
methods for elliptic interface problems. Journal of Computational Physics, 334:327–348,
2017.
[28] L. Chen and J. Xu. Optimal Delaunay triangulations. Journal of Computational Mathemat-
ics, 22(2):299–308, 2004.
[29] Z. Chen and J. Zou. Finite element methods and their convergence for elliptic and parabolic
interface problems. Numerische Mathematik, 79(2):175–202, 1998.
[30] S.-W. Cheng, T. K. Dey, H. Edelsbrunner, M. A. Facello, and S.-H. Teng. Silver exudation.
J. ACM, 47(5):883–904, Sept. 2000.
[31] L. P. Chew. Guaranteed-quality delaunay meshing in 3d (short version). In Proceedings of
the thirteenth annual symposium on Computational geometry, pages 391–393. ACM, 1997.
[32] M. H. Cho, H. G. Choi, and J. Y. Yoo. A direct reinitialization approach of level-set/splitting
finite element method for simulating incompressible two-phase flows. International Journal
for Numerical Methods in Fluids, 67(11):1637–1654, Dec. 2011.
[33] C. C. Chu, I. G. Graham, and T. Y. Hou. A new multiscale finite element method for high-
contrast elliptic interface problems. Math. Comp, 79:1915–1955, 2010.
[34] M. Cruchaga, D. Celentano, and T. Tezduyar. A moving Lagrangian interface technique
for flow computations over fixed meshes. Computer Methods in Applied Mechanics and
Engineering, 191(6-7):525–543, Dec. 2001.
[35] F. Dassi, S. Perotto, L. Formaggia, and P. Ruffo. Efficient geometric reconstruction of
complex geological structures. Mathematics and Computers in Simulation, 106:163–184,
2014.
[36] T. S. Developers. SageMath, the Sage Mathematics Software System (Version 7.2), 2016.
http://www.sagemath.org.
[37] R. G. Duran. Error estimates for 3–D narrow finite elements. Math. Comp., 68(225):187–
199, 1999.
[38] R. G. Durán and A. L. Lombardi. Error estimates for the Raviart-Thomas interpolation
under the maximum angle condition. SIAM Journal on Numerical Analysis, 46(3):1442–
1453, 2008.
[39] H. Edelsbrunner. Triangulations and meshes in computational geometry. Acta Numerica
2000, 9:133–213, jan 2000.
[40] H. Edelsbrunner. Triangulations and meshes in computational geometry. Acta Numerica
2000, 9:133–213, 2000.
[41] H. Edelsbrunner, X.-Y. Li, G. Miller, A. Stathopoulos, D. Talmor, S.-H. Teng, A. Üngör,
and N. Walkington. Smoothing and cleaning up slivers. In Proceedings of the 32nd Annual
ACM Symposium on the Theory of Computing, pages 273–277. ACM, 2000.

68
[42] H. Edelsbrunner and R. Seidel. Voronoi diagrams and arrangements. Discrete & Computa-
tional Geometry, 1(1):25–44, 1986.

[43] T.-P. Fries and T. Belytschko. The intrinsic xfem: a method for arbitrary discontinuities
without additional unknowns. International journal for numerical methods in engineering,
68(13):1358–1385, 2006.

[44] T.-P. Fries and T. Belytschko. The extended/generalized finite element method: An overview
of the method and its applications. International Journal for Numerical Methods in Engi-
neering, 84(April):253–304, 2010.

[45] P. Frolkovič. Flux-based method of characteristics for contaminant transport in flowing

groundwater. Computing and Visualization in Science, 5(2):73–83, 2002.

[46] P. Frolkovič and K. Mikula. High-Resolution Flux-Based Level Set Method. SIAM Journal
on Scientific Computing, 29(2):579–597, Mar. 2007.

[47] S. FUJIMA, Y. FUKASAWA, and M. TABATA. Finite element formulation of periodic

conditions and numerical observation of three-dimensional behavior in a flow. 1993.

[48] J. Glimm, J. W. Grove, X. L. Li, and N. Zhao. Simple front tracking, volume 238 of Con-
temporary Mathematics. American Mathematical Society, Providence, Rhode Island, 1999.

[49] Y. Gong, B. Li, and Z. Li. Immersed-interface finite-element methods for elliptic interface
problems with nonhomogeneous jump conditions. SIAM Journal on Numerical Analysis,
46(1):472–495, 2008.

[50] S. Goswami, A. Gillette, and C. Bajaj. Efficient delaunay mesh generation from sampled
scalar functions. In Proceedings of the 16th International Meshing Roundtable, pages 495–
512. Springer, 2008.

[51] G. Guyomarc’h, C.-O. Lee, and K. Jeon. A discontinuous galerkin method for elliptic inter-
face problems with application to electroporation. Communications in numerical methods
in engineering, 25(10):991–1008, 2009.

[52] J. Guzman, M. Sánchez, and M. Sarkis. On the accuracy of finite element approximations
to a class of interface problems. Mathematics of Computation, 2015.

[53] A. Hansbo and P. Hansbo. An unfitted finite element method, based on Nitsche’s method,
for elliptic interface problems. Computer Methods in Applied Mechanics and Engineering,
191(47-48):5537–5552, 2002.

[54] A. Hansbo and P. Hansbo. A finite element method for the simulation of strong and weak
discontinuities in solid mechanics. Comput. Methods Appl. Mech. Engrg., 193:3523–3540,
2004.

[55] P. Hansbo, M. G. Larson, and S. Zahedi. A cut finite element method for a Stokes interface
problem. Applied Numerical Mathematics, 85:90–114, 2014.

69
[56] F. H. Harlow and J. E. Welch. Numerical calculation of time-dependent viscous incom-
pressible flow of fluid with free surface. Physics of fluids, 1965.

[57] X. He, T. Lin, and Y. Lin. Immersed finite element methods for elliptic interface problems
with non-homogeneous jump conditions. Int. J. Numer. Anal. Model, 8(2):284–301, 2011.

[58] J. S. Hesthaven and T. Warburton. Nodal discontinuous Galerkin methods: algorithms,

analysis, and applications. Springer Science & Business Media, 2007.

[59] J. S. Hesthaven and T. Warburton. Nodal Discontinuous Galerkin Methods, volume 54 of

Texts in Applied Mathematics. Springer New York, New York, NY, 2008.

[60] C. W. Hirt and B. D. Nichols. Volume of fluid (VOF) method for the dynamics of free
boundaries. Journal of Computational Physics, 39(1):201–225, Jan. 1981.

[61] S. Hou and X.-D. Liu. A numerical method for solving variable coefficient elliptic equation
with interfaces. Journal of Computational Physics, 202(2):411–445, 2005.

[62] S. Hou, P. Song, L. Wang, and H. Zhao. A weak formulation for solving elliptic interface
problems without body fitted grid. Journal of Computational Physics, 249:80–95, 2013.

[63] T. Y. Hou, Z. Li, S. Osher, and H. Zhao. A hybrid method for moving interface problems
with application to the hele–shaw flow. Journal of Computational Physics, 134(2):236–252,
1997.

[64] T. Y. Hou, Z. L. Li, S. Osher, and H. K. Zhao. A hybrid method for moving interface
problems with application to the Hele-Shaw flow. Journal of Computational Physics,
134(2):236–252, 1997.

[65] J. Huang and J. Zou. Some new a priori estimates for second-order elliptic and parabolic
interface problems. Journal of Differential Equations, 184(2):570–586, 2002.

[66] J. Huang and J. Zou. Uniform A Priori Estimates for Elliptic and Static Maxwell Interface
Problems. Discrete and Continuous Dynamic Systems-Series B, 7(1):145–170, 2007.

[67] J.-S. Huh and J. A. Sethian. Exact subgrid interface correction schemes for elliptic interface
problems. Proc. Natl. Acad. Sci., 105:9874–9879, 2008.

[68] L. N. T. Huynh, N. C. Nguyen, J. Peraire, and B. C. Khoo. A high-order hybridizable

discontinuous Galerkin method for elliptic interface problems. International Journal for
Numerical Methods in Engineering, 93(June 2012):183–200, 2013.

[69] G. Iaccarino and R. Verzicco. Immersed boundary technique for turbulent flow simulations.
Applied Mechanics Reviews, 56(3):331–347, 2003.

[70] H. Ji and J. Dolbow. On strategies for enforcing interfacial constraints and evaluating jump
conditions with the extended finite element method. International Journal for Numerical
Methods in Engineering, 61(14):2508–2535, 2004.

70
[71] A. Johansson and M. G. Larson. A high order discontinuous Galerkin Nitsche method for
elliptic problems with fictitious boundary. Numerische Mathematik, 123:607–628, 2013.

[72] R. Kafafy, T. Lin, Y. Lin, and J. Wang. Three-dimensional immersed finite element methods
for electric field simulation in composite materials. International journal for numerical
methods in engineering, 64(7):940–972, 2005.

[73] B. Khoo, Z. Li, and P. Lin. Interface Problems and Methods in Biological and Physical
Flows. World Scientific, 2009.

[74] G. Kossioris, C. Makridakis, and P. E. Souganidis. Finite volume schemes for Hamil-
ton–Jacobi equations. Numerische Mathematik, 83(3):427–442, 1999.

[75] M. Krizek. On the maximum angle condition for linear tetrahedral elements. SIAM Journal
on Numerical Analysis, 29(2):513–520, 1992.

[76] M. Krizek. On the maximum angle condition for linear tetrahedral elements. SIAM journal
on numerical analysis, 29(2):513–520, 1992.

[77] D.-T. Lee and B. J. Schachter. Two algorithms for constructing a delaunay triangulation.
International Journal of Computer & Information Sciences, 9(3):219–242, 1980.

[78] R. J. Leveque and Z. Li. The immersed interface method for elliptic equations with
discontinuous coefficients and singular sources. SIAM Journal on Numerical Analysis,
31(4):1019–1044, 1994.

[79] R. J. Leveque and Z. Li. Immersed Interface Methods for Stokes Flow with Elastic Bound-
aries or Surface Tension. SIAM J. Sci. Comput., 18(3):709–735, 1997.

[80] J. Li, J. M. Melenk, B. Wohlmuth, and J. Zou. Optimal Convergence of Higher Order
Finite Element Methods for Elliptic Interface Problems. Applied Numerical Mathematics,
60:19–37, 2010.

[81] X.-Y. Li and S.-H. Teng. Generating well-shaped delaunay meshed in 3d. In Proceedings
of the Twelfth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’01, pages
28–37, Philadelphia, PA, USA, 2001. Society for Industrial and Applied Mathematics.

[82] Z. Li. Immersed interface methods for moving interface problems. Numerical Algorithms,
14(4):269–293, 1997.

[83] Z. Li and K. Ito. The immersed interface method: numerical solutions of PDEs involving
interfaces and irregular domains, volume 33. SIAM, 2006.

[84] Z. Li, T. Lin, and X. Wu. New Cartesian grid methods for interface problems using the finite
element formulation. Numer. Math., 96(1):61–98, 2003.

[85] Z. Li, T. Lin, and X. Wu. New cartesian grid methods for interface problems using the finite
element formulation. Numerische Mathematik, 96(1):61–98, 2003.

71
[86] T. Lin, Y. Lin, and X. Zhang. Partially penalized immersed finite element methods for
elliptic interface problems. SIAM Journal on Numerical Analysis, 53(2):1121–1144, 2015.

[87] X.-D. Liu and T. C. Sideris. Convergence of the ghost fluid method for elliptic equations
with interfaces. Math. Comp.., 72(244):1731–1746, 2003.

[88] R. Löhner, J. R. Cebral, F. E. Camelli, S. Appanaboyina, J. D. Baum, E. L. Mestreau, and

O. A. Soto. Adaptive embedded and immersed unstructured grid techniques. Computer
Methods in Applied Mechanics and Engineering, 197(25):2173–2197, 2008.

[89] P. Macklin and J. S. Lowengrub. A New Ghost Cell/Level Set Method for Moving Boundary
Problems: Application to Tumor Growth. Journal of Scientific Computing, 35(2-3):266–
299, 2008.

[90] E. Marchandise, P. Geuzaine, N. Chevaugeon, and J.-F. Remacle. A stabilized finite element
method using a discontinuous level set approach for the computation of bubble dynamics.
Journal of Computational Physics, 225(1):949–974, July 2007.

[91] E. Marchandise and J.-F. Remacle. A stabilized finite element method using a discontinuous
level set approach for solving two phase incompressible flows. Journal of Computational
Physics, 219(2):780–800, Dec. 2006.

[92] R. Massjung. An Unfitted Discontinuous Galerkin Method Applied to Elliptic Interface

Problems. SIAM J. Numeri. Analysis., 50(6):3134–3162, 2012.

[93] J. M. Melenk and I. Babuška. The partition of unity finite element method: basic theory and
applications. Computer methods in applied mechanics and engineering, 139(1):289–314,
1996.

[94] R. Mittal and G. Iaccarino. IMMERSED BOUNDARY METHODS. dx.doi.org, 37(1):239–

261, Jan. 2005.

[95] N. Moes, J. Dolbow, and T. Belytschko. A finite element method for crack growth
without remeshing. International Journal for Numerical Methods in Engineering,
150(February):131–150, 1999.

[96] N. Moës, J. Dolbow, and T. Belytschko. A finite element method for crack growth without
remeshing. Int. J. Numer. Meth. Eng, 46(1):131–150, 1999.

[97] R. Moore and S. Saigal. Eliminating slivers in three-dimensional finite element models.
CMES: Computer Modeling In Engineering & Sciences, 7(3):283–291, 2005.

[98] L. Mu, J. Wang, X. Ye, and S. Zhao. A new weak Galerkin finite element method for elliptic
interface problems. Journal of Computational Physics, 325:157–173, 2016.

[99] J. Nitsche. Uber ein Yariationsprinzip zur Liisung yon Dirichlet-Problemen bei Verwendung
yon Teilr˜iumen, die keinen Randbedingungen unterworfen sind. Abh. Math. Sem. Univ.
Hamburg, 36(2):9–15, 1971.

72
[100] E. Olsson, G. Kreiss, and S. Zahedi. A conservative level set method for two phase flow II.
Journal of Computational Physics, 225(1):785–807, July 2007.

[101] S. Osher and J. A. Sethian. Fronts propagating with curvature-dependent speed: Algorithms
based on Hamilton-Jacobi formulations. Journal of Computational Physics, 79(1):12–49,
Nov. 1988.

[102] S. Osher and J. A. Sethian. Fronts propagating with curvature-dependent speed: algorithms
based on hamilton-jacobi formulations. Journal of computational physics, 79(1):12–49,
1988.

[103] J. Ovadia and Q. Nie. Numerical methods for two-dimensional stem cell tissue growth.
Journal of scientific computing, 58(1):149–175, 2014.

[104] D. Peng, B. Merriman, S. Osher, H. Zhao, and M. Kang. A pde-based fast local level set
method. Journal of computational physics, 155(2):410–438, 1999.

[105] P.-O. Persson and G. Strang. A simple mesh generator in matlab. SIAM review, 46(2):329–
345, 2004.

[106] C. S. Peskin. Numerical analysis of blood flow in the heart. Journal of computational
physics, 25(3):220–252, 1977.

[107] C. S. Peskin. The immersed boundary method. Acta numerica, 11:479–517, 2002.

[108] C. S. Peskin. The immersed boundary method. Acta Numer., 11:479–517, 2002.

[109] C. Pflaum. Subdivision of boundary cells in 3d, 2000.

[110] C. Pflaum. Semi-unstructured grids. computing, 67(2):141–166, 2001.

[111] S. B. Pillapakkam and P. Singh. A Level-Set Method for Computing Solutions to Viscoelas-
tic Two-Phase Flow. Journal of Computational Physics, 174(2):552–578, Dec. 2001.

[112] J. Qian, Y.-T. Zhang, and H.-K. Zhao. Fast sweeping methods for eikonal equations on
triangular meshes. SIAM Journal on Numerical Analysis, 45(1):83–107, 2007.

[113] R. Rangarajan and A. J. Lew. Universal meshes: A new paradigm for computing with
nonconforming triangulations. arXiv preprint arXiv:1201.4903, 2012.

[114] J. A. Sethian. Level set methods and fast marching methods: evolving interfaces in com-
putational geometry, fluid mechanics, computer vision, and materials science, volume 3.
Cambridge university press, 1999.

[115] C. Shu. WENO methods. Scholarpedia, 6(5):9709, 2011. revision #91939.

[116] A. Smolianski. Finite-element/level-set/operator-splitting (FELSOS) approach for comput-

ing two-fluid unsteady flows with free moving interfaces. International Journal for Numer-
ical Methods in Fluids, 48(3):231–269, May 2005.

73
[117] J. Strain. Tree Methods for Moving Interfaces. Journal of Computational Physics,
151(2):616–648, May 1999.

[118] G. Strang and G. J. Fix. An analysis of the finite element method, volume 212. Prentice-Hall
Englewood Cliffs, NJ, 1973.

[119] M. Sussman and E. Fatemi. An Efficient, Interface-Preserving Level Set Redistancing Al-
gorithm and Its Application to Interfacial Incompressible Fluid Flow. SIAM Journal on
Scientific Computing, 20(4):1165–1191, July 2006.

[120] M. Sussman, P. Smereka, and S. Osher. A Level Set Approach for Computing Solutions to
Incompressible Two-Phase Flow. Journal of Computational Physics, 114(1):146–159, Sept.
1994.

[121] O. J. Sutton. The virtual element method in 50 lines of matlab. Numerical Algorithms,
75(4):1141–1159, 2017.

[122] C. Talischi, G. H. Paulino, A. Pereira, and I. F. Menezes. Polymesher: a general-purpose

mesh generator for polygonal elements written in matlab. Structural and Multidisciplinary
Optimization, 45(3):309–328, 2012.

[123] A.-K. Tornberg and B. Engquist. A finite element based level-set method for multiphase
flow applications. Computing and Visualization in Science, 3(1-2):93–101, 2000.

[124] G. Tryggvason, B. Bunner, A. Esmaeeli, D. Juric, N. Al-Rawahi, W. Tauber, J. Han, S. Nas,

and Y. J. Jan. A Front-Tracking Method for the Computations of Multiphase Flow. Journal
of Computational Physics, 169(2):708–759, May 2001.

[125] S. P. Van der Pijl, A. Segal, C. Vuik, and P. Wesseling. A mass-conserving Level-Set method
for modelling of multi-phase flows. International Journal for Numerical Methods in Fluids,
47(4):339–361, Feb. 2005.

[126] S. P. Van der Pijl, A. Segal, C. Vuik, and P. Wesseling. Computing three-dimensional two-
phase flows with a mass-conserving level set method. Computing and Visualization in Sci-
ence, 11(4-6):221–235, 2008.

[127] E. Wadbro, S. Zahedi, and G. Kreiss. A uniformly well-conditioned, unfitted Nitsche

method for interface problems. BIT Numer Math, 53:791–820, 2013.

[128] F. Wang, Y. Xiao, and J. Xu. High-Order Extended Finite Element Methods for Solving
Interface Problems. ArXiv, pages 1–25, 2016.

[129] H. Wei, L. Chen, Y. Huang, and B. Zheng. Adaptive mesh refinement and superconver-
gence for two-dimensional interface problems. SIAM Journal on Scientific Computing,
36(4):A1478–A1499, 2014.

[130] J. E. Welch, F. H. Harlow, and J. P. Shannon. The MAC method, 1966.

74
[131] A. Wiegmann and K. P. Bube. The explicit-jump immersed interface method: finite dif-
ference methods for pdes with piecewise smooth solutions. SIAM Journal on Numerical
Analysis, 37(3):827–862, 2000.

[132] D. E. Womble. A Front-Tracking Method for Multiphase Free Boundary Problems. SIAM
Journal on Numerical Analysis, 26(2):380–396, July 2006.

[133] H. Wu and Y. Xiao. An unfitted hp-interface penalty finite element method for elliptic
interface problems. arXiv preprint arXiv:1007.2893, 2010.

[134] K. Xia, M. Zhan, and G.-W. Wei. MIB Galerkin method for elliptic interface problems.
Journal of Computational and Applied Mathematics, 272(7):195–220, 2014.

[135] H. Xie, Z. Li, and Z. Qiao. A finite element method for elasticity interface problems with
locally modified triangulations. International journal of numerical analysis and modeling,
8(2):189, 2011.

[136] J. Xu. Error estimates of the finite element method for the 2nd order elliptic equations with
discontinuous coefficients. J. Xiangtan University, 1:1–5, 1982.

[137] J. Xu. Estimate of the Convergence Rate of Finite Element Solutions to Elliptic Equations
of Second Order with Discontinuous Coefficients. Natural Science Journal of Xiangtan
University, 1(1):1–5, 1982.

[138] J.-J. Xu and H.-K. Zhao. An Eulerian Formulation for Solving Partial Differential Equations
Along a Moving Interface. Journal of Scientific Computing, 19(1-3):573–594, 2003.

[139] J. Yang and E. Balaras. An embedded-boundary formulation for large-eddy simulation of

turbulent flows interacting with moving boundaries. Journal of Computational Physics,
215(1):12–40, June 2006.

[140] S. Yu and G. W. Wei. Three-dimensional matched interface and boundary (MIB) method for
treating geometric singularities. Journal of Computational Physics, 227(1):602–632, 2007.

[141] K. Zhang, L. Zhang, H. Song, and D. Zhang. Reinitialization-free level set evolution via
reaction diffusion. IEEE Transactions on Image Processing, 22(1):258–271, 2013.

[142] X. Zhang. Nonconforming immersed finite element methods for interface problems. PhD
thesis, Virginia Polytechnic Institute and State University, 2013.

[143] X. Zheng and J. Lowengrub. An interface-fitted adaptive mesh method for elliptic problems
and its application in free interface problems with surface tension. Advances in Computa-
tional Mathematics, 42(5):1225–1257, 2016.

[144] Y. Zhou and T. H. North. Kinetic modelling of diffusion-controlled, two-phase moving inter-
face problems. Modelling and Simulation in Materials Science and Engineering, 1(4):505–
516, July 1993.

75
[145] Y. Zhou and G.-W. Wei. On the fictitious-domain and interpolation formulations of
the matched interface and boundary (mib) method. Journal of Computational Physics,
219(1):228–246, 2006.

[146] Y. Zhou, S. Zhao, M. Feig, and G.-W. Wei. High order matched interface and boundary
method for elliptic equations with discontinuous coefficients and singular sources. Journal
of Computational Physics, 213(1):1–30, 2006.

[147] Y. C. Zhou, S. Zhao, M. Feig, and G. W. Wei. High order matched interface and bound-
ary method for elliptic equations with discontinuous coefficients and singular sources. J.
Comput. Phys., 213(1):1–30, 2006.

76
Appendix A

Appendices

A.1 The Elliptic Interface Equation and Weak Formulation

Elliptic interface problems arise in many physical applications, including electromagnetism, fluids
dynamics, materials science, free boundary problems, and biological applications [49, 69, 114]. In
this work we study a typical elliptic interface problem which is described as piece-wisely defined
elliptic partial differential equations (PDE) in different regions which are coupled together with
non-homogeneous jump conditions, such as solution and flux jump across the interface.

A.1.1 The Elliptic Interface Equation

Ω is an open bounded domain in Rd , d = 2, 3 and Γ is a continuous simple connected interface

which separates the domain Ω into two disjoint regions Ω+ and Ω− . Ω+ denotes the exterior
domain and Ω− denotes the interior domain enclosed by Γ. The equation of the typical elliptic

77
interface problem is as follows:

−∇ · (β(x)∇u(x)) = f (x), x ∈ Ω\Γ (A.1)

with solution and flux jump conditions across the interface Γ:

[u]Γ = u+ − u− = q0 , [βun ]Γ = β + u+ − −
n − β un = q1 , u=g on ∂Ω, (A.2)

where un is the normal flux (∇u) · n and n is the unit norm direction of the interface Γ pointing
from Ω− to Ω+ (outwards). The superscripts + and − stand for the constraint of a function on Ω+
and Ω− , respectively. The coefficient β(x) is assumed to be uniformly positive and smooth on each
disjoint subdomain, but may be discontinuous across the interface Γ [27].

A.1.2 The Weak Formulation

Given functions q0 and q1 along the interface, we prescribe the non-homogeneous jump condi-
tions in Eq. (A.2). The flux jump q1 is imposed weakly in the weak formulation. Elliptic
interface problems with such jump conditions arise in many application fields. For example,
BurtonCabreraFrank-type model [7] solves the diffusion equation on terraces with jump condi-
tion across the interface representing atomic steps. Another example is that the reaction potential
model of electrostatics of a solvation energy has a non homogeneous flux jump condition [49].

Proposition A.1. The homogeneous jump conditions are q0 = 0 and q1 = 0 on Γ((A.2)), then the
problem (A.1)-(A.2) is to find u ∈ H 1 (Ω) with u = g on ∂Ω such that

Z Z
β∇u · ∇v dx = f vdx, ∀v ∈ H01 (Ω)
Ω Ω

Proof. Multiply the equation by a smooth test function v ∈ H01 (Ω) and integrate over the whole

78
domain Ω,

Z Z
−∇ · (β∇u)vdx = f vdx, ∀v ∈ H01 (Ω) (A.3)
Ω Ω

then use integration by parts

Z Z Z Z
−∇ · (β∇u)vdx = − β∇u · nvds + β∇u · ∇vdx = β∇u · ∇vdx (A.4)
Ω ∂Ω Ω Ω

Thus combining (A.3) and (A.4), we get

Z Z
β∇u · ∇vdx = f vdx, ∀v ∈ H01 (Ω)
Ω Ω

R
Note that ∂Ω
β∇u · nvds = 0, since v = 0 on ∂Ω and q1 = 0 on Γ.

Proposition A.2. If q0 = 0, q1 6= 0 on Γ, then the corresponding weak formulation is :

(β∇u, ∇v)Ω = (f, v)Ω − hq1 , viΓ , ∀v ∈ H01 (Ω) (A.5)

Proof. From (A.1), we could write in this way:

−∇ · (β − (x)∇u(x)) = f (x), x ∈ Ω− , −∇ · (β + (x)∇u(x)) = f (x), x ∈ Ω+

For the first equation, using integration by parts,

Z Z Z
− −
−β 4u · nvds + β ∇u · ∇vdx = f vdx
∂Ω− Ω− Ω−

for all v ∈ H01 (Ω). Since ∂Ω− = Γ, then

Z Z Z
− −
−β 4u · nvds + β ∇u · ∇vdx = f vdx. (A.6)
Γ Ω− Ω−

79
Similarly, for the second equation, using integration by parts,

Z Z Z
+ +
−β 4u · ñvds + β ∇u · ∇vdx = f vdx
∂Ω+ Ω+ Ω+

for all v ∈ H01 (Ω). Since ∂Ω+ = ∂Ω ∪ Γ, v = 0 on ∂Ω and ñ = −n, then

Z Z Z
+ +
β 4u · nvds + β ∇u · ∇vdx = f vdx. (A.7)
Γ Ω+ Ω+

Adding (A.6) and (A.7), we could get

Z Z
+ + − −
(β 4u − β 4u ) · nvds + β∇u · ∇vdx = (f, v)Ω , ∀v ∈ H01 (Ω)
Γ Ω

Since q1 = (β + 4u+ − β − 4u− ) · n, then

Z Z
q1 vds + β∇u · ∇vdx = (f, v)Ω , ∀v ∈ H01 (Ω)
Γ Ω

Removing the first term to the right hand side, then

(β∇u, ∇v)Ω = (f, v)Ω − hq1 , viΓ , ∀v ∈ H01 (Ω)

On the other direction, if we choose v ∈ H01 (Ω+ ) ,then we get

Z Z
+
β ∇u · ∇vdx = f vdx, ∀v ∈ H01 (Ω+ ).
Ω Ω+

Similarly, if we choose v ∈ H01 (Ω− ), then we get

Z Z
−
β ∇u · ∇vdx = f vdx, ∀v ∈ H01 (Ω− )
Ω Ω−

1
(β + 4u+ − β − 4u− ) · nvds =
R R
Finally, we get Γ
q vds
Γ 1
when choosing v ∈ H 2 (Γ).

80
Proposition A.3. If q0 6= 0 on Γ, we can find w− : Ω− → R with w− = q0 on ∂Ω− and w− ∈
H 1 (Ω− ). The zero extension of w− ,denote by w satisfies w ∈ H 1 (Ω− ∪ Ω+ ) with w = w− on Ω−
and w = 0 on Ω+ . The model is equivalent to: find u = p − w with p ∈ Hg1 (Ω) such that:

(β∇p, ∇v)Ω = (f, v)Ω − hq1 , viΓ + (β∇w, ∇v)Ω− , ∀v ∈ H01 (Ω) (A.8)

Proof. Since p satisfies [p]Γ = 0 and [βpn ]Γ = q1 + [βwn ]Γ , then (A.5) holds for p, that is

(β∇p, ∇v)Ω = (f − ∇ · (β∇w), v)Ω − hq1 + [βwn ]Γ , viΓ , ∀v ∈ H01 (Ω)

then

(β∇p, ∇v)Ω = (f, v)Ω − hq1 , viΓ − (∇ · (β∇w), v)Ω − h[βwn ]Γ , viΓ , ∀v ∈ H01 (Ω)

Since ∀v ∈ H01 (Ω)

Z Z
−(∇ · (β∇w), v)Ω = −(∇ · (β∇w), v)Ω− = − β∇w · n · vds + β∇u · ∇vdx,
Γ Ω−

and

Z
h[βwn ]Γ , viΓ = − β∇w · n · vds,
Γ

then we get

(β∇p, ∇v)Ω = (f, v)Ω − hq1 , viΓ + (β∇w, ∇v)Ω− , ∀v ∈ H01 (Ω)

81
T HEOREM A.4. The weak formulation of the elliptic interface equation ((A.1)-(A.2)) is stated as:
find u = p − w with p ∈ Hg1 (Ω) = {v ∈ H 1 (Ω) : v|∂Ω = g} such that

(β∇p, ∇v)Ω = (f, v)Ω − hq1 , viΓ + (β∇w− , ∇v)Ω− , ∀v ∈ H01 (Ω). (A.9)

A.2 Efficient Implementation of Virtual Element Methods in

Two Dimensions

In this section, we mainly present how to implement the linear virtual element method for the two
dimensional Poisson equation on a polygonal domain. In [121], it aims to short implementation
of algorithms for the education purpose. But the codes could be more efficient especially for the
large size of matrix. Here we focus on efficient implementation in Matlab.

A.2.1 Introduction

Virtual Element Method ( VEM ) [10] is developed by Brezzi’s group, which can be viewed as
an extension of Finite Element Methods to general polygonal and polyhedral elements. In two
dimensions, elements in finite element methods are either triangles or quadrilateral. But in VEM,
polygonal elements can be of very general shape, which may even be non convex. Theoretical
results and generalize matrix formulation are discussed in [10, 11]. [11]is about the coding. Here
we mainly focus on the efficient implementation of linear virtual element methods. At first, we
will introduce the concrete matrix formulation for the linear virtual element methods. Then we
will demonstrate how to implement each component of the matrix in the next section.

82
First we study the model problem of the two dimensional Poisson equation:


−∆u = f in Ω,

(A.10)

u = g on ∂Ω.


Here Ω is a polygonal domain in R2 , f ∈ L2 (Ω) and g ∈ H 1/2 (∂Ω). Define H01 (Ω) := {v ∈
H 1 (Ω) : v = 0 on ∂Ω} and Hg1 (Ω) := {v ∈ H 1 (Ω) : v = g on ∂Ω}. The weak formulation will
be derived using the same strategy of the classical finite element methods: find u ∈ Hg1 (Ω) such
that
a(u, v) = L(v), ∀v ∈ H01 (Ω),

R
where the bilinear form a(u, v) = (∇u, ∇v) = Ω ∇u · ∇v dx, the linear form L(v) = (f, v) =
f v dx and (·, ·) is the inner product in L2 . By Lax-Milgram Lemma, it has a unique solution.
R
Ω

Let Th be the collection of partitions of the domain Ω into polygonal elements with not self-
intersecting boundary. h denotes the maximum diameter of all elements in Th . In our numerical
results, h is approximated by the square root of the area of each element. A local finite-dimensional
vector space Vh (E) for a polygon E ∈ Th can be written as

Vh (E) := {v ∈ H 1 (E) : ∆v|E = 0, v|∂E is continuous and piecewise linear polynomial}.

Since a piecewise linear function will be uniquely determined by its value on vertices, dim Vh (E) =
nvE , where nvE is the number of vertices of E.

The global virtual element space is defined as

Vh = {vh ∈ H 1 (Ω) : vh |E ∈ Vh (E) for all E ∈ Th }.

N (Th ) is the set of vertices of mesh Th and N = |N (Th )| denotes the number of vertices. We
define the operator dof i from Vh to R as dof i (vh ) = vh (xi ), for a vertex xi ∈ N (Th ). The

83
canonical basis {φ1 , · · · , φN } ⊂ Vh is chosen as dof i (φj ) = δij , i, j = 1, · · · , N . And the nodal
interpolation Ih : C(Ω̄) → Vh is defined as Ih u = N
P
i=1 u(xi )φi and denoted by uI = Ih u.

In finite element methods, we use discrete solutions uh to approximate the solution u. The linear
virtual element approximation of (A.10) is: finding uh ∈ Vh ∩ Hg1 (Ω) such that:

(∇uh , ∇vh )Ω = (f, vh )Ω , ∀vh ∈ Vh ∩ H01 (Ω).

PN
Suppose uh = j=1 uj φj , by linearity, we have for i ∈ 1, · · · , N ,

N
X
( ∇φj , ∇φi )Ω uj = (f, φi )Ω . (A.11)
j=1

We denote the matrix Aij = (∇φj , ∇φi )Th and the vector b = (b1 , · · · , bN )t by bi = (f, φi )Ω .
Equation (A.11) is written in the matrix form as

Ah uh = b. (A.12)

Here Ah is an N × N matrix and uh = (u1h , · · · , uN t

h ) . As the matrix Ah is symmetric and positive

definite, the algebraic system (A.12) could be solved stably and efficiently by using algebraic
multigrid methods.

For finite element methods, it suffices to compute the local stiffness matrix in each element at first,
and then the matrix A is assembled by summing the contribution from each element. Therefore,
the major task is to compute (∇φj , ∇φi )E , where (·, ·)E is the L2 inner product on E.

In order to do so, we define some projection operators at first. In each polygonal element E, the
operator Π∇ : Vh (E) → P1 (E) is represented as the H 1 projection to P1 (E) space, i.e.,:

(∇pk , ∇Π∇ vh )E = (∇pk , ∇vh )E for all pk ∈ P1 (E).

84
Here P1 (E) is the space of linear polynomials. It is obvious that the above condition defines
Π∇ vh only up to a constant. This can be resolved by mapping a projection operator onto constants
P0 : Vh (E) → P0 (E) and requiring

P0 (Π∇ vh − vh ) = 0.

PnvE PnvE
One simple choice is P0 vh = i=1 vh (xi )/nvE = i=1 dof i (vh )/nvE .

We rewrite the basis function φi ∈ Vh (E) as Π∇ φi + (I − Π∇ )φi using the projection Π∇ , and then
split the entry of the local stiffness matrix as

(∇Π∇ φi , ∇Π∇ φj )E + (∇(I − Π∇ )φi , ∇(I − Π∇ )φj )E .

Again the second term is not computable since the basis φi is not known point-wise. Now we
replace by a so-called stabilization term S E (·, ·)

(∇Π∇ φi , ∇Π∇ φj )E + S E ((I − Π∇ )φi , (I − Π∇ )φj ).

A scaled l2 inner product was chosen in the stabilization term

nEv
X
∇ ∇
E
S ((I − Π )φi , (I − Π )φj ) = dof r ((I − Π∇ )φi ) dof r ((I − Π∇ )φj )
r=1

to satisfy the assumption of S E

c1 (∇v, ∇v) ≤ S E (v, v) ≤ c2 (∇v, ∇v), ∀v ∈ Vh (E) and Π∇ v = 0,

where some positive constants c1 and c2 are independent of E [10]. We write the explicit expression

85
of the local stiffness matrix of the virtual element method as follows:

v
nE
X
∇ ∇
(AE
h )ij := (∇Π φi , ∇Π φj )E + dof r ((I − Π∇ )φi ) dof r ((I − Π∇ )φj ). (A.13)
r=1

The matrix representation of the operator Π∇ is computed using the following concrete formulae.
xE = (xE , yE ) denotes the centroid of E. A scaled monomial basis of P1 (E) is chosen as m1 =
1, m2 = (x − xE )/hE , m3 = (y − yE )/hE , where hE = |E|1/2 , which is approximated by the
square root of the area of each element E. Let the matrix G3×3 be defined as

v v
 
  nE nE
1
P xi −xE 1
P yi −yE
 P0 m 1 P0 m2 P0 m3   1 nvE hE nvE hE 
 
  i=1 i=1 
G := 
 0 =
(∇m2 , ∇m2 )0,E (∇m3 , ∇m2 )0,E 

0 1 0 
  



0 (∇m2 , ∇m3 )0,E (∇m3 , ∇m3 )0,E 0 0 1

Here B 3×nvE is a matrix defined as:

 
 P0 φ1 ··· P0 φnvE 
 
B := (∇m 2 , ∇φ1 ) E · · · (∇m2 , ∇φ nv )E
.
 E 
 
(∇m3 , ∇φ1 )E · · · (∇m3 , ∇φnvE )E

The first row of B is P0 φ1 = P0 φ2 = . . . = P0 φnvE = 1/nvE . For the other components

R R ∂m
(∇mj , ∇φi )E , j = 2, 3, we have (∇mj , ∇φi )E = − E ∆mj φi + ∂E ∂nj φi by integration by
parts. Here the first term is zero as ∆mj = 0 for linear polynomials and the second term need to
be only computed.

nje
Z P
∂mj node i∈edge e
φi = , (A.14)
∂E ∂n 2hE

where n = (nxe , nye ) = (n2e , n3e ) denotes an outward unit normal direction to the edge e.

86
In order to compute the stabilization term, one more matrix D nvE ×3 is needed.

 
hE x1 − xE y1 − yE 
 
  hE x2 − xE y 2 − y E

−1 
D := dof i (mj ) = hE  ,

· · · ··· ··· 
 
 
hE xnvE − xE ynvE − yE

where (xi , yi ), i = 1, · · · , nvE mean the vertices in each polygon E.

By definition,

3
X
∇
Π vh = sα mα (A.15)
α=1

and the coefficients (sα ) are determined based on the following linear systems

(∇mα , ∇(Π∇ vh − vh ))E = 0 α = 1, . . . , 3. (A.16)

The matrix representation of Π∇ : Vh (E) → P1 (E) relative to the basis (mα ) is Π∇ = G−1 B.

Furthermore, we also need the matrix representation of Π∇ in the canonical basis {φi }. Let
PnvE ∇
Π∇ φi = ∇ v
j=1 dof j (Π φi )φj , i = 1, · · · , nE , then the matrix representation Π∗ of the opera-

tor Π∇ : Vh (E) → Vh (E) in the canonical basis is written as Π∇ −1 ∇

∗ = DG B = DΠ .

Finally the matrix formulation of AE

h could be given by

∇ T ∇ ∇ T ∇
AE
h = [Π ] G̃Π + [I − Π∗ ] [I − Π∗ ],

where G̃ is the same with G except that the elements in the first row are all zero.Then it is easily
to get the local stiffness matrices (A− +
h |E ) and (Ah |E ).

For the right hand side vector bi in (A.12), we simply approximate f by a piecewise constant and

87
approximate
X X
(f, φi )Ω = (f, φi )E ≈ |E|f (xE , yE )/nvE .
E∈Th E∈Th

A.2.2 Implementation

In this subsection, we discuss how to implement the code efficiently in Matlab. In [121], it shows
how to implement the code in Matlab based on the matrix form G, B and D, but it is not efficient
for two reasons. One reason is that there are large for loops. Vectorization should be taken into
account to avoid for loop as much as possible. Another one is that it does not take advantage of
sparsity. Because the stiffness matrix we obtained via the discretization is very sparse. We shall
think about using sparse matrix algorithms, which require less computational time and computer
memory [26].

The polygonal mesh we used is generated from PolyMesher.m, which is written in MATLAB [122].
The output contains a matrix named node which represents the coordinates of vertices and a cell
array named element which records the vertices of each element with a counter-clockwise order.
Given node and element, the stiffness matrix A could be implemented as follows:

function A = assemblematrix(node,element)
N = size(node,1);
Nelement = size(element,1);
% initialize large enough array
ii = zeros(Nelement*49,1);
jj = zeros(Nelement*49,1);
ss = zeros(Nelement*49,1);
curIdx = 1; % current index
for t = 1:Nelement
v = element{t};
Nv = length(v);
x1 = node(v,1);% the coordinate of node

88
 y1 = node(v,2);
x2 = circshift(x1,-1);
y2 = circshift(y1,-1);
bdIntegral = x1.*y2 - y1.*x2;
area = sum(bdIntegral)/2;% the area
h = sqrt(abs(area));
cx = sum((x1+x2).*bdIntegral)/(6*area); % the centroid
cy = sum((y1+y2).*bdIntegral)/(6*area);
edgeVecx = x2 - x1; %v(k-1)-v(k) for x (-x)
edgeVecy = y2 - y1; %v(k-1)-v(k) for y (-y)
normVecx = edgeVecy; % the first normal derivative (-y)
normVecy = -edgeVecx; % the second normal derivative (x)
H = zeros(2,Nv);
H(1,:) = (normVecx + circshift(normVecx,1))/(2*h);
H(2,:) = (normVecy + circshift(normVecy,1))/(2*h);
% Part 1: matrix of projections
A1 = H’*H;
% Part 2: matrix of difference
D = ones(Nv,3);
D(:,2) = (x1 - cx)’./h; % cx, cy, h computed before
D(:,3) = (y1 - cy)’./h;
c1 = (ones(1,Nv) - sum(D(:,2:3))*H)/Nv;
GinvB = [c1; H];
IminusP = eye(Nv)- D*GinvB;
A2 = IminusP’*IminusP;
% Record nonzero entries (i,j,s) for the sparse matrix
[loci,locj,s] = find(A1 + A2);
nnz = length(s);
ii(curIdx:curIdx+nnz-1) = v(loci); % change to global index
jj(curIdx:curIdx+nnz-1) = v(locj);
ss(curIdx:curIdx+nnz-1) = s;
curIdx = curIdx + nnz;
end

89
A = sparse(ii(1:curIdx-1),jj(1:curIdx-1),ss(1:curIdx-1),N,N);

In the subroutine assemblematrix, we use build in function sparse to form the sparse matrix
A for the sake of memory. In this way, we avoid updating a sparse matrix inside a large loop. It
is much faster than the code in [121]. Because the stiffness matrix in [121] is stored as a full
matrix. It will be out of memory for a large number N , where N is the number of nodes. The code
in lines 1 − 8 is some initialization step. After line 9, we will get the vectorization form of the
stiffness matrix for each element. The number of vertices for each element denotes N v in the code.
circshift is used for getting the neighbor for each node. Given node and element, it is easy

to get the area, the diameter and centroid of each element. The centroid of a non-self-intersecting
closed polygon defined by N v vertices (x1 , y1 ), . . . , (xN v , yN v ) is the point (Cx , Cy ), then

Nv
1 X
Cx = (xi + xi+1 )(xi yi+1 − xi+1 yi )
6E i=1
Nv
1 X
Cy = (yi + yi+1 )(xi yi+1 − xi+1 yi )
6E i=1

1
PN v
where E denotes the polygon’s signed area and is calculated using the formula E = 2 i=1 (xi yi+1 −
xi+1 yi ) [15]. The diameter of each element h could be approximated by the square root of the
area. From lines 10−24, we get the area, diameter, centroid and normal direction for each element.
The local stiffness matrix is given in lines 25 − 37. For the first part in Equation (A.13), we can
compute accurately. For each basis function ϕi , we define sαi as the coefficients of Π∇ ϕi in the
basis mα :

3
X
Π∇ ϕi = sαi mα , i = 1, . . . , 3. (A.17)
α=1

90
 x−Cx y−Cy
where m1 = 1, m2 = h
, m3 = h
. From equation (A.17) we have

3 X
X 3 3 X
X 3
∇ ∇
(∇Π ϕi , ∇Π ϕj )0,E = sαi sβj (∇mα , ∇mβ )0,E = sαi sβj .
α=1 β=1 α=2 β=2

Note that (∇m1 , ∇m1 )E = 0 and (∇mα , ∇mβ )E = 1, α = 2, 3, β = 2, 3, then the second equality
holds. In addition, we can get

1 1
Z
∂m2
s2i = (∇m2 , ∇ϕi )E = ϕi = (n (ei ) + n1 (ei+1 ))
∂E ∂n 2h

from Eq.(A.14) and similarly,

1 2
Z
∂m3
s3i = (∇m3 , ∇ϕi )E = ϕi = (n (ei ) + n2 (ei+1 )),
∂E ∂n 2h

where n = (n1 , n2 ) is the normal direction for each edge and the intersection of two edges ei and
ei+1 is the vertex i.

Combing equation (A.15) and (A.17), we get sα = (sα1 , . . . , sαN v ), α = 1, . . . , 3. Therefore we

define a matrix H with size 2 × N v such that H(1, :) = s2 , H(2, :) = s3 , then (∇Π∇ φi , ∇Π∇ φj )E
could be written as [H T H]ij .

PN v
For the second part in Eq. (A.13), we know r=1 dofr ((I − Π∇ )ϕi )dofr ((I − Π∇ )ϕj ) = [(I −
Π∇ )T (I − Π∇ )]ij , where Π∇ = DG−1 B,matrix D is N v × 3 and G−1 B is 3 × N v. It is straight-
forward to implement matrix D in code. The first column of D is 1, the second column of D
is the value of m2 in different vertices and the third column of D is the value of m3 in differ-
ent vertices. Finally we need to compute G−1 B. Based on deep understanding each component
of the matrices formed from Eq. (A.16), we could rewrite G−1 B = [s1 ; s2 ; s3 ] = [s1 ; H]. The
only thing is solving s1 . By the definition of P0 and P0 vh = N
P v
i=1 dof i (vh )/N v,then we obtain

(Π ϕi )(V ) = ϕi (V ), finally N
P ∇ P P v 1 2 3
i=1 (si +si m2 (Vi )+si m3 (Vi )) = 1 from equation (A.17),that

is, each component of s1 could be s1i = (1 − (s2i m2 (Vi ) + s3i m3 (Vi )))/N v, see line 34. Then a list

91
of index and nonzero entries of local stiffness matrix are recorded in line 38 − 45, which are used
for updating a sparse matrix outsider the inner for loop. The global stiffness matrix is stored as
sparse matrix in order to save the memory.

But there exists a large for loop in the subroutine assemblematrix. When the number of
elements is large, it can quickly add significant overhead. We then use the vectorization technique
to avoid the large for loop for the sake of efficiency. Since the length of each element is different,
the vectorization could also be used with the same length of the elements. Here is the most efficient
code for the assembling of stiffness matrix in two dimensions.

function A = assemblevem(node,elem)
%% Assemble the matrix equation
N = size(node,1);
elemVertexNumber = cellfun(’length’,elem);
nnz = sum(elemVertexNumber.ˆ2);
ii = zeros(nnz,1); %initialization
jj = zeros(nnz,1);
ss = zeros(nnz,1);
index = 0;
for Nv = min(elemVertexNumber):max(elemVertexNumber)
% find polygons with Nv vertices
idx = find(elemVertexNumber == Nv); % index of elements
NT = length(idx); % the # of elements
% vertex index and coordinates
vertex = cell2mat(elem(idx));
x1 = reshape(node(vertex,1),NT,Nv);
y1 = reshape(node(vertex,2),NT,Nv);
x2 = circshift(x1,[0,-1]);
y2 = circshift(y1,[0,-1]);
% Compute geometry quantity: edge, normal, area, center
bdIntegral = x1.*y2 - y1.*x2;
area = sum(bdIntegral,2)/2; % the area per element

92
 h = repmat(sqrt(abs(area)),1,Nv); % h = sqrt(area)
cx = sum((x1+x2).*bdIntegral,2)./(6*area); % the centroid
cy = sum((y1+y2).*bdIntegral,2)./(6*area); % the centroid
normVecx = y2 - y1; % normal vector
normVecy = x1 - x2;
% matrix B, D, I - P
Bx = (normVecx + circshift(normVecx,[0,1]))./(2*h);
By = (normVecy + circshift(normVecy,[0,1]))./(2*h);
Dx = (x1 - repmat(cx,1,Nv))./h; % m(x) = (x - cx)/h
Dy = (y1 - repmat(cy,1,Nv))./h;
c1 = (1 - (repmat(sum(Dx,2),1,Nv).*Bx + repmat(sum(Dy,2),1,Nv).*By))/Nv;
IminusP = zeros(NT,Nv,Nv);
for i = 1:Nv
for j = 1:Nv
IminusP(:,i,j) = - c1(:,j) - Dx(:,i).*Bx(:,j) - Dy(:,i).*By(:,j);
end
IminusP(:,i,i) = ones(NT,1) + IminusP(:,i,i);
end
% assemble the matrix
for i = 1:Nv
for j = 1:Nv
ii(index+1:index+NT) = vertex(:,i);
jj(index+1:index+NT) = vertex(:,j);
ss(index+1:index+NT) = Bx(:,i).*Bx(:,j) + By(:,i).*By(:,j) + dot(
IminusP(:,:,i),IminusP(:,:,j),2);
index = index + NT;
end
end
end
A = sparse(ii,jj,ss,N,N);

The out for loop goes over elements with the same the number of vertices,which is a big improve-

93
ment compared with all the elements. The vectorization code is similar to assemblematrix. The
major difference is vectorizing the matrix I − DG−1 B. We introduce three dimensional array to
store, see lines 34 − 40.

For the right hand side b, we have

X
b≈ |E|f (Cx , Cy )/N v.
E∈Th

Here we use the build in command accumarray in Matlab to avoid the for loop over all elements.
The following codes are the whole assembling procedure:

function [u,A] = PoissonVEM(node,elem,pde)

%% Assemble the matrix equation
N = size(node,1);
elemVertexNumber = cellfun(’length’,elem);
nnz = sum(elemVertexNumber.ˆ2);
ii = zeros(nnz,1); %initialization
jj = zeros(nnz,1);
ss = zeros(nnz,1);
b = zeros(N,1);
edge = zeros(sum(elemVertexNumber),2);
index = 0;
edgeIdx = 1;
for Nv = min(elemVertexNumber):max(elemVertexNumber)
% find polygons with Nv vertices
idx = find(elemVertexNumber == Nv); % index of elements
NT = length(idx); % the # of elements
% vertex index and coordinates
vertex = cell2mat(elem(idx));
x1 = reshape(node(vertex,1),NT,Nv);
y1 = reshape(node(vertex,2),NT,Nv);
x2 = circshift(x1,[0,-1]);

94
y2 = circshift(y1,[0,-1]);
% record edges
nextIdx = edgeIdx + NT*Nv;
newEdgeIdx = edgeIdx:nextIdx-1;
edge(newEdgeIdx,1) = vertex(:); % get edge per element
vertexShift = circshift(vertex,[0,-1]);
edge(newEdgeIdx,2) = vertexShift(:);
edgeIdx = nextIdx;
% Compute geometry quantity: edge, normal, area, center
bdIntegral = x1.*y2 - y1.*x2;
area = sum(bdIntegral,2)/2; % the area
h = repmat(sqrt(abs(area)),1,Nv); % h = sqrt(area)
cx = sum((x1+x2).*bdIntegral,2)./(6*area); % the centroid
cy = sum((y1+y2).*bdIntegral,2)./(6*area); % the centroid
normVecx = y2 - y1; % normal vector
normVecy = x1 - x2;
% matrix B, D, I - P
Bx = (normVecx + circshift(normVecx,[0,1]))./(2*h);
By = (normVecy + circshift(normVecy,[0,1]))./(2*h);
Dx = (x1 - repmat(cx,1,Nv))./h; % m(x) = (x - cx)/h
Dy = (y1 - repmat(cy,1,Nv))./h;
c1 = (1 - (repmat(sum(Dx,2),1,Nv).*Bx + repmat(sum(Dy,2),1,Nv).*By))/Nv;
IminusP = zeros(NT,Nv,Nv);
for i = 1:Nv
for j = 1:Nv
IminusP(:,i,j) = - c1(:,j) - Dx(:,i).*Bx(:,j) - Dy(:,i).*By(:,j);
end
IminusP(:,i,i) = ones(NT,1) + IminusP(:,i,i);
end
% assemble the matrix
for i = 1:Nv
for j = 1:Nv
ii(index+1:index+NT) = vertex(:,i);

95
 jj(index+1:index+NT) = vertex(:,j);
ss(index+1:index+NT) = Bx(:,i).*Bx(:,j) + By(:,i).*By(:,j) + dot(
IminusP(:,:,i),IminusP(:,:,j),2);
index = index + NT;
end
end
% compute the right hand side
ft = area.*pde.f([cx cy])/Nv;
b = b + accumarray(vertex(:),repmat(ft,Nv,1),[N,1]);
end
A = sparse(ii,jj,ss,N,N);
%% Find boundary edges and nodes
totalEdge = sort(edge(:,1:2),2);
[i,j,s] = find(sparse(totalEdge(:,2),totalEdge(:,1),1));
bdEdge = [j(s==1), i(s==1)]; % find the boundary edge
isBdNode = false(N,1);
isBdNode(bdEdge) = true;
bdNode = find(isBdNode); % get the boundary node
%% Impose boundary conditions
u = zeros(N,1);
u(bdNode) = pde.g_D(node(bdNode,:));
b = b - A*u;
%% Solve Au = b
freeNode = find(˜isBdNode); % get the interior node
u(freeNode) = A(freeNode,freeNode)\ b(freeNode);

Note that we impose the Dirichlet boundary conditions, which is realized by lines 74 − 76. Prob-
lems involving Neumann boundary part can also be implemented in a similar way. The boundary
integral involving the Neumann boundary part is added into the right hand side. It is also vectorized
using accumarray.

96
A.2.3 Numerical Results

In Figure A.1, it shows the irregular mesh from [122] in the rectangle domain [0, 1] × [0, 1]. We
shall define the following errors:

p
kuI − uh kA = (uI − uh )T A(uI − uh ),

kuI − uh k∞ = max |uI − uh |,

where uh is the numerical solution obtained by the linear virtual element methods; uI is the nodal
interpolation of the exact solution u in Th and A is the corresponding stiffness matrix.

We choose the exact solution is u = cos (πx) cos (πy) − 1 and f = 2π 2 cos (πx) cos (πy). The
errors of ||uI − uh ||A and ||uI − uh ||∞ are presented in Table A.1. We could conclude that the
order of accuracy in energy norm is near 1 and in maximum norm is almost 2. It verifies that
virtual element method is feasible for solving Poisson equations with polygonal meshes in two
dimensions. Finally, we also compare the corresponding assembling CPU time along with that
in [121] in Table A.2, which reveals that our assembling time is not increasing too much for the
large number of nodes. While it is more expensive to assemble the stiffness matrix due to the large
for loop as the number of nodes is large enough in [121].

#Dofs # of elements ||uI − uh ||A ||uI − uh ||∞

34 16 8.2e-02 2.7e-02
130 64 5.9e-02 2.1e-02
514 256 2.4e-02 6.9e-03
2,044 1024 1.0e-02 1.6e-03
8,168 4096 4.9e-03 5.2e-04
32,652 20014 2.3e-03 1.2e-04
Rate 1.1 1.9

Table A.1: Errors measured in H 1 and L∞ norms

97
 15 24 30 33 34

14 23 29 31
13 22 28 32

21 25 27
12
5 20 26
11

4 9 16 18
3 10 17 19

1 2 6 7 8

Figure A.1: The mesh when the number of the element is 16.

#Dofs Assemble(our algorithm) Assemble(other algorithm)

34 0.030971 0.0398962
130 0.018579 0.0416274
514 0.011572 0.194252
2,044 0.016255 0.521451
8,168 0.049402 3.4407
32,652 0.19699 40.2361

Table A.2: The comparison of assembling CUP time ( in seconds)

98