i

Nonconforming Least-Square Spectral Element Method for

Parabolic Partial Differential Equation with Non-Smooth Data
and Application to Finance

by
Arbaz Khan

Department of Mathematics and Statistics

Indian Institute of Technology, Kanpur

August, 2014
 Nonconforming Least-Square Spectral Element Method for
Parabolic Partial Differential Equation with Non-Smooth Data
and Application to Finance

A Thesis Submitted

in Partial Fulfilment of the Requirements

for the Degree of

DOCTOR OF PHILOSOPHY

by

ARBAZ KHAN

to the
Department of Mathematics and Statstics
Indian Institute of Technology, Kanpur
August, 2014
Synopsis

Name of the Student : Arbaz Khan

Roll Number : Y8108065
Degree for which
submitted : Ph.D.
Department : Mathematics and Statistics
Thesis Title : Nonconforming Least-Square Spectral Element Method
for Parabolic Partial Differential Equation with Non
-Smooth Data and Application to Finance
Thesis Supervisor : Prof. Pravir Dutt and Prof. C. S. Upadhyay
Month and year of
submission : August, 2014

Several methods have been proposed in the literature for solving the Black-Scholes
equation for European options. Recently, Spectral methods have been used to solve Option
Pricing problems. In 2000, Bunin et. al. [24] proposed Chebyshev Collocation methods to
solve the European Call Option problem on parallel computers. After this, Greenberg [47]
solved American Options problem by Chebyshev Tau method. For smooth initial condi-
tions, De Frutos [45] has presented a Laguerre-Galerkin Spectral Method to price bonds.
More recently, Zhu et. al. [77] have used a Spectral element method to price European
Options. These methods give quadratic accuracy in time, while being spectrally accurate
in space. Dominik et. al. [69] proposed hp-version of the Discontinuous Galerkin Finite
Element Method to solve parabolic problems.
In this thesis we develop a Non-Conforming Least-Squares Spectral Element Method
(LSSEM) to solve the Black-Scholes equation for European Options on parallel computers.

vii
LSSEM are spectrally accurate in both time and space. Sobolev spaces of different orders
in space and time are used for the results, as presented in [57]. If the data belong to certain
Gevrey spaces then the solution also belongs to a Gevrey space [58].

The proposed method is a Least-Squares method as presented in [37]. The space domain
is an interval which is divided into a number of sub-intervals. The functional is the sum
of the squares of the residuals in the partial differential equation and initial condition in
different Sobolev norms, and a term which measures the jump in the function and its
derivatives across inter-element boundaries in appropriate fractional Sobolev norms. We
minimize the functional on a given time interval. Hermite mollifiers, as described in [73, 74],
are used to resolve the difficulty of non-smooth initial data.
The whole thesis is divided into five chapters. Now we briefly discuss the content of all
chapters.

Chapter 1 In this chapter, we give a brief description of different methods, namely spec-
tral and finite element methods, with advantages and disadvantages. Basic properties
of these methods are discussed. An overview of the existing work is also provided.

Chapter 2 In this chapter the main concern is to recover the discontinuous function from
its spectral representation. Here we mainly concentrate on two aspects of processing
such piecewise smooth data:

Edge detection. Detecting the location and amplitude of the edges. Often, these
are essential features sought in piecewise smooth data. Moreover, they are the
boundaries of the region of smoothness and are therefore essential for the second
aspect. For our problem the region of smoothness is known.

Reconstruction. Recovering the underlying function f inside its different regions

of smoothness.

Chapter 3 discusses the non- conforming least-squares spectral element method for Eu-
ropean options with single asset on parallel computers. The method proposed in

viii
 the current study achieves spectral accuracy in both space and time. The method
is based on minimization of a functional given in terms of the sum of squares of the
residuals in the partial differential equation and initial condition in different Sobolev
norms, and a term which measures the jump in the function and its derivatives across
inter-element boundaries in appropriate fractional Sobolev norms. To obtain values
of the solution and its derivatives the initial condition is mollified and the computed
solution is post processed. Error estimates are obtained for this method. Specific
numerical examples are given to show the efficacy of this method.

Chapter 4 presents the non- conforming least-squares spectral element method for Eu-
ropean options with multi asset on parallel computers.

Chapter 5 briefly outlines our plan for future work.

ix
 Dedicated
to

My Family
(Dr. Pravir Dutt, Dr. C. S. Upadhyay,
Mr. Asif Khan, Mrs. Usha Khan,
Dr. Zeba Khan, Mr. Shahbaj Khan,
Dr. Hina Vasishtha,
Mr. Arif Siddiqui, Mr. Shahnawaz Khan)
&
Mr. (Late) Anand Narayan Pandey
 Acknowledgements

Firstly, I thank Almighty Allah for the uncountable blessings that He has bestowed on me.
I thank Him for all the opportunities showered on me in my life.

Completing my PhD perhaps has been the most challenging task in my life. I shared
the best and worst moments of my doctoral journey with many people. It has been a
privilege to spend several years in the Department of Mathematics, IIT Kanpur and its
members will always remain dear to me.

The list of the people I need to thank would not fit in a single Acknowledgement. I
just mention some people whose contributions are significant.

I would like to express my deepest gratitude, regards and thanks to my supervisor Prof.
Pravir Dutt for his support, ideas and encouragement throughout my doctoral research.
His mathematical expertise, excellent guidance, availability and wisdom proved very stim-
ulating and helpful. I profoundly thank and express deep regards to my co-advisor Dr. C.
S. Upadhyay for his discussions, invaluable suggestions, comments, motivation and help at
all stages of my research period.

I sincerely thank all the faculty members of the Department of Mathematics and Statis-
tics. In particular, I am grateful of Dr. Akash Anand, Dr. A. K. Lal, Dr. B. V. Rathish
Kumar, Dr. G. Santhanam, Dr. M. K. Kadalbajoo, Dr. Peeyush Chandra, Dr. Prawal
Sinha, Dr. Sameer Chavan, Dr. Shobha Madan, Dr. S. Dutta, Dr. T. Muthukumar and
Dr. V. Raghavendra for their teaching and guidance in completing the academic course
work and their advise and cooperation in many ways.

I would like to express my sincere thanks to Dr. Akhlaq Husain for helping me in

xiii
developing the numerical code. I also owe many thanks to our badminton team with
whom I played at Prof. Pravir’s house. My heartiest thanks to all my friends at Hall IV,
who made my stay so enjoyable.
I express my gratitude to my undergraduate teacher and my friend Dr. Abhishek
Agrawal for their inspiration, motivation and teachings in academics. Hardiest thanks to
my childhood friends Imran, Kuldip, Surendra and Tabrej.
I am indebted to the CSIR (Council of Scientific and Industrial Research, New Delhi)
and IIT Kanpur for providing me financial support that facilitated the complition of his
dissertation.
I thank all my friends for their companionship and I will cherish their memories for long.
Special thanks to Dr. Abdur Raheem, Dr. Abdullah, Alpesh, Ambuj, Arshad, Ashiq, Arun,
Awanish, Imran, Lokpati, Dr. Manoj, Dr. Maqbul, Musheer, Noman, Peeyush, Pradip,
Rashid, Dr. Ravi, Saurabh, Shariq, Dr. Shiblee, Dr. Subhashani, Sumit, Surjit, Shweta,
Trivesh, Waqur, Yousuf.
I would thank my family from the bottom of my heart but for them my heart has no
bottom. This thesis is dedicated to them.
The persons who I always looked up to for inspiration, advice and guidance are my
brother Shahbaj Khan and my sister Dr. Zeba Khan. Many thanks to all my relatives,
especially, Arif Siddiqui, Neha and Shahnawaz Khan for their unconditional love.
None of this could have happened without the love and support of Chandni. Her
patience and unwavering love were undeniably the bedrock upon which the past six months
of my life have been built. Her tolerance of my occasional moods is a testament in itself
of her unyielding devotion and love.
I would not be where I am today if not for my parents who sacrificed their so much
for me. I am forever in their debt and I hope this thesis is a worthy reflection of their
commitment. They have always given me the strength and wisdom to be sincere in my
work, setting high moral standards and supporting me through their hard work.

August, 2014 Arbaz Khan

xiv
 Contents

Acknowledgements xiii

Contents xv

List of Figures xix

List of Tables xxi

1 Introduction 1
1.1 European Call and Put Options . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Black-Scholes Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Spectral Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3.1 Galerkin Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.2 Collocation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.3 Tau Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.4 Advantages of Spectral Methods . . . . . . . . . . . . . . . . . . . . 7
1.3.5 Drawback of Spectral Methods . . . . . . . . . . . . . . . . . . . . 8
1.4 Finite Element Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4.1 Rayleigh-Ritz Method . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4.2 Galerkin Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4.3 Least-Squares Method . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4.4 Drawback of Least-Squares Formulation . . . . . . . . . . . . . . . 11
1.5 Least-Squares Spectral Element Methods . . . . . . . . . . . . . . . . . . . 11

xv
 1.6 Non-conforming methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6.1 Iterative Patching . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6.2 Constrained Approximation . . . . . . . . . . . . . . . . . . . . . . 13
1.6.3 Mortar Element Method . . . . . . . . . . . . . . . . . . . . . . . 13
1.6.4 Discontinuous Galerkin method . . . . . . . . . . . . . . . . . . . . 13
1.6.5 Interior Penalty Methods . . . . . . . . . . . . . . . . . . . . . . . 13
1.7 Review of Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.8 Organization of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2 Methods for the reconstruction of non smooth data 17

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Concentration Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Mollifiers and Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4 Spectral Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.5 Reconstruction of Piecewise Smooth Data . . . . . . . . . . . . . . . . . . 22
2.6 Spectral Mollifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.7 Spectral Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3 Nonconforming LSSEM for European Options with Single Asset 29

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Function Space For the Formulation . . . . . . . . . . . . . . . . . . . . . 31
3.3 Discretization and Stability Estimates . . . . . . . . . . . . . . . . . . . . . 34
3.3.1 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3.2 Stability Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.4 Numerical Scheme and Parallelization . . . . . . . . . . . . . . . . . . . . 37
3.4.1 Symmetric Formulation . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4.2 Parallelization and Preconditioning . . . . . . . . . . . . . . . . . . 45
3.5 Estimates in Negative Norms . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.6 Error Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

xvi
 3.7 Computational Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

4 Nonconforming LSSEM for European Options with Multi Assets 67

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Function Space For the Formulation . . . . . . . . . . . . . . . . . . . . . 69
4.2.1 Definitions and Prelimineries . . . . . . . . . . . . . . . . . . . . . . 69
4.2.2 Black-Scholes Equation with Two Assets . . . . . . . . . . . . . . . 70
4.3 Discretization and Stability Estimates . . . . . . . . . . . . . . . . . . . . . 73
4.3.1 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.3.2 Stability Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4 Numerical Scheme and Parallelization . . . . . . . . . . . . . . . . . . . . 76
4.4.1 Symmetric Formulation . . . . . . . . . . . . . . . . . . . . . . . . 79
4.4.2 Parallelization and Preconditioning . . . . . . . . . . . . . . . . . . 80
4.5 Estimates in Negative Norms . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.6 Error Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.7 Computational Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

5 Conclusion and Future Work 99

5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

Bibliography 103

xvii
 List of Figures

1.1 Put Option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Call Option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Least-Squares Spectral Element Methods [67] . . . . . . . . . . . . . . . . . . 12

3.1 Inter Element Boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.2 Parallelization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3 Numerical and Exact solution at t=1 . . . . . . . . . . . . . . . . . . . . . . . 55
3.4 Derivative (ux ) of Numerical and Exact solution . . . . . . . . . . . . . . . . 56
3.5 Point-wise error between derivative of Numerical and Exact solution . . . . . . 56
3.6 Second Derivative (uxx ) of Numerical and Exact solution . . . . . . . . . . . . 57
3.7 Point-wise error between second derivative (uxx ) of Numerical and Exact solution 57
3.8 Numerical solution and Exact solution at t=1 . . . . . . . . . . . . . . . . . . 59
3.9 Derivative (∆) of Numerical and Exact solution . . . . . . . . . . . . . . . . . 59
3.10 Point-wise error between derivative (∆) of Numerical and Exact solution . . . 60
3.11 Second derivative (Γ) of Numerical and Exact solution . . . . . . . . . . . . . 60
3.12 Point-wise error between second derivative (Γ) of Numerical and Exact solution 61
3.13 Numerical solution and Exact solution at t=1 . . . . . . . . . . . . . . . . . . 63
3.14 Derivative (∆) of Numerical and Exact solution at t=1 . . . . . . . . . . . . . 63
3.15 Point-wise error between derivative (∆) of Numerical and Exact solution . . . 64
3.16 Second derivative (Γ) of Numerical and Exact solution at t=1 . . . . . . . . . 64
3.17 Point-wise error between second derivative (Γ) of Numerical and exact solution 65

xix
4.1 Max. point-wise error of Numerical and Exact solution at t=1 . . . . . . . . . 90
4.2 Max. point-wise error of first derivative (∆x ) of Numerical and Exact solution 90
4.3 Max. point-wise error of first derivative (∆y ) of Numerical and Exact solution 91
4.4 Max. point-wise error of second derivative (Γxx ) of Numerical and Exact solution 91
4.5 Max. point-wise error of second derivative (Γyy ) of Numerical and Exact solution 92
4.6 Max. point-wise error of mixed derivative (Γxy ) of Numerical and Exact solution 92
4.7 Max. point-wise error of Numerical and Exact solution at t=1 . . . . . . . . . 94
4.8 Max. point-wise error of first derivative (∆x ) of Numerical and Exact solution 94
4.9 Max. point-wise error of first derivative (∆y ) of Numerical and Exact solution 95
4.10 Max. point-wise error of second derivative (Γxx ) of Numerical and Exact solution 95
4.11 Max. point-wise error of second derivative (Γyy ) of Numerical and Exact solution 96
4.12 Max. point-wise error of mixed derivative (Γxy ) of Numerical and Exact solution 96

xx
 List of Tables

3.1 Point-wise error as function of p . . . . . . . . . . . . . . . . . . . . . . . . . . 55

3.2 Put Option problem : Variable value from Zhu et al [77] . . . . . . . . . . . . 58
3.3 Put Option problem : Point-wise error as function of p for LSSEM . . . . . . 58
3.4 Call Option problem : Variable value from Bunnin et al [24] . . . . . . . . . . 62
3.5 Call Option : Point-wise error as function of p for LSSEM . . . . . . . . . . . 62
3.6 Call Option problem : Point-wise error, as reported in Bunnin et al [24] . . . . 62

4.1 Maximum Point-wise error as function of p . . . . . . . . . . . . . . . . . . . . 89

4.2 Rainbow Option problem : Variable value from Zhu et al [78] . . . . . . . . . 93
4.3 Rainbow Put Option problem : Point-wise error as function of p for LSSEM . 93

xxi
 Chapter
1
Introduction
There are many fields in which we need to solve parabolic initial boundary value problems
with non-smooth data. In this thesis we shall examine some problems in Computational
Finance and specifically parabolic initial value problems for European Options.

1.1 European Call and Put Options

Now we briefly discuss the European Call Option and European Put Option. In European
Call Option the holder has the right, but not the obligation, to buy an asset at a prescribed
price (strike price) at maturity time in future. In European Put Option the holder has the
right, but not the obligation, to sell an asset at a prescribed price (strike price) at maturity
time in future. But the American Option can be exercised before maturity time. This is
the basic difference between European Option and American Option.
Assume that the stock price of European Call Option at time t is St , expiration date T
and the strike price K. Then two cases arise:

• If K > ST , then the European Call Option will not be exercised.

• If K < ST , then the European Call Option will be exercised and the holder will make
the profit of St − K.

1
 Chapter 1. Introduction

Hence the nature of the payoff function of European Call Option is

CT = (ST − K)+ = max(ST − K, 0).

Similarly, suppose that the stock price of European Put Option at time t is St , expiration
date T and the strike price K. Then two cases arise:
• If K < ST , then the European Put Option will not be exercised.
• If K > ST , then the European Put Option will be exercised and the holder will make
the profit of K − St .
Then the payoff function of the European Put Option is

PT = (K − ST )+ = max(K − ST , 0).

Figure 1.1: Put Option

1.2 Black-Scholes Model

In the early 1970s, Fisher Black, Myron Scholes and Robert Martin achieved a break-
through in the field of Option Pricing [18, 48]. This breakthrough is known as Black-Scholes
(Black-Scholes-Martin) model.

2
 1.3. Spectral Methods

Figure 1.2: Call Option

1.3 Spectral Methods

Spectral Methods are a class of techniques which are used to solve differential equations
numerically. They are one of the most accurate methods for solving partial differential
equations. Others methods which are used are Finite Difference Methods (FDM) and Fi-
nite Element Methods (FEM).
Spectral Methods are also considered as high order finite element methods due to their
so called spectral/exponential accuracy and use of high order polynomials for computing
numerical solution. The very high accuracy of spectral methods allows us to treat prob-
lems which would require an enormous number of grid points by finite difference or finite
element methods with much fewer degrees of freedom.
Bilinova proposed spectral methods in 1940 [19]. The first implementation of spectral meth-
ods was given by Silberman [71], but abandoned in the mid-1960s. Orszag [63] and Eliasen,
Machenhauer and Rasmussen [40] resurrected it again. Gottlieb and Orszag [46] provided
the first unified mathematical formulation of the theory of modern spectral methods for
the numerical solution of partial differential equations. Multi-dimensional discretizations

3
 Chapter 1. Introduction

were formulated as tensor products of one-dimensional constructs in separable domains.

Since then spectral methods were extended to a broader class of problems. They were
thoroughly analyzed in the 1980s and entered the mainstream of scientific computation
in the 1990s. The text book of Canuto, Hussaini, Quarteroni and Zang [25] focuses on
fluid dynamics algorithms and includes both practical as well as theoretical aspects of
global spectral methods. A companion book Spectral Methods, Fundamentals in Single
Domains by Canuto, Hussaini, Quarteroni and Zang [26] is focused on the essential aspects
of spectral methods on separable domains. The book Spectral/hp Element Methods for
Computational Fluid Dynamics, by Karniadakis and Sherwin [52], deals with many impor-
tant practical aspects of computations using spectral methods and summarizes the recent
research in the subject. In the latest book by Bochev and Gunzburger [20] the least-squares
finite element method (LSFEM) for partial differential equations has been described.
The first practitioners of spectral methods were meteorologists studying global weather
modelling and fluid dynamicists investigating isotropic turbulence. The original idea was
to use truncated Fourier series to approximate the solution when the problem was specified
with periodic boundary conditions. In order to tackle problems with more general bound-
ary conditions (Dirichlet or Neumann type), the set of algebraic polynomials replaced the
set of truncated series, but the characterization of the unique discrete function that would
provide the numerical solution was still achieved following the original strategy.
The key components for spectral methods are the trial functions (also called the expansion
or approximating functions) and the test functions (also known as weight functions). The
trial functions, which are linear combinations of suitable trial basis functions, are used to
provide the approximate representation of the solution. The test functions are used to
ensure that the differential equation and some boundary conditions are satisfied as closely
as possible by the truncated series expansion. This is achieved by minimizing, with respect
to a suitable norm, the residual produced by using the truncated expansion instead of the
exact solution. For this reason they may be viewed as a special case of the method of
weighted residuals (Finlayson and Scriven, [43]). An equivalent requirement is that the
residual satisfy a suitable orthogonality condition with respect to each of the test functions.

4
 1.3.1. Galerkin Method

From this perspective, spectral methods may be viewed as a special case of Petrov-Galerkin
methods (Zienkiewicz and Cheung [79], Aziz and Babuška [5]).
The most frequently used approximation functions (trial functions) are trigonometric poly-
nomials, Chebyshev polynomials and Legendre polynomials. Generally, trigonometric poly-
nomials are used for periodic problems whereas Chebyshev and Legendre polynomials are
used for non-periodic problems. Laguerre polynomials are used for problems on semiinfi-
nite domains and Hermite polynomials for problems on infinite domains.
Boyd [22] contains a wealth of detail and advise on spectral algorithms and is an espe-
cially good reference for problems on unbounded domains and in cylindrical and spherical
co-ordinate systems. A thorough analysis of the theoretical aspect of spectral methods for
elliptic equations was provided by Bernardi and Maday [13].

1.3.1 Galerkin Method

The Galerkin approach [50] has the feature that the test functions and the trial function
are the same. The test functions which are infinitely smooth functions, individually satisfy
some or all of the boundary conditions. The differential equation is enforced by requiring
that the integral of the residual times each test function be zero, after some integration
by parts, accounting in the process for any remaining boundary conditions. The first
serious application of spectral methods to PDE’s - that of Silberman [71] for meteorological
modeling - was a Galerkin method. Suppose that the unknown function u(x) can be
approximated by a sum of the N + 1 basis functions φn (x):
N
X
uN (x) = ai φi (x). (1.3.1)
i=0

Then the approximated function is substituted in the following differential equation

Lu(x) = f (x), (1.3.2)

where L is the differential operator and f (x) is the forcing function, and the result is called
the residual function which is defined by

RN = LuN (x) − f (x). (1.3.3)

5
 Chapter 1. Introduction

For the exact solution, the residual function RN (x; ai ) must be equal to zero. In the
Galerkin approach, the series coefficients {an } are chosen in a way in which the residual is
minimized in some norm.

1.3.2 Collocation Method

The collocation approach appears to have been first used by Slater [72] and by Kantorovic
[51] in specific applications. The collocation approach as a general method for solving
ordinary differential equation is developed by Frazer et. al. [44]. In this approach, Frazer
et. al. [44] used a variety of trial functions and arbitrary distribution of collocation points.
The work of Lancoz [55] first established the approach for the proper choice of the trial
function and the distribution of collocation points which is very crucial to the accuracy
of the solution. For partial differential equations, the earliest applications of spectral
collocation method were made for spatially periodic problems by Kreiss and Oliger [54]
(who called it Fourier method) and Orszag [64], who termed it pseudo-spectral. The
main feature of collocation method is the approach of choosing the numerical solution
which satisfies the original partial differential equation at some suitably chosen collocation
points. The choice of a set of collocation points is of fundamental importance for the
method and the number of collocation points must be equal to the dimension of the space
of the approximation. Otherwise, the problem could, in general, be over or under specified.
In the collocation method, we generalize the galerkin method by using different spaces of
trial functions and test functions. The approximation of u is written as

N
X
N
u (x, t) = ai (t)φi (x), (1.3.4)
i=0

where {φi }i is space of trial functions. Now consider the space of the test functions {ψi }i
which is different from the space of trial function and impose the orthogonality condition

(LuN (x, t), ψi ) = 0; f or all i = 0, · · · , N. (1.3.5)

6
 1.3.3. Tau Method

∂ ∂
where L = ∂t
+ ∂x . Now we choose ψi = δ(x − xi ) for a suitably chosen set of points {xi }i
and obtain the set of equations

(LuN (xi , t)) = 0 f or all i = 0, · · · , N. (1.3.6)

The points xi are chosen as the quadrature points of a Gaussian integration formula.

1.3.3 Tau Method

The spectral tau method was introduced by Lanczos [55]. This method is similar to
Galerkin method in the way the differential equation is enforced. In this method the trial
basis function need not satisfy the boundary conditions. Hence some equations are needed
to ensure that the global expansion satisfies the boundary conditions. Tau method which
may be viewed as a special case of Petrov-Galerkin, is applicable to non-periodic boundary
value problem.

1.3.4 Advantages of Spectral Methods

In this section we discuss some important features of spectral methods [46]. The important
features are the following :

1. Rate of convergence : If the solution of the given problem is infinitely smooth then
spectral methods give spectral accuracy. If the solution is analytic then spectral
methods give exponential accuracy. FDM and FEM give finite-order rates of conver-
gence. It is important to note that spectral methods can achieve high accuracy with
moderate resolution.

2. Boundary conditions: The mathematical features of spectral methods follow very

closely those of the partial differential equation being solved. Thus the boundary con-
ditions imposed on spectral approximations are normally the same as those imposed
on the differential equation. In contrast, FDM of higher order than the differential
equation require additional “boundary conditions”. Many of the complications of
finite-order FDM disappear with the infinite-order-accurate spectral methods.

7
 Chapter 1. Introduction

Another aspect of the treatment of boundary conditions by spectral methods is their

high resolution of boundary layers. For details we refer to [46, 65].

3. Discontinuities: Spectral methods do a better job of localizing error than difference

schemes [46, 65].

4. Bootstrap estimation of accuracy: It is often possible to estimate the accuracy of

spectral computations by examination of the shape of the spectrum. On the other
hand, similar criteria for finite-difference methods can be misleading.

1.3.5 Drawback of Spectral Methods

The drawback of spectral methods is their inability to handle complex geometries. But
these difficulties can be overcome by combining domain decomposition techniques with
spectral discretization (S. A. Orszag [65]).

1.4 Finite Element Methods

From mid 20th century, finite element methods (FEM) have emerged as one of the most
powerful methods for obtaining the approximate solution to a large variety of engineering
problems. The finite element method has gained great popularity due to its flexibility in
dealing with complex geometries. It has three versions which are as follows: h version, p
version and h − p version. In h version, the degree of polynomial is fixed and the mesh size
is reduced for increasing accuracy. In p version, the mesh size h is fixed and the degree of
polynomial is increased for gaining high accuracy. h − p version FEM is the combination of
h version and p version. In this approach, the degree of polynomial is increased and mesh
size is reduced to obtain high accuracy.
Finite element method which is based on variational method of approximation, can
be divided into three major groups: The Rayleigh-Ritz Method, Galerkin Method and
Least-Squares Method.

8
 1.4.1. Rayleigh-Ritz Method

1.4.1 Rayleigh-Ritz Method

The basic idea of Rayleigh-Ritz method is seeking to minimize the total potential energy,
and hence the Rayleigh-Ritz finite element method possesses the best approximation prop-
erty. It means that the difference between the finite element solution and exact solution is
minimized with respect to a certain energy norm. Another feature of Rayleigh-Ritz finite
element method is that it produces a symmetric and positive-definite system of linear al-
gebraic equation. The widest use of Rayleigh-Ritz method is in engineering analysis where
commercial codes are based on this method.

1.4.2 Galerkin Method

The Galerkin method which is also known as Bubnov-Galerkin method, is based on the
weighted residual form. To describe the basic idea of the Galerkin method, consider the
model problem which is defined by a set of partial differential equations

Au = f in Ω,
(1.4.1)
Bu = 0 on Γ.
Here A, B, u, f, Ω and Γ are respectively the linear differential operator, the boundary
operator, the dependent unknown vector, the force vector, the domain and the boundary
of the domain.
In this method, the function is written in terms of unknown parameters uj and the trial
(basis) functions Φj (x) in approximation sense.
n
X
u≈ Φj uj . (1.4.2)
j=1

The algebraic equation permitting a numerical solution is formed as a “weighted residual”,

Z Z
vi (Au − f )dΩ + v̄iT Bu dΓ = 0, i = 1, 2, ......, n.
T
(1.4.3)
Ω Γ

where vi and v̄i are “suitably chosen” test functions, and T denotes transpose. In the
conventional Galerkin method

vi = v̄i = Φi . (1.4.4)

9
 Chapter 1. Introduction

1.4.3 Least-Squares Method

The basic idea of least-squares method is due to Gauss. The least-squares finite element
method (LSFEM) is based on the minimization of the residuals in a least-squares sense.
The least-squares approach has the following significant advantages.

Universality : To use different numerical schemes for different types of differential equa-
tions has become a standard practice. In finite difference methods there are various
difference schemes for a particular problems. In finite element methods, there are
many approaches such as classical Galerkin, Petrov-Galerkin, Taylor-Galerkin, etc
which are used for different problems and have different structures and principles.
On the contrary, the LSFEM has a unified formulation for numerical solution of
partial differential equations [50].

Efficiency : In case of linear PDEs, the LSFEM always leads to symmetric positive-
definite matrices [21] which can be efficiently solved by matrix-free iterative methods,
for example the preconditioned conjugate gradient method (PCGM) [28]. Moreover,
we can solve large-scale problems with reasonable computer memory and in a rel-
atively small amount of time [33]. The methods can be easily parallelized [29]. In
spite of this, the method can be implemented with an efficient element-by-element
approach and it does not require global assembly of the local matrices [28, 30].

Robustness : For the solution of those problems which arise in pure convection and high-
speed compressible flows, the LSFEM inherently contains a mechanism to capture
shocks or discontinuities [21]. Moreover, the LSFEM can be applied without the
limitation of the Ladyzhenskaya Babuska-Brezzi (LBB) condition; that is equal order
elements, which make the programming much easier, can be used [21].

Optimality : There are many cases where it can be rigorously proven that LSFEM solu-
tion is the best approximation. That means, the error of the LSFEM solution has the
same order as the interpolation error. It meets the need for a posteriori error anal-
ysis by supplying an error indicator in the form of the residual which is minimized

10
 1.4.4. Drawback of Least-Squares Formulation

by the procedure. To achieve optimal solution, the error indicator can be used for
adaptive refinement. This information can not be supplied in other methods without
additional calculation.

1.4.4 Drawback of Least-Squares Formulation

A disadvantage of the least-squares method is that the governing equation must be trans-
formed into first-order system to mitigate the continuity requirement between neighbouring
elements and to keep the condition number of the resulting discrete system under control
[6]. Another drawback is the requirement of higher regularity of the finite element spaces
in the basic least-squares formulation.

1.5 Least-Squares Spectral Element Methods

Least-squares spectral element methods, which are based on two important and successful
methods: spectral/hp element methods and least-squares finite element methods, seem
very powerful because of the accuracy of spectral methods with the generality of finite
element methods. The LSSEM also have the theoretical and computational advantage in
the algorithmic design and implementation of the least-squares methods. To obtain the
prescribed level of accuracy, spectral element methods need fewer degrees of freedom in
comparison with finite element methods, but the amount of work that needs to be done
per degree of freedom is higher.

1.6 Non-conforming methods

Domain decomposition methods, which provide a powerful tool for approximating solution
of partial differential equations, decompose the complex domain into sub-domains with
or without overlap and then different numerical techniques can be applied appropriately.
The process of discretizing of sub-domains independently may lead to non-matching grids
across inter-element boundaries. The main goal of studying non-conforming methods is to
achieve the continuity of the traces of the solutions across the inter-element boundaries.

11
 Chapter 1. Introduction

Figure 1.3: Least-Squares Spectral Element Methods [67]

Here, we discuss the major kinds of non-conformities.

Geometrical non-conformities : The situation, where the intersection of two sub-

domains during division of domain is neither an empty set nor a face/edge nor a
vertex, is called Geometrical non-conformities.

Overlapping non-conformities : Domain decomposition may allow overlap between

sub-domains with different discretization in each.

Functional non-conformities : In finite element methods and spectral element meth-

ods, discrete space on each sub-domain lie in local variational space and may not
be in global one. Moreover, the matching conditions, which are employed along the
inter element boundary, may not ensure conformity. Then this situation is called
functional non-conformity.

Now we describe some non-conforming methods which are as follows :

12
 1.6.1. Iterative Patching

1.6.1 Iterative Patching

This formulation employs geometrically non-conforming elements but maintains C 0 conti-
nuity of the global polynomial expansion.

1.6.2 Constrained Approximation

The method of constrained approximation was introduced by Oden and his associates [32,
62, 68] to deal with geometrically non-conforming discretizations introduced by refinement.
The basic idea is to maintain C 0 continuity across elemental interfaces by modifying the
unconstrained basis functions appropriately.

1.6.3 Mortar Element Method

In this method C 0 continuity is no longer imposed and new weak forms of the problem are
developed. Patera and co-workers [14] first introduced this method. They coined the term
“mortar element methods” because the discretization introduces a set of functions that
mortar the brick-like elements together. The method generalizes the SEM to geometrically
nonconforming partitions, to sub-domains with different resolutions (polynomial degrees)
on sub-domain interfaces and allows for the coupling of variational discretizations of differ-
ent types in non-overlapping domains, that is, the non-conformity may be due to geometry,
approximation spaces, or both.

1.6.4 Discontinuous Galerkin method

In this method, we do not require C 0 continuity as in mortar methods. While the main
application of most discontinuous Galerkin methods (DGM) was in solving hyperbolic
problems, more recent work has led to formulations for parabolic and elliptic problems
[31].

1.6.5 Interior Penalty Methods

The origin of interior penalty method can be traced back to the use of a penalty formulation
for enforcing Dirichlet boundary condition weakly in the late 1960’s. In 1973, Babuška

13
 Chapter 1. Introduction

[8] used this approach in finite element methods. The motivation for enforcing inter-
element continuity for the space of discontinuous piecewise polynomials was provided by
this approach.
For fourth order problem, the use of penalties in formulations of non-conforming methods is
discussed by Zeinkiewicz and Cheung [79]. Here the trial functions, though continuous, are
not contained in H 2 (Ω). Babuska and Zlámal [9] have presented a scheme implementing
this idea to solve the biharmonic equation. Douglas and Dupont [34] have analyzed interior
penalty procedure for elliptic and parabolic Galerkin methods.

1.7 Review of Previous Work

For a bounded domain Ω in Rn , Bramble and Thomee [23] proposed least-squares methods
for the heat equation with Dirichlet boundary conditions. They used implicit and Crank-
Nicholson methods for time differencing. The resulting elliptic problem at each time step
was solved by least-squares approach. In [41], Tal-Ezer has used a pseudo-spectral explicit
scheme, to solve linear periodic parabolic problems. This scheme has infinite accuracy in
both time and in space. The key point of this approach is that the time resolution parameter
M (M = O(1/∆t) for time marching algorithm) and the space resolution parameter N
(N = O(1/∆x)) must satisfy M = O(N 1+ǫ ), where ǫ > 0, compared to common stability
condition M = O(N 2 ), which must be satisfied in any explicit finite-order time algorithm.
Finite element method of lines (FEMOL) for the solution of parabolic PDEs was given
by Babuska and Luskin in 1981. After this Bieterman and Babuska [15, 16] discussed
computable a posteriori estimate of the space discretization error in the finite element
method of lines (FEMOL) for the parabolic equation for time-dependent space meshes.
The importance of the error estimator is in conditions on the solution, regularity, mesh
family type and asymptotic range for the mesh size.
Jamet [49] proposed a discontinuous Galerkin finite element (DGFE) time stepping
method for parabolic partial differential equation. In this approach, Jamet considered a
time-dependent domain and continuous approximations with respect to the space variables
for each fixed time, but the discontinuities come with respect to time variables at each time

14
 1.8. Organization of Thesis

step. After this Markidakis and Babuska [59] describe the quasi-optimality of the DGFEM
in certain mesh-dependent norms. The above method is “h-version” DGFEM.

In the 1980s, Babuska and their coworkers (for more details see [10], the monograph [70]
and the references therein) introduced the p- and hp- versions of the finite element method
(FEM). In this work, they acheived the exponential convergence for elliptic problems with
piecewise analytic solutions by hp-FEM. In [11, 12], Babuska and Janik began work in
parabolic problems, which have singularity induced by incompatibility of initial data but
the solution is smoothed, and solved by p- and hp- version of a Petrov- Galerkin method
in time.

D. Schotzau and Schwab [69] analyzed a DGFEM for the time discretization of parabolic
problems in the sense of hp-version with Galerkin approach. Spectral convergence has been
shown for parabolic problem with smooth time dependence.

Aziz et. al. [7] proposed a method, which is based on a weighted least-squares method,
for the computed solution of parabolic PDE where the diffusion coefficient changes sign.
Here the main idea is the transformation of the second order equation into a first order
system of symmetric-positive differential equations in the sense of Friedrichs and the system
is then solved by using least-squares approach.

Recently, the least-squares spectral element method for parabolic problem with smooth
data and on smooth domains has been examined in [17, 37] . Two methods for parabolic
PDE have been analyzed - the h-version and p-version. If data belongs to Gevrey spaces
then the p-version gives exponential convergence. A preconditioner for the parabolic partial
differential equation has also been proposed [38]. Using separation of variables technique
this preconditioner can be diagonalized in a new set of basis functions which is given in
[38].

1.8 Organization of Thesis

The whole thesis is divided into five chapters. Now we briefly discuss the content of these
chapters.

15
 Chapter 1. Introduction

Chapter 1 In this chapter, we give a brief description of different methods, namely spec-
tral and finite element methods, with advantages and disadvantages. Basic properties
of these methods are discussed. An overview of the existing work is also provided.

Chapter 2 is devoted to the description of methods for the reconstruction of non smooth
data.

Chapter 3 discusses the non- conforming least-squares spectral element method for Eu-
ropean options with single asset on parallel computers. The method proposed in
the current study achieves spectral accuracy in both space and time. The method
is based on minimization of a functional given in terms of the sum of squares of the
residuals in the partial differential equation and initial condition in different Sobolev
norms, and a term which measures the jump in the function and its derivatives across
inter-element boundaries in appropriate fractional Sobolev norms. To obtain values
of the solution and its derivatives the initial condition is mollified and the computed
solution is post processed. Error estimates are obtained for this method. Specific
numerical examples are given to show the efficacy of this method.

Chapter 4 presents the non- conforming least-squares spectral element method for Eu-
ropean options with multi assets on parallel computers.

Chapter 5 briefly outlines our plan for future work.

16
 Chapter
2
Methods for the reconstruction of
non smooth data

In this chapter we describe the methods for the reconstruction of non smooth data. This
methods we outline are given in [73] and [74].

2.1 Introduction
In this chapter the main concern is to recover the discontinuous function from its spectral
representation. Here we mainly concentrate on two aspects of processing such piecewise
smooth data:

Edge detection. Detecting the location and amplitude of the edges. Often, these are
essential features sought in piecewise smooth data. Moreover, they are the boundaries
of the region of smoothness and are therefore essential for the second aspect.

Reconstruction. Recovering the underlying function f inside its different regions of

smoothness.

17
 Chapter 2. Methods for the reconstruction of non smooth data

Many classical algorithms exist to detect edges and reconstruct the data in between
those edges by local information. Let f be a one dimensional function and suppose that the
values of the function are given at equidistant points, fν = f (ν∆x). Then, the first-order
differences, ∆fν := fν+1 − fν , can detect edges where ∆fν = O(1), by separating them
from smooth regions where ∆fν = O(∆x). We can also recover the point-wise value of
f (x) up to order O((∆x)2 ) by piecewise linear interpolators.
P
If f is sufficiently smooth then the Fourier projection SN f = |k|≤N fˆ(k)eikx gives highly
accurate approximation of f . But, if f is only piecewise smooth then Fourier projection
loses high accuracy due to spurious oscillations which are formed around the edges of
f . This spurious oscillation became known as the Gibbs phenomenon, originating with
Gibbs’ letter (Gibbs 1899). It is caused by the global nature of Fourier projection SN f
which extracts smoothness information across the internal edges of f . The effect of Gibbs
phenomenon is not only local but global. Hence the highly accurate content in spectral
data, {fˆ(k)}|K|≤N , is lost in the Fourier projection, SN f .
Here the main aim is the computation of the Gibbs phenomenon to regain the high accuracy
by detecting edges and reconstructing piecewise smooth functions. For edge detection and
reconstruction, we discuss two main approaches :

1. Concentration Kernels

2. Mollifiers and Filters.

2.2 Concentration Kernels

For detecting edges, we consider a general framework which is based on partial sum of the
form   Z 1
πi X |k| ˆ σ(ξ)
KNσ := sgn(k)σ f (k)eikx , cσ := dξ. (2.2.1)
cσ N 0 ξ
|k|≤N

σ(ξ)
Here cσ is normalization constant and ξ
∈ C 2 [0, 1]. In [73] it is proved that KNσ f (x)
approximates the local jump function, KNσ f (x) ≈ f (x+) − f (x−). The main concentration
of KNσ f (x) is near the edges, where f (x+) − f (x−) 6= 0. Otherwise, the region between

18
 2.3. Mollifiers and Filters

any two successive edges is smooth, where KNσ f (x) ≈ 0. KNσ f (x) can be written in terms
of convolution with the Fourier projection of f . The form is
N  
πi X |k|
KNσ = σ
KN ∗ (SN f )(x), σ
KN (x) := − σ sin kx. (2.2.2)
cσ k=1 N

σ
Here KN (x) is the corresponding concentration kernel. The main feature of concen-
tration kernels is to convert the global moment of SN (f ) into local information about its
 
edges, their location and their amplitude. The concentration factor, σ |k| N
, is chosen so
that KNσ f detects the O(1)-edges, [f ](cj ), j = 1, ..., m, by separating them from the much
smaller scale of KNσ f (x) in regions of smoothness,where [f ](cj ) = f (cj +) − f (cj −). It
means that all σ’s can serve admissible concentration factors. Concentration kernels are
also useful to deal with noisy data and incomplete data. From [73], we can improve the
edge detector by enhancement of this separation of scales.
Now we switch our attention to highly accurate, Gibbs-free reconstruction of f inside its
regions of smoothness from its spectral content.

2.3 Mollifiers and Filters

In this section, we discuss two interchangeable processes to recover the values of a piecewise
smooth function f (x) with high accuracy. First is mollification which is carried out in
physical space and the other is filtering which works in Fourier space, i.e.,
 
X |k| ˆ
Φp,δ ∗ (SN f )(x) ←→ φp,δ f (k)eikx . (2.3.1)
N
|k|≤N

 
Here, the convergence of filtering is based on a rapidly decaying function φp,δ |k|
N
which
 
is a pre-multiplier of Fourier coefficients. The rapid decay of φp,δ |k|
N
fˆ(k) as |k| ↑ N in
Fourier space is also responsible for convergence of highly localized mollifiers, Φp,δ (x), in
physical space.
In section 2.6, we will discuss spectral mollifiers briefly. Now we give some details of
mollifiers. Mollifiers have two free parameters δ and p. The first parameter δ is chosen
so that the essential support of Φp,δ ∗ (SN f )(x) does not cross edges of f . Thus δ =

19
 Chapter 2. Methods for the reconstruction of non smooth data

dx = dist{x, sing sup f }/π and (x − πdx , x + πdx ) is the largest interval of smoothness
enclosing x. Here we use the information on the location of the edges of f . This leads to
adaptive mollifiers Φp,dx . Here p is responsible for the accuracy of the mollifiers. Define
root-exponential accurate mollifiers
   
1 πx x
ΦpN ,dx (x) := φ Dp N .
dx dx dx

Here dx = π1 dist{x, singsupp f }, pN ∼ dx N and

sin(pN + 1/2)x
DpN (x) := ,
2π sin(x/2)

is the Dirichlet kernel of order pN . Now we define φ = φ2q which is a proper C0∞ (−1, 1)
cut-off function, e.g.,  
y 2q
y 2 −1
φ2q (y) := e 1(−1,1) (y).

After this we will discuss the Hermite mollifier, which was introduced by J. Tanner [74]
and also examined by Tadmor [73], in section 2.6. i.e.
P  
2
− x2
X (−1)j x
Φ(x) = e H2j √ ,
j=0
4j j! 2

where H2j are Hermite polynomials of order 2j.

Section 2.7 is devoted to filters which are of the form
 
φ
X |k| ˆ
SpN := φp N f (k)eikx .
N
|k|≤N
√
Choosing pN ∼ dx N, we can show that the filter is accurate to order pN . Here the
choice of filter φN yields a highly localized mollifier (root-exponential mollifiers) which is
essentially supported in the smoothness interval (x − πdx , x + πdx ).

2.4 Spectral Accuracy

Consider the fourier projection SN f (x) of a 2π periodic function f (x)
Z π
X 1
SN f (x) = fˆ(k)e , fˆ(K) =
ikx
f (y)e−iKy dy.
2π −π
|k|≤N

20
 2.4. Spectral Accuracy

It is well known that the error SN f − f which depends on the global smoothness of f (x),
decays as rapidly as the global smoothness of f (x) permits. The truncation error
X
TN f (x) = fˆ(k)eikx ,
|k|>N

which is spectrally small, satisfies

X 1
|SN f (x) − f (x)| ≤ |fˆ(k)| ≤ ||f ||C s , f or all s > 1, (2.4.1)
N s−1
|k|>N

where
||f ||C s := max ||f (k)(.)||L∞ .
k≤s

Now, we define the Sobolev norms

X Z π
||f ||H s = 2π (1 + |k| )|fˆ(k)|2 ,
2s
||f ||2H s := ((f (y))2 + (f s (y))2)dy. (2.4.2)
k −π

The above expression can be written in terms of spectral decay of fourier coefficient
1
|fˆ(k)| . ||f ||C s , s > 1. (2.4.3)
1 + |k|s
Now, we define the actual decay rate of error which depends on the growth of ||f ||C s−1 .
First we define Gevrey classes Gα , α ≥ 1, as
(s!)α
Gα = {f | ||f ||C s−1 . , s = 1, 2...} (2.4.4)
ηfs
We now discuss two examples of Gevery classes.

1. Analytic functions are example of G1 . The estimate come from Cauchy integral
formula with 2ηf being the width of their analytic strip.

2. Certain C0∞ cut off functions belong to G2 . Consider a cut-off function

cxp
( )
ρp (x) := e x2 −π 2 1(−π,π) (x), c > 0, p even. (2.4.5)

The computation of derivatives of cut-off function is as follows:

s! cxp
 
|ρ(s)
p | . e x2 −π 2 ,
(λρ |x − π 2 |)s
2

21
 Chapter 2. Methods for the reconstruction of non smooth data

where λρ may depends on p but is otherwise independent of s. The maximum of upper

bound on the right is at x = xmax with x2max − π 2 ∼ −π 2 c/s. Then
 s
(s) s (s!)2
sup |ρp | . s! e−s . s , s = 1, 2, . . . (2.4.6)
x∈(−π,π) η η

The above estimate implies the G2 -bound 2.4.4 with η = cλρ π 2 .

Combining the results (2.4.3) and (2.4.4), we have
 α s
ˆ s 1/α
|f (k)| . min α
∼ e−α(η|k|) ,
s ηe |k|
for f ∈ Gα , α ≥ 1.
Moreover the truncation error between the function and its fourier projection satisfies
1/α
|SN f (x) − f (x)| . Ne−α(ηN ) , f ∈ Gα≥1 .

If f (x) is analytic then α = 1, and

|SN f (x) − f (x)| . Ne−(ηN ) , f ∈ G1 , (2.4.7)

and if f (x) ∈ G2 then α = 2, and

√
|SN f (x) − f (x)| . Ne− ηN
, f ∈ G2 . (2.4.8)

2.5 Reconstruction of Piecewise Smooth Data

In this section we discuss how can we reconstruct the function f from its spectral projection
{fˆ(k)}|k|≤N . The classical F ej́er sum, which are useful to avoid Gibbs oscillation, is given
as :
X  
|k| ˆ
F
SN f (x) = 1− f (K)eikx ;
N
|k|≤N

and the above equation is also written as convolution against the F ej́er Kernel.
   2
F 1 X |k| iKy 1 sin(Ny/2)
SN f (x) = (FN ∗ f )(x), FN = 1− e = .
2π N 2πN sin(y/2)
|k|≤N

Here F ej́er partial sums not only avoids spurious oscillation but also they are monotone
F
and converge uniformly whenever f is continuous. But SN is at most second order accurate

22
 2.6. Spectral Mollifier

for every family of linear positive operators by a classical theorem of Korovkin which states
that
F 1
|SN f (x) − f (x)| ≤ , f ∈ C 1,
N2
and this second order convergence rate does not improve for more regular functions f for
example f (x) = 1, x, x2 .
To improve the convergence, consider the partial sum
 
X |k|
φ
SN f (x) = φ fˆ(k)eikx ,
N
|k|≤N

where 
   1, N
|k| |k| ≤ 2
,
φ =
N  2− 2|k|
, N
≤ |k| ≤ N.
N 2

φ
Here SN is no longer positive because of the difference of two positive F ej́er sums sums,

φ F F
SN f (x) ≡ 2SN f (x) − SN/2 f (x).

The converge rate is still uniform whenever f is merely continuous. We can also improve
φ
the convergence rate of SN if f is globally smooth and the following estimate holds
 
1 − 2|k|  .|fˆ(k)| + 1
X   X
φ
|SN f (x) − f (x)| ≤ |fˆ(k)| . ||f ||C s s−1 , f or all s > 1.
N
 N  N
≤|k|≤N |k|>N
2

φ
But what happens when we apply SN f to piecewise smooth f ’s ? We will discuss this in
the next section.

2.6 Spectral Mollifier

First, we shall discuss the classical compactly supported mollifier. Choose p < q and
suppose Φ = Φp ∈ C0q (−π, π) is a unit mass kernel which possesses p−1 vanishing moments.

Z π  1, n = 0,
xn Φn (x) dx = (2.6.1)
−π  0, n = 1, . . . , p.

23
 Chapter 2. Methods for the reconstruction of non smooth data

Here the construction of some Φ’s which satisfy the moment constraints for small p’s is
easy. We choose arbitrary p and set Φp to be the ωα -weighted Gegenbauer polynomial of
degree p.
  x 2 α− 12 x
Φp (x) = cα,p 1 − Cp(α) 1(−π,π) (x), α < q.
π π
Clearly Φp (x) is a C0p -function, which can be normalized to have a unit mass by a proper
choice of cα,p and Φαp satisfies the moments condition (2.6.1) by the ωα -orthogonality of the
Gegenbauer polynomial Ckα ’s.
Consider a mollifier Φ(x) = Φp (x), which satisfies the condition (2.6.1) and construct the
family of dilated mollifiers
1 x
Φp,δ (x) := Φp ,
δ δ
where δ is a free dilation parameter at our disposal and the support of Φp,δ is (−πδ, πδ). We
can adjust the support of Φp,δ by changing δ. Now consider the modified Fourier projection

Φp,δ ∗ (SN f )(x) ≈ f (x).

The error is

Φp,δ ∗ (SN f )(x) − f (x) = SN (Φp,δ ) ∗ f (x) − f (x), (by orthogonality)

and
Φp,δ ∗ (SN f )(x) − f (x) ≡ (SN (Φp,δ ) − Φp,δ ) ∗ f (x) + (Φp,δ ∗ f (x) − f (x)) .
| {z } | {z }
T runcation error (TN (Φp,δ )) Regularization error

The estimate of the truncation error is as follow :

1
|TN (Φp,δ ) ∗ f (x)| . ||f ||L1 ||Φp,δ ||C q , (2.6.2)
N q−1
1
. αp,q , (2.6.3)
δ q+1 N q−1

where αp,q = ||Φp,δ ||C q .

Now we estimate the regularization error,
Z π
Φp,δ ∗ f (x) − f (x) = [f (x − δy) − f (x)]Φp (y)dy. (2.6.4)
−π

24
 2.6. Spectral Mollifier

Using Taylor expansion, we have

p
X (−1)n 1
f (x − δy) − f (x) = δ n f (n) (x)y n + δ p+1 f p+1 (. . .)y p+1.
n=1
n! p + 1!

Then the bound for regularization error is

|Φp,δ ∗ f (x) − f (x)| . βp ||f ||C p+1(x−δ,x+δ) δ p+1 , (2.6.5)

1
Rπ
where βp = p! −π
|y|pΦp (y)dy.

1. If p ∼ q and we take δ = δN ∼ √1 then f (x) can be recovered from the δN -

N

neighborhood of it spectral projection SN f (x),

1 1
Φp,δN ∗ (SN f )(x) − f (x)| . γp (1 + ||f ||C p(x−δN ,x+δN ) ) , δN = √ .
N p/2 N

Here γp = αp,p + βp .

2. Take δ as
1
δ = dx , dx = dist{x, {c1 , ..., cj }}[modπ] ,
π
here c1 , ...., cj are points of singularity around x and (x − πdx , x + πdx ) is the largest
neighborhood of analyticity around x.

Now we discuss some important results on mollifier.

Theorem 2.6.1. (Root-exponential accurate mollifiers) (Tadmor and Tanner 2002)

Given the Fourier projection, SN f (·) of a piecewise smooth function f (·) ∈ piecewise − G2 ,
we consider the 2-parameter family of spectral mollifiers

1 x x cx2

 
Φp,δ (x) := ρ2 Dp , ρ2 := e x2 −π 2 1(−π,π) (x), c > 0. (2.6.6)
δ δ δ

Fix x inside one of the smoothness intervals of f and set the adaptive parameterization

1
δ = dx = dist{x, {c1 , ..., cj }}[modπ], (2.6.7)
π
√
p = pN (x) ∼ dx N/ e. (2.6.8)

25
 Chapter 2. Methods for the reconstruction of non smooth data

Then there exists a constant η = ηρ,f such that ΦpN ,dx ∗SN f recovers f (x) with the following
root-exponential accuracy:
√
|Φp,δ ∗ (SN f )(x) − f (x)| . dx Ne−0.84 ηdx N
. (2.6.9)

Proof. Proof of the above Theorem is given in page 355 of [73].

Theorem 2.6.2. (Exponential accurate mollifiers) (Tanner 2006)

2N −1
Given the 2N equidistant samples, of a piecewise analytic function, {f ( Nπ (ν − N))})ν=0 ,
the function can be recovered within the exponential bound
 
 π X 
fπ (yν )φopt (N(x − yν )) . constf .N 9/4 τ −N dx
 
f (x) − (2.6.10)
 N yν ∈Ix


κ κ 1/2α
e √
where τ := min(βfk , 2eα/2 ( αe ) , (1+ 2)κ
), and the mollifier is given by

  ⌊κN dx ⌋ r !
2 −N
1 −Nx X 4 N
φopt (N, x) := √ exp H2n x . (2.6.11)
αNdx 2αdx n=0
n! 2αNdx

Here H2n is the Hermite polynomial.

Proof. Proof of the above Theorem is given in page 11 of [74].

√
Convergence is ensured if parameter constant satisfy ακ < 1/2 ln(1 + 2) ≈ 0.56.
Moreover, we can choose α = 1 and κ = 1/15 as optimal values.

2.7 Spectral Filter

In this section, we discuss some results on filters which are root exponential accurate and
exponential accurate.

φ
Theorem 2.7.1. Assume that f (·) is piecewise analytic and let SNp denote the filtered sum
   p 
φp
X |k| ˆ ikx
ξ
SN f (x) := φp f (k)e , φp (ξ) = e ξ2 −1 1(−1,1) (ξ) .
N
|k|≤N

26
 2.7. Spectral Filter

√
We set the order p = pN ∼ dx N (pN even) where, as usual,
1
dx = dist {x, {c1 , . . . , cj }} [modπ],
π
so that (x − πdx , x + πdx ) is the largest interval of analyticity enclosing x. Then, the
φp
adaptive filter SN N f recovers the point values f (x) within the following root-exponential
accuracy:
√
φp
|SN N f (x) − f (x)| . dx Ne−η dx N
. (2.7.1)

Here, the constant η = ηφ,f is dictated by the specific Gevrey and piecewise analyticity
properties of φ and f .

Proof. Proof of the above Theorem is given in page 369 of [73].

Theorem 2.7.2. Assume that f (·) is piecewise analytic, and let

 
φp,δ
X |k| ˆ
SN f (x) := φp,δ f (k)eikx
N
|k|≤N

denote the filtered Fourier projection, based on the quadratic exponential filter
p  j
− (δξ)
2 X
1 (δξ)2
φp,δ (ξ) = φp (δξ) := e 2 , (2.7.2)
j=0
j! 2
√
of degree p = pN := θ2 dx N, with adaptive scaling δ = δx := θdx N. Here,
1
dx = dist {x, {c1 , . . . , cj }} [modπ],
π
defines a πdx -neighborhood of analyticity around x. Then, for sufficiently small θ < 1,
φp ,δN
there exists η = ηθ,f > 0 such that the adaptive filter SN N f (x) recovers f (x) with the
following exponential accuracy:
r
φp ,δ N −ηdx N
|SN N N f (x) − f (x)| . e . (2.7.3)
dx
The constant η = ηθ,f > 0 is dictated by the specific piecewise analyticity properties of f .
The exponential adaptive filter takes the final form
 2 
[θ dx N ]  
2 j
φp ,δ X X 1 θdx k θdx k2
SN N N f (x) =   e− 2N fˆ(k)eikx .
j=0
j! 2N
|k|≤N

27
 Chapter 2. Methods for the reconstruction of non smooth data

Proof. Proof of the above Theorem is given in page 373 of [73].

φ
Remark 2.7.3 (Exponential accurate mollifier). Observe that the mollifier SNp,δ (x) as-
sociated with the filter φp,δN (ξ) in (2.7.2) is a exponentially close to Φp,δ (Ny). Accord-
√
ingly, we find the exponentially accurate mollifier Φp,δ (Ny): with δ = δx := θdx N and
p = pN := θ2 dx N
[θ dx N ] 2 √ !
1 Ny 2 X (−1)j Ny
− 2θd
ΦpN ,δx (Ny) = √ e x ×
j j!
H2j √ . (2.7.4)
θdx N j=0
4 2θdx

We also have
Z x+πdx
r
N −ηdx N
| ΦpN ,δx (Ny)f (x − y)dy − f (x)| ≤ e .
x−πdx dx

Then from Theorem 4.2 of [74]

2N r
X N −ηdx N
|h ΦpN ,δx (N(x − yν ))f (yν ) − f (x)| . e .
ν=0
dx

28
 Chapter
3
Nonconforming LSSEM for European
Options with Single Asset

3.1 Introduction
Consider the Black-Scholes (BS) equation [18] for European Option
1
Vτ + σ 2 S 2 VSS + rSVS − rV = 0 in (0, ∞) × [0, T ] (3.1.1)
2
where V, S, r and σ are respectively option price, underlying asset price, risk-free interest
rate and volatility.
Now, we define the European Call option and European Put option.

Definition 3.1.1. In European Call Option the holder has the right, but not the obligation,
to buy an asset at a prescribed price K (strike price) at maturity time T in future. The
payoff function for European Call Option is

VC (S, T ) = max(S − K, 0). (3.1.2)

Definition 3.1.2. In European Put Option the holder has the right, but not the obligation,
to sell an asset at a prescribed price K (strike price) at maturity time T in future. The

29
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

payoff function for European Put Option is

VP (S, T ) = max(K − S, 0). (3.1.3)

Recently, Spectral methods have been used to solve Option Pricing problems. In 2000,
Bunin et. al. [24] proposed Chebyshev Collocation methods to solve the European Call
Option problem on parallel computers. After this, Greenberg [47] solved American Op-
tions problem by Chebyshev Tau method. For smooth initial conditions, De Frutos [45]
has presented a Laguerre-Galerkin Spectral Method to price bonds. More recently, Zhu et.
al. [77] have used a Spectral element method to price European Options. These methods
give quadratic accuracy in time, while being spectrally accurate in space. Dominik et. al.
[69] proposed hp-version of the Discontinuous Galerkin Finite Element Method to solve
parabolic problems.

In this chapter we develop a Non-Conforming Least-Squares Spectral Element Method

(LSSEM) to solve the Black-Scholes equation for European Options on parallel computers.
LSSEM are spectrally accurate in both time and space. Sobolev spaces of different orders
in space and time are used for the results, as presented in [57]. If the data belong to certain
Gevrey spaces then the solution also belongs to a Gevrey space [58].

The proposed method is a Least-Squares method as presented in [37]. The space do-
main is an interval which is divided into a number of sub-intervals. The functional is the
sum of the squares of the residuals in the partial differential equation and initial condi-
tion in different Sobolev norms, and a term which measures the jump in the function and
its derivatives across inter-element boundaries in appropriate fractional Sobolev norms.
We minimize the functional on a given time interval. Hermite mollifiers, as described in
[73, 74], are used to resolve the difficulty of non-smooth initial data.

Now we give the details of this chapter. In Section 3.2 the function spaces and a priori
estimates for parabolic initial value problem, as presented in [35, 57, 58], are given. In Sec-

30
 3.2. Function Space For the Formulation

tion 3.3 discretization of the domain and stability estimate are discussed. In Section 3.4 we
discuss the numerical scheme, symmetric formulation, parallelization and preconditioning
for our method. In Section 3.5 the estimate for non smooth initial condition is presented
in negative Sobolev norms. In Section 3.6 error estimates are obtained for this method.
Finally, in Section 3.7 specific numerical examples are given to show the effectiveness of
this method.

3.2 Function Space For the Formulation

T −τ
We consider Ω = R as the domain of logarithmic price x = log(S/K) and define t = T

on the time interval I = [0, 1]. We shall focus here on the Black -Scholes equation for the
European Call with the assumption that the rate of interest r and volatility σ are smooth
(or even analytic) functions of x and t with bounded derivatives. The coefficients a, b and
c belong to D2,1 (Ω × I) as defined in [58] and satisfy

||Dxi Dtj a(x, t)||L∞ (Ω×I) ≤ AB i+j i!(j!)2 ,

where A and B are positive numbers.

The price u(x, t) has to satisfy the BS equation

Lu = ut − auxx − bux − cu = 0 in Ω × I ,
(3.2.1)
u(x, 0) = f (x) in Ω × {0} .

f (x) may not be in L2 (Ω) as, for example

f (x) = (Kex − K)+ .

To resolve this difficulty, let

v(x, t) = u(x, t)sech(ηx) , (3.2.2)

where η > 0 is sufficiently large so that

v(x, 0) = u(x, 0) sech(ηx) , (3.2.3)

31
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

is such that veµx , ve−µx ∈ L1 (Ω) ∩ L2 (Ω) for some µ > 0.

Clearly v satisfies the partial differential equation

Lv = vt − α vxx − β vx − γ v = 0 in Ω × I ,
(3.2.4)
v(x, 0) = f (x) sech(ηx) = g(x) in Ω × {0} .

We assume the coefficients a, b and c in (3.2.1) are smooth or even analytic and all deriva-
tives are bounded. Clearly the same assumption will continue to hold for the coefficients
α = a , β = 2 a η tanh ηx + b and γ = η 2 a + bη tanh ηx + c, since tanh ηx has bounded
derivatives of all orders. Moreover the coefficients belong to D2,1 (Ω̄ × [0, 1]).
The initial data g(x) = f (x)sechηx is not smooth. To resolve this difficulty we use the
Hermite mollifier [73]

P  
− x2
2 X (−1)j x
Φ(x) = e H2j √ . (3.2.5)
j=0
4j j! 2

Let
1 x
Φδ (x) = Φ .
δ δ
Define

 
N N Nx
θN/δ (x) = Φδ (Nx) = Φ . (3.2.6)
2π 2πδ δ
Then
(Nx)2 P  
e− 2δ 2 X (−1)j Nx
Φδ (Nx) = H2j √ ,
δ j=0
4j j! 2δ
√
where δ = β1 N, β1 = θ1 dx and P = θ12 dx N. Here

1
dx = dist{x, {c1 , ..., cj }}[modπ] ,
π
and c1 , ...., cj are points around x where the initial function is not regular and (x − πdx , x +
πdx ) is a neighborhood of analyticity around x and 0 < θ1 < 1.
We use the above mollifiers to replace the initial function g(x) by its mollified version gδ (x)

32
 3.2. Function Space For the Formulation

where
Z
gδ (x) = (g ∗ θN/δ )(x) = g(x − y)θN/δ (y)dy, ∀ x ∈ Ω . (3.2.7)
|y|≤πdx

Then we define vδ (x, t) to be a solution of the following IVP

L vδ = 0 in Ω × I , (3.2.8)

vδ = gδ on Ω × {0} . (3.2.9)

Let us now define some Sobolev spaces. Let ω(x, t) be a smooth function. Define the norm
Z Z X X β
2
||ω||H r,s(Ω×I) = ( |∂xα w|2 + |∂t w|2 ) dx dt . (3.2.10)
I Ω α≤r 0<β≤s

Next, let h(x) be a smooth function, with

Z X
||h||2H r (Ω) = |∂xα h|2 dx . (3.2.11)
Ω α≤r

Now, consider the initial value problem (IVP)

Lω = F in Ω × I ,
(3.2.12)
ω = h on Ω × {0} ,

where F and h are smooth. Then the following a priori estimate holds

||ω||2H 2r+2,r+1(Ω×I) + ||ω||2H 2r+1(Ω×{1}) ≤ Cr (||L ω||2H 2r,r (Ω×I) + ||ω||2H 2r+1(Ω×{0}) ) , (3.2.13)

where Cr is a constant which depends on r .

Define the negative Sobolev norms

|(ω, Φ)Ω |
||ω||H −m(Ω) = sup , (3.2.14)
Φ∈H m (Ω) ||Φ||H m (Ω)

and

|(ω, Φ)Ω×I |
||ω||H −r,−s(Ω×I) = sup . (3.2.15)
Φ∈H r,s (Ω×I) ||Φ||H r,s (Ω×I)

We now define some Gevrey Spaces which are needed in the error analysis.

33
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

Definition 3.2.1. Let

D1 (Ω̄) ={Φ(x) : Φ is an infinitely differentiable function in Ω̄ such that there exist two
positive numbers A1 and B1 such that

||DxαΦ(x)||L2 (Ω̄) ≤ A1 (B1 )i i! , |α| = i, f or all i = 0, 1, 2, 3 · ··} .

Definition 3.2.2. Let

D2,1 (Ω̄ × [0, 1])
= {ψ(x, t) : ψ is an infinitely differentiable function in Ω̄ × [0, 1] such that there exist two
positive numbers A1 and B1 such that

||DxαDtj ψ(x, t)||L2 (Ω̄×[0,1]) ≤ A1 (B1 )i+j i!(j!)2 , |α| = i, f or all i, j ≥ 0} .

3.3 Discretization and Stability Estimates

3.3.1 Discretization
Let N and p be integers and p be proportional to N. We solve the initial value problem
(3.2.8) for IN = [−N, N]. First we divide the interval IN = [−N, N] into a number of
N −1 S N −1
sub-intervals {Ωl }l=−N , where Ωl = (l, l + 1) and IN = {Ω̄l }l=−N . Each of these in-
N −1
tervals {Ωl }l=−N is mapped to the standard element S = (0, 1) by a set of smooth maps
{Ml−1 }l=−N
N −1
, where Ml is a map from S = (0, 1) to Ωl = (l, l +1). The map Ml has the form

Ml (ξ) = l + ξ ,

where ξ ∈ (0, 1).

Here the size of the mesh interval is h = 1 and the time step k is proportional to h2 = 1.
S
Let {xl }l=−N,N be the inter-element boundaries and boundary of IN = {Ω̄l }l=−N,N −1 .
Hence xl = l.

34
 3.3.2. Stability Estimates

Figure 3.1: Inter Element Boundary

3.3.2 Stability Estimates

Let v̌lp (ξ, t) be the spectral element function which is defined to be a polynomial of degree
p in the space variable ξ and of degree q in the time variable t, i.e.
p q
X X
v̌lp (ξ, t) = l
δi,j ξ i tj
i=0 j=0

l
for ξ ∈ (0, 1), t ∈ [0, 1]. Here δi,j denote the coefficients and q is proportional to p2 .
Then
vlp (x, t) = v̌lp (Ml−1 (x), t) .

Define v p (x, t) as

v p (x, t) = vlp (x, t), for (x, t) ∈ Ωl × I, for all − N ≤ l ≤ N − 1, (3.3.1)

= 0, f or (x, t) ∈ (IN )c × I . (3.3.2)

Thus v p (x, t) = 0 for |x| ≥ N.

Using the chain rule, we can write

∂vlp
= (v̌lp )ξ (ξ)x . (3.3.3)
∂x
ˆ x to be the polynomial of orthogonal projection of (ξ)x into the space of poly-
Assume (ξ)
nomials of degree p with respect to the inner product in H 2(0, 1). Then the polynomial
approximation of the above equation (3.3.3) at the interior point is defined by
 p a
∂v̌l ˆx .
= (v̌lp )ξ (ξ) (3.3.4)
∂x

35
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

Let xl be the common interior point of the subintervals Ωl−1 and Ωl , which are the image
of ξ = 1 under the map Ml−1 and the image of ξ = 0 under the map Ml respectively. Now
we define the jump in the derivative at the inter element boundary xl as follows :
 p a 2  p a  p a 2
 ∂v̌   ∂v̌l ∂v̌l−1 
  = 
 (0, t) − (1, t) .
 ∂x 
H s ({xl }×I) ∂x ∂x H s ({xl }×I)

Next
Z Z
||vlp ||2H 1 (Ωl ×{0}) = |v̌lp |2 dξ + |Dv̌lp |2 dξ = ||v̌lp ||2H 1 ((0,1)×{0}) , (3.3.5)
(0,1)×{0} (0,1)×{0}

and
Z Z
||L vlp ||2 dxdt = ||Ll v̌lp ||2 dξdt , (3.3.6)
Ωl ×(0,1) (0,1)×(0,1)

where Ll is the differential operator L in ξ and t coordinates. Now we take the orthogonal
projection of the coefficients of the differential operator Ll into the space of polynomials
with respect to the usual inner product in H 2,1 ((0, 1) × (0, 1)) and define a new differential
operator Lla . The coefficients of the differential operator Lla are polynomials of degree p
in ξ and of degree q in t. Hence
Z Z
p 2
||L vl || dxdt = ||Lla v̌lp ||2dξdt , (3.3.7)
Ωl ×(0,1) (0,1)×(0,1)

upto a negligible error term [17, 37].

We now state the stability theorem which is needed to formulate the numerical scheme.
Define the quadratic form
N
X −1 X
V p ({v̌lp (ξ, t)}−N ≤l≤N −1) = ||(Lla v̌lp )||2L2 ((0,1)×I) + (||[v̌ p ]||2H 3/4 ({xl }×I)
l=−N xl ∈int(IN )
X
+||[(v̌xp )a ]||2H 1/4 ({xl }×I) ) + (||v̌ p||2H 3/4 ({xl }×I) (3.3.8)
xl ∈∂IN
N
X −1
+||(v̌xp )a ||2H 1/4 ({xl }×I) ) + ||v̌lp ||2H 1 ((0,1)×{0}) .
l=−N

Here int(IN ) denotes interior of IN and ∂IN denotes the boundary of IN , where IN =
[−N, N].
Then from Theorem 11 in [37] the following result holds.

36
 3.4. Numerical Scheme and Parallelization

Theorem 3.3.1. There exists a constant C such that the estimate

N
X −1
||v̌lp ||2H 2,1 ((0,1)×I) ≤ C (ln p)2 V p ({v̌lp (ξ, t)}−N ≤l≤N −1) (3.3.9)
l=−N

holds.

We now define a modified version of the quadratic form W p (v̌lp ) as:

N
X −1 X
W p ({v̌lp (ξ, t)}−N ≤l≤N −1) = ||(Lla v̌lp )||2L2 ((0,1)×I) + (||[v̌ p ]||2H 3/4 ({xl }×I)
l=−N xl ∈int(IN )∪∂IN
N
X −1
+||[(v̌xp )a ]||2H 1/4 ({xl }×I) ) + ||v̌lp ||2H 1 ((0,1)×{0}) .
l=−N

(3.3.10)

Then the following result follows immediately.

Theorem 3.3.2. There exists a constant C such that the estimate

N
X −1
||v̌lp ||2H 2,1 ((0,1)×I) ≤ C (ln p)2 W p ({v̌lp (ξ, t)}−N ≤l≤N −1) (3.3.11)
l=−N

holds.

3.4 Numerical Scheme and Parallelization

Let (gδ )l (ξ) = gδ (Ml (ξ)), where gδ is as defined in (3.2.7) and let (g̃δ )l (ξ) be the orthogonal
projection of gδ (ξ) into the space of polynomials of degree p in ξ with respect to the usual
inner product in H 1 (S).
We define our approximate solution to be the unique {ωlp }l=−N,N −1 which minimizes the
functional
N
X −1 N
X
Rp ({v̌lp (ξ, t)}−N ≤l≤N −1) = ||(L )al v̌lp )||2L2 ((0,1)×I) + (||[v̌ p ]||2H 3/4 ({xl }×I)
l=−N l=−N
N −1
(3.4.1)
X
+||[(v̌xp )a ]||2H 1/4 ({xl }×I) ) + ||(v̌lp − (g̃δ )l )||2H 1 ((0,1)×{0}) ,
l=−N

37
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

over all {v̌lp }−N ≤l≤N −1 . Here [v̌ p ] denotes the jump in v̌ p across x = xl , t ∈ I.
Now consider the I.V.P.

Lvδ = 0 in Ω × I , (3.4.2)

vδ = gδ on Ω × {0} . (3.4.3)

Clearly, gδ ∈ D1 (Ω × {0}). Hence vδ ∈ D2,1 (Ω × I). Now

vδ (x, 0) = gδ (x) = g(x) ∗ θN/δ (x) . (3.4.4)

And

||vδ ||H s (Ω×{0}) = ||g(x) ∗ θN/δ (x)||H s (Ω) , (3.4.5)

≤ C||g||L1(Ω) ||θN/δ (x)||H s (Ω) , (3.4.6)

where C is a generic constant.

Lemma 3.4.1. The estimate

|θN/δ |s ≤ BN s! , (3.4.7)

holds . Here θN/δ (x) is the Hermite mollifier which is defined in (3.2.6) and BN =
√
√ ( N e/β1 )
N e (N e3 /β1 ) 4 .

Proof. We know that

 
N N Nx
θN/δ (x) = Φδ (Nx) = Φ ,
2π 2πδ δ
√
where δ = β1 N.
Let φ(ξ) be the Fourier transform of Φ(x). Then
N
!
− ξ2
2 X ξ 2j
Φ̂(ξ) = φ(ξ) = e . (3.4.8)
j=0
2j j!

We have
N
X
− ξ2
2
ξ 2j+s
|φ(ξ)|s ≤ ||e ||L2 (R) . (3.4.9)
j=0
2j j!

38
 3.4. Numerical Scheme and Parallelization

Now
Z ∞ 1/2
− ξ2
2
ξ 2j+s 1 −ξ 2 4j+2s
||e ||L2 (R) = j e ξ dξ . (3.4.10)
j
2 j! 2 j! −∞

Substituting η = ξ 2 we obtain
Z ∞ 1/2
ξ 2j+s− ξ2
2
1 −η 2j+s−1/2
||e ||L2(R) = j e η dη , (3.4.11)
2j j! 2 j! 0
1 p
= j Γ(2j + s + 1/2) . (3.4.12)
2 j!
√ √
Γ(2j+s+1/2) (2j+s−1/2)
Maximum of 2j j!
is achieved when 2j
∼ 1 . Then
√
s
j∼ .
2

Hence
√
|θN/δ |s N (( s + s)!)1/2 p
≤ √s √ ( N/β1 )s . (3.4.13)
s! s
2 ( 2 )!
2 s!

|θN/δ |s
Let As = s!
. Now by Stirling’s formula

√
n! ∼ 2πn nn e−n .

And so
√ √ √
|θN/δ |s N (( s + s)( s+s) e−( s+s) )1/2 p
As = ∼ √s √ √s √s s −s
( N/β1 )s .
s! s −
22(2)2e 2 se

Hence √ √ √
s(1+ s) ( s+s)
(1 + √1 ) −
e
N s
2 2 p
As ∼ √
s
√ √ ( N/β1 )s .
− 2s s
− 2s
s 4 e s 2 e−s
Therefore

√ p √
s
√
s
As ∼ N e ( N e/(β1 s) )s e 2 s 4 . (3.4.14)

39
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

Ne Ne
Now the maximum of As is achieved when β1 s
∼ 1 or s ∼ β1
. Hence
r
Ne √
√ β1 ( N e/β1 )
(3.4.15)
As ∼ N e (e) 2 (N e/β1 ) 4 .

Thus
√
√ 3
( N e/β1 )
As . N e (N e /β1 ) 4 = BN . (3.4.16)

Hence

|θN/δ |s ≤ BN s! . (3.4.17)

Now by Lemma 3.4.1

√
√ ( N e/β1 )
||gδ ||H s (Ω×{0}) ≤ C N e (N e3 /β1 ) 4 (s)! .

Hence

√
||gδ ||H s (Ω×{0}) . BN s! = K N da N log N
s! (3.4.18)
√
where K, d and a are constants. Here BN . KN da N log N
.
Then by results on Gevrey spaces in Lions and Magenes [57, 58]

||Dxi Dtj vδ ||L2 (Ω×I) . C BN i!(j!)2 (α1 )i+j , (3.4.19)

where C and α1 are constants.

Now

||gδ eµx ||H 1 (Ω×{0}) ≤ C||g eµx ||L1 (Ω×{0}) ||θN/δ (x) eµx ||H 1 (Ω×{0}) , (3.4.20)

and
||θN/δ (x) eµx ||H 1 (Ω×{0}) = ||(1 + |ξ + iη|)θ̂N/δ (ξ + iη)||L2 (Ω×{0}) .
| {z }
I

40
 3.4. Numerical Scheme and Parallelization

Here C is a generic constant.

Moreover
    2  N 2 2
 δ(ξ + iη)   − δ (ξ+iη)2  X δ (η + ξ 2 )j
|θ̂N/δ (ξ + iη)| = φ  ≤ e 2N 2 
2 2j j!
,
N   
j=0
N
N
!
2 2
− δ ξ2
2 2
+ δ η2
X δ 2 (η 2 + ξ 2 )j
≤e 2N 2N .
j=0
N 2 2j j!

And so
Z N  2 2 2j !
2 2 2 −δ
2 ξ2 δ2 η2 X δ (η + ξ 2 ) 1
|I| ≤ 2 (1 + ξ + η )Ne N2 e
N2 dξ,
R j=0
N2 (2j j!)2 j
Z N X j   2k 2(2j−k)  4j−2k !
δ2 η2 δ 2 ξ2 X 2j δξ η δ
≤ 2N(1 + η 2 )e N 2 e− N 2 j 2
dξ
R j=0 k=0
k N (2 j!) N
N j
!
δ2 η2
Z
δ 2 ξ2 X X 2j   δξ 2k η 2(2j−k)  δ 4j−2k
+ 2Ne 2N 2 ξ 2 e− N 2 dξ,
R j=0 k=0
k N (2j j!)2 N
N X j
N
!
2j 1 4j−2k
δ2 η2 X Γ(k + )(δη)
≤ 2N(1 + η 2 )e N 2 k δ 2

j=0 k=0
(2 j!)2
j

!
N X j N 3 2j 3 4j−2k
δ2 η2 X
δ k
Γ(k + 2
)(δη)
+ 2Ne N 2 j j!)2
.
j=0 k=0
(2

Thus
j
N X 
    3 !
δ2 η2 X (δ 2 η 2 )2j−k N 1 N
|I|2 ≤ 2Ne N2 (1 + η 2 ) + k+ ,
j=0 k=0
(2j − k)! δ 2 δ
     3 !
2δ 2 η 2 N 1 N
≤ 2N 2 e N 2 (1 + η 2 ) + N+ .
δ 2 δ

Moreover  3/2  1/2 2 2

µx N 1 δ η
||θN/δ (x) e ||H 1 (Ω×{0}) ≤ CN N+ e N2
δ 2
Hence

||gδ eµx ||H 1 (Ω×{0}) ≤ C , (3.4.21)

41
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

for some µ > 0.

Hence

||vδ eµx ||H 1 (Ω×{0}) ≤ C . (3.4.22)

Moreover

||gδ e−µx ||H 1 (Ω×{0}) ≤ C . (3.4.23)

Hence

||vδ e−µx ||H 1 (Ω×{0}) ≤ C . (3.4.24)

Thus we can conclude that

||vδ eµ|x| ||H 2,1 (Ω×I) ≤ C . (3.4.25)

Consequently

||vδ ||H 2,1 (INc ×I) ≤ Ke−ρN . (3.4.26)

Here K and ρ are generic constants and INc = R \ IN , IN = [−N, N].

Let spl (x, t) be the approximate representation of vδ (x, t) on Ωl defined in Theorem 13 in
[37] and Theorem 4.2.1 in [17]. Then
N
X −1
||vδ − spl ||H 2,1 (Ωl ×I) ≤ K e−ρ N , (3.4.27)
l=−N

provided q is proportional to p2 and N is proportional to p.

Now sp (x, t) = 0 for (x, t) ∈ INc × I. Then from equation (3.4.19) and (3.4.25)

||[(vδ )x − (sp )x ]||H 1/4 ({xl }×I) ≤ K e−ρ N , f or xl = − N,

(3.4.28)
||[vδ − sp ]||H 3/4 ({xl }×I) ≤ K e−ρ N , f or xl = − N.

And

||[(vδ )x − (sp )x ]||H 1/4 ({xl }×I) ≤ K e−ρ N , f or xl = + N,

(3.4.29)
||[vδ − sp ]||H 3/4 ({xl }×I) ≤ K e−ρ N , f or xl = + N.

42
 3.4.1. Symmetric Formulation

Hence

Rp ({spl (x, t)}−N ≤l≤N −1 ) ≤ K e−ρ N , (3.4.30)

provided p is proportional to N and N is large enough.

Hence

Rp ({ωlp (x, t)}−N ≤l≤N −1 ) ≤ Rp ({spl (x, t)}−N ≤l≤N −1 ) ≤ K e−ρ N , (3.4.31)

where Rp ({ωlp (x, t)} is defined in (3.4.1).

Thus

Rp ({(ωlp − spl )(x, t)}−N ≤l≤N −1 ) ≤ K e−ρ N . (3.4.32)

Therefore by Theorem 3.3.2 we can conclude that

N −1
!1/2
X p p 2
||ωl − sl ||H 2,1 (Ωl ×I) ≤ K e−ρ N . (3.4.33)
l=−N

Moreover, ω p ((x, t) = 0 and sp (x, t) = 0 for (x, t) ∈ INc × I.

Combining the above with (3.4.26), (3.4.27) and (3.4.33) we obtain
N
X −1
||vδ − ωlp ||H 2,1 (Ωl ×I) + ||vδ − ω p ||H 2,1 (INc ×I) ≤ K e−ρ N . (3.4.34)
l=−N

Here INc = R \ IN and IN = [−N, N].

3.4.1 Symmetric Formulation

As defined in (3.4.1) we choose our approximate solution to be the unique {ωlp }l=−N
N −1
which
minimizes the functional Rp ({v̌lp (ξ, t)}−N ≤l≤N −1) over all {v̌lp (ξ, t)}−N ≤l≤N −1 Let the above
overdetermined system [37, 75], of equations (3.4.1), be of the form

AW = G . (3.4.35)

Then the Normal Equations are

AT AW = AT G , (3.4.36)

43
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

where W is a vector assembled from the values of {v̌lp (ξ, t)}−N ≤l≤N −1, and G is assembled
from the data. Here A is a matrix.
Our method is a Least-Squares method. So we use the preconditioned conjugate gradient
method for solving the Normal Equations. Now from [75]

Rp (U + ǫW ) = Rp (U) + 2ǫ(W )T (SU − T G) + O(ǫ2 ) ,

for all W . Here U is the vector assembled from the values of {ωlp }l=−N,N −1 and S, T are
matrices which contain valuation of integrals in equation (3.4.1) using quadrature rule.
Define
p,q
Ul,(p+1)k+i = ωlp (ξip , tqk ) f or 0 ≤ i ≤ p, 0 ≤ k ≤ q .

Similarly
2p,2q
Ul,(2p+1)k+i = ωlp (ξi2p , t2q
k ) f or 0 ≤ i ≤ 2p, 0 ≤ k ≤ 2q .

Those integrals, which arise in the above minimization formulation, are computed by the
Gauss-Lobatto-Legendre (GLL) quadrature formula. Then the minimization formulation
is represented as :

(Vl2p,2q )T Ol2p,2q , (3.4.37)

where Ol2p,2q is a (2p + 1)(2q + 1) vector which can be easily computed. Now we can always
find a matrix Flp,q such that
Vl2p,2q = Flp,q Vlp,q ,

and

(Vl2p,2q )T Ol2p,2q = (Vlp,q )T (Flp,q )T Ol2p,2q . (3.4.38)

Hence the residuals

Rp = (Flp,q )T Ol2p,2q .

Note that there is no need to compute any mass and stiffness matrices and we can calculate
the residuals in the normal equations inexpensively and efficiently. Nor do we need to filter
the coefficients and data. A detailed description can be found in [17, 75].

44
 3.4.2. Parallelization and Preconditioning

3.4.2 Parallelization and Preconditioning

From (3.3.10) and (3.4.1), we can conclude that the quadratic form W p ({v̌lp (ξ, t)}l=−N
N −1
)
(which is defined in (3.3.10)) is obtained from the functional Rp ({v̌lp (ξ, t)}−N ≤l≤N −1) with
zero data. For the quadratic form W p ({v̌lp (ξ, t)}−N ≤l≤N −1), we define the preconditioner
which is denoted by U p ({v̌lp (ξ, t)}−N ≤l≤N −1). Here
N
X −1
U p
({v̌lp (ξ, t)}−N ≤l≤N −1) = ||v̌lp ||2H 2,1 ((0,1)×I) . (3.4.39)
l=−N

Then from [37] the following result holds

W p ({v̌lp (ξ, t)}−N ≤l≤N −1) ≤ K U p ({v̌lp (ξ, t)}−N ≤l≤N −1) , (3.4.40)

where K is a constant. By Theorem 3.3.2, we get the following result

1
2
U p ({v̌lp (ξ, t)}−N ≤l≤N −1) ≤ W p ({v̌lp (ξ, t)}−N ≤l≤N −1) . (3.4.41)
C(log p)

By (3.4.40) and (3.4.41), we conclude that the condition number of the preconditioned
system is O((log p)2 ). Assume that v̌lp is defined in terms of Legendre polynomial in ξ of
degree upto p and in t of degree upto q for each element −N ≤ l ≤ N − 1. Then v̌lp can
be written as
p q
X X
v̌lp (ξ, t) = ai,j Li (2ξ − 1)Lj (2t − 1) , (3.4.42)
i=0 j=0

where the coefficients ai,j are arranged lexicographically in i and j.

Then ||v̌lp ||2H 2,1 ((0,1)×I) is a ((p+1)(q+1)×(p+1)(q+1)) matrix. Using separation of variables
technique this preconditioner can be diagonalized in a new set of basis functions which is
given in [38]. In section 3.3.1, the discretization of domain has already been discussed.
Each element is mapped to a single processor for ease of parallelism. During the PCGM
process, communication between neighbouring processors is confined to the interchange of
information of the value of function and its derivatives at inter-element boundaries on which
v̌lp is defined. Moreover we need to compute two global scalars to update the approximate
solution and the search direction. Hence inter-processor communication is quite small.

45
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

Figure 3.2: Parallelization

3.5 Estimates in Negative Norms

Lemma 3.5.1. Assume g is piecewise analytic in Ω and gδ is the mollified representation
of g such that
gδ (x) = (g ∗ θN/δ )(x) .

Here
N N
θN/δ (x) = Φδ (Nx) = Φ(Nx/δ) ,
2π 2πδ
√
where δ = β1 N. Then the estimate

C β1N +1
||g − gδ ||H −2N−2 (Ω) ≤ (3.5.1)
2N +1 (N + 1)!N N +1

holds, where C is a positive constant.

Proof. Let φ(ξ) be the Fourier Transform of Φ(x). Then

N
!
2
− ξ2
X ξ 2j
Φ̂(ξ) = φ(ξ) = e . (3.5.2)
j=0
2j j!

We have
hg − gδ , ψi = hĝ − ĝδ , ψ̂i .

Now
ĝδ = g\
∗ θN/δ = ĝ θ̂N/δ .

Hence
hg − gδ , ψi = hĝ − ĝ θ̂N/δ , ψ̂i = hĝ, (1 − θ̂N/δ )ψ̂i .

Here
θ̂N/δ (ξ) = φ(δξ/N) .

46
 3.6. Error Estimates

Now
µN µN +1
eµ = 1 + µ + · · · + + eη ,
N! (N + 1)!
where 0 ≤ η ≤ µ. Hence
µN µN +1
|1 − e−µ (1 + µ + · · · + )| ≤ .
N! (N + 1)!
And so we obtain
β1N +1 ξ 2(N +1)
|1 − θ̂N/δ (ξ)| ≤ .
2N +1 (N + 1)!N N +1
Finally, we have
β1N +1 ||ξ 2(N +1) ψ̂||L2
|hg − gδ , ψi| = |hĝ, (1 − θ̂N/δ )ψ̂i| ≤ ||ĝ|| N +1
L2 ,
2 (N + 1)! N N +1
C β N +1 |ψ|H 2N+2
≤ N +1 1 .
2 (N + 1)!N N +1
Hence
|(g − gδ , ψ)|Ω
||g − gδ ||H −2N−2 (Ω) ≤ sup ,
ψ∈H 2N+2 (Ω) ||ψ||H 2N+2 (Ω)

C β1N +1
≤ .
2N +1 (N + 1)!N N +1

3.6 Error Estimates

In this section we recover point-wise values with spectral accuracy. We use the expo-
nentially accurate mollifier which was proposed by Tanner in his seminal paper [74] and
obtain the error estimate for the solution at a point (x0 , 1). Tadmor has also examined the
exponentially accurate mollifier in his erudite exposition [73].
√
Lemma 3.6.1. Let ǫ = γ1 N , where γ1 = ǫ1 dx . Here 0 < ǫ1 < 1 and dx = 1. The
estimate

|(v − v ∗ θN/ǫ )(x0 , 1)| ≤ Ce−ρN (3.6.1)

holds . Here C and ρ are constants and θN/ǫ (x) is the Hermite mollifier which is defined
in (3.2.6).

47
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

Proof. Note that v(x, t) is analytic for t > 0 and satisfies [69]

Dt Dx v(x, t) ≤ √ K
 j α 
α+2j
(j!)2 α! ,
( t)

where K, j and α are positive numbers.

Moreover

||v(x, t)eµ|x| ||L2 (Ω×{to }) ≤ C . (3.6.2)

for some µ > 0 and all t0 > τ > 0.

Now
Z ∞
|v − v ∗ θN/ǫ | = |v(x0 , 1) − v(x0 − y, 1)θN/ǫ (y)dy| ,
Z−∞
π Z
≤ |v(x0 , 1) − v(x0 − y, 1)θN/ǫ (y)dy| + | v(x0 − y, 1)θN/ǫ (y)dy| .
−π |y|≥π
| {z } | {z }
J1 J2

Here
Z π
J1 = v(x0 , 1) − v(x0 − y, 1)θN/ǫ (y)dy ,
Z−π
π
= v(x0 , 1) − v(x0 − y, 1)ΨN/ǫ (y)dy
−π
| {z }
L1
Z π Z π
+ v(x0 − y, 1)ΨN/ǫ (y)dy − v(x0 − y, 1)θN/ǫ (y)dy .
−π −π
| {z }
L2

Here as in [73] and (3.2.6) we define

+∞
N X
ΨN/ǫ (x) = Φǫ (N(x + 2πj)) . (3.6.3)
2π j=−∞

Then
|J1 | ≤ |L1 | + |L2 | .

48
 3.6. Error Estimates

Using the bound on regularization error (I1 + I2 ) in Theorem 11.6 in [73], where
Z
I1 = | ΨN/ǫ (y) (v(x0 , 1) − v(x0 − y, 1)) dy| ≤ Ce−ηN
ǫ1 dx ≤|y|≤π

and Z
I2 = | ΨN/ǫ (y) (v(x0 , 1) − v(x0 − y, 1)) dy| ≤ Ce−ηN ,
|y|≤ǫ1 dx
we have

|L1 | . Ce−ηN , (3.6.4)

where η is a positive constant.

Now we estimate L2 .
Z π Z π
L2 = v(x0 − y, 1)ΨN/ǫ(y)dy − v(x0 − y, 1)θN/ǫ (y)dy ,
−π −π
Z j=+∞
!
π X N
= v(x0 − y, 1) Φ(N(y + 2πj)/ǫ) dy .
−π j=−∞,j6=0
2πǫ

Now by (2.14a) of Lemma 2.2 in [73]

j=+∞ ∞
X N 2P X − ((2j−1)πN) 2
| Φ(N(y + 2πj)/ǫ)| . e 4ǫ2 .
j=−∞,j6=0
2πǫ ǫ j=1

Hence
j=+∞
X N 2P −η2 N/γ1
| Φ(N(y + 2πj)/ǫ)| . √ e , |x| ≤ π .
j=−∞,j6=0
2πǫ γ1 N
Here P = ǫ21 dx N = ǫ21 N. Hence we have 2P ≤ exp(κǫ21 N) with κ := log(2). Then
j=+∞
N 1
X  
η2 N
κǫ21 N −
| Φ(N(y + 2πj)/ǫ)| . √ e γ1
, |x| ≤ π .
j=−∞,j6=0
2πǫ γ1 N

And so
j=+∞
N 1
X  
η
κǫ21 − ǫ 2 N
| Φ(N(y + 2πj)/ǫ)| . √ e 1 , |x| ≤ π .
j=−∞,j6=0
2πǫ γ1 N
Since γ1 = ǫ1 dx and dx = 1.
For sufficiently small ǫ1 < 1 the above estimate is exponentially accurate.
Hence
′
|L2 | . Ce−η N , (3.6.5)

49
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

′
where η is a positive constant.
′
Choose ρ1 = min{η, η }. Then |J1 | . |L1 | + |L2 | satisfies

|J1 | . Ce−ρ1 N . (3.6.6)

By (2.14b) in [73], we have

2P −η1 N
|θN/ǫ (y)| . √ e f or |y| ≥ π , (3.6.7)
γ1 N

where η1 is a positive constant. And

||v(x, t)||L1(Ω×{t0 }) ≤ ||v(x, t)eµ|x| ||L2 (Ω×{t0 }) ||e−µ|x| ||L2 (Ω×{t0 }) ≤ C . (3.6.8)

Now from (3.6.7) and (3.6.8), an estimate for J2 is given as

|J2 | . Ce−ρ2 N , (3.6.9)

provided ǫ1 is small enough. Here ρ2 is a positive constant.

Choose ρ = min{ρ1 , ρ2 }. Hence

|v − v ∗ θN/ǫ | . Ce−ρN . (3.6.10)

Lemma 3.6.2. The estimate

|(v ∗ θN/ǫ − vδ ∗ θN/ǫ )(x0 , 1)| ≤ Ce−ρN , (3.6.11)

holds . Here C and ρ are generic constants and θN/ǫ (x) is the Hermite mollifier which is
√
defined in (3.2.6) and ǫ = γ1 N .

Proof. To verify the above bound, we define

I2 = v ∗ θN/ǫ − vδ ∗ θN/ǫ = (v − vδ ) ∗ θN/ǫ .

Hence
|I2 | ≤ ||v − vδ ||H −2N−2 ||θN/ǫ ||H 2N+2 .

50
 3.6. Error Estimates

Consider the adjoint problem

L∗ ψ = 0 in Ω × I , (3.6.12)

with initial conditon

N Nx
ψ = θN/ǫ (x) = Φ( ) on Ω × {1} .
2πǫ ǫ
Then
(v ∗ ψ)(x0 , 1) = (v ∗ ψ)(x0 , 0) .

Hence
((v − vδ ) ∗ ψ)(x0 , 1) = ((g − gδ ) ∗ ψ)(x0 , 0) .

Now from Lemma 3.4.1

||ψ||H s(Ω×{1}) ≤ BN s! α1s .

Here

√
√ 3
( N e/γ1 ) ′
√
′ ′ N
BN = C N e (N e /γ1 ) 4 ∼ K Nd a log N
.

Hence
||ψ||H s(Ω×{0}) ≤ CBN s! α1s ,

for some constants C and α1 .

By Lemma 3.5.1

|I2 | = |((g − gδ ) ∗ ψ)(x0 , 0)| ≤ ||g − gδ ||H −2N−2 ||ψ||H 2N+2 , (3.6.13)
C β1N +1
≤ N +1 BN (2N + 2)! α12N +2 . (3.6.14)
2 (N + 1)!N N +1

By Stirling’s Formula
√
n! ∼ 2 π n nn e−n .

Hence

2
|I2 | ∼ C(2β1 α12 )N +1 BN , (3.6.15)
e(N +1)
51
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

∼ C1 e−ρN . (3.6.16)

4β1 (α1 )2
 √ √
provided β1 is small enough and satisfies e
< 1. Here δ = β1 N and ǫ = γ1 N .

Theorem 3.6.3. Define

ω p = ωlp in Ωl × I f or − N ≤ l ≤ N − 1 , (3.6.17)

= 0, otherwise. (3.6.18)

√
Let ǫ = γ1 N , where γ1 = ǫ1 dx . Here 0 < ǫ1 < 1 and dx = 1. If vδ ∈ D2,1 (Ω̄ × [0, 1]) then
the following error estimate holds

|v(x0 , 1) − (ω p ∗ ψ)(x0 , 1)| ≤ C1 e−ρ N , (3.6.19)

for any x0 ∈ IN = [−N, N], provided q is proportional to p2 , as p tends to infinity and N

is proportional to p. Here C1 and ρ are generic constants and ψ = θN/ǫ (x) is the Hermite
mollifier which is defined in (3.2.6).

Proof. Now

|(v − ω p ∗ ψ)(x0 , 1)| = |((v − v ∗ ψ) + (v ∗ ψ − vδ ∗ ψ) + (vδ ∗ ψ − ω p ∗ ψ))(x0 , 1)| ,

(3.6.20)

= |I1 + I2 + I3 | , (3.6.21)

≤ |I1 | + |I2 | + |I3 | . (3.6.22)

Here

I1 = (v − v ∗ ψ)(x0 , 1) .

Then from Lemma 3.6.1 the following result holds

|I1 | = |(v − v ∗ ψ)(x0 , 1)| ≤ C1 e−ρN . (3.6.23)

52
 3.6. Error Estimates

Now

I2 = (v ∗ ψ − vδ ∗ ψ)(x0 , 1) , (3.6.24)

= ((v − vδ ) ∗ ψ)(x0 , 1) . (3.6.25)

Using Lemma 3.6.2, we have

|I2 | ≤ C1 e−ρN . (3.6.26)

From (3.4.34) and Lemma 3.4.1 the following result holds

|I3 | = |((vδ − ω p ) ∗ ψ)(x0 , 1)| ≤ ||(vδ − ω p )||L2 ||ψ||L2 ≤ C1 e−ρ1 N . (3.6.27)

Using (3.6.23), (3.6.26) and (3.6.27) we obtain

|(v − ω p ∗ ψ)(x0 , 1)| ≤ C1 e−ρ N .

Now we want to recover point-wise values at an interior point (x0 , t0 ) with spectral
accuracy. Assume that ω p (x, t) ∈ D2,1 (O), where the set O is

O = {(x, t) : |x − x0 | ≤ δ1 , |t − t0 | ≤ ǫ1 } ⊆ R × (0, 1).

Here we use the Hermite mollifier, which is defined in (3.2.6), to recover the value in space
direction. We use the root exponential accurate mollifier [73] to recover the value in time
direction. Define the root exponential accurate mollifier
   
1 t t ( ct2
)
ΘQ,δ2 (t) = η1 DQ , η1 := e t2 −π 2 1(−π,π) (t), c > 0 , (3.6.28)
δ2 δ2 δ2
1 √
with adaptive parameterization, δ2 = dt := π
dist{t, {0, 1}}[modπ] and Q ∼ dt N 2 / e.
Here DQ (t) denotes the Dirichlet kernel

 sin(Q+1/2)t t 6= 2mπ ,
2πsin(t/2)
DQ (t) := (3.6.29)
 2Q + 1 t = 2mπ .

53
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

Now we define the regularized version of ω p at (x0 , t0 ) as:

Z π Z πdt
p
Rω (x0 , t0 ) = θN/ǫ (x)ΘQ,δ2 (t)ω p (x0 − x, t0 − t)dxdt ,
−π −πdt

and the regularized version of the Dxα Dtj ω p at (x0 , t0 ) as:

Z π Z πdt
α j p
RDx Dt ω (x0 , t0 ) = (−1) α+j
Dxα θN/ǫ (x) Dtj ΘQ,δ2 (t) ω p (x0 − x, t0 − t)dxdt .
−π −πdt

Once again it can be shown that this regularized version of w p (x0 , t0 ) approximates v(x0 , t0 )
with exponential accuracy. Moreover the regularized version of the derivative approximates
the actual value of the derivative with exponential accuracy.

3.7 Computational Results

The efficacy of the proposed computational strategy is established through numerical ex-
amples. All computations have been done on 372-node HPC cluster which is based on
n Intel Xeon Quadcore processors with a total of 2944 cores and high-speed Infiniband
network and it has a peak performance of 34.5 TF. The details of the configuration of Intel
Xeon CPU [email protected] are as follows : Number of CPU (Physically)-2, Cores per
CPU (Physically and after Hyper-Threading)-4, Total CPU cores (Physically)-8, Number
of CPU (after Hyper-Threading)-4, Total CPU cores (after Hyper-Threading)-16, RAM
-24 GB, HDD Capacity-2X500 GB.

Example 3.7.1 (Nonsmooth initial data). Consider the problem

ut − uxx = 0 in Ω × (0, 1) , (3.7.1)

u(x, 0) = f (x) on Ω × {0} . (3.7.2)

Here 
 1 x ∈ (0, 1) ,
f (x) =
 0 otherwise .

54
 3.7. Computational Results

Table 3.1: Point-wise error as function of p

p q Error(1, 1) Error(0, 1) Error(−1, 1) Iterations No. of cores Cpu(s)

5 25 7.25 × 10−5 7.34 × 10−5 6.98 × 10−5 144 10 1.9
6 36 1.88 × 10−5 1.87 × 10−5 9.27 × 10−6 163 12 3.1
7 49 1.22 × 10−6 1.14 × 10−6 9.35 × 10−7 175 14 26.7
8 64 1.99 × 10−7 1.99 × 10−7 9.97 × 10−8 186 16 36.9
9 81 9.87 × 10−9 9.75 × 10−9 8.75 × 10−9 195 18 49.8
10 100 9.92 × 10−10 9.92 × 10−10 8.87 × 10−10 204 20 61.2

0.35
exact solution
numerical solution
0.3 of p=10

0.25

0.2
u(x,1)

0.15

0.1

0.05

0
−5 0 5
x axis

Figure 3.3: Numerical and Exact solution at t=1

From the numerical results given in Table 3.1, Figures 3.3, 3.4, 3.5, 3.6 and 3.7 , it
can be seen that the point-wise errors of solution and its derivative decay rapidly with
polynomial order p. Further, from Table 3.1 it is observed that number of iterations from
the PCGM method increases marginally with p, though the computational time increases
due to increased matrix size as p is increased. This example validates the efficacy of the
proposed method (i.e. LSSEM). In the next two examples, the European options problem
is dealt with.

55
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

0.06
exact derivative
numerical derivative
0.04

0.02
ux(x,1)

−0.02

−0.04

−0.06
−5 0 5
x axis

Figure 3.4: Derivative (ux ) of Numerical and Exact solution

−3
10

−4
10
Error=|(ux−(u ∗ Dθδ/N))(x0; 1)|

−5
10

−6
10
p

−7
10 p=5
p=6
p=7
−8
10 p=8
p=9
p=10
−9
10
−5 0 5
x

Figure 3.5: Point-wise error between derivative of Numerical and Exact solution

56
 3.7. Computational Results

0.1
second exact derivative
second numerical derivative
for p=10
0.05

0
uxx(x,t)

−0.05

−0.1

−0.15
−5 0 5
x axis

Figure 3.6: Second Derivative (uxx ) of Numerical and Exact solution

−2
10

−3
10
Error=|(uxx−(u ∗ D θδ/N))(x0; 1)|

−4
10
2

−5
10
p

−6
10 p=5
p=6
p=7
−7
10 p=8
p=9
p=10
−8
10
−5 0 5
x

Figure 3.7: Point-wise error between second derivative (uxx ) of Numerical and Exact solution

57
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

Example 3.7.2 (European Black-Scholes Put Options problem). Here a problem of the
“European Black-Scholes Put option” is considered. The method is used to solve this prob-
lem and the results are compared with those due to Zhu et al [77]. Consider the problem
:
1
Vτ − σ 2 S 2 VSS − rSVS + rV = 0 in (0, ∞) × [0, T ] ,
2
V (S, 0) = max(K − S, 0) on Ω × {0} .

Here V, S, K, r and σ are respectively option price, underlying asset price, strike price,
risk-free interest rate and volatility .

Table 3.2: Put Option problem : Variable value from Zhu et al [77]

K r σ T
100 0.05 0.15 0.25

Table 3.3: Put Option problem : Point-wise error as function of p for LSSEM

p q Error(0, 1) Error(−1, 1) Error(−2, 1) Iteration No. of cores CP U(s)

5 25 6.81.10−5 5.98.10−5 5.92.10−5 151 10 2.1
6 36 6.12.10−6 5.32.10−6 5.29.10−6 178 12 3.2
7 49 5.87.10−7 5.23.10−7 5.14.10−7 190 14 27.5
8 64 5.96.10−8 4.99.10−8 4.88.10−8 202 16 38.1
9 81 6.67.10−9 5.87.10−9 5.57.10−9 213 18 51.3
10 100 6.24.10−10 5.22.10−10 5.22.10−10 226 20 61.6

The results obtained using the proposed method are given in Table 3.3. From Tables
3.3, it can be observed that :

1. In order to achieve an accuracy of 10−6 , LSSEM requires p = 6, q = 36 and the

computational time required is only 3.2 s..

58
 3.7. Computational Results

100

Exact Solution
Numerical Solution for p=10
80

60

u(K exp(x), 1)
40

20

−20
0 100 200 300 400 500 600 700 800
Underlying Asset Price (S= K exp(x))

Figure 3.8: Numerical solution and Exact solution at t=1

−0.2

−0.4

−0.6
Delta (∆)

−0.8

−1
Exact Derivative
Numerical Derivative for p=10
−1.2

−1.4
0 100 200 300 400 500 600 700 800
Underlying Asset Price (S= K exp(x))

Figure 3.9: Derivative (∆) of Numerical and Exact solution

2. LSSEM can easily obtain high accuracies. For examples, an accuracy of 10−10 is
obtained with only q = 100.

59
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

−3
10

−4
10

Error=|(ux−(u ∗ Dθδ/N))(x0; 1)|

−5
10

−6
10
p

−7
10 p=5
p=6
p=7
−8
10 p=8
p=9
p=10
−9
10
−2 −1.5 −1 −0.5 0
x

Figure 3.10: Point-wise error between derivative (∆) of Numerical and Exact solution

0.06

Second Exact Derivative

0.05 Second Numerical Derivative for p=10

0.04
Gamma (Γ)

0.03

0.02

0.01

−0.01
0 100 200 300 400 500 600 700 800
Underlying Asset Price (S= Kexp(x))

Figure 3.11: Second derivative (Γ) of Numerical and Exact solution

3. From Figures 3.9, 3.10, 3.11 and 3.12, we observe that the errors of derivatives also
decay exponentially with polynomial order p.

60
 3.7. Computational Results

−2
10

−3
10

Error=|(uxx−(u ∗ D θδ/N))(x0; 1)|

−4
10
2

−5
10
p

−6 p=5
10
p=6
p=7
−7 p=8
10
p=9
p=10
−8
10
−2 −1.5 −1 −0.5 0
x

Figure 3.12: Point-wise error between second derivative (Γ) of Numerical and Exact solution

4. Number of iterations for PCGM increases marginally with p.

5. LSSEM is exponentially accurate theoretically as well as numerically .

Example 3.7.3 (European Black-scholes Call Options problem (Put-Call Parity)). Usu-
ally, in the literature, the “European Black-Scholes Put option” problem is solved. Few
researchers, e.g. Bunnin et. al. [24] have addressed the “European Black-Scholes Call
option” problem. The difficulty is due to an unbounded initial state. In the following the
Call option problem is solved and the results are compared with those due to Bunnin et. al.
[24]. Consider the problem

1
Vτ − σ 2 S 2 VSS − rSVS + rV = 0 in (0, ∞) × [0, T ] ,
2

V (S, 0) = max(S − K, 0) on Ω × {0} .

Here V, S, K, r and σ are respectively option price, underlying asset price, strike price,
risk-free interest rate and volatility.

61
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

Table 3.4: Call Option problem : Variable value from Bunnin et al [24]

K r σ T
10 0.1 0.4 1

Table 3.5: Call Option : Point-wise error as function of p for LSSEM

p q Error(1, 1) Error(2, 1) Error(3, 1) Iteration No. of cores Cpu(s)

5 25 7.12.10−5 7.10.10−5 7.03.10−5 246 10 2.5
6 36 7.09.10−6 7.09.10−6 7.14.10−6 283 12 3.9
7 49 6.96.10−7 6.96.10−7 6.98.10−7 319 14 29.6
8 64 6.03.10−8 6.03.10−8 6.06.10−8 356 16 40.3
9 81 7.18.10−9 7.18.10−9 7.23.10−9 389 18 54
10 100 7.96.10−10 7.96.10−10 7.92.10−10 412 20 65.7

Table 3.6: Call Option problem : Point-wise error, as reported in Bunnin et al [24]

N StockP rice Error

100 3 −0.1059
100 6 −0.0021
100 9 0.0020
100 12 0.0012
100 15 0.0014
100 20 0.0043

In Tables 3.5 and 3.6 the results are presented. From these results it can be seen that

1. In [24] an accuracy of 10−3 is achieved for N = 100, while LSSEM achieves an

accuracy of 10−5 with p = 5, q = 25.

2. LSSEM achieves an accuracy of 10−10 for p = 10, q = 100.

3. In Figures 3.14, 3.15, 3.16 and 3.17, the errors in derivatives also decay rapidly.

62
 3.7. Computational Results

70

60

50

Option Price u(K exp(x), 1)

40

30

20

10 Exact Solution
Numerical Solution of p=10
0

−10
0 10 20 30 40 50 60 70 80
Underlying Asset Price (S= K exp(x))

Figure 3.13: Numerical solution and Exact solution at t=1

1.2

0.8

0.6
Delta (∆)

0.4

0.2

0 Exact Derivative
Numerical Derivative for p=10

−0.2
0 10 20 30 40 50 60 70 80
Underlying Asset Price (S= K exp(x))

Figure 3.14: Derivative (∆) of Numerical and Exact solution at t=1

4. LSSEM achieves exponential accuracy.

63
 Chapter 3. Nonconforming LSSEM for European Options with Single Asset

−3
10

−4
10

Error=|(ux−(u ∗ Dθδ/N))(x0; 1)| 10

−5

−6
10
p

−7
10 p=5
p=6
p=7
−8
10 p=8
p=9
p=10
−9
10
0 0.5 1 1.5 2 2.5 3
x

Figure 3.15: Point-wise error between derivative (∆) of Numerical and Exact solution

0.14
Second Exact Derivative
0.12 Second Numerical Derivative of p=10

0.1

0.08
Gamma (Γ)

0.06

0.04

0.02

−0.02
0 10 20 30 40 50 60 70 80
Underlying Asset Price (S= K exp(x))

Figure 3.16: Second derivative (Γ) of Numerical and Exact solution at t=1

64
 3.7. Computational Results

−2
10

−3
10
Error=|(uxx−(u ∗ D θδ/N))(x0; 1)|

−4
10
2

−5
10
p

p=5
−6
10 p=6
p=7
p=8
−7
10 p=9
p=10

−8
10
0 0.5 1 1.5 2 2.5 3
x

Figure 3.17: Point-wise error between second derivative (Γ) of Numerical and exact solution

65
 Chapter
4
Nonconforming LSSEM for European
Options with Multi Assets

4.1 Introduction

In the previous Chapter 3, we discussed how Black-Scholes equation for European Put
Option and Call Option can be solved numerically, where Call Option gives the right to
holder, but not the obligation, to buy an asset at a prescribed price K (strike price) at
maturity time T in future (The Payoff Function for European Call Option is VC (S, T ) =
max(S − K, 0)) and Put Option gives the right to holder, but not the obligation, to sell
an asset at a prescribed price K (strike price) at maturity time T in future (The Payoff
Function for European Put Option is VP (S, T ) = max(K − S, 0)). Now we can easily
extend our method, the Nonconforming Least-Squares Spectral Element Method, to multi
assets, for example - European Basket Options and Rainbow Put Options, etc.
In European Basket Options where analytic solution does not exist, whose Payoff Function
depends on the Value of a Portfolio of Risky Assets, like the Basket Put Option on two

67
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

Assets

Vp (S1 , S2 , T ) = max(K − a1 S1 − a2 S2 , 0), (4.1.1)

where a1 , a2 > 0.
Now we discuss another example, where an analytical solution exists. European rainbow
option is a benchmark problem for which there is an analytic solution [78]. The Pay-off
Function is

Vp (S1 , S2 , T ) = max(K − max(S1 , S2 ), 0). (4.1.2)

Recently, Zhu et al. [78] solved the European Rainbow Option by Spectral Element Method
in 2010. This method gives quadratic convergence in time and is spectrally accurate in
space. But they do not have rigorous error estimates. Our method which is a Nonconform-
ing Least-Squares Spectral Element Method, is exponentially accurate in time and space.
We have used Sobolev spaces of different orders in space and time to formulate our results,
as given in [57].

Now we give the details of this paper. In Section 4.2 the function spaces and a priori
estimates for parabolic initial value problems, as presented in [46, 53, 57], are given. In
Section 4.3 discretization of the domain and stability estimates are discussed. In Section
4.4 we discuss the numerical scheme, symmetric formulation, parallelization and precon-
ditioning for our method. In Section 4.5 the estimate for non smooth initial condition
is presented in negative Sobolev norms. In Section 4.6 error estimates are obtained for
this method. Finally, in Section 4.7 specific numerical examples are given to show the
effectiveness of this method.

68
 4.2. Function Space For the Formulation

4.2 Function Space For the Formulation

4.2.1 Definitions and Prelimineries

Let X be a Hilbert space. Define the Hilbert space H 0 (a, b; X) of functions f , strongly
measurable on [a, b] with range in X, with the norm
Z b
||f ||2H 0(a,b;X) = ||f (t)||2X dt < ∞. (4.2.1)
a

where ||.||X is the Hilbert norm of X [56]. Let m be an integer and define

df dm f
m
H (a, b; X) = {f |f, f = , ...., f = m ∈ H 0 (a, b; X)}.
m 1
dt dt

Then H m (a, b; X) is a Hilbert space with the scalar product defined by [56, 57]
m Z
X b
(f, g)H m = (f j (t), g j (t))X dt.
j=0 a

Let σ(0 < σ < 1) be a non-negative real number and define

Z b Z bZ b
||f (t1) − f (t2 )||2X
||f ||2H σ (a,b;X) = ||f (t)||2X dt + dt1 dt2 . (4.2.2)
a a a |t1 − t2 |1+2σ

Here H σ (a, b; X) is also a Hilbert space with the norm defined in (4.2.2).
Let r, s be non-negative real numbers and ω(x1 , x2 , t) be a smooth function. Now we define
some Sobolev spaces as in [57].

H r,s (Ω × I) = H 0 (I; H r (Ω)) ∩ H s (I; H 0 (Ω))

which is a Hilbert space with norm

Z
||ω||2H r,s(Ω) = ||ω(t)||2H r (Ω) dt + ||ω||2H s(I;H 0 (Ω)) . (4.2.3)
I

Next, let h(x1 , x2 ) be a smooth function, with

Z X X
2
||h||H r (Ω) = |∂xα1 ∂yα2 h|2 dxdy < ∞. (4.2.4)
Ω |α|≤r α +α ≤α
1 2

69
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

Define the negative Sobolev norms

|(ω, Φ)Ω |
||ω||H −m(Ω) = sup (4.2.5)
Φ∈H m (Ω) ||Φ||H m (Ω)

and

|(ω, Φ)Ω×I |
||ω||H −r,−s(Ω×I) = sup . (4.2.6)
Φ∈H r,s (Ω×I) ||Φ||H r,s (Ω×I)

We now define some Gevrey Spaces which are needed for our error analysis.

Definition 4.2.1. Let

D1 (Ω̄) ={Φ(x) : Φ is an infinitely differentiable function in Ω̄ such that there exist two
positive numbers A1 and B1 such that

||DxαΦ(x)||L2 (Ω̄) ≤ A1 (B1 )i i! , |α| = i, f or all i = 0, 1, 2, 3 · ··} .

Definition 4.2.2. Let

D2,1 (Ω̄ × [0, 1])
= {ψ(x, t) : ψ is an infinitely differentiable function in Ω̄ × [0, 1] such that there exist two
positive numbers A1 and B1 such that

||DxαDtj ψ(x, t)||L2 (Ω̄×[0,1]) ≤ A1 (B1 )i+j i!(j!)2 , |α| = i, f or all i, j ≥ 0} .

4.2.2 Black-Scholes Equation with Two Assets

Consider the initial value problem
1 1
Vτ + σ12 S12 VS1 S1 + ρσ1 σ2 S1 S2 VS1 S2 + σ22 S22 VS2 S2 + rS1 VS1 + rS2 VS2 − rV = 0
2 2 (4.2.7)
in R2+ × (0, T ),

V (S1 , S2 , T ) = F (S1 , S2 ) on R2+ × {T } .

70
 4.2.2. Black-Scholes Equation with Two Assets

For example F (S1 , S2 ) may be max(K − S1 − S2 , 0) or max(K − max(S1 − S2 ), 0).

Now we use the transformation x1 = log(S1 /K), x2 = log(S2 /K) for Ω = R2 as the domain
T −τ
of logarthimic price and define t = T
on the time interval I = [0, 1]. We shall focus here
on Black -Scholes equation for the European Call with the assumption that rate of interest
r and volatility σ1 , σ2 of two assets are smooth (or even analytic) functions of x = (x1 , x2 )
and t with bounded derivatives. The coefficients a1 , a2 , ... and a6 belong to D2,1 (Ω × I), i.e.

||Dxα Dtj an (x, t)||L∞ (Ω×I) ≤ AB i+j i!(j!)2 f or all n = 1, 2, · · · , 6.

where A and B are positive numbers and |α| = i, for all j = 0, 1, 2, 3 · · · .

The price u(x1 , x2 , t) has to satisfy the BS equation

Lu = ut − a1 ux1x1 − a2 ux2 x2 − a3 ux1 x2 − a4 ux1 − a5 ux2 − a6 u = 0 in Ω × I ,

(4.2.8)
u(x1 , x2 , 0) = f (x1 , x2 ) in Ω × {0} .

f (x1 , x2 ) may not be in L2 (Ω) for example f (x1 , x2 ) = (K − Kmax(ex1 , ex2 ))+ . To remove
this difficulty, we multiply by a weight function in u and construct a new partial differential
equation. Let

v(x1 , x2 , t) = u(x1 , x2 , t)sech(ηx1 )sech(ηx2 ) (4.2.9)

where η > 0 is sufficiently large so that

v(x1 , x2 , 0) = u(x1 , x2 , 0) sech(ηx1)sech(ηx2 ) (4.2.10)

is such that veµ|x| at t = 0 belongs to L1 (Ω) ∩ L2 (Ω) for some µ > 0.

Hence v satisfies the new partial differential equation

Lv = vt − α1 vx1 x1 − α2 vx2 x2 − α3 vx1 x2 − α4 vx1 − α5 vx2 − α6 v = 0 in Ω × I,

(4.2.11)
v(x1 , x2 , 0) = f (x1 , x2 ) sech(ηx1 ) sech(ηx2 ) in Ω × {0} .

We assume the coefficients a1 , a2 , ...a6 in (4.2.8) are smooth or even analytic and all deriva-
tives are bounded. So the same assumption will continue to hold for the coefficients
α1 , α2 , ...α6 . Moreover the coefficients belong to D2,1 (Ω̄ × [0, 1]).

71
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

The initial data g(x1 , x2 ) = f (x1 , x2 )sech(ηx1 )sech(ηx2 ) is not smooth. To resolve this
difficulty we use the Hermite mollifiers [73, 74]
P  
− x2
2 X (−1)j x
Φ(x) = e H2j √ . (4.2.12)
j=0
4j j! 2

Define

   
N N N Nx1 N Nx2
θN/δ (x1 , x2 ) = Φδ (Nx1 ) Φδ2 (Nx2 ) = Φ Φ , (4.2.13)
2π 1 2π 2πδ1 δ1 2πδ2 δ2
Then
(Nx1 )2
− P1  
2δ 2
e 1 X (−1)j Nx1
Φδ1 (Nx1 ) = H2j √
δ1 j=0
4j j! 2δ1
and
(Nx2 )2
− P2  
2δ 2
e 2 X (−1)j Nx
Φδ2 (Nx2 ) = H2j √ 2 ,
δ2 j=0
4j j! 2δ2
√ √
where δ1 = β1 N , δ2 = β2 N , β1 = θ1 dx1 , β2 = θ2 dx2 , P1 = θ12 dx1 N and P2 = θ22 dx2 N.
Here

1 1
d x1 = dist{x1 , {c1 , ..., cj }}[modπ], dx2 = dist{x2 , {d1 , ..., dj }}[modπ].
π π
(c1 , d1 ), ...., (cj , dj ) are singularity points around (x1 , x2 ). The neighborhood of analyticity
around (x1 , x2 ) is (x1 − πdx1 , x1 + πdx1 ) × (x2 − πdx2 , x2 + πdx2 ) and 0 < θ1 , θ2 < 1.
First we mollify the initial function g(x1 , x2 ) by using Hermite Mollifier which is given in
equation (4.2.13). The mollified representation of initial function g(x1 , x2 ) is denoted as
gδ (x1 , x2 ), where
Z Z
gδ (x1 , x2 ) = (g ∗ θN/δ )(x1 , x2 ) = g(x1 − x0 , x2 − y0 )θN/δ (x0 , y0 )dx0 dy0
|x0 |≤πdx1 |y0 |≤πdx2

∀ x1 , x2 ∈ Ω.
(4.2.14)

Then we can find vδ (x1 , x2 , t) which satisfies the following Initial value problem

Lvδ = 0 in Ω × I ,

72
 4.3. Discretization and Stability Estimates

vδ = gδ on Ω × {0} . (4.2.15)

Here L is the differential operator given in (4.2.11).

Now consider the initial value problem (IVP)

Lω = F in Ω × I ,

ω = h on Ω × {0} ,

where F and h are smooth. Then the following a priori estimate holds

||ω||2H 2r+2,r+1(Ω×I) + ||ω||2H 2r+1(Ω×{1}) ≤ Cr (||Lω||2H 2r,r (Ω×I) + ||ω||2H 2r+1(Ω×{0}) ), (4.2.16)

where Cr is a constant which depends on r .

4.3 Discretization and Stability Estimates

4.3.1 Discretization
Choose integers N and p and let p be proportional to N. The truncated domain ΩN =
[−N, N] × [−N, N] where we solve the problem (4.2.15), is divided into a number of sub-
N −1 S N −1
domains {Ωl,m }l,m=−N . Here Ωl,m = (l, l + 1) × (m, m + 1) and ΩN = {Ω̄l,m }l,m=−N .
N −1
Each of these sub-domains {Ωl,m }l,m=−N is mapped to the unit square S = (0, 1) × (0, 1)
−1 N −1
by a set of smooth maps {Ml,m }l,m=−N , where Ml,m is a map from S = (0, 1) × (0, 1) to
Ωl,m = (l, l + 1) × (m, m + 1) as in [37, 75]. The map Ml,m has the form

 x = (1 − ξ )(1 − ξ )X + (1 − ξ )ξ X + (1 − ξ )ξ X + ξ ξ X ,
1 1 2 1 1 2 4 2 1 2 1 2 3
Ml,m =
 x = (1 − ξ )(1 − ξ )Y + (1 − ξ )ξ Y + (1 − ξ )ξ Y + ξ ξ Y ,
2 1 2 1 1 2 4 2 1 2 1 2 3

where ξ1 , ξ2 ∈ (0, 1). Here Xi and Yi are constants for i = 1, · · · , 4.

The mesh size h of the sub-domain and time step k both are 1 because k is proportional
to h2 = 1. Now the boundary of ∂Ωl,m consists of four stright line segments. Let γs denote
an enumeration of these straight line segments which are contained in ΩN .

73
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

4.3.2 Stability Estimates

p
Define the spectral element function v̌l,m (ξ1 , ξ2 , t) which is a polynomial of degree p in each
of the space variables ξ1 and ξ2 and of degree q in the time variable t, i.e.
p p q
p
X X X l,m i j k
v̌l,m (ξ1 , ξ2, t) = δi,j,k ξ1 ξ2 t
i=0 j=0 k=0

l,m
for ξ1 , ξ2 ∈ (0, 1), t ∈ [0, 1]. Here δi,j,k denote the coefficients and q is proportional to p2 .
Then
p p −1
vl,m (x1 , x2 , t) = v̌l,m (Ml,m (x1 , x2 ), t).

Define v p (x1 , x2 , t) as

p
v p (x1 , x2 , t) = vl,m (x1 , x2 , t), f or (x1 , x2 , t) ∈ Ωl,m × I,

= 0, f or (x1 , x2 , t) ∈ (ΩN )c × I, (4.3.1)

for all −N ≤ l, m ≤ N − 1.
Using the chain rule, we can write
p
∂vl,m p p
= (v̌l,m )ξ1 (ξ1 )x1 + (v̌l,m )ξ2 (ξ2 )x1 (4.3.2)
∂x1
and
p
∂vl,m p p
= (v̌l,m )ξ1 (ξ1 )x2 + (v̌l,m )ξ2 (ξ2 )x2 . (4.3.3)
∂x2
Assume (ξˆ1 )x1 , (ξˆ1 )x2 , (ξˆ2 )x1 and (ξˆ2 )x2 are the polynomials of orthogonal projection of
(ξ1 )x1 , (ξ1 )x2 , (ξ2 )x1 and (ξ2 )x2 into the space of polynomials of degree p with respect to the
inner product in H 2 (S). Then the polynomial approximation of the above equation (4.3.2)
and (4.3.3) at the interior point is defined by
 p a
∂v̌l,m p
= (v̌l,m )ξ1 (ξˆ1 )x1 + (v̌l,m
p
)ξ2 (ξˆ2 )x1 (4.3.4)
∂x1
and
 p a
∂v̌l,m p
= (v̌l,m )ξ1 (ξˆ1 )x2 + (v̌l,m
p
)ξ2 (ξˆ2 )x2 . (4.3.5)
∂x2

74
 4.3.2. Stability Estimates

Assume γs is a side common to Ωl,m and Ωn,o which is the image of ξ1 = 1 under the map
Ml,m and the image of ξ1 = 0 under the map Mn,o . Then we define the approximation to
the jump in the derivative at the inter element boundary γs as
 p a 2   p a 2
 ∂v̌   ∂v̌ p a ∂v̌ 
 n,o l,m 

 ∂x1
 =  (0, ξ2, τ ) − (1, ξ2, τ ) .
 r,s
H ({γs }×I)
 ∂x1 ∂x1  r,s
H ((0,1)×I)

Now
Z Z
p p 2 p 2 p 2

||vl,m ||2H 1 (Ωl,m ×{0}) = |v̌l,m | Jl,m dξ1 dξ2 + |∂ξ1 v̌l,m | + |∂ξ2 v̌l,m | Jl,m dξ1 dξ2
S×{0} S×{0}

(4.3.6)
and
Z Z
p 2 p 2
|L vl,m | dx1 dx2 dt = |Ll,mv̌l,m | Jl,m dξ1dξ2 dt, (4.3.7)
Ωl,m ×(0,1) S×(0,1)

where Ll,m is the differential operator L in ξ1 , ξ2 and t coordinates and Jl,m is the Jacobian
of the map Ml,m from S to Ωl,m . Now we take the orthogonal projection of the coefficients of
the differential operator Ll,m into the space of polynomials with respect to the usual inner
product in H 2,1 (S × (0, 1)) and define a new differential operator Ll,m
a
. The coefficients
a
of the differential operator Ll,m are polynomials of degree p in (ξ1 , ξ2 ) and of degree q in
t. The orthogonal projection of Jl,m is denoted by Jˆl,m into the space of polynomials of
degree p in (ξ1 , ξ2) coordinates with respect to the usual inner product in H 2 (S). Hence
Z Z
p 2 a p 2 ˆ
|L vl,m | dxdt = |Ll,m v̌l,m | Jl,m dξ1dξ2 dt (4.3.8)
Ωl,m ×(0,1) S×(0,1)

upto a negligible error term [17, 37].

Define the quadratic form
N
X −1 N
X −1 q X
p
W p ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1 ) = a
||(Ll,m p
v̌l,m ) Jˆl,m ||2L2 (S×I) + (||[v̌ p]||2H 0,3/4 (γs ×I)
l=−N m=−N γs ⊆ΩN
N
X −1 N
X p
+||[(v̌xp1 )a ]||2H 1/2,1/4 (γs ×I) + ||[(v̌xp2 )a ]||2H 1/2,1/4 (γs ×I) ) + ||v̌l,m ||2H 1 (Ωl,m ×{0}) .
l=−N m=−N −1

(4.3.9)
Here ΩN = [−N, N] × [−N, N].
Then from Theorem 11 in [37] the following result holds.

75
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

Theorem 4.3.1. ( Stability estimate) There exists a constant C such that the estimate
N
X −1 N
X −1
p p
||v̌l,m ||2H 2,1 (S×I) ≤ C (ln p)2 W p ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1) (4.3.10)
l=−N m=−N

holds.

4.4 Numerical Scheme and Parallelization

Define (gδ )l,m (ξ1 , ξ2) = (gδ )l,m (Ml,m (ξ1 , ξ2 )) and let (ĝδ )l,m (ξ1 , ξ2 ) be the orthogonal projec-
tion of (gδ )l,m (ξ1 , ξ2) into the space of polynomials of degree p in (ξ1 , ξ2 ) with respect to
the usual inner product in H 1 (S).
p
We define our approximate solution to be the unique {ωl,m }l,m=−N,N −1 which minimizes
the functional
N
X −1 N
X −1 q
R p p
({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1) = a
||(Ll,m p
v̌l,m ) Jˆl,m ||2L2 (S×I)
l=−N m=−N
X
+ (||[v̌ p ]||2H 0,3/4 (γs ×I) + ||[(v̌xp1 )a ]||2H 1/2,1/4 (γs ×I) + ||[(v̌xp2 )a ]||2H 1/2,1/4 (γs ×I) ) (4.4.1)
γs ⊆ΩN
N
X −1 N
X −1
p
+ ||(v̌l,m − (ĝδ )l,m )||2H 1 (Ωl,m ×{0})
l=−N m=−N
p
over all {v̌l,m }−N ≤l,m≤N −1 . Here ΩN = [−N, N] × [−N, N].
Now consider the I.V.P.

Lvδ = 0 in Ω × I,

vδ = gδ on Ω × {0}. (4.4.2)

Clearly, gδ ∈ D1 (Ω × {0}). Hence vδ ∈ D2,1 (Ω × I). Now

vδ (x1 , x2 , 0) = gδ (x1 , x2 ) = g(x1 , x2 ) ∗ θN/δ (x1 , x2 ). (4.4.3)

And

||vδ ||H s (Ω×{0}) = ||g(x1, x2 ) ∗ θN/δ (x1 , x2 )||H s (Ω) ,

(4.4.4)
≤ C||g||L1(Ω) ||θN/δ (x1 , x2 )||H s(Ω) .
Here C is a generic constant.

76
 4.4. Numerical Scheme and Parallelization

Lemma 4.4.1. The estimate

|θN/δ |s ≤ 3AN BN s! (4.4.5)

holds . Here
 2  2
N N 1
θN/δ (x1 , x2 ) = Φδ1 (Nx1 )Φδ2 (Nx2 ) = Φ(Nx1 /δ1 )Φ(Nx2 /δ2 )
2π 2π δ1 δ2
√
√ ( N e/β1 )
is the Hermite mollifier which is defined in (4.2.13) and AN = N e (N e3 /β1 ) 4 ,
√
√ 3
( N e/β2 )
BN = N e (N e /β2 ) 4 .

Proof. We know that

 2  2
N N 1
θN/δ (x1 , x2 ) = Φδ1 (Nx1 )Φδ2 (Nx2 ) = Φ(Nx1 /δ1 )Φ(Nx2 /δ2 ),
2π 2π δ1 δ2
√ √
where δ1 = β1 N and δ2 = β2 N .
Using Lemma 3.4.1, we have
 
N 
 Φδ (Nx1 ) ≤ AN s! (4.4.6)
 2π 1 
s

and
 
N 
 Φδ2 (Nx2 ) ≤ BN s!, (4.4.7)
 2π 
s
√ √
√ ( N e/β1 ) √ ( N e/β2 )
where AN = N e (N e3 /β1 ) 4 and BN = N e (N e3 /β2 ) 4 .
Hence
!
  X
θN/δ  ≤ AN BN i!j! . (4.4.8)
s
i+j=s

Combining (4.4.6), (4.4.7) and (4.4.8), we obtain

 
θN/δ  ≤ 3AN BN s!. (4.4.9)
s

77
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

From Lemma 4.4.1, we obtain

||gδ ||H s (Ω×{0}) ≤ C AN BN s!.

Hence

√
||gδ ||H s(Ω×{0}) ≤ EN s! ∼ K da N log N
s!, (4.4.10)
√
where K, d and a are constants. Here EN = K da N log N
.
Then by results on Gevrey spaces in Lions and Magenes [57, 58]

||Dxβ Dtj vδ ||L2(Ω×I) ≤ C EN i!(j!)2 (α1 )i+j , (4.4.11)

where α1 is a constant, for |β| = i. Here x = (x1 , x2 ).

Moreover

||gδ e±µx1 ||H 1 (Ω×{0}) ≤ C . (4.4.12)

||gδ e±µx2 ||H 1 (Ω×{0}) ≤ C . (4.4.13)

Here µ > 0 is a constant.

Hence from [53]

||vδ e±µx1 ||H 2,1 (Ω×I) ≤ C . (4.4.14)

||vδ e±µx2 ||H 2,1 (Ω×I) ≤ C . (4.4.15)

Thus

||vδ ||H 2,1 (ΩcN ×I) ≤ Ke−ρN . (4.4.16)

Here K and ρ are generic constants and ΩcN = R2 \ ΩN , where ΩN = [−N, N] × [−N, N].
Let spl,m (x1 , x2 , t) be the approximate representation of vδ (x1 , x2 , t) on Ωl,m defined in The-
orem 13 in [37]. Then
N
X −1 N
X −1
||vδ − spl,m ||H 2,1 (Ωl,m ×I) ≤ K e−ρ N (4.4.17)
l=−N m=−N

78
 4.4.1. Symmetric Formulation

provided q is proportional to p2 and N is proportional to p.

Here spl,m (x1 , x2 , t) = 0 for (x1 , x2 , t) ∈ ΩcN ×I. Now using (4.4.11) and (4.4.14), we conclude
the following estimates

||(vδ )x1 − (spl,m )x1 ||H 1/2,1/4 (γs ×I) ≤ K e−ρ N , f or γs ∈ ΩN ,

||(vδ )x2 − (spl,m )x2 ||H 1/2,1/4 (γs ×I) ≤ K e−ρ N , f or γs ∈ ΩN , (4.4.18)

||vδ − spl,m ||H 0,3/4 (γs ×I) ≤ K e−ρ N , f or γs ∈ ΩN .

Hence

Rp ({spl,m (x1 , x2 , t)}−N ≤l,m≤N −1 ) ≤ K e−ρ N (4.4.19)

provided p is proportional to N and N is large enough.

Therefore

p
Rp ({ωl,m (x1 , x2 , t)}−N ≤l,m≤N −1 ) ≤ Rp ({spl,m (x1 , x2 , t)}−N ≤l,m≤N −1 ) ≤ K e−ρ N . (4.4.20)

Thus

p
Rp ({(ωl,m − spl,m )(x1 , x2 , t)}−N ≤l,m≤N −1) ≤ K e−ρ N . (4.4.21)

From Theorem 4.3.1 we obtain the following estimate

N −1 N −1
!1/2
X X p
||ωl,m − spl,m ||2H 2,1 (Ωl,m ×I) ≤ K e−ρ N . (4.4.22)
l=−N m=−N

Combining (4.4.16), (4.4.17) and (4.4.22), we conclude

N
X −1 N
X −1
p
||vδ − ωl,m ||H 2,1 (Ωl,m ×I) + ||vδ − ω p ||H 2,1 (ΩcN ×I) ≤ K e−ρ N . (4.4.23)
l=−N m=−N

Here ΩcN = R2 \ ΩN and ΩN = [−N, N] × [−N, N].

4.4.1 Symmetric Formulation

p
As defined in (4.3.1) we choose our approximate solution to be the unique {ωl,m }−N ≤l,m≤N −1
p p
which minimizes the functional Rp ({v̌l,m (ξ1, ξ2 , t)}−N ≤l,m≤N −1 ) over all {v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1
Let the above system, of equations (4.4.1), be of the form

AW = G. (4.4.24)

79
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

Then the normal equations are

AT AW = AT G, (4.4.25)

p
where W is a vector assembled from the values of {v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1 and G is
assembled from the data.
The approach which is used to solve the problem, is based on least squares. The solution
to the least-squares problem can be found using the PCGM for the normal equations. The
Guass-Lobatto-Legendre (GLL) quadrature formula will be used to evaluate integrals in
minimization formulation. The detailed discussion of symmetric formulation is given in
Chapter 5 and [53]. There is no need to evaluate any mass and stiffness matrices and the
residuals in the normal equation can be computed inexpensively and efficiently.

4.4.2 Parallelization and Preconditioning

p
Now we define the preconditioner of the the quadratic form W p ({v̌l,m (ξ1 , ξ2, t)}−N ≤l,m≤N −1 )
p
which is given in (4.3.9). This quadratic form W p ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1 ) is equivalent
p
to the functional Rp ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1) with zero data. Define a quadratic form
p
U p ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1), where
N
X −1 N
X −1
p p p
U ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1 ) = ||v̌l,m ||2H 2,1 (S×I) . (4.4.26)
l=−N m=−N
p p
Here U p ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1 ) is the preconditioner for W p ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1 ).
p N −1
Now we find the estimate of condition number of this preconditioner U p ({v̌l,m (ξ1 , ξ2 , t)}l,m=−N ).
Then from [37] the following result holds

p p
W p ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1 ) ≤ K U p ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1 ) , (4.4.27)

where K is a constant. By Theorem 4.3.1, we get the following result

1 p p
2
U p ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1) ≤ W p ({v̌l,m (ξ1 , ξ2 , t)}−N ≤l,m≤N −1) . (4.4.28)
C(log p)

From (4.4.27) and (4.4.28), we observe that the condition number of the preconditioned
p
system is O((log p)2 ). Now v̌l,m is defined in terms of Legendre polynomials and the form

80
 4.5. Estimates in Negative Norms

p
of v̌l,m is

p p q
p
X X X
v̌l,m (ξ1 , ξ2 , t) = ai,j,k Li (2ξ1 − 1)Lj (2ξ2 − 1)Lk (2t − 1). (4.4.29)
i=0 j=0 k=0

The coefficients ai,j,k are arranged lexicographically in i j and k.

Here
Z
p p p
||v̌l,m ||2H 2,1 (S×I) = ||v̌l,m (t)||2H 2 (S) dt + ||v̌l,m ||2H 1 (I;H 0 (S)) . (4.4.30)
I

The quadratic form P defined in (4.4.30) is essentially equivalent to the quadratic form
Z X
p p 2 p
P(v̌l,m ) = ( |Dξα1 ,ξ2 v̌l,m | + |(v̌l,m )t |2 )dξ1dξ2 dt. (4.4.31)
S×I |α|≤2

p
Now we define the new quadratic form Q(v̌l,m ), where
Z
p p p p p p p
Q(v̌l,m ) = (|(v̌l,m )ξ1 ξ1 |2 + |(v̌l,m )ξ2 ξ2 |2 + |(v̌l,m )ξ1 |2 + |(v̌l,m )ξ2 |2 + |(v̌l,m )t |2 + |(v̌l,m )|2 )dξ1 dξ2dt.
S×I

(4.4.32)

p
From Theorem 2.1 of [38], it immediately follows that the above quadratic forms P(v̌l,m )
p
and Q(v̌l,m ) are spectrally equivalent. We use separation of variables technique to diag-
p
onalize the quadratic form Q(v̌l,m ) in a new set of basis function. The details of new
basis function and the computation of preconditioner are given in [38]. During the PCGM
process, communication between neighbouring processors is confined to the interchange of
information consisting of the value of function and its derivatives at inter-element bound-
aries. In addition we need to compute two global scalars to update the approximate solution
and the search direction. Hence inter-processor communication is quite small.

4.5 Estimates in Negative Norms

Lemma 4.5.1. Assume g is piecewise analytic in Ω and gδ is the mollified representation
of g such that
gδ (x1 , x2 ) = (g ∗ θN/δ )(x1 , x2 ).

81
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

Here
 2  2
N N 1
θN/δ (x1 , x2 ) = Φδ1 (Nx1 )Φδ2 (Nx2 ) = Φ(Nx1 /δ1 )Φ(Nx2 /δ2 ),
2π 2π δ1 δ2
√ √
where δ1 = β1 N and δ2 = β2 N. Then the estimate

C β N +1
||g − gδ ||H −2N−2 (Ω) ≤ (4.5.1)
2N +1 (N + 1)!N N +1

holds where C is a positive constant and β = max{β1 , β2 }.

Proof. Let φ(ξ) be the Fourier Transform of Φ(x). Then

N
!
2
− ξ2
X ξ 2j
Φ̂(ξ) = φ(ξ) = e . (4.5.2)
j=0
2j j!

We have
hg − gδ , ψi = hĝ − ĝδ , ψ̂i.

Now
ĝδ = g\
∗ θN/δ = ĝ θ̂N/δ .

Then
hg − gδ , ψi = hĝ − ĝ θ̂N/δ , ψ̂i = hĝ, (1 − θ̂N/δ )ψ̂i.

Here
θ̂N/δ (ξ1 , ξ2 ) = φ(δ1 ξ1 /N)φ(δ2 ξ2 /N).

Now
µN µN +1
eµ = 1 + µ + · · · + + eη ,
N! (N + 1)!
where 0 ≤ η ≤ µ. Hence

µN µN +1
|1 − e−µ (1 + µ + · · · + )| ≤ .
N! (N + 1)!

Now

1 − θ̂N/δ (ξ1 , ξ2 ) = 1 − φ(δ1 ξ1 /N)φ(δ2 ξ2 /N),

= (1 − φ(δ1 ξ1 /N)) + (1 − φ(δ2 ξ2 /N))φ(δ1 ξ1 /N).

82
 4.6. Error Estimates

Moreover
|φ(δ1 ξ1 /N)| ≤ 1.

Hence
|1 − θ̂N/δ (ξ1 , ξ2)| ≤ |1 − φ(δ1 ξ1 /N)| + |1 − φ(δ2 ξ2 /N)|.

And so we obtain
2(N +1) 2(N +1)
β N +1 (ξ1 + ξ2 )
|1 − θ̂N/δ (ξ1 , ξ2 )| ≤ N +1 N +1
.
2 (N + 1)!N
Finally, we have
β N +1 ||((ξ12)N +1 + (ξ22 )N +1 )ψ̂||L2
hg − gδ , ψi = hĝ, (1 − θ̂N/δ )ψ̂i ≤ ||ĝ||L2 ,
2N +1 (N + 1)! N N +1
C β N +1 |ψ|H 2N+2
≤ N +1 .
2 (N + 1)!N N +1
Hence
|(g − gδ , ψ)|Ω
||g − gδ ||H −2N−2 (Ω) ≤ sup ,
ψ∈H 2N+2 (Ω) ||ψ||H 2N+2 (Ω)
(4.5.3)
C β N +1
≤ N +1 .
2 (N + 1)!N N +1

4.6 Error Estimates

In this section we recover point-wise values with spectral accuracy. We use the exponen-
tially accurate mollifier which was proposed by J. Tanner (Tanner 2006, Theorem 4.2 )[74]
and obtain the error estimate for the solution at a point (a1 , a2 , 1) .

Lemma 4.6.1. The estimate

|(v − v ∗ θN/ǫ )(a1 , a2 , 1)| ≤ Ce−ρN (4.6.1)

holds . Here C and ρ are generic constants and θN/ǫ (x1 , x2 ) is the Hermite mollifier which
is defined as       
N Nx1 N Nx2
θN/ǫ (x1 , x2 ) = Φ Φ ,
2πǫ1 ǫ1 2πǫ2 ǫ2
√ √
where ǫ1 = γ1 N and ǫ2 = γ2 N .

83
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

Proof. The L.H.S. of above equation is written as

 Z ∞Z ∞ 
 
|v − v ∗ θN/ǫ | = v(a1 , a2 , 1) − v(a1 − x1 , a2 − x2 , 1)θN/ǫ (x1 , x2 )dx1 dx2  , (4.6.2)
−∞ −∞
 Z ∞     
 N Nx 1 
≤ v(a1 , a2 , 1) − v(a1 − x1 , a2 , 1) Φ dx1 
−∞ 2πǫ1 ǫ1
| {z }
J1
 
 
Z ∞  Z ∞     
 N Nx2 
+  v(a1 − x1 , a2 , 1) − v(a1 − x1 , a2 − x2 , 1) Φ dx2 
−∞  −∞ 2πǫ 2 ǫ2 
| {z }
 J2

   
 N Nx1 

 2πǫ1 Φ dx1 .
ǫ1 
(4.6.3)
Here
 Z ∞     
 N Nx1 
|J1 | = v(a1 , a2 , 1) −
 v(a1 − x1 , a2 , 1) Φ dx1  (4.6.4)
−∞ 2πǫ1 ǫ1

and
 Z ∞     
 N Nx2 
|J2 | = v(a1 − x1 , a2 , 1) −
 v(a1 − x1 , a2 − x2 , 1) Φ dx2  . (4.6.5)
−∞ 2πǫ2 ǫ2

Using Lemma 3.6.1 we get

|J1 | . Ce−ρN (4.6.6)

and

|J2 | . Ce−ρN . (4.6.7)

Hence
Z ∞   
N  Nx 1 
|v − v ∗ θN/ǫ | . |J1 | + |J2 | Φ  dx1 .
−∞ 2πǫ1  ǫ1  (4.6.8)
| {z }
J3

Let Z    
∞
N Φ Nx1  dx1 .
 
|J3 | = |J2 |
−∞ 2πǫ1  ǫ1 

84
 4.6. Error Estimates

Now from equation (2.14b) of [73]

 Z s !
∞
N 2P −η1 N x21 /γ1 |J2 | N
|J3 | . |J2 | e dx1 = 2P .
2πǫ1 −∞ ǫ1 2 ǫ1 η1 π

Here P = γ12 dx1 N = γ12 N. Hence we have 2P ≤ exp(κγ12 N) with κ := log(2). Then

|J3 | . Ce(κγ1 N −ρN ) .

2

For sufficiently small γ1 < 1 the above estimate is exponentially accurate.

Hence
|J3 | . e−ρN .

Thus

|v − v ∗ θN/ǫ | . |J1 | + |J3 | . Ce−ρN . (4.6.9)

Lemma 4.6.2. The estimate

|(v ∗ θN/ǫ − vδ ∗ θN/ǫ )(a1 , a2 , 1)| ≤ Ce−ρN (4.6.10)

holds . Here C and ρ are generic constants and θN/ǫ (x1 , x2 ) is the Hermite mollifier which
is defined as       
N Nx1 N Nx2
θN/ǫ (x1 , x2 ) = Φ Φ ,
2πǫ1 ǫ1 2πǫ2 ǫ2
√ √
where ǫ1 = γ1 N and ǫ2 = γ2 N .

Proof. To verify the above bound, we rearrange the L.H.S. of the equation as:

I2 = v ∗ θN/ǫ − vδ ∗ θN/ǫ = (v − vδ ) ∗ θN/ǫ .

Hence
|I2 | ≤ ||v − vδ ||H −2N−2 ||θN/ǫ ||H 2N+2 .

Consider the adjoint problem

L∗ ψ = 0 in Ω × I, (4.6.11)

85
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

with initial conditon

      
N Nx1 N Nx2
ψ = θN/ǫ (x1 , x2 ) = Φ Φ on Ω × {1}.
2πǫ1 ǫ1 2πǫ2 ǫ2
Then
(v ∗ ψ)(a1 , a2 , 1) = (v ∗ ψ)(a1 , a2 , 0).

Hence
((v − vδ ) ∗ ψ)(a1 , a2 , 1) = ((g − gδ ) ∗ ψ)(a1 , a2 , 0).

Now from Lemma 4.4.1

||ψ||H s(Ω×{1}) ≤ FN s! α1s .

Hence
||ψ||H s(Ω×{0}) ≤ CFN s! α1s .

where C and α1 are constants. Here

′√
FN ∼
′ ′
=K da N log N
,

′ ′
for some constants K , d and a .
By Lemma 4.5.1

|I2 | = |((g − gδ ) ∗ ψ)(a1 , a2 , 0)| ≤ ||g − gδ ||H −2N−2 ||ψ||H 2N+2 ,

C β N +1
≤ FN (2N + 2)! α12N .
2N +1 (N + 1)!N N +1

By Stirling’s approximation
√
n! ∼ 2 π n nn e−n .

Hence

|I2 | ∼ C(2βα12)N +1 FN e−(N +1) ∼ C1 e−ρN (4.6.12)

provided β is small enough.

86
 4.6. Error Estimates

Theorem 4.6.3. Define

p
ω p = ωl,m in Ωl,m × I f or − N ≤ l, m ≤ N − 1, (4.6.13)

=0 otherwise. (4.6.14)

If vδ ∈ D2,1 (Ω̄ × [0, 1]) then the following error estimate holds

|v(a1 , a2 , 1) − (ω p ∗ ψ)(a1 , a2 , 1)| ≤ C1 e−ρ N (4.6.15)

provided q is proportional to p2 , as p tends to infinity and N is proportional to p. Here C1

and ρ are generic constants.

Proof. We can also write the L.H.S. of above equation

|(v − ω p ∗ ψ)(a1 , a2 , 1)| = |((v − v ∗ ψ) + (v ∗ ψ − vδ ∗ ψ)

(4.6.16)
+(vδ ∗ ψ − ω p ∗ ψ))(a1 , a2 , 1)|,

= |I1 + I2 + I3 |,

≤ |I1 | + |I2 | + |I3 |.

Here

I1 = (v − v ∗ ψ)(a1 , a2 , 1).

Then from Lemma 4.6.1 the following result holds

|I1 | = |(v − v ∗ ψ)(a1 , a2 , 1)| ≤ C1 e−ρN . (4.6.17)

Now

I2 = (v ∗ ψ − vδ ∗ ψ)(a1 , a2 , 1),

= ((v − vδ ) ∗ ψ)(a1 , a2 , 1).

Using Lemma 4.6.2, we have

|I2 | ∼ C1 e−ρN . (4.6.18)

87
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

From (4.4.23) and Lemma (4.4.1) the following result holds

|I3 | = |((vδ − ω p ) ∗ ψ)(a1 , a2 , 1)| ≤ ||(vδ − ω p )||L2 ||ψ||L2 ≤ C1 e−ρN . (4.6.19)

Now using (4.6.17), (4.6.18) and (4.6.19)

|(v − ω p ∗ ψ)(a1 , a2 , 1)| ≤ C1 e−ρ N .

Now we want to recover the point-wise value of the solution at an interior point
(a1 , a2 , τ ) with spectral accuracy. Assume that ω p (x1 , x2 , t) is smooth and belong to D2,1
in the set O which is a neighborhood of (a1 , a2 , τ ), where

O = {(x1 , x2 , t) : |x1 − a1 | ≤ δ1 , |x2 − a2 | ≤ δ2 , |t − τ | ≤ ǫ1 }.

Here the Hermite mollifier, which is defined in equation (4.2.12), is used in both space di-
rections (x1 , x2 ) to recover the value in space domain. Similarly we use the root exponential
accurate mollifier [73, 74] to recover the value in time direction. Define root exponential
accurate mollifier
   
1 t t ( ct2
)
ΘP,δ3 (t) = η1 DP , η1 := e t2 −π 2 1(−π,π) (t), c > 0, (4.6.20)
δ3 δ3 δ3
√
with adaptive parameterization, δ3 = dt := π1 dist{t, {c1 , ..., cj }}[modπ] and P ∼ dt N 2 / e.
Here Dp (t) denote the Dirichlet kernel

 sin(P +1/2)t t 6= 2mπ,
2πsin(t/2)
DP (t) := (4.6.21)
 2P + 1 t = 2mπ.
Now the regularized version of ω p at (a1 , a2 , τ ) is defined as:
Z π Z π Z πdt
p
Rw (a1 , a2 , τ ) = θN/ǫ (x1 , x2 )ΘP,δ3 (t)w p (a1 − x1 , a2 − x2 , τ − t)dx1 dx2 dt
−π −π −πdt

and
Z π Z π Z πdt
α
RD(x 1 ,x2 )
Dtj w p (a1 , a2 , τ ) = (−1) |α|+j
Dxα11 Dxα22 θN/ǫ (x1 , x2 ) Dtj ΘQ,δ3 (t)
−π −π −πdt

w p (a1 − x1 , a2 − x2 , τ − t)dx1 dx2 dt .

Where |α| = α1 + α2 . Once again it can be shown that this regularized version of
w p (a1 , a2 , τ ) approximates v(a1 , a2 , τ ) with exponential accuracy.

88
 4.7. Computational Results

4.7 Computational Results

The efficiency of the proposed computational strategy is established through numerical
examples.

Example 4.7.1 (Unsmooth initial data). Consider the problem

u t − u x1 x1 − u x2 x2 = 0 in Ω × (0, 1),

u(x1 , x2 , 0) = f (x1 , x2 ) on Ω × {0}.

Where 
 1 (x , x ) ∈ S = (0, 1) × (0, 1),
1 2
f (x1 , x2 ) =
 0 otherwise.

Table 4.1: Maximum Point-wise error as function of p

p q Maximum point − wise Error Between [−5, 5] × [−5, 5] at t = 1

5 25 5.151279362854173 × 10−4
6 36 8.668633639307025 × 10−5
7 49 1.047492316950638 × 10−5
8 64 3.864413350240621 × 10−6
9 81 5.844791402492613 × 10−7
10 100 5.199656919842266 × 10−8

The numerical results are given in Table 4.1, and Figures 4.1, 4.2, 4.3, 4.4, 4.5 and
4.6. We observe that the maximum point-wise errors of the solution and its derivative
decay rapidly with polynomial order p. This example confirms the efficacy of the proposed
method (i.e. LSSEM). In the other example, the European Rainbow Options problem is
dealt with.

89
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

−3
10

Maximum of point−wise error of solution at t=1

−4
10

−5
10

−6
10

−7
10

−8
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.1: Max. point-wise error of Numerical and Exact solution at t=1

−2
10
Max. error of derivative of solution in x direction at t=1

−3
10

−4
10

−5
10

−6
10

−7
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.2: Max. point-wise error of first derivative (∆x ) of Numerical and Exact solution

90
 4.7. Computational Results

−2
10

Max. error of derivative of solution in y direction at t=1

−3
10

−4
10

−5
10

−6
10

−7
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.3: Max. point-wise error of first derivative (∆y ) of Numerical and Exact solution
Max. error of second derivative of solution in x direction at t=1

−2
10

−3
10

−4
10

−5
10

−6
10

−7
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.4: Max. point-wise error of second derivative (Γxx ) of Numerical and Exact solution

91
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

Max. error of second derivative of solution in y direction at t=1

−2
10

−3
10

−4
10

−5
10

−6
10

−7
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.5: Max. point-wise error of second derivative (Γyy ) of Numerical and Exact solution

−2
10
Max. error of mixed second derivative of solution at t=1

−3
10

−4
10

−5
10

−6
10

−7
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.6: Max. point-wise error of mixed derivative (Γxy ) of Numerical and Exact solution

92
 4.7. Computational Results

Example 4.7.2 (European Rainbow Options problem). Here a problem of the “European
Black-Scholes European Rainbow Put option” is considered. The method is used to solve
this problem and the results are compared with those due to Zhu et al [78]. Consider the
problem :
1 1
Vτ + σ12 S12 VS1 S1 +ρσ1 σ2 S1 S2 VS1 S2 + σ22 S22 VS2 S2 +rS1 VS1 +rS2 VS2 −rV = 0 in R2+ ×(0, T ),
2 2
V (S1 , S2 , T ) = max(K − max(S1 , S2 ), 0) on Ω × {T },

where V, S, K, r and σ are respectively option price, underlying asset price, strike price,
risk-free interest rate and volatility .

Table 4.2: Rainbow Option problem : Variable value from Zhu et al [78]

K r σ1 σ2 ρ T
100 0.15 0.15 0.20 −0.20 0.25

Table 4.3: Rainbow Put Option problem : Point-wise error as function of p for LSSEM

p q Maximum point − wise Error Between [−5, 5] × [−5, 5] at T = 1

5 25 4.332887695828502 × 10−4
6 36 8.097912513572733 × 10−5
7 49 9.787329968113090 × 10−6
8 64 2.088472253202587 × 10−6
9 81 4.552934625526420 × 10−7
10 100 3.157985444430164 × 10−8

The results obtained using the proposed method are given in Table 4.3. Comparing
Tables 4.3 and Table 3 of Zhu et al [78], it can be observed that :

1. Zhu et al [78] could not achieve exponential accuracy in time. As accuracy of 10−5
is obtained with 320 time steps.

93
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

−3
10

Maximum of point−wise error of solution at t=1

−4
10

−5
10

−6
10

−7
10

−8
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.7: Max. point-wise error of Numerical and Exact solution at t=1

−3
10
Max. error of derivative of solution in x direction at t=1

−4
10

−5
10

−6
10

−7
10

−8
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.8: Max. point-wise error of first derivative (∆x ) of Numerical and Exact solution

2. In order to achieve an accuracy of 10−5 , LSSEM requires p = 7, q = 49.

94
 4.7. Computational Results

−3
10

Max. error of derivative of solution in y direction at t=1

−4
10

−5
10

−6
10

−7
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.9: Max. point-wise error of first derivative (∆y ) of Numerical and Exact solution
Max. error of second derivative of solution in x direction at t=1

−2
10

−3
10

−4
10

−5
10

−6
10

−7
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.10: Max. point-wise error of second derivative (Γxx ) of Numerical and Exact solution

3. While [78] could achieve a maximum accuracy of 10−6 with N = 30 and 640 timestep
, LSSEM can easily obtain much higher accuracies. For examples, an accuracy of

95
 Chapter 4. Nonconforming LSSEM for European Options with Multi Assets

Max. error of second derivative of solution in y direction at t=1

−2
10

−3
10

−4
10

−5
10

−6
10

−7
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.11: Max. point-wise error of second derivative (Γyy ) of Numerical and Exact solution

−2
10
Max. error of mixed second derivative of solution at t=1

−3
10

−4
10

−5
10

−6
10

−7
10
5 6 7 8 9 10
Polynomial order (p)

Figure 4.12: Max. point-wise error of mixed derivative (Γxy ) of Numerical and Exact solution

10−8 is obtained with q = 100 only.

96
 4.7. Computational Results

4. From Figure 4.8, 4.9, 4.10, 4.11 and 4.12, we observe that the errors of derivatives
also decay exponentially with polynomial order p.

5. LSSEM is exponentially accurate theoretically as well as numerically .

97
 Chapter
5
Conclusion and Future Work

5.1 Conclusion
The main focus of Chapter 2 is on how to recover the pointwise values of a discontinuous
function from its spectral representation. Here we mainly concentrate on two aspects of
processing such piecewise smooth data:

Edge detection. Detecting the location and amplitude of the edges. Often, these are
essential features sought in piecewise smooth data. Moreover, they are the boundaries
of the region of smoothness and are therefore essential for the second aspect. For our
problem this is not an issue since the solution is analytic for t > 0.

Reconstruction. Recovering the underlying function f inside its different regions of

smoothness. We may choose the region of smoothness to be a fixed interval with
the spatial point where we wish to recover the pointwise value of the function as the
centre of the interval of smoothness.

In Chapter 3 we have presented a non-conforming least squares spectral-element method

for Black-Scholes equation. Hermite mollifier has been used to resolve the difficulty of

99
 Chapter 5. Conclusion and Future Work

nonsmooth initial conditions. We have given a priori error estimates to establish the
exponential accuracy of the method theoretically. Specific numerical examples have been
given to validate the error estimate. In the first example we have shown the point-wise
exponential accuracy of the proposed method. The second example is the European Black-
scholes Put Option problem. LSSEM can easily obtain very high accuracies. European
Black-Scholes Call Option problem has been choosen as the third example. The numerical
solution of this problem has been compared to that obtained by Bunnin et. al. [24].
Bunnin et. al. [24] have achieved a maximum accuracy of 10−3 for N = 100, while LSSEM
achieves an accuracy of 10−10 with p = 10, q = 100. From the three examples and the
theoretical results, it has been demonstrated that LSSEM is a spectrally accurate method
in space and time. Further, the method is non-conforming and hence is parallelizable.

In Chapter 4 we have introduced a non-conforming least squares spectral-element

method for Black-Scholes equation with two assets. The difficulty of nonsmooth initial
condition is resolved by Hermite mollifier which is given in [73, 74]. In general, Option
Pricing problems do not use the information about boundary conditions. In our method
all the information we need is to know the growth condition at infinity which is satisfied
by the initial condition. The validation of the error estimate have been confirmed by spe-
cific examples. In the first example we have shown the point-wise exponential accuracy
of the proposed method. The second example is European Black-Scholes Rainbow Put
Option problem. Here we compare the results of Rainbow Options problem with those
of Zhu et.al.[77]. Zhu et. al.[77] could not achieve exponential accuracy in time. While
[77] could achieve a maximum accuracy of 10−5 with N = 20 and Nt = 320, LSSEM can
easily obtain much higher accuracies. It is demonstrated by the theoretical and numerical
results that LSSEM is a spectrally accurate method in space and time. The method is
easily parallelizable because it is nonconforming.

5.2 Future Work

The method can also be used to solve the following problems.

100
 5.2. Future Work

1. Nonconforming Least-Squares Spectral Element Method for Parabolic

Initial Boundary Value Problems With Non-Smooth Data using Parallel
Computers.

In Chapters 3, 4, we have discussed Nonconforming Least-Squares Spectral Element

Method for Parabolic Initial Value Problem which has non-smooth initial data. The
approach of regularization of initial data in [53] is mollification by Hermite mollifier,
which was introduced by J. Tanner [74] and also discussed by Tadmor [73]. But this
approach is not applicable for Parabolic Initial Boundary Value Problems. There
are many issues. First the mollification technique is workable on periodic data and
data which is defined on the whole of R. The second issue is the compatibility
conditions at the space time corner for Parabolic Initial Boundary Value Problem.
In real life problems it is not necessary that the boundary data and initial data match
at the space time corners. Moreover the solution of the adjoint problem need not
satisfy these compatibility conditions at the space time corners. Hence it is no longer
true that the error between the computed and exact solution in a negative Sobolev
norm decays at a rate which depends only on the order of the norm. We want to
develop a Nonconforming Least-Squares Spectral Element Method (LSSEM) [36] to
solve Parabolic Initial Boundary Value Problem with nonsmooth data on Parallel
Computers.

2. Nonconforming h-p spectral element methods for parabolic stochastic

partial differential equation using parallel computers.

Parabolic SPDEs arise in many fields of science and engineering such as finance,
weather prediction, hetrogenity of materials with microstructure, well roughness in
fluid dynamics, the variability of soil permeability in subsurface aquifers and ocean
dynamics etc. Due to the presence of randomness in data the numerical simulation
becomes difficult and inefficient and the accuracy of the solution decreases. A natural
framework to describe uncertainty would be to apply probability theory. Here all
uncertain inputs are treated as random variables or random fields.

101
 Chapter 5. Conclusion and Future Work

3. Nonconforming h-p spectral element methods for parabolic problems on

non-smooth domain using parallel computers.

Parabolic partial differential equations arise in many fields of science and engineer-
ing such as distribution of heat, fluid dynamics, structural/mechanical engineering,
aerospace engineering, medical science and seismology etc. Due to the presence of
singularities in non-smooth domains such as domains with corners and edges the nu-
merical simulation becomes difficult to obtain and the accuracy of the solution may
be low. p and h − p Spectral Element Methods (SEM) have recently emerged as a vi-
able alternatives to the finite difference and finite element methods for the numerical
solution of partial differential equations. A geometric mesh is imposed on the spatial
domain to resolve the singularities at the corners. Spectral methods have proved par-
ticularly useful for computational fluid dynamics where large spectral hydrodynamics
codes are now regularly used to study turbulence and transition, numerical weather
prediction, and ocean dynamics. In numerical computation time complexity is one
of the important issues which needs to be addressed. Parallel computers can reduce
this time up to a great extent. We intend to examine this in future work, within the
frame work of the least squares approach [37, 75] which may be nonconforming or
conforming at a selected set of points.

102
 Bibliography

[1] S. Abarbanel, D. Gottlieb and E. Tadmor (1986): Spectral methods for discontinuous
problems, Numerical methods for fluid dynamics, II (Reading, 1985) Inst. Math.
Appl. Conf. Ser. New Ser., Vol. 7, Oxford Univ. Press, New York, pp. 129-153.

[2] Yves Achdou and Olivier Pironneau (2005): Computational Methods for Option
Pricing, Society for Industrial and Applied Mathematics (SIAM), Philadelphia .

[3] Robert A. Adams and John J. Fournier (2003): Sobolev spaces, Vol. 140, Academic
press.

[4] G. Anagnostou (1990): Nonconforming sliding spectral element methods for the
unsteady incompressible Navier-Stokes equations, Ph.D., Massachusetts Institute of
Technology.

[5] A. K. Aziz and I. B. Babuška (1972): The Mathematical foundations of the finite
element method with applications to partial differential equations, Academic Press,
New York.

[6] A. K. Aziz, R. B. Kellogg and A. B. Stephens (1985): Least squares methods for
elliptic systems, Mathematics of Computation, 44(169), 53-70.

[7] A. K. Aziz and J. L. Liu (1991): A weighted least squares method for the backward-
forward heat equation, SIAM Journal on Numerical Analysis, 28(1), 156-167.

[8] I. Babuška (1973): The finite element method with penalty, Mathematics of Com-
putation, 27(122), 221-228.

103
 Bibliography

[9] I. Babuška and M. Zlámal (1973): Nonconforming elements in the finite element
method with penalty, SIAM Journal on Numerical Analysis, 10(5), 863-875.

[10] I. Babuška and M. Suri (1994): The p and h-p versions of the finite element method,
basic principles and properties, SIAM review, 36(4), 578-632.

[11] I. Babuška and T. Janik (1989): The h-p version of the finite element method for
parabolic equations. Part I. The p-version in time, Numerical Methods for Partial
Differential Equations, 5(4), 363-399.

[12] I. Babuška and T. Janik (1990): The h-p version of the finite element method for
parabolic equations. II. The h-p version in time, Numerical Methods for Partial
Differential Equations, 6(4), 343-369.

[13] C. Bernardi and Y. Maday (1997): Spectral methods, Handbook of numerical anal-
ysis, 5, 209-485.

[14] C. Bernardi Y. Maday, C. Mavriplis and A. T. Patera (1989): The mortar element
method applied to spectral discretizations, In Proceedings of the Seventh Interna-
tional Conference on Finite Element Methods in Flow Problems, Huntsville.

[15] M. Bieterman and I. Babuška (1982): The finite element method for parabolic equa-
tions. I: A Posteriori Error Estimation, Numerische Mathematik, 40(3), 339-371.

[16] M. Bieterman and I. Babuška (1982): The finite element method for parabolic equa-
tions, II: A Posteriori Error Estimation and Adaptive Approach, Numerische Math-
ematik, 40(3), 373-406.

[17] P. Biswas (1998): Least-Squares Spectral Element Methods for Parabolic

Problems, Ph.D. thesis, Indian Institute of Technology Kanpur.
http://172.28.64.70:8080/jspui/bitstream/123456789/10905/1/Y220868.pdf.

[18] Fischer Black and Myron S. Scholes (1973): The pricing of options and corporate
liabilities, Journal of Political Economy, 81(3), 637-654.

104
 Bibliography

[19] E. N. Blinova (1943): The Hydrodynamical Theory of Pressure Waves, Temperature

Waves and Centers of Activity in the Atmosphere, Proc.(Dokl.) Acad. Sci. USSR.
Vol. 39.

[20] P. B. Bochev and M. D. Gunzburger (2009): Least-squares: finite element methods,

Vol. 166, Springer.

[21] P. B. Bochev and M. D. Gunzburger (1998): Finite element methods of least-squares

type, SIAM review, 40(4), 789-837.

[22] J. P. Boyd (2001): Chebyshev and Fourier spectral methods, Courier Dover Publi-
cations.

[23] J. H. Bramble and V. Thomée (1972): Semidiscrete least-squares methods for a

parabolic boundary value problem, Mathematics of Computation, 26(119), 633-648.

[24] F. O. Bunnin, Y. Guo, Y. Ren and J. Darlington (2000): Design of high perfomance
financial modelling environment, Parallel Computing, vol. 26, 601-622.

[25] C. Canuto, M. Y. Hussaini, A. Quarteroni and T. A. Zang (1988): Spectral methods

in fluid dynamics (Scientific Computation), Springer-Verlag, New York-Heidelberg-
Berlin.

[26] C. Canuto, M. Y. Hussaini, A. Quarteroni and T. A. Zang (2006): Spectral Methods:

Fundamentals in Single Domains, Springer-Verlag, New York-Heidelberg-Berlin.

[27] G. F. Carey and B. N. Jiang (1986): Element-by-element linear and nonlinear so-
lution schemes, Communications in Applied Numerical Methods, 2(2), 145-153.

[28] G. F. Carey and B. N. Jiang (1987): Least-squares finite element method and pre-
conditioned conjugate gradient solution, International journal for numerical methods
in engineering, 24(7), 1283-1296.

105
 Bibliography

[29] G. F. Carey, Y. Shen and R. T. Mclay (1998): Parallel conjugate gradient perfor-
mance for least-squares finite elements and transport problems, International journal
for numerical methods in fluids, 28(10), 1421-1440.

[30] C. L. Chang and J. J. Nelson (1997): Least-squares finite element method for the
Stokes problem with zero residual of mass conservation, SIAM Journal on Numerical
Analysis, 34(2) 480-489.

[31] Bernardo Cockburn and Chi-Wang Shu (1998): The local discontinuous Galerkin
method for time-dependent convection-diffusion systems, SIAM Journal on Numer-
ical Analysis, (35(6)), 2440-2463.

[32] L. Demkowicz, J. T. Oden, W. Rachowicz and O. Hardy (1989): Toward a universal

hp adaptive finite element strategy, Part 1. Constrained approximation and data
structure, Computer Methods in Applied Mechanics and Engineering, 77(1)5, 79-
112.

[33] X. Ding and T. T. Tsang (2001): Large eddy simulation of turbulent flows by a
least-squares finite element method, International journal for numerical methods in
fluids, 37(3), 297-319.

[34] J. Douglas and T. Dupont (1976): Interior penalty procedures for elliptic and
parabolic Galerkin methods, Computing methods in applied sciences, Springer Berlin
Heidelberg, 207-216.

[35] P. K. Dutt and A. K. Singh (1993): Spectral methods for periodic intial value prob-
lems with non smooth data, Mathematics of Computation, vol. 61, 645-658.

[36] P. Dutt (1999): Spectral methods for initial boundary value problems with nonsmooth
data, Numerische Mathematik, 81(3), 323-344.

[37] P. Dutt, P. Biswas and S.Ghorai (2007): Spectral element methods for parabolic
problems, Journal of Computational and Applied Mathematics, vol. 203, , 461-486.

106
 Bibliography

[38] P. Dutt, P. Biswas and G. N. Raju (2008): Preconditioners for Spectral element
methods for elliptic and parabolic problems, Journal of Computational and Applied
Mathematics, vol. 215 , 152-166.

[39] P. Dutt and S. Bedekar (2001): Spetral methods for Hyperbolic Intial Boundary
Value Problem on Parallel Computers, Journal of Computational and Applied Math-
ematics , vol. 134 , 164-190.

[40] E. Eliasen, B. Machenhauer and E. Rasmussen (1970): On a numerical method

for integration of the hydrodynamical equations with a spectral representation of the
horizontal fields, Kobenhavns Universitet, Institut for Teoretisk Meteorologi.

[41] H. Tal-Ezer (1989): Spectral methods in time for parabolic problems, SIAM Journal
on Numerical Analysis, 26(1), 1-11.

[42] F. Fang and C. W. Oosterlee (2008): A novel pricing method for European options
based on Fourier-cosine series expansions, SIAM Journal on Scientific Computing,
31(2), 826-848.

[43] B. A. Finlayson and L. E. Scriven (1966): The method of weighted residuals: A

review, Applied Mechanics Reviews, 19(9), 735-748.

[44] R. A. Frazer, W. P.Jones and S. W. Skan (1937): Approximations to functions and

to the solutions of differential equations, HM Stationery Office.

[45] J. De Frutos (2008): A spectral method for bonds, Computers & Operations Re-
search, 35, 64-75.

[46] D. Gottlieb and Steven A. Orszag (1977): Numerical Analysis of Spectral Methods:
Theory and Application, SIAM.

[47] A. Greenberg (2003): Chebyshev spectral method for singular moving boundary prob-
lems with application to finance, Ph.D. thesis, California Institute of Technology.

107
 Bibliography

[48] John. Hull (2010): Options, Futures, and Other Derivatives, 7/e (With CD), Pear-
son Education India.

[49] P. Jamet (1978): Galerkin-type approximations which are discontinuous in time for
parabolic equations in a variable domain, SIAM Journal on Numerical Analysis,
15(5), 912-928.

[50] B. N. Jiang (1998): On the least-squares method, Computer methods in applied

mechanics and engineering, 152(1), 239-257.

[51] L. V. Kantorovich (1934): On a new method for solving partial differential equations,
In Doklady Acad. Sci. USSR, Vol. 2, No. 8-9, pp. 532-536.

[52] G. E. Karniadakis and S. J. Sherwin (1999): Spectral/hp element methods for CFD,
Oxford University Press.

[53] Arbaz Khan, P. Dutt and C. S. Upadhyay Nonconforming Least-Squares Spectral

Element Method for European options, Under Review.

[54] H. O. Kreiss and J. Oliger (1972): Comparison of accurate methods for the integra-
tion of hyperbolic equations, Tellus, 24, l99-215.

[55] C. Lanczos (1938): Trigonometric interpolation of empirical and analytical func-

tions, Journal of Mathematical Physics, 17, 123-199.

[56] J. L. Lions and E. Magenes (1972): Non-Homogeneous Boundary Value Problems

and Applications I, Springer, Berlin.

[57] J. L. Lions and E. Magenes (1973): Non-Homogeneous Boundary Value Problems

and Applications II, Springer, Berlin.

[58] J. L. Lions and E. Magenes (1973): Non-Homogeneous Boundary Value Problems

and Applications III, Springer, Berlin.

108
 Bibliography

[59] C. G. Makridakis and I. Babuska (1997): On the stability of the discontinuous

Galerkin method for the heat equation, SIAM journal on numerical analysis, 34(1),
389-401.

[60] C. Mavriplis (1989): Nonconforming discretizations and a posteriori error esti-

mators for adaptive spectral element techniques, Ph.D., Massachusetts Institute of
Technology.

[61] J. T. Oden, I. Babuška and Carlos Erik Baumann (1998): A Discontinuous h-p
Finite Element Method for Diffusion Problems, Journal of computational physics
146(2), 491-519.

[62] J. T. Oden, L. Demkowicz, W. Rachowicz and T. A. Westermann (1989): Toward a

universal hp adaptive finite element strategy, Part 2. A posteriori error estimation,
Computer Methods in Applied Mechanics and Engineering, 77(1), 113-180.

[63] S. A. Orszag (1969): Numerical methods for the simulation of turbulence, Physics
of Fluids, 12(12), II-250.

[64] S. A. Orszag (1972): Comparison of pseudospectral and spectral approximation,

Studies in Applied Mathematics, 51(3), 253-259.

[65] S. A. Orszag (1980): Spectral methods for problems in complex geometries, Journal
of Computational Physics, 37(1), 70-92.

[66] A. T. Patera (1984): A spectral element method for fluid dynamics: laminar flow
in a channel expansion, Journal of computational Physics, 54(3), 468-488.

[67] M. M. J. Proot and M. I. Gerrtisma (2002): Least-squares spectral elements applied

to the Stokes problem, Journal of Computational Physics, 181(2), 454-477.

[68] W. Rachowicz, J. T. Oden and L. Demkowicz (1989): Toward a universal hp adap-

tive finite element strategy part 3. design of hp meshes, Computer Methods in Ap-
plied Mechanics and Engineering, 77(1), 181-212.

109
 Bibliography

[69] Dominik Schötzau and Christoph Schwab (2000): Time discretization of parabolic
problems by the hp-version of the discontinuous Galerkin finite element method,
SIAM Journal on Numerical Analysis, 38(3), 837-875.

[70] C. Schwab (1998): p-and hp-finite element methods: Theory and applications in
solid and fluid mechanics, Oxford: Clarendon Press.

[71] I. Silberman (1954): Planetary waves in the atmosphere, Journal of Atmospheric

Sciences, 11, 27-34.

[72] J. C. Slater (1934): Electronic energy bands in metals, Physical Review, 45(11),
794.

[73] E. Tadmor (2007): Filters, mollifiers and the computation of the Gibbs phenomenon
, Acta Numerica, 16, 305-378.

[74] J. Tanner (2006): Optimal filter and mollifier for piecewise smooth spectral data,
Mathematics of Computation, 75(254), 767-790.

[75] S. K. Tomar (2006): h-p Spectral Element Method for Elliptic Problems on Non-
smooth Domains Using Parallel Computers, Computing, 78, 117-143. 2552, High
Performance Computing, Springer Verlag.

[76] L. N. Trefethen (2013): Approximation theory and approximation practice, Siam.

[77] Zhu Wuming and David A. Kopriva (2009): A Spectral Element Method to Price
European Options. I. Single Asset with and without Jump Diffusion, Journal of
Scientific Computing (39), 222-243.

[78] Zhu Wuming and David A. Kopriva (2009): A spectral element approximation to
price European options. II. The Black-Scholes model with two underlying assets,
Journal of Scientific Computing, 39(3), 323-339.

[79] O. C. Zienkiewicz and Y.K. Cheung (1967): The Finite Element Method in Con-
tinuum and Structural Mechanics, McGraw Hill, London.

110