Marshall University

Marshall Digital Scholar

Theses, Dissertations and Capstones

1-1-2012

A Local Radial Basis Function Method for the

Numerical Solution of Partial Differential
Equations
Maggie Elizabeth Chenoweth
[email protected]

Follow this and additional works at: http://mds.marshall.edu/etd

Part of the Numerical Analysis and Computation Commons

Recommended Citation
Chenoweth, Maggie Elizabeth, "A Local Radial Basis Function Method for the Numerical Solution of Partial Differential Equations"
(2012). Theses, Dissertations and Capstones. Paper 243.

This Thesis is brought to you for free and open access by Marshall Digital Scholar. It has been accepted for inclusion in Theses, Dissertations and
Capstones by an authorized administrator of Marshall Digital Scholar. For more information, please contact [email protected].
A LOCAL RADIAL BASIS FUNCTION METHOD FOR THE
NUMERICAL SOLUTION OF PARTIAL DIFFERENTIAL
EQUATIONS

A thesis submitted to

the Graduate College of

Marshall University

In partial fulfillment of

the requirements for the degree of

Master of Arts in Mathematics

by

Maggie Elizabeth Chenoweth

Approved by

Dr. Scott Sarra, Committee Chairperson

Dr. Anna Mummert
Dr. Carl Mummert

Marshall University
May 2012
 Copyright by

Maggie Elizabeth Chenoweth

2012

ii
 ACKNOWLEDGMENTS

I would like to begin by expressing my sincerest appreciation to my thesis advisor, Dr. Scott

Sarra. His knowledge and expertise have guided me during my research endeavors and the

process of writing this thesis. Dr. Sarra has also served as my teaching mentor, and I am

grateful for all of his encouragement and advice. It has been an honor to work with him.

I would also like to thank the other members of my thesis committee, Dr. Anna Mummert

and Dr. Carl Mummert. Their feedback and counsel have been exceedingly beneficial while

finalizing my thesis. The leadership of the chair of the Mathematics Department, Dr. Alfred

Akinsete, and formerly Dr. Ralph Oberste-Vorth, as well as the guidance of the graduate

advisor, Dr. Bonita Lawrence, have been outstanding. I am lucky to have been a teaching

assistant in a department with these individuals.

If it had it not been for numerous teachers, I would not be where I am today. Specif-

ically, I would like to thank the professors I had the privilege of taking mathematics

courses from while at Marshall: Dr. John Drost, Dr. Judith Silver, Dr. Karen Mitchell, Dr.

Ariyadasa Aluthge, Dr. Evelyn Pupplo-Cody, Dr. Yulia Dementieva, Dr. Scott Sarra, Dr.

Ralph Oberste-Vorth, Dr. Anna Mummert, Dr. Carl Mummert, and Dr. Bonita Lawrence.

You have taught me not only mathematics, but a love for a subject that I hope to instill in

the hearts and minds of future students.

Finally, I would like to thank my family and friends for all of their continued love and

support. This thesis is dedicated to them.

iii
CONTENTS

ACKNOWLEDGMENTS iii

ABSTRACT viii

1 INTRODUCTION 1

2 RADIAL BASIS FUNCTION METHODS 3

2.1 GLOBAL RBF INTERPOLATION . . . . . . . . . . . . . . . . . . . . . . 5

2.2 CATEGORIES OF RBFS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.3 OTHER PROPERTIES OF RBFS . . . . . . . . . . . . . . . . . . . . . . . 9

2.4 USING RBFS TO APPROXIMATE DERIVATIVES . . . . . . . . . . . . . 12

2.4.1 GLOBAL DERIVATIVE EXAMPLE . . . . . . . . . . . . . . . . . 13

3 EFFICIENT AND ACCURATE IMPLEMENTATION OF RBF METH-

ODS FOR TIME-DEPENDENT PDES 17

3.1 THE LOCAL METHOD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.1.1 LOCAL DERIVATIVE EXAMPLE . . . . . . . . . . . . . . . . . . 19

3.1.2 THE LOCAL METHOD USING SINGULAR VALUE DECOMPO-

SITION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.2 GLOBAL VERSUS LOCAL RBF APPROXIMATION . . . . . . . . . . . . 22

3.3 1D ADVECTION DIFFUSION . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.4 2D ADVECTION DIFFUSION REACTION . . . . . . . . . . . . . . . . . 26

4 PATTERN FORMATION IN MATHEMATICAL BIOLOGY 31

iv
 4.1 APPLICATION 1: TURING PATTERNS . . . . . . . . . . . . . . . . . . . 31

4.2 APPLICATION 2: CHEMOTAXIS . . . . . . . . . . . . . . . . . . . . . . . 34

5 CONCLUSION 40

A MATLAB CODE 42

REFERENCES 47

CURRICULUM VITAE 49

v
LIST OF FIGURES

2.1 Global RBF interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2 Algebraic versus spectral convergence . . . . . . . . . . . . . . . . . . . . . 8

2.3 Relationship between  and " for the MQ . . . . . . . . . . . . . . . . . . . 11

2.4 Derivative calculated using RBF methods . . . . . . . . . . . . . . . . . . . 16

3.1 Local RBF approximation of the derivative of f (x) = esin(⇡x) . . . . . . . . . 20

3
3.2 Global RBF approximation of the derivative of f (x) = ex cos(2x). . . . . 22

3.3 Eigenvalue stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.4 One-dimensional linear advection-di↵usion results . . . . . . . . . . . . . . . 25

3.5 Circle with stencil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.6 The initial condition of Equation (3.16). . . . . . . . . . . . . . . . . . . . . 28

3.7 Error in the numerical solution of Equation (3.16) at t=5. . . . . . . . . . . 29

3.8 Local method solution at t=5. . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.9 Local method error at t=5. . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4.1 Turing patterns observed on a leopard and in galaxies . . . . . . . . . . . . 32

4.2 Turing patterns on a fish . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4.3 Turing patterns on a butterfly domain. . . . . . . . . . . . . . . . . . . . . . 34

4.4 Density plots for chemotaxis experimental results . . . . . . . . . . . . . . . 38

4.5 Chemotaxis experimental results . . . . . . . . . . . . . . . . . . . . . . . . 39

vi
LIST OF TABLES

2.1 Global, infinitely di↵erentiable RBFs . . . . . . . . . . . . . . . . . . . . . . 8

3.1 Error versus stencil size from Equation (3.11). . . . . . . . . . . . . . . . . . 23

3.2 Error versus stencil size from Equation (3.16). . . . . . . . . . . . . . . . . . 28

4.1 Parameters for Equations (4.4) and (4.5) . . . . . . . . . . . . . . . . . . . . 36

vii
 ABSTRACT

A LOCAL RADIAL BASIS FUNCTION METHOD FOR THE NUMERICAL

SOLUTION OF PARTIAL DIFFERENTIAL EQUATIONS

Maggie Elizabeth Chenoweth

Most traditional numerical methods for approximating the solutions of problems in science,

engineering, and mathematics require the data to be arranged in a structured pattern and

to be contained in a simply shaped region, such as a rectangle or circle. In many important

applications, this severe restriction on structure cannot be met, and traditional numerical

methods cannot be applied.

In the 1970s, radial basis function (RBF) methods were developed to overcome the

structure requirements of existing numerical methods. RBF methods are applicable with

scattered data locations. As a result, the shape of the domain may be determined by the

application and not the numerical method.

Radial basis function methods can be implemented both globally and locally. Compar-

isons between these two techniques are made in this work to illustrate how the local method

can obtain very similar accuracy to the global method while only using a small subset of

available points, and thus using substantially less computer memory.

Finally, radial basis function methods are applied to solve systems of nonlinear partial

di↵erential equations (PDEs) that model pattern formation in mathematical biology. The

local RBF method will be used to evaluate Turing pattern and chemotaxis models that are

both modeled by advection-reaction-di↵usion type PDEs.

viii
Chapter 1

INTRODUCTION

Radial basis function (RBF) methods were first studied by Roland Hardy, an Iowa State

geodesist, in 1968, when he developed one of the first e↵ective methods for the interpolation

of scattered data [8]. Polynomial methods had previously been used, but they do not have

a unisolvency property for two-dimensional and higher dimensional scattered data. After

much investigation, Hardy developed what would later be known as the multiquadric (MQ)

radial basis function [9]. This is only one of many existing RBFs.

Then, in 1979, Richard Franke published a study of all known methods of scattered data

interpolation and concluded that the MQ RBF method was the best method. Because of

Franke’s extensive numerical experiments with the MQ, he is often credited for introducing

the MQ into the field of mathematical science [6].

The next significant event in RBF history was in 1986 when Charles Micchelli, an

IBM mathematician, developed the theory behind the MQ method. He proved that the

system matrix for the MQ method was invertible, which means that the RBF scattered

data interpolation problem is well-posed [15]. Four years later, physicist Edward Kansa

first used the MQ method to solve partial di↵erential equations [12]. In 1992, results from

Wolodymyr Madych and Stuart Nelson [14] showed the spectral convergence rate of MQ

interpolation. Since Kansa’s discovery, research in RBF methods has rapidly grown, and

RBFs are now considered an e↵ective way to solve partial di↵erential equations [30]. All

RBF methods using an infinitely di↵erentiable RBF have been proven to be generalizations

of the polynomial based pseudospectral methods [22].

1
 Over the years, RBF interpolation has been shown to work in many cases where poly-

nomial interpolation has failed [20]. RBF methods overcome the limitation of polynomial

interpolation because they do not need to be implemented on tensor product grids or in

rectangular domains. RBF methods are frequently used to represent topographical surfaces

as well as other intricate three-dimensional shapes [23], having been successfully applied in

such diverse areas as climate modeling, facial recognition, topographical map production,

auto and aircraft design, ocean floor mapping, and medical imaging. RBF methods have

been actively developed over the last 40 years and the RBF research area remains very

active as many open questions still remain.

In this work, comparisons will be made between global and local RBF approximation

methods. It will be shown how the local method can obtain very similar accuracy to that of

the global method while using only a small subset of available points. Hence, less computer

memory is required. This will be illustrated with several numerical examples including

those that involve pattern formation obtained from Turing and chemotaxis models.

2
Chapter 2

RADIAL BASIS FUNCTION METHODS

In order to e↵ectively understand RBF methods, several definitions are first required.

Definition 1. A function : Rd ! R is called radial if there exists a univariate function

' : [0, 1) ! R such that

(x) = '(r),

where r = kxk, and k · k is a norm on Rd . (k · k is typically the Euclidean norm.)

Definition 2. A radial basis function, (r), is a one-variable, continuous function

defined for r 0 that has been radialized by composition with the Euclidean norm on Rd .

RBFs may have a free parameter, the shape parameter, denoted by ".

Definition 3. The scattered data interpolation problem states that given data (xj , fj ),

with j = 1, . . . , N, xj 2 Rd , and fj 2 R, find a smooth function s such that s(xj ) = fj , for

j = 1, . . . , N .

Given a set of N centers, xc1 , . . . , xcN , in Rd , a radial basis function interpolant is of the

form
N
X
s(x) = ↵j (k x xcj k2 ). (2.1)
j=1

The ↵j coefficients in the RBF are determined by enforcing the interpolation condition

s(xi ) = f (xi )

3
at a set of points that usually coincides with the N centers. Enforcing the interpolation

condition at the N centers results in the N ⇥ N linear system

B↵ = f (2.2)

to be solved for the expansion coefficients ↵. The matrix B, called the interpolation matrix

or the system matrix, has entries

bij = (kxci xcj k2 ), i, j = 1, . . . , N.

In Section 2.3 it will be shown that the system matrix is always invertible.

For the distance matrix r (as defined in Definition 1) that is used in the calculation of

the system matrix, it may seem as though two loops should be created in Matlab. This can

be naively coded as a double loop:

N = length(xc);

for i=1:N

for j=1:N

r(i,j) = abs( xc(i) - xc(j) );

end

end

However, when loops are used in Matlab, the program can take a long time to run, and

it is not as efficient as other approaches. For instance, a summation, which in a computer

language is coded as a loop, can be replaced by a more efficient dot product with a vector

of ones.

In Matlab, the distance matrix is formed as follows where xc represents the vector of N

distinct centers.

o = ones(1,length(xc));

r = abs( xc*o - (xc*o)’ );

The interpolant is evaluated using (2.1) at the M points, s(xi ), by forming an M ⇥ N

evaluation matrix H with entries

4
 hij = (kxi xcj k2 ), i = 1, . . . , M and j = 1, . . . , N.

In Matlab, the evaluation matrix is formed in a manner similar to the system matrix where

xc represents the vector of N distinct centers and x is the vector of the M points at which

to evaluate the RBF interpolant.

xc = xc(:);

x = x(:);

r = abs( x*ones(1,length(xc)) - ones(length(x),1)*xc’ );

H = mqRbf(r,shape);

The interpolant is then evaluated at the M points by the matrix-vector product

fa = H↵.

2.1 GLOBAL RBF INTERPOLATION

RBF approximation methods may be either global or local. The global approach uses

information from every center in the domain to approximate a function value or derivative

at a single point. In contrast, the local method only uses a small subset of the available

centers. In this section, the mechanics of the global interpolation method is illustrated in

the following example.

For this example, the MQ RBF, as defined in Table 2.1, is used to interpolate a function.

Let f (x) = esin(⇡x) be restricted to the interval [0,1]. This function is interpolated using

the following three centers that are not evenly spaced: xc1 = 0, xc2 = 0.6, and xc3 = 1.

The interpolant is evaluated at the five evenly spaced evaluation points x1 = 0, x2 = 0.25,

x3 = 0.5, x4 = 0.75, and x5 = 1. For the MQ, a shape parameter is also required. We will

let " = 1.25. A discussion of how how to choose the best value of the shape parameter can

be found in Section 2.3.

Let

↵ = [↵1 ↵2 ↵3 ]

5
be the unknown vector of the expansion coefficients, let
⇥ ⇤T
f = [f (xc1 ) f (xc2 ) f (xc3 )]T = 1 esin(0.6⇡) 1 ,

and let
2 3
c xc1 k2 ) (kxc1 xc2 k2 ) (kxc1 xc3 k2 )7
6 (kx1
6 7
B=6
6 (kx2
c xc1 k2 ) (kxc2 xc2 k2 ) (kxc2 xc3 k2 )77
4 5
(kxc3 xc1 k2 ) (kxc3 xc2 k2 ) (kxc3 xc3 k2 )
be the 3 ⇥ 3 system matrix. The linear system B↵ = f can now be solved resulting in the

following expansion coefficients:

↵ = [9.5029 26.3975 15.3013]T .

The 5 ⇥ 3 evaluation matrix, H, is used to evaluate the interpolant.

2 3
(kx1 xc1 k2 ) (kx1 xc2 k2 ) (kx1 xc3 k2 )
6 7
6 7
6 (kx2 xc1 k2 ) (kx2 xc2 k2 ) (kx2 xc3 k2 )7
6 7
6 7
H=6 6 (kx3 xc1 k2 ) (kx3 xc2 k2 ) (kx3 xc3 k2 )77
6 7
6 7
6 (kx4 xc1 k2 ) (kx4 xc2 k2 ) (kx4 c
x3 k2 )7
4 5
(kx5 xc1 k2 ) (kx5 xc2 k2 ) (kx5 c
x 3 k2 )
The next step is to calculate fa = H↵. The interpolated values are:
 T
fa = 1 2.1167 2.6473 2.1994 1 .

The exact solutions are 1, 2.0281, 2.7183, 2.0281, and 1. The point-wise errors are as follows:

0.000000000000000

0.088631801660779

0.070973556874207

0.171280064870869

0.000000000000000.

This is illustrated in Figure 2.1. The errors at the endpoints are zero, as the evaluation

points coincide with centers, and the interpolation conditions dictate that the interpolant

agree with the function values at the centers. This is just a basic example to demonstrate

how MQ RBFs are computed, but in further examples, there are multiple strategies that

can be used to minimize point-wise errors.

6
 2.8

2.6

2.4

2.2

2
f(x)

1.8

1.6

1.4

1.2

1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x

Figure 2.1: Global RBF interpolation of the function f (x) = esin(⇡x) . Red asterisks denote
centers and open black circles denote evaluation points.

2.2 CATEGORIES OF RBFS

Three primary categories of radial basis function methods exist: compactly supported,

global with finite smoothness, and global infinitely di↵erentiable. Compactly supported

RBFs have an algebraic convergence rate.

Definition 4. Algebraic convergence rates occur when error decays at the rate O(N m)

for some constant m.

Second, we have the globally supported, finitely di↵erentiable radial basis functions,

which also have algebraic convergence rates [28]. Finally, there are the globally supported,

infinitely di↵erentiable radial basis functions. This category can achieve spectral conver-

gence rates for smooth functions [10, 17].

Definition 5. Spectral convergence rates occur when error decays at the rate of O(⌘ N )
1
as N increases where 0 < ⌘ < 1. In particular with RBF methods, N is equal to "h , where

7
 0
10

−2
10

−4
10

−6
10
error
−8
10

−10
10

−12
10

−14
10
0 1 2
10 10 10
N

Figure 2.2: Algebraic convergence (represented by blue asterisks) versus spectral conver-
gence (represented by red dots).

" is the shape parameter and h is the minimum separation distance between centers. Error

decays as both the shape parameter and minimum separation distance decrease.

The di↵erence between algebraic convergence and spectral convergence is illustrated in

Figure 2.2.

Four common RBFs that are globally supported and infinitely di↵erentiable are con-

tained in Table 2.1. The multiquadric (MQ) is arguably the most popular RBF that is used

in applications and is representative of the class of global infinitely di↵erentiable RBFs. For

that reason the MQ has been used in all examples throughout.

RBF Name Equation

2 2
Gaussian (GA) (r, ") = e " r
Inverse Quadratic (IQ) (r, ") = 1/(1 2 2
p +" r )
Inverse Multiquadric (IMQ) (r, ") = 1/(p 1 + "2 r2 )
Multiquadric (MQ) (r, ") = 1 + "2 r2

Table 2.1: Global, infinitely di↵erentiable RBFs

8
2.3 OTHER PROPERTIES OF RBFS

Radial basis function methods have numerous properties. Several of those properties will

be highlighted in this section.

Definition 6. A real symmetric matrix A is called positive semi-definite if its associated

quadratic form is non-negative, i.e.,

N X
X N
cj ck Ajk 0 (2.3)
j=1 k=1

for c = [c1 , ..., cN ]T 2 RN . If the inequality is zero only for c ⌘ 0, then A is called positive

definite.

A positive definite matrix is non-singular since all of the eigenvalues of a positive definite

matrix are positive. In other words, a well-posed interpolation problem exists if the basis

functions generate a positive definite system matrix. Definition 6 relates positive definite

functions to positive semi-definite matrices. To make sure that the interpolation problem

is well-posed, this definition can be refined to that of a strictly positive definite function.

Definition 7. A complex-valued continuous function : Rd ! C is called positive defi-

nite on Rd if
N X
X N
cj ck (xj xk ) 0 (2.4)
j=1 k=1

for any N pairwise di↵erent points x1 , ..., xN 2 Rd , and c = [c1 , ..., cN ]T 2 CN . The function

is called strictly positive definite on Rd if the inequality is zero only for c.

The inverse quadratic, inverse multiquadric, and Gaussian functions are all strictly

positive definite and thus have a positive definite system matrix which is invertible.

After many experiments, Franke concluded that the system matrix was always uniquely

solvable because the linear system resulting from the scattered data interpolation problem

was nonsingular [6]. However, Franke never provided a proof. In 1986, Micchelli provided

the conditions necessary for the proof of Theorem 1, and he generalized the idea of uniquely

solvable systems to all radial basis functions [15].

9
 In order to understand Theorem 1, the definition of a function being completely mono-

tone is necessary.

Definition 8. A function is completely monotone on [0, 1) if

i. 2 C[0, 1)

ii. 2 C 1 (0, 1)

iii. ( 1)l l (r) 0 where r > 0 and l = 0, 1, ...

p 0 (r)
Theorem 1. Let (r) = ( r) 2 C[0, 1) and (r) > 0 for r > 0. Let be completely

monotone and nonconstant on (0, 1). Then for any set of N distinct centers {xcj }N
j=1 , the

N ⇥ N matrix B with entries bjk = (k xkj xck k2 ) is invertible. Such a function is said to

be conditionally positive definite of order one.

The MQ is an example of a global, infinitely di↵erentiable RBF that is a conditionally

positive definite function of order one.

To show this, let

p p
(r) = ( r) = 1 + "2 r (2.5)

and evaluate the first several derivates with respect to r:

0 "2
(r) = p
2 1 + "2 r
00 "4
(r) =
4 (1 + "2 r)3/2
(3) 3"6
(r) = (2.6)
8 (1 + "2 r)5/2
(4) 15"8
(r) =
16 (1 + "2 r)7/2
.. ..
. = .

From this pattern, it can be seen that the sign alternates for even and odd derivatives [4].

Thus, by Definition 8 and Theorem 1, the MQ is conditionally positive definite of order

one.

10
 The MQ contains a free variable, ", known as the shape parameter. The shape parameter

a↵ects both the conditioning of the system matrix and the accuracy of the RBF method

when given a set of fixed centers. The condition number

1 max
(B) = kBk2 B 2
= (2.7)
min

is a measure of how difficult a problem is to be accurately approximated by a numerical

algorithm. The condition number is defined using any matrix norm. In the particular case

of the 2-norm, it is the ratio of the maximum to minimum singular values. The singular

values of B are represented by . As a rule of thumb, when the condition number is 10n , one

would expect to lose approximately n accurate digits when solving a general linear system

Ax = b.
N = 100 N = 100
−5 22
10 10

20
10
−6
10
18
10
Condition Number
|error|

−7 16
10 10

14
10
−8
10
12
10

−9 10
10 10
0 2 4 6 8 10 0 2 4 6 8 10

Figure 2.3: MQ RBF interpolation of the function, f (x) = ex . Left: Error versus the shape
parameter. Right: Condition number versus the shape parameter.

As shown is Figure 2.3, the condition number exponentially increases as the shape pa-

rameter decreases. In order for the system matrix to be well-conditioned, the shape param-

eter must not be too small. However, small shape parameters are required to obtain good

accuracy for the RBF method, but this results in the system matrix being ill-conditioned.

Hence, having the best accuracy and conditioning can obviously not occur at the same time.

11
This is known as the Uncertainty Principle, which indicates the more favorably valued one

quantity is the less favorably valued the other is [24].

For small ", the error improves, but the condition number grows. There are several

approaches that can be taken to find the best shape parameter. These approaches include

using the power function [29], “leave-one-out” cross validation [19], or the Contour-Pade

Algorithm [5]. When the system matrix is ill-conditioned, RBF methods are most accurate.

The shape parameter can be selected so that the resulting system matrix has a condition

number, (B), in the range 1013  (B)  1015 in order to determine the corresponding

value for ". These bounds for the condition number are valid when using a computer that

implements double precision floating point arithmetic, but the bounds will be di↵erent when

using other floating point number systems [16].

2.4 USING RBFS TO APPROXIMATE DERIVATIVES

The derivative of a function f (x) is approximated by the RBF method as

N
X
@ @
f (x) = ↵j (k x xcj k2 ). (2.8)
@xi @xi
j=1

n oN
When evaluated at the N centers, xcj , (2.8) becomes
j=1

@ @
f (x) = H↵. (2.9)
@xi @xi

The N ⇥ N evaluation matrix, H, has entries

@
hi,j = (kxci xcj k2 ), i, j = 1, . . . , N. (2.10)
@xi

From Equation (2.2), it is known that ↵ = B 1f . The di↵erentiation matrix can be defined

as
@ 1
D= HB . (2.11)
@xi

12
The derivative is then approximated as

@ @
⇡ f (x) = Df. (2.12)
@xi @xi

(Note: In Matlab, the inverse of B is not actually formed. Instead, D = H/B.) Using the

chain rule to di↵erentiate the radial basis function, [r(x)], the result is

@ d @r
= (2.13)
@xi dr @xi

for the first derivative, with

@r xi
= . (2.14)
@xi r

The second derivative is calculated as follows

✓ ◆2
@2 d @2r d2 @r
= + , (2.15)
@x2i dr @x2i dr2 @xi

with h i2
@r
@2r 1 @xi
= . (2.16)
@x2i r

In particular, for the MQ,

d "2 r
=p (2.17)
dr 1 + "2 r 2

and

d2 "2
= . (2.18)
dr2 [1 + "2 r2 ]3/2

2.4.1 GLOBAL DERIVATIVE EXAMPLE

For this example, f 0 (x) is approximated for f (x) = esin(⇡x) on the interval [0,1] using the

following three center locations xc1 = 0, xc2 = 0.6, and xc3 = 1 to approximate the derivative.

The shape parameter is taken to be " = 1.25.

13
 To begin, let
2 3
 6 1 7
T 6 7
f = f (xc1 ) f (xc2 ) f (xc3 ) = 6esin(0.6⇡) 7
6
7,
4 5
1

and let
2 3
6 (kxc1 xc1 k2 ) (kxc1 xc2 k2 ) (kxc1 xc3 k2 )7
6 7
B=6
6 (kx2
c xc1 k2 ) (kxc2 xc2 k2 ) (kxc2 xc3 k2 )7
7
4 5
(kxc3 xc1 k2 ) (kxc3 xc2 k2 ) c c
(kx3 x3 k2 )
2 3
6 (k0 0k2 ) (k0 0.6k2 ) (k0 1k2 ) 7
6 7
=6
6 (k0.6 0k2 ) (k0.6 0.6k2 ) (k0.6 1k2 )7 7
4 5
(k1 0k2 ) (k1
0.6k2 ) (k1 1k2 )
2 3
6 1 1.25 1.60087
6 7
=6
6 1.25 1 1.11807
7
4 5
1.6008 1.1180 1

be the 3 ⇥ 3 system matrix.

The 3 ⇥ 3 evaluation matrix, H, is formed to evaluate our interpolant:

2 3
c xc1 k2 ) c xc2 k2 ) c xc3 k2 )7
6 x (kx1 x (kx1 x (kx1
@ 6 7
H=6
6 x (kx2
c xc1 k2 ) c
x (kx2 xc2 k2 ) x (kx2
c xc3 k2 )7
7
@x 4 5
c xc1 k2 ) c xc2 k2 ) c ck )
x (kx3 x (kx3 x (kx 3 x 3 2
2 3
6 x (k0 0k2 ) x (k0 0.6k2 ) x (k0 1k2 ) 7
6 7
=6
6 x (k0.6 0k2 ) x (k0.6 0.6k2 ) x (k0.6 1k )
2 7
7
4 5
x (k1 0k2 ) x (k1 0.6k2 ) x (k1 1k2 )
2 3
6 0 0.75 0.97617
6 7
=6
6 0.75 0 0.55907
7.
4 5
0.9761 0.5590 0

14
 In Matlab, the di↵erentiation matrix is formed as D = H/B, and this is calculated to

be
2 3
6 2.0783 3.1217 1.13937
6 7
D=6
6 0.7439 0.8939 1.6313 7
7.
4 5
0.5809 3.4902 2.9723

The derivative can now be approximated by

fa0 = Df
2 32 3
6 2.0783 3.1217 1.13937 6 1 7
6 76 7
=6
6 0.7439 0.8939 7 6
1.6313 7 6e sin(0.6⇡) 7
7
4 54 5
0.5809 3.4902 2.9723 1
2 3
6 4.8628 7
6 7
=6 7
6 1.42657 .
4 5
5.4810

The exact derivative of f (x) = esin(⇡x) is f 0 (x) = ⇡esin(⇡x) cos (⇡x). The results for this

example are represented in Figure 2.4. If more centers are added, it is possible to obtain

better accuracy.

15
 6

2
f (x)

−2

−4

−6
0 0.2 0.4 0.6 0.8 1
x

Figure 2.4: The green line represents the derivative of f (x) = esin(⇡x) . The red asterisks are
the exact values at the centers. The black open circles are the approximate values at the
evaluation points.

16
Chapter 3

EFFICIENT AND ACCURATE IMPLEMENTATION OF

RBF METHODS FOR TIME-DEPENDENT PDES

3.1 THE LOCAL METHOD

Up until this point in this thesis, the main focus has been on the global method for radial

basis functions. In many cases, the local method can be as accurate as the global method.

The main advantage to using the local method is that less computer storage and flops are

needed. By definition, a flop refers to one addition, subtraction, multiplication, or division

of two floating point numbers.

The local RBF interpolant takes the form

X
In f (x) = ↵k (k x xck k2 ) (3.1)
k2Ii

at each of the N centers xc1 , . . . , xcN in Rd . In the above equation, Ii is a vector associated

with center i that contains the center number and the indices of the n 1 nearest neighboring

centers, ↵ is a vector of expansion coefficients, and is a RBF. Each center and its n 1

neighbors are called a stencil. The ↵k coefficients in Equation (3.1) are chosen by enforcing

the interpolation condition so that

In f (xk ) = fk (3.2)

with k 2 Ii on each stencil. This gives N linear systems each with dimension n ⇥ n of the

17
form

B↵ = f. (3.3)

This equation will be solved for the expansion coefficients. Just like the global method, the

matrix B is called the interpolation matrix or the system matrix with entries

bjk = (kxcj xck k2 ), j, k = Ii (1), . . . , Ii (n).

Because the system matrix B is always invertible, the expansion coefficients are uniquely

defined on each stencil.

The derivative of a function can be approximated at the center locations by applying a

linear di↵erential operator L to the local interpolant of the RBF. Depending on the problem

at hand, L will either be a single derivative operator or a linear combination of derivative

operators. The equation

N
X
Lf (x) = ↵j L (k x xcj k2 ) (3.4)
j=1

is then obtained by evaluating it at the center where the stencil is based. The equation can

be simplified to a dot product of the form

Lf (x) = h↵. (3.5)

In this case, h is a 1 ⇥ n vector containing the elements

hi = L (k x xcj k2 ), (3.6)

and ↵ is the n ⇥ 1 vector of RBF expansion coefficients. The equation can be simplified by

recognizing that
1 1
Lf (xi ) = hB f (Ii ) = (hB )f (Ii ) = w · f (Ii ) (3.7)

with the stencil weights as

1
w = hB . (3.8)

18
The weights are the solution of the linear system

wB = h. (3.9)

Derivatives are approximated by multiplying the weights by the function values at the

centers.

3.1.1 LOCAL DERIVATIVE EXAMPLE

For this example, f 0 (x) is approximated for f (x) = esin(⇡x) on the interval [0,1] using the

local RBF method with " = 0.43. Let x0 , . . . , x10 be N = 11 evenly spaced points on the

interval [0,1]. For convenience, a stencil size of n = 3 will be chosen. In order to approximate

the derivative at x5 = 0.5, theoretically the equation f 0 (x5 ) ⇡ w4 f (x4 ) + w5 f (x5 ) + w6 f (x6 )

would need to be solved. This is accomplished by calculating

2 3
6 (k0.4 0.4k2 ) (k0.4 0.5k2 ) (k0.4 0.6k2 )7
6 7
B=6
6 (k0.5 0.4k2 ) (k0.5 0.5k2 ) (k0.5 0.6k2 )77
4 5
(k0.6 0.4k2 ) (k0.6 0.5k2 ) (k0.6 0.6k2 )
2 3
6 1 1.0009 1.00377
6 7
=6
6 1.0009 1 1.0009 7
7
4 5
1.0037 1.0009 1

and


h = L (k0.5 0.4k2 ) L (k0.5 0.5k2 ) L (k0.5 0.6k2 )

= 0.0185 0 0.0185

to find that
2 3
 6 1 1.0009 1.00377 
6 7
61.0009 1.00097
w4 w5 w6 6 1 7 = 0.0185 0 0.0185 .
4 5
1.0037 1.0009 1

19
 5

1
f (x)

−1

−2

−3

−4

−5
0 0.2 0.4 0.6 0.8 1
x

Figure 3.1: The green line represents the derivative of f (x) = esin(⇡x) . The red asterisks are
the exact values at the centers. The black open circles are the approximate values at the
evaluation points using local RBF approximation.

Solving this system results in

 
w4 w5 w6 = 5 0 5 .

The derivative can now be approximated at x5 = 0.5 by

f 0 (x5 ) ⇡ w4 f4 + w5 f5 + w6 f6

⇡ 5(2.5884) + 0(2.7183) + 5(2.5884)

15
⇡ 1.7764 ⇥ 10 .

The exact derivative at x5 = 0.5 is 0, so this approximation is extremely accurate.

20
 The Matlab results for this example when the local RBF method is applied at each data

point are represented in Figure 3.1. The red asterisks are the exact derivatives at each of

the 11 points, and the open circles represent the approximated solutions found using the

local method.

3.1.2 THE LOCAL METHOD USING SINGULAR VALUE DECOM-

POSITION

In order to calculate wB = h, the singular value decomposition (SVD) of the system matrix
P
B is used. The SVD of B = U V T . Since B is an N ⇥ N matrix, U and V are N ⇥ N
P
orthogonal matrices [26]. The matrix is a diagonal N ⇥ N matrix having N singular

values of B as its nonzero elements. Because U and V are orthogonal, the inverse of each

these two matrices is the transpose of the matrix. The inverse of the system matrix is
P 1 T
V U , and by multiplying both sides of Equation (3.9) by B 1 , the weights utilized in

the local method are calculated as follows:

P 1
w = h(V U T ).

Condition numbers can also be calculated using the SVD. In the following pseudocode,

the shape parameter is adjusted until the condition number is in the desired range of

1013  (B)  1015 :

condtionNumber = 1

K = conditionNumber

while (K < minConditionNumber OR K > maxConditionNumber)

construct B

[U,S,V] = svd(B)

condtionNumber = maximum(S)/minimum(S)

if K < minimumConditionNumber

shape = shape - shapeIncrement

elseif K > maximumConditionNumber

shape = shape + shapeIncrement

21
3.2 GLOBAL VERSUS LOCAL RBF APPROXIMATION

As previously mentioned, the local RBF method on various stencils can be just as accurate

and efficient as the global RBF method. This is especially true for derivative approximation,

as the derivative is a local property.

In Section 3.1, a small example was examined that utilized global RBF interpolation.

For this section, an example with a larger N that uses the global method to approximate the

derivative of a function will be analyzed. It will then be compared to the local derivative

method. Matlab code for the local method is provided in Appendix A.

As a motivating example, the derivative of the function

3
f (x) = ex cos(2x) (3.10)

will be approximated on the interval [ 1, 1] using both the global and local RBF method.

10

6
f (x)

−2
−1 −0.5 0 0.5 1
x
3
Figure 3.2: The green line represents the derivative of f (x) = ex cos(2x). The red
asterisks are the exact values at the centers. The black open circles are the approximate
values at the evaluation points using global RBF approximation.

22
 The exact derivative of Equation (3.10) is

3
f 0 (x) = 3ex x2 + 2 sin(2x). (3.11)

3
In Figure 3.2, the green line represents the derivative, f 0 (x) = 3ex x2 + 2 sin(2x), and

the red asterisks are the exact values of the derivative at 100 data points. The black open

circles are the approximate solutions at each of the corresponding 100 data points using the

global RBF method. The error ranges from 8.7128 ⇥ 10 11 to 0.0562 at the right endpoint.

On average, the error is 8.3183 ⇥ 10 4.

A typical result with RBF methods is that the accuracy of the global method can

be matched with the local method with a problem dependent n < N . This can be seen in

Table 3.1 where the local method was used to approximate the derivative of Equation (3.11).

When the stencil size was equal to 7, it was found that the average error resembled the

average error obtained when using the global RBF method.

Stencil Size Average Error

3 1.6124 ⇥ 10 3
7 5.1948 ⇥ 10 4
15 1.8976 ⇥ 10 3

Table 3.1: Error versus stencil size from Equation (3.11).

3.3 1D ADVECTION DIFFUSION

After the space derivatives of a PDE have been discretized by the RBF method, a system

of ODEs

ut = F (u) (3.12)

remains to be advanced in time. Any numerical ODE method can be used. This approach

is called the method of lines. In all numerical examples, the following fourth-order Runge-

23
Kutta method

k1 = tF (un , tn )

k2 = tF (un + 0.5k1 , tn + 0.5 t)

k3 = tF (un + 0.5k2 , tn + 0.5 t) (3.13)

k4 = tF (un + k3 , tn + t)
1
un+1 = un + (k1 + 2k2 + 2k3 + k4 )
6

has been used [1]. The eigenvalues (scaled by t) of the discretized linearized operator must

lie within the stability region of the ODE methods. A Runge-Kutta stability region, along

with the scaled eigenvalues of a 1d advection di↵usion operator, is shown in Figure 3.3.

As a rule of thumb, the RBF method is stable if the eigenvalues of the linearized spatial

discretization operator, scaled by t, lie in the stability region of the time-discretization

operator [25]. As can be seen in Figure 3.3, several eigenvalues are located outside of

the Runge-Kutta stability region when " = 1.5. Hence, instability occurs for this one-

dimensional advection-di↵usion equation when the shape parameter equals 1.5. When the

shape parameter was changed to " = 6, all of the eigenvalues were contained in the region,

and as a result, the method was stable.

The solution of the one-dimensional advection-di↵usion equation

ut + ux ⌫uxx = 0 (3.14)

will be approximated on ⌦ = [0, 1]. The exact solution for ⌫ > 0 is

 ✓ ◆ ⇣x⌘ ✓ ◆
1 x t x+t
u(x, t) = erfc p + exp erfc p . (3.15)
2 2 ⌫t ⌫ 2 ⌫t

Initial and boundary conditions are prescribed according to the exact solution. Let N =

51 evenly spaced centers, " = 6 be the shape parameter, and ⌫ = 0.002. The fourth-order

Runge-Kutta method is used with a time-step of 0.005 to advance the problem until the

final time is 0.5. At t = 0.5, the maximum point-wise error was 4.72 ⇥ 10 4. These results

24
 3 3

2 2

1 1
t)

t)
0 0
imag(

imag(
−1 −1

−2 −2

−3 −3
−3 −2 −1 0 1 −3 −2 −1 0 1
real( t) real( t)

Figure 3.3: Runge-Kutta stability region with the eigenvalues from Equation (3.14). Left:
Instability with " = 1.5 and (B) = 1.3482 ⇥ 1019 . Right: Stability with " = 6 and
(B) = 7.5254 ⇥ 1013 .

are illustrated in Figure 3.4, and Matlab code is provided in Appendix A.

Because this is a one-dimensional problem on a small interval, the global method can

be used. However, this will not be the case later in two-dimensional problems that may

require as many as N = 25,000 centers.

−4
x 10
1.2 6

1
5

0.8
4

0.6
3
0.4

2
0.2

1
0

−0.2 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 3.4: Left: Exact solution versus MQ solution of Equation 3.14. Right: Point-wise
error.

25
3.4 2D ADVECTION DIFFUSION REACTION

This next example is of an advection di↵usion reaction equation with an exact solution

available. This equation has all the elements of the more complicated chemotaxis and

Turing equations that will be used later. It is defined as

ut = ⌫(uxx + uyy ) ↵(ux + uy ) + u2 (1 u) (3.16)

where ⌫ is the viscosity coefficient, ↵ is the advection coefficient, and is the reaction

coefficient. The partial di↵erential equation is linear with a nonlinear reaction term.
1
Consider Equation (3.16) with ⌫ = 0.5, ↵ = 1, and = ⌫. The analytical solution is

given by

1
u(x, y, t) = (3.17)
1+ ea(x+y bt)+c
q p
where a = 4⌫ , b = 2↵ + ⌫, and c = a(b 1). Dirichlet boundary conditions are

prescribed for t > 0 by using the exact solution. As a domain, take the circle of radius 1.5

centered at (1.5, 1.5).

When locating the boundary points on a circle, the results may not be very precise.

By simply setting the distance from the circle’s center to the boundary point equal to the

radius, the computer would be storing the points as floating point numbers. This problem

can be eliminated by using the distance between adding a factor of ‘100 ⇥ machine epsilon’

to the radius and subtracting a factor of ‘100 ⇥ machine epsilon.’ Machine epsilon refers to

the smallest positive number that when added to 1 results in a number that the computer

recognizes as di↵erent from 1. For double-precision, machine epsilon is equal to 2 52 [7].

bi = find((sqrt(x.^2+y.^2)<(R+100*eps)) & ...

(sqrt(x.^2+y.^2)>(R-100*eps)));

For this example, the global method is implemented on 4000 points in a circle beginning

at a time of t = 0 and moving by a time step of 0.0005 until the time equals 5. The

condition number was equal to 6.9955 ⇥ 1014 and the error was 7.4401 ⇥ 10 6. The results

26
 1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1

Figure 3.5: Left: Circle with N = 500 centers. Right: Stencil size of n = 21. The center
of the stencil is marked with blue asterisk, and the other points in stencil are marked with
red squares.

are illustrated in Figures 3.6 and 3.7. It took two minutes for the condition number to be

computed and seventeen minutes for the entire program to run. When trying to increase

the number of centers, out of memory errors occurred.

On an average desktop computer, the global method cannot be implemented with a large

number of centers, such as 15,000, due to memory restraints when the 15,000 ⇥ 15,000 dense

matrix is formed and every element has to be stored. However, it can implement the local

method on these problems. Using the local method, sparse matrices are stored, and we

can obtain desired accuracy. This is illustrated in Table 3.2 using Equation (3.16). When

the local method is executed with n = 15, it is just as accurate as when n = 300. As the

stencil size decreases, not only is less computer memory needed, but the faster the execution

time of the program. An example of a circle with N = 500 centers with a stencil size of

n = 21 is illustrated in Figure 3.5. A problem with a large number of centers that would be

impossible to run on an average desktop computer using the global method (due to memory

restraints) can execute in a matter of seconds using the local method.

27
0.06

0.05

0.04

0.03

0.02

0.01

0
0

4 1 0
3 2

Figure 3.6: The initial condition of Equation (3.16).

Stencil Size Time in Seconds Max Error

300 30.693051 2.1994 ⇥10 5
150 16.567052 1.4822 ⇥10 5
80 9.766413 2.3395 ⇥10 5
40 5.9083 1.2764 ⇥10 5
15 2.744234 2.9272 ⇥10 5
8 2.09120 2.9 ⇥10 3

Table 3.2: Error versus stencil size from Equation (3.16).

28
 t=5

−6
x 10
8

0
3

1 3
2.5
2
1.5
y 1
0 0.5
0
x

Figure 3.7: Error in the numerical solution of Equation (3.16) at t=5.

0.8

0.6

0.4

0.2

0
0

5
10
8
6
4
10 2
0

Figure 3.8: Numerical solution from the local method at t=5.

29
 −5
x 10

1.4

1.2

0.8

0.6

0.4

0.2

0
−5
0
5
10
0 −5
15 10 5
15

Figure 3.9: Local method error at t=5.

30
Chapter 4

PATTERN FORMATION IN MATHEMATICAL BIOLOGY

Many phenomena that occur in nature can be modeled using partial di↵erential equa-

tions. More specifically, two such biological examples that utilize time-dependent advection-

di↵usion-reaction equations are Turing patterns and chemotaxis. RBF methods can be im-

plemented efficiently on these types of problems with large data sets on complexly shaped

domains to approximate the solutions. However, as it will be seen, the global RBF method

is not always the best approach.

4.1 APPLICATION 1: TURING PATTERNS

Alan Turing (1912–1954) is perhaps most notably remembered as an early computer science

visionary who conceptualized a universal computing device now known as a Turing machine.

Later in his life, Turing became interested in modeling how genes of an organism can be

seen in physical traits, such as spots and stripes. Examples of Turing patterns can be seen

in Figures 4.2 and 4.1.

Turing patterns are modeled by a reaction-di↵usion system. In addition to having

a method for di↵using chemicals, this system has two primary parts. First, there is an

activator that can generate more of itself. Second, there is an inhibitor. The inhibitor slows

down the activator. Patterns are formed when the chemicals are a↵ected by the activator

and inhibitor and spread across a given region. In the 1980s, these patterns were able to

be reproduced on a layer of gel in a petri dish by chemists and simulated on computers to

show that they were indeed Turing patterns [18, 3].

31
Figure 4.1: Left: Many animals, such as the leopard, have prints that resemble Turing
patterns. Right: Galaxies are also believed to exhibit properties of Turing patterns [13].

Turing equations are a reaction-di↵usion model that have the form

ut = DO2 u + f (u, v) (4.1)

vt = O2 v + g(u, v).

The evolution of the chemical concentration is given by u(x,y,t) and v(x,y,t) at spatial

position (x,y) at time t. D is a constant di↵usion coefficient that is a ratio of the di↵usion

coefficients of the two chemical concentrations. The functions f and g model the reaction.

They are nonlinear reaction terms of the chemical. Under certain conditions, various pat-

terns, such as dots or stripes, form. In addition to parameter values, these patterns vary

depending on boundary conditions and the overall shape of the domain.

Consider the following Turing system

ut = D O2 u + ↵u(1 ⌧1 v 2 ) + v(1 ⌧2 u) (4.2)

↵⌧1
vt = O2 v + v(1 + uv) + u(↵ + ⌧2 v)

on a domain shaped like a butterfly. The boundary of the domain is outlined by the

32
Figure 4.2: Left: Turing patterns observed on a fish. Right: Computer generated Turing
pattern [13].

parametric curve

r(✓) = 3esin ✓ sin2 (2✓) + 3ecos ✓ cos2 (2✓), 0  ✓  2⇡. (4.3)

The parameters in the equations were chosen according to the results found in [18, 3]

where spotted patterns were formed with D = 0.516, = 0.0045, ↵ = 0.899, ⌧1 = 0.02,

= 0.91, ⌧2 = 0.2, and = ↵. Zero Dirichlet boundary conditions are applied to u and

v where u and v have initial conditions that were randomly chosen between -0.5 and 0.5.

On the butterfly domain, there are 8,125 centers, and the local RBF method is applied on

a stencil size of 100. The results demonstrate that spotted patterns can be formed from the

Turing system. This is illustrated in Figure 4.3. These findings are in qualitative agreement

with numerical and experimental results from [18, 3].

33
 t = 120
6

2
y

−2

−4
−5 0 5 10
x

Figure 4.3: A butterfly shaped domain with the solution to the Turing system from Equation
(4.2) at time t = 120.

4.2 APPLICATION 2: CHEMOTAXIS

Chemotaxis is an essential process in a variety of organisms ranging from the smallest

bacteria to the most complex vertebrates. The single-celled prokaryotic bacteria move away

from hostile conditions and toward nutrient clusters using chemotaxis. Similarly, the DNA

carrying eukaryotes use chemotaxis for immune responses and wound healing. So what

exactly is chemotaxis?

At its most basic level, chemotaxis is cell movement. The prefix “chemo” means “com-

bining form” whereas the suffix “taxis” is Greek for “arrange, turning.” By definition,

chemotaxis is the movement of a motile cell or organism, or part of one, in a direction cor-

responding to a gradient of increasing or decreasing concentration of a particular substance

[11].

34
 Budrene and Berg’s biological experiments on the patterns that Escherichia coli and

Salmonella typhimurium form provide the basis for our research [2]. Simply stated, when

bacteria were exposed to a liquid medium, patterns in the bacteria materialize and rearrange

before they eventually disappear. Their experiments showed that the biological processes

that caused the patterns formed by the bacteria were the results of random migration and

chemotaxis.

In order for the patterns to form, the bacteria were exposed to tricarboxylic acid (TCA).

Succinate and fumarate accounted for the most e↵ective results. In response to the TCA,

the bacteria released the strong chemoattractant aspartate. This chemoattractant is what

causes the cells to move in a direction corresponding to a particular gradient. This movement

increases cell density whereas di↵usion has the opposite e↵ect. These two combatting

forces is the primary reason behind the formation of the patterns. Hence, when studying

chemotaxis, we must focus on the cells, the stimulate, and the chemoattractant. In Budrene

and Berg’s experiments, these were respectively the bacteria, succinate and fumarate, and

aspartate [2].

Tyson, Lubkin, and Murray ultimately analyzed a second order ordinary di↵erential

equation modeled after Budrene and Berg’s experiments in order to understand the numer-

ical properties of chemotaxis [27]. Their research focused on the patterns formed by E. coli

and salmonella bacteria as they undergo chemotaxis in liquids. The dimensionless form of

the mathematical model that was used for the experiments as defined in [27] is


u
2
ut = du r u ↵r ⇧ rv (4.4)
(1 + v)2

u2
v t = r2 v + w , (4.5)
µ + u2

where u represents the cell density, v is the chemoattractant concentration, and w represents

the succinate concentration. For the experiments, the succinate is not consumed, and it is

uniformly distributed in the beginning. Hence, w is constant, and let u = 1, v = 0, and

w = 1. Also, r is the gradient operator, and r2 is the Laplacian operator.

35
 u
In order to simplify the calculations, let f = (1 v 2 )
and let v = hvx , vy i. Then Equa-

tion (4.4) becomes

ut = du (uxx + uyy ) ↵r ⇧ hf vx , f vy i
⇣ ⌘
= du (uxx + uyy ) ↵ (f vx )x + (f vy )y (4.6)

= du (uxx + uyy ) ↵ (fx vx + f vxx + fy vy + f vyy )

and Equation (4.5) becomes

✓ ◆
u2
vt = (vxx + vyy ) + w . (4.7)
µ + u2

In Matlab, this is coded as:

fp = zeros(N,2);

fp(:,1) = du*(d2*V(:,1)) ...

- alpha*((d2*V(:,2)).*(V(:,1)./(1+V(:,2)).^2)...

+ (d1x*V(:,2)).*(d1x*(V(:,1)./(1+V(:,2)).^2)))...

- alpha*((d1y*V(:,2)).*(d1y*(V(:,1)./(1+V(:,2)).^2)));

fp(:,2) = d2*V(:,2) + (w*V(:,1).^2)./(mu + V(:,1).^2);

Paramater Suggested Value Value Used in Experiment

↵ 80-90 80
du 0.25-0.5 0.33
µ Unknown 1

Table 4.1: Parameters for Equations (4.4) and (4.5)

The goal of this two dimensional chemotaxis experiment is to find solutions to (4.4) and

(4.5) that oscillate, grow, and finally decay as time progresses. To simplify the results, like

the experiments of Tyson, Lubkin, and Murray, we assumed that the bacteria do not die

36
[27]. Therefore, a death variable was not included in Equation (4.4). Similarly, even though

a chemoattractant is produced, there is no deterioration variable in Equation (4.5). Hence,

in the experiments the chemoattractant increases continually, and di↵usion eventually dom-

inates the e↵ects of chemotaxis.

As can be seen from Figures 4.4 and 4.5, the initially random bacteria arrange themselves

into high-density collections. It is obvious that the collections of cells are very dense as

compared to the regions between each collection. When t = 1, there is one cluster in the

center. However, when t = 75, it is evident that a set of continuous rings has developed.

If we continued to increase the amount of time that passes, it is likely that the clusters

will reach a steady-state pattern. This is exactly what happened in Tyson, Lubkin, and

Murray’s numerical experiments and Budrene and Berg’s biological experiments [2, 27].

37
 t = 20
t = 43
25 25

20 20

15 15

10 10

5 5

0
y

y
−5 −5

−10 −10

−15 −15

−20 −20

−25 −25
−30 −20 −10 0 10 20 30 −30 −20 −10 0 10 20 30
x x

t = 55 t = 65
25 25

20 20

15 15

10 10

5 5

0 0
y

−5 −5

−10 −10

−15 −15

−20 −20

−25 −25
−30 −20 −10 0 10 20 30 −30 −20 −10 0 10 20 30
x x

Figure 4.4: Density plots of the chemotaxis experimental results when Top Left: t=20, Top
Right: t=43, Bottom Left: t=55, Bottom Right: t=65

38
 t = 43
t = 20

30
1.5
25

20
1
15

10
0.5
5

0
0 40
40
20 40
20 40
0 20
0 20
0 0
−20 −20
−20 −20
y −40 −40 y −40 −40
x x

t = 55

t = 65

30

25 30

20 25

15 20

15
10
10
5
5
0
40 0
40
20 40
20 20 40
0 20
0 0
−20 0
−20 −20 −20
y −40 −40 −40
x y −40 x

Figure 4.5: Chemotaxis experimental results when Top Left: t=20, Top Right: t=43,
Bottom Left: t=55, Bottom Right: t=65

39
Chapter 5

CONCLUSION

Due to the extreme flexibility of the RBF methods, they have attracted much attention from

mathematicians and scientists. RBF methods overcome some of the deficiencies of polyno-

mial methods and can be applied on scattered data sites and in regions with complicated

boundaries to interpolate data and to solve di↵erential equations.

As seen in this thesis, both global and local methods can be used when working with

radial basis functions. Specifically, these methods were applied to Turing equations and

chemotaxis models, but the applications are endless. Similar to these examples, other situ-

ations arise that require tens of thousands of data points for centers, and desktop computers

cannot efficiently execute such large problems in a timely manner with the global method.

Not only is time an issue, but so is computer memory. The local method overcomes these

issues by allowing the user to select small stencils of points that can provide very similar

accuracy to the global method while using substantially less computer memory. As a result,

the local method can execute e↵ectively in a considerable less amount of time.

In today’s modern scientific world, numerical computing is vital. Today’s computing is

often done heterogeneously with both CPUs and GPUs. For a future direction of this work,

a high performance computer (HPC) cluster could be used to dramatically increase the

accuracy and efficiency of radial basis function methods. By using a HPC cluster, it would

be possible to carry out extensive calculations that require large amounts of data storage

and memory allocation in a vastly shorter time frame than the typical desktop computer.

Marshall University has recently installed a HPC cluster. Known as “BigGreen,” it

40
currently features 276 central processing unit cores, 552 gigabytes of memory, and more than

10 terabytes of storage. It currently has 8 NVIDIA Tesla M2050 GPU computing modules

installed with 448 cores each. This configuration provides support for extremely large

parallel computation with an estimated peak of six Teraflops. This is equal to six trillion

floating point operations per second. Because RBF methods are an active research area

with widespread application in engineering and science, formulating the methods to work

efficiently on multi-core processors will further enhance the popularity and applicability of

the RBF methods.

41
 Appendix A

MATLAB CODE

1 %
% d r i v e r D .m
3 %

5 CENTERS = 1 ; % step 1: load centers

STENCILS = 0 ; % s t e p 2 : s e t up t h e s t e n c i l s
7 WEIGHTS = 0 ; % step 3 : f i n d the s t e n c i l weights
GO = 0 ; % s t e p 4 : graph t h e s o l u t i o n and d e t e r m i n e t h e e r r o r
9

11 i f CENTERS
disp ( ’ Centers ’ )
13 xc = l i n s p a c e ( 1 ,1 ,100) ’;
xp = l i n s p a c e ( 1 ,1 ,200) ’;
15 end

17 i f STENCILS
disp ( ’ S t e n c i l s ’ )
19 ns = 17 % stencils size
s t = s t e n c i l s D ( xc , ns ) ;
21 end

23 i f WEIGHTS
d i s p ( ’ Weights ’ )
25 shape = 0 . 4 3 ; % i n i t i a l shape p a r a m e t e r
dc = 0 . 0 0 1 ; % shape p a r a m e t e r i n c r e m e n t
27 minK = 1 e5 ; % minimum c o n d i t i o n number o f t h e system m a t r i x
maxK = 1 e15 ; % maximum c o n d i t i o n number o f t h e system m a t r i x
29 D = weightsD ( s t , xc , shape , minK , maxK, dc ) ;

42
 end
31

i f GO
33 d i s p ( ’ Graphing ’ )
u = go ( xc , D, xp ) ;
35 end

1 %
% s t e n c i l s D .m
3 %

5 f u n c t i o n s t = s t e n c i l s D ( xc , ns )

7 N = l e n g t h ( xc ) ;

9 f o r i =1:N % s t e n c i l s f o r d e r i v a t i v e approximation
x0 = xc ( i ) ;
11 r = abs ( xc ( : ) x0 ) ; %d i s t a n c e between c e n t e r i and t h e r e s t o f t h e c e n t e r s
[ r , ix ] = sort ( r ) ;
13 s t ( i , : ) = i x ( 1 : ns ) ;
end

1 %
% weightsD .m
3 %

5 f u n c t i o n D = weightsD ( s t , xc , shape , minK , maxK, dc )

% INPUTS
7 % st indexes of the s t e n c i l c e n t e r s
% xc centers
9 % shape i n i t i a l shape p a r a m e t e r
% minK min c o n d i t i o n number o f t h e RBF m a t r i x ( 1 e +13)
11 % maxK max c o n d i t i o n number o f t h e RBF m a t r i x ( 1 e +15)
% dc shape p a r a m e t e r i n c r e m e n t
13 % OUTPUTS
% D weights to d i s c r e t i z e
15

warning o f f
17 N = l e n g t h ( xc ) ; % total centers
n = length ( st ( 1 , : ) ) ;
19 o = ones (1 , n ) ;

43
 D = s p a r s e (N, n ) ;
21

f o r i =1:N % interior centers

23

pn = s t ( i , : ) ;
25 r x = xc ( pn ) ⇤ o ( xc ( pn ) ⇤ o ) ’ ;
r = abs ( r x ) ;
27 K = 1;

29 w h i l e (K<minK | | K>maxK)
B = mq( r , shape ) ;
31 [ U, S ,V] = svd (B ) ;
K = S(1 ,1)/S(n , n ) ;
33 i f K<minK , shape = shape dc ;
e l s e i f K>maxK, shape = shape + dc ; end
35 end

37 Bi = V⇤ d i a g ( 1 . / d i a g ( S ) ) ⇤U ’ ;

39 h = m q D e r i v a t i v e s ( s q r t ( ( xc ( i ) xc ( pn ) ) . ˆ 2 ) , xc ( i ) xc ( pn ) , shape , 1 ) ;

41 D( i , pn ) = h ’ ⇤ Bi ;

43 end , warning on

%
2 % go .m
%
4

f u n c t i o n u = go ( xc , D, xp )
6

f = exp ( xc . ˆ 3 ) c o s ( 2 . ⇤ xc ) ;
8

f D e r i v a t i v e E x a c t = 3 . ⇤ exp ( xc . ˆ 3 ) . ⇤ xc . ˆ 2 + 2 . ⇤ s i n ( 2 . ⇤ xc ) ;
10

f D e r i v a t i v e A p p r o x = D⇤ f ;
12

f o r m a t l o n g , f o r m a t compact
14 p o i n t W i s e E r r o r s = abs ( f D e r i v a t i v e A p p r o x fDerivativeExact )
format
16

44
 mean ( p o i n t W i s e E r r o r s )
18

twoNormError = norm ( f D e r i v a t i v e A p p r o x f D e r i v a t i v e E x a c t , 2 )
20

p l o t ( xp , 3 . ⇤ exp ( xp . ˆ 3 ) . ⇤ xp . ˆ 2 + 2 . ⇤ s i n ( 2 . ⇤ xp ) , ’ g ’ , xc , f D e r i v a t i v e E x a c t , ’ r ⇤ ’ , ...
22 xc , f D e r i v a t i v e A p p r o x , ’ ko ’ )
xlabel ’x ’ , ylabel ’ f ˆ\ prime ( x ) ’

1 function advectionDiffusion1D ()

3 N = 51; %number o f c e n t e r s
dt = 0 . 0 0 5 ; %t i m e s t e p
5 finalTime = 0 . 5 ; %f i n a l time
a = 1; %v a r i a b l e i n e q u a t i o n advection c o e f f i c i e n t
7 nu = 0 . 0 0 2 ; %v a r i a b l e i n e q u a t i o n diffusion coefficient
shape = 6 ; %shape p a r a m e t e r
9

x = l i n s p a c e ( 0 , 1 ,N ) ’ ;
11 o = ones (1 , length ( x ) ) ;
r x = x⇤ o ( x⇤ o ) ’ ; % s i g n e d d i s t a n c e m a t r i x
13 r = abs ( r x ) ; % d i s t a n c e m a t r i x

15 H = z e r o s (N,N ) ; % e v a l u a t i o n m a t r i x

17 H( 1 , : ) = mq( r ( 1 , : ) , shape ) ; % D i r i c h l e t boundary c o n d i t i o n a t x=0

H( 2 : N 1 , : ) = nu . ⇤ m q D e r i v a t i v e s ( r ( 2 : N 1 , : ) , r x ( 2 : N 1 , : ) , shape , 2 ) . . .
19 a . ⇤ m q D e r i v a t i v e s ( r ( 2 : N 1 , : ) , r x ( 2 : N 1 , : ) , shape , 1 ) ;

21 B = mq( r , shape ) ; % system m a t r i x

dm = H/B ; % d i s c r e t i z a t i o n D, o f l i n e a r o p e r a t o r L
23

U = e x a c t S o l u t i o n ( x , 0 ) ; %i n i t i a l c o n d i t i o n
25

t =0;
27

while t < finalTime %w h i l e our time i s l e s s than our f i n a l time

29 u = r k 4 (U, t , dt ,@F ) ;
t = t + dt ; %advance i n time
31 u(1) = 1;
u (N) = e x a c t S o l u t i o n ( 1 , t ) ;
33 U=u ;

45
 end
35

exact = exactSolution (x , finalTime ) ;

37

p l o t ( x , exact , ’ r ’ , x , u , ’ b ’ )
39

figure
41 p l o t ( x , abs (U e x a c t ) )

43 f u n c t i o n f p = F( u , t )
u(1) = 1;
45 u (N) = e x a c t S o l u t i o n ( 1 , t ) ;
f p = dm⇤u ;
47 end

49 f u n c t i o n ex = e x a c t S o l u t i o n ( x , t )
i f t<dt
51 i f l e n g t h ( x)>1
ex ( 1 ) = 1 ;
53 ex ( 2 : l e n g t h ( x ) ) = 0 ;
ex = ex ( : ) ;
55 e l s e , ex = 0 ;
end
57 else
den = 2 . 0 ⇤ s q r t ( nu⇤ t ) ;
59 f r a c 1 = ( x t ) / den ;
f r a c 2 = ( x+t ) / den ;
61 ex = 0 . 5 ⇤ ( e r f c ( f r a c 1 ) + exp ( x/nu ) . ⇤ e r f c ( f r a c 2 ) ) ;
end
63 end

65 end

Matlab code used for the 2d advection-di↵usion-reaction problem, the Turing system,

and the chemotaxis problem was modified from reference [21].

46
REFERENCES

[1] J. P. Boyd, Chebyshev and Fourier spectral methods, second ed., Dover Publications,
Mineola, New York, 2000.

[2] E.O. Budrene and H.C. Berg, Complex patterns formed by motile cells of Escheria Coli,
Nature (1991), 349:630–633.

[3] J. L. Aragon C. Varea and R. A. Barrio, Turing patterns on a sphere, The American
Physical Society 60 (1999), no. 4, 4588–4592.

[4] B. Fornberg and N. Flyer, Accuracy of radial basis function interpolation and deriva-
tive approximations on 1-d infinite grids, Advances in Computational Mathematics 23
(2005), 37–55.

[5] B. Fornberg and G. Wright, Stable computation of multi quadratic interpolants for
all values of the shape parameter, Computers and Mathematics with Applications 47
(2004), 497–523.

[6] R. Franke, A critical comparison of some methods for interpolation of scattered data,
Technical Report NPS (1979), 53–79.

[7] D. Goldberg, What every computer scientist should know about floating-point arith-
metic, Computing Surveys (1991), 171–264.

[8] R. L. Hardy, Multiquadric equations of topography and other irregular surfaces, Journal
of Geophysical Research 76 (1971), no. 8, 1905–1915.

[9] , Theory and applications of the multiquadric-biharmonic method: 20 years of

discovery, Computers and Mathematics with Applications 19 (1990), no. 8/9, 163–208.

[10] I. R. H. Jackson, Convergence properties of radial basis functions, Constructive Ap-

proximation 4 (1988), no. 1, 243–264.

[11] T. Jin and D. Hereld, Chemotaxis methods and protocols, Humana Press, New York,
2009.

[12] E. J. Kansa, Multiquadrics – A scattered data approximation scheme with applications

to computational fluid dynamics I: Surface approximations and partial derivative esti-
mates, Computers and Mathematics with Applications 19 (1990), no. 8/9, 127–145.

[13] Brandon Keim, Alan Turing’s patterns in nature, and beyond, http://www.wired.
com/wiredscience/2011/02/turing-patterns/.

47
[14] W. R. Madych and S. A. Nelson, Bounds on multivariate interpolation and exponential
error estimates for multiquadric interpolation, Journal of Approximation Theory 70
(1992), 94–114.
[15] C. Micchelli, Interpolation of scattered data: Distance matrices and conditionally pos-
itive definite functions, Constructive Approximation 2 (1986), 1122.
[16] M. L. Overton, Numerical computing with IEEE floating point arithmetic, Society for
Industrial and Applied Mathematics, Philadelphia, 2001.
[17] R. B. Platte, How fast do radial basis function interpolants of analytic functions con-
verge?, IMA Journal of Numerical Analysis 31 (2011), no. 4, 1578–1597.
[18] J. L. Aragon R. A. Barrio, C. Varea and P. K. Maini, A two-dimensional numerical
study of spatial pattern formation in interacting turing systems, Bulletin of Mathemat-
ical Biology 61 (1999), 483–505.
[19] S. Rippa, An algorithm for selecting a good value for the parameter c in radial basis
function interpolation, Advances in Computational Mathematics 11 (1999), 193–210.
[20] S. A. Sarra, Radial basis function approximation methods with extended precision float-
ing point arithmetic, Engineering Analysis with Boundary Elements 35 (2011), no. 1,
68–76.
[21] , A local radial basis function method for advection-di↵usion-reaction equations
on complexly shaped domains., To appear in Applied Mathematics and Computation
(2012).
[22] S. A. Sarra and E. J. Kansa, Multiquadric radial basis function approximation methods
for the numerical solution of partial di↵erential equations, vol. 2, Advances in Compu-
tational Mechanics, 2009.
[23] T. Sauer, Numerical analysis, Pearson Education, Inc., Boston, 2006.
[24] R. Schaback, Error estimates and condition numbers for radial basis function interpo-
lation, Advances in Computational Mathematics 3 (1995), 251–264.
[25] L. N. Trefethen, Spectral methods in Matlab, SIAM, Philadelphia, 2000.
[26] L. N. Trefethen and III D. Bau, Numerical linear algebra, Society for Industrial and
Applied Mathematics, Philadelphia, Pennsylvania, 1997.
[27] R. Tyson, S.R. Lubkin, and J.D. Murray, Model and analysis of chemotactic bacterial
patterns in a liquid medium, Journal of Mathematical Biology (1998), 38:359–375.
[28] H. Wendland, Scattered data approximation, Cambridge University Press, Cambridge,
2005.
[29] Z. Wu and R. Schaback, Local error estimates for radial basis function interpolation of
scattered data, IMA Journal of Numerical Analysis 13 (1993), 13–27.
[30] J. R. Xaio and M. A. McCarthy, A local heaviside weighted meshless method for two-
dimensional solids using radial basis functions, Computational Mechanics 31 (2003),
301–315.

48
 Maggie Elizabeth Chenoweth
[email protected]

Education
• Master of Arts. Mathematics. Marshall University, May 2012. Thesis Advisor: Scott
Sarra.

• Bachelor of Arts. Double major in Applied Mathematics and Math Education 5-

Adult. Minor in Computer Science. Marshall University, May 2010, Summa cum
laude.
Teaching Experience
• Student Teaching - Chesapeake Middle School (John Gibson) and Chesapeake High
School (Dee Rucker) - Spring 2010

• MTH 099 - Mathematics Skills II - Spring 2011

• MTH 121 - Concepts and Applications (Critical Thinking) - Fall 2010 and Fall 2011

• MTH 127 - College Algebra (Expanded) - Summer 2011 and Fall 2011

• MTH 225 - Introductory Statistics - Spring 2012

Publications
1. A Numerical Study of Generalized Multiquadric Radial Basis Function Interpolation.
Maggie Chenoweth. SIURO 2 n. 2 (October 2009), 58–70.

2. A Local Radial Basis Function Method for the Numerical Solution of Partial Di↵er-
ential Equations. Master’s thesis, Marshall University, May 2012.
Professional Affiliations
• Pi Mu Epsilon - Former Chapter President at Marshall University

• National Society of Collegiate Scholars

• Golden Key International

• Gamma Beta Phi

• Sigma Xi

• Society for Industrial and Applied Mathematics

• Phi Kappa Phi

• Kappa Delta Pi

• Mathematical Association of America

• American Mathematical Society

49
Awards and Recognitions

• John Marshall Scholar

• Graduate of the Marshall University Honors Program

• Dean’s List

• National Dean’s List

• 200 Level Book Award for the Marshall University Honors Program

• NASA Space Grant Recipient

• Presenter at the West Virginia Undergraduate Research Day at the Capitol

• Poster Presentation Winner at Marshall’s Cyberinfrastructure Day

• Marshall Math Department Junior Student of the Year

• Outstanding Graduate in Mathematics Award

• Outstanding Graduate in the College of Education Award

50