Higher Order Smoothing Schemes For Inhomogeneous Parabolic Problems With Applications To Nonsmooth Payoff in Option Pricing

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 27

SUBMITTED MANUSCRIPT, JUNE 16, 2005, pp.

1–27

Higher Order Smoothing Schemes for Inhomogeneous


Parabolic Problems with Applications to Nonsmooth
Payoff in Option Pricing
B.A. Wade 1

Center for Industrial Mathematics, University of Wisconsin–Milwaukee


Milwaukee, Wisconsin 53201-0413, USA.
E-mail: wade@uwm.edu

A.Q.M. Khaliq
Department of Mathematical Sciences, Middle Tennessee State University
Murfreesboro, TN 37132-0001, USA.

M. Yousuf
Department of Mathematics and Physics, Alfred State College
Alfred, NY 14802, USA

J. Vigo–Aguiar 2

Departamento de Matemática Aplicada, Universidad de Salamanca


37003 Salamanca, Spain
E-mail: jvigo@usal.es

Abstract — A new family of numerical schemes for inhomogeneous parabolic partial


differential equations is developed. The new family of algorithms utilizes diagonal
Padé schemes combined with positivity–preserving Padé schemes as damping devices.
We also utilize partial fraction decomposition to address problems with accuracy and
computational efficiency in solving the higher order methods and to implement the
algorithms in parallel. Numerical experiments are presented for various examples from
financial mathematics, especially pricing options with nonsmooth payoffs.
2000 Mathematics Subject Classification: 65M12, 65M15, 65Y05, 65Y20.
Keywords: Padé Scheme, parabolic problem, nonsmooth data, positivity, nonsmooth
pay-off, Black-Scholes PDE.

1
Supported in part by grant number No. SAB2003-0266 from the Ministerio de Educación, Cultura y
Deporte, Spain, and by the U.S. National Security Agency under Grant Agreement Number H98230-05-1-
0062. The United States Government is authorized to reproduce and distribute reprints notwithstanding
any copyright notation herein.
2
Supported by grants No. MTM2004-00295 from the Ministerio de Ciencia y Tecnologı́a and No.
SAB2003-0266 from the Ministerio de Educación, Cultura y Deporte, Spain, as well as no. SA024/04
from Junta Castilla y León, Spain.
2 Wade, Khaliq, Yousuf, Vigo-Aguiar

1. Introduction
For homogeneous parabolic partial differential equations (PDE) with nonsmooth initial data,
a family of higher order accurate smoothing schemes has recently been developed by Wade
et al. [22]. Convergence results and numerical experiments show that these schemes can
be more robust than the well known Rannacher smoothing schemes [16, 17] with respect to
spurious oscillations generated through high frequency components in nonsmooth initial or
boundary data. Other papers related to the subject are [2, 3, 4, 5, 9, 16, 17, 20, 24], which,
however, treat cases of nonsmooth data only for the homogeneous case.
Robust schemes for inhomogeneous parabolic partial differential equations with nons-
mooth initial, boundary, or forcing terms have only been lightly treated, for example, in
[3, 4, 9]. In this article we develop a useful family of numerical schemes for advection–
diffusion–reaction equations under conditions of low regularity by extending the aforemen-
tioned techniques of [22]. We then focus on some practical illustrations of the new family of
schemes using important problems in the area of financial mathematics.
Khaliq and Wade [9] have worked on a smoothing strategy for the Crank-Nicolson scheme
using variable time steps for inhomogeneous equations. Using a rational approximation of the
exponential function e−z , designed to have three distinct real poles, Serbin [19] has developed
a fourth order A-stable scheme for inhomogeneous case, although only for smooth data. Voss
and Khaliq [21] have also developed a fourth order scheme for inhomogeneous problems with
smooth data using a rational approximation with four distinct real poles. In these papers the
authors use partial fraction decomposition to deal with the difficulties in computing higher
order matrix polynomials in the denominators, and to allow a parallel implementation. These
schemes are designed to be implemented on serial or parallel machines. For the scheme given
by Serbin [19], three systems must be solved, whereas the scheme developed Khaliq and Voss
[21] requires four systems at each time level.
In contrast to the aforementioned rational approximations, only one algebraic system
arises if the diagonal (2, 2)-Padé is employed instead. However, one disadvantage of using
diagonal Padé schemes is that they are not L-stable, in particular, their symbols converge
to magnitude one at infinity, so that when the initial data is not smooth high frequency
components in the error are not damped out. Here, we develop damping schemes, cf. [22], to
the inhomogeneous case, being careful to preserve the essential properties of parallelizability
and robustness for nonsmooth initial data enjoyed by the original family of schemes.
We shall begin by mentioning a class of single step fully discrete numerical schemes
described by Brenner et al. [3] and summarized in the book of Thomée [20, Ch. 9] since this
will lead us most easily to the introduction of the new family of numerical schemes. This
class of numerical methods achieves optimal order of convergence for smoothing initial data
while avoiding imposing unsatisfactory boundary conditions on the forcing term f and some
of its derivatives, cf. [20]. The inhomogeneous problem considered in [3, 20] satisfies certain
regularity assumptions, and the problem of nonsmoothness of the initial data has not been
sufficiently addressed.
In the presence of nonsmooth data, the diagonal Padé schemes need to have some damping
device to get optimal order convergence, cf. [11, 12, 16, 17, 22]. Our aim is to develop
an implementation strategy to approximate more efficiently and accurately the solution of
the linear inhomogeneous parabolic equation with nonsmooth initial data by synthesizing
various existing ideas. Similar to the approach adopted by Wade et al. [22], our first
Smoothing Schemes for Inhomogeneous Parabolic PDEs 3

basic strategy is to use a positivity preserving Padé scheme at the start as a damping
device followed by a diagonal Padé scheme. The second component of our strategy is a
partial fraction decomposition of the Padé approximants using the algorithm developed by
Gallopoulos and Saad [6] and Khaliq et al. [8]. With these approaches, we shall use partial
fraction decomposition form of the schemes at each time step, which needs only the work of
solving certain robust linear algebraic systems on parallel or serial machines. Even in serial,
it is necessary to implement this split version in order to avoid conditioning problems from
the higher matrix polynomials that would have to be inverted. That complex arithmetic is
necessary adds only a small inconvenience and a fractional amount of extra computational
time, which is compensated for by the robust performance of the higher order schemes.

2. The Abstract PDE and Standard Numerical Schemes


We shall assume an unbounded operator as infinitesimal generator of an analytic semigroup,
whose generality affords an easier analysis for convergence estimates that work without
dependence on the spatial mesh size. We adopt the treatment of V. Thomée, concisely
summarized in [20, Ch. 3, 7, 8, 9]. Consider the following PDE:

ut (x, t) + A(x)u = f (x, t) x ∈ Ω, t ∈ 0, t = J, (1)
u(x, t) = 0 x ∈ ∂Ω, t ∈ J, u(·, 0) = v on Ω,

where Ω is a bounded domain in Rd with lipschitz boundary and A denotes the uniformly
elliptic operator

d   X d
X ∂ ∂ ∂
A := − aj,k (x) + bj (x) + b0 (x).
j,k=1
∂x j ∂x k j=1
∂x j

The coefficients aj,k and bj are to be C ∞ (or sufficiently smooth) functions on Ω, aj,k = ak,j ,
b0 > 0, and for some c0 > 0

d
X
aj,k (·)ξj ξk > c0 |ξ|2 , on Ω, for all ξ ∈ Rd .
j,k=1

The initial value problem (2) is reset to be posed in a Hilbert space X, as follows, for the
reason that it makes the analysis simpler. Consider now A to be a linear, selfadjoint, positive
definite closed operator with a compact inverse T , defined on a dense domain D(A) ⊂ X.
The operator A could represent any of {Ah }0<h6h0 , obtained from a spatial discretization
and X could be X = Sh , an appropriate finite dimensional subspace of L2 (Ω), cf. [20].
We assume the resolvent set ρ(A) of A satisfies, for some α ∈ (0, π2 ),

ρ(A) ⊃ Σα , Σα := {z ∈ C : α < | arg(z)| 6 π, z 6= 0}.

Also, assume there exists M > 1 such that

k(zI − A)−1 k 6 M |z|−1 , z ∈ Σα .


4 Wade, Khaliq, Yousuf, Vigo-Aguiar

It follows that −A is the infinitesimal generator of an analytic semigroup {e−tA }t>0 which is
the solution operator for (2) below, cf. [14, 20]. There is a standard representation:
1
Z
−tA
E(t) := e = e−tz (zI − A)−1 dz,
2πi
Λ

where Λ := {z ∈ C : | arg(z)| = θ}, oriented so that Im(z) decreases, for any θ ∈ (α, π2 ).
The abstract initial value problem is as follows:
ut + Au = 0, t > 0, u(0) = v ∈ X. (2)
By the Duhamel principle the exact solution of (2) can be written as
Zt
u(t) = E(t)v + E(t − s)f (s)ds. (3)
0

The solution then satisfies:


Zt Zt+k
u(t + k) = E(k)E(t)v + E(k) E(t − s)f (s)ds + E(t + k − s)f (s)ds,
0 t
Z1
= E(k)u(t) + k E(k − kτ )f (t + kτ )dτ,
0

Thus, to develop a numerical scheme we may utilize the representation


Z1
u(tn+1 ) = e−kA u(tn ) + k e−kA(1−τ ) f (tn + τ k) dτ, (4)
0

where 0 < k 6 k, for some k, is the time step and tn = nk with 0 6 n 6 n = bt/kc. One
must be careful to respect any lack of regularity in the data and design a computational
procedure that is robust in its convergence properties.
To approximate the solution of (4) in time using vn to denote the computed approximation
of u(tn ), the scheme we consider has the following form:
s
X
vn+1 = r(kA)vn + k Pi (kA)f (tn + τi k), 0 6 n 6 n, v0 = v, (5)
i=1
Ps
which, by writing Rk f (tn ) = i=1 Pi (kA)f (tn + τi k), can be rewritten as
vn+1 = r(kA)vn + kRk (tn ), 0 6 n 6 n, v0 = v.
The functions r (z) and {Pi (z)}si=1 are to be rational functions bounded on the spectrum of
kA, uniformly in k. The distinct real numbers {τi }si=1 would typically be taken as Gaussian
Quadrature points in the interval [0, 1].
For the time stepping scheme (5) to be accurate of order q, it is required that the scheme
satisfy conditions developed by Brenner et al. [3]. The reader may consult [20, Ch. 9] to
fill in various details omitted here for the sake brevity. The following result describes the
accuracy and establishes some equivalence relations, which we then use later.
Smoothing Schemes for Inhomogeneous Parabolic PDEs 5

Proposition 1. [20, L. 9.1] The time discretization scheme (5) is accurate of order q if
and only if
r(z) = e−z + O(k q+1 ), z → 0, (6)
and for 0 6 l 6 q,
s l
!
X l! X (−z)j
τil Pi (z) = e−z − + O(z q−l ), z → 0. (7)
i=1
(−z)l+1 j=0
j!

Equation (7) is equivalent to


s
X Z1
τil Pi (z) = sl e−z(1−s) ds + O(z q−l ), z → 0. (8)
i=1 0

It is computationally efficient to have r(z) and {Pi (z)}si=1 such that they share the same
poles, which puts an additional constraint on the scheme. By considering, from [20],
N (z) Ni (z)
r(z) = and Pi (z) = , i = 1, 2, ..., s,
D(z) D(z)
with N (z), D(z), and Ni (z) as polynomials, the scheme (5) formally can be rewritten as
s
X
D(kA)vn+1 = N (kA)vn + Ni (kA)f (tn + τi k).
i=1

It is shown in [20] that for the case s = q (s is the number of quadrature points and q is the
accuracy of the scheme) this form as well as the conditions of the proposition can be achieved
by choosing the rational functions r(z) satisfying (6), selecting distinct real numbers, say,
by Gaussian Quadrature, {τi }m i=1 , and finally solving the system
q l
!
j
X l! X (−z)
τil Pi (z) = r(z) − , l = 0, 1, ..., q − 1, (9)
i=1
(−z)l+1 j=0
j!

for the coefficients of the Pi (z). This system (9) is of Vandermonde type (whose determinant
is not zero, the τi are distinct [1]), which gives the rational functions {Pi (z)}qi=1 as linear
combinations of the terms on the right hand side of (9). The only singularities are then
those of r(z). Thus the denominators of the {Pi }qi=1 have the same factors as that of r(z).
If r(z) is bounded for large z, then the right hand sides of (9) are small for large z, and the
numerator of Pi (z) would be of lower degree than its denominator for each i.
For the case when the number of quadrature points s is less than the order of the scheme q,
an alternative formula similar to (9) is given in [20]. The accuracy conditions are reformulated
by defining
l m
(−z)j
  X
l! X
δl (z) = r(z) − − τil qi (z), l = 0, 1, ..., q − 1
(−z)l+1 j=0
j! i=1
and
q
(−z)j
 
q! X
δq (z) = r(z) − .
(−z)q+1 j=0
j!
6 Wade, Khaliq, Yousuf, Vigo-Aguiar

and requiring that


δl (z) = O(z q−l ), as z → 0, for l = 0, 1, ..., q.
For the purpose of construction of the schemes when s < q, it is shown in [20, Lemma 9.2]
that this is equivalent to the condition (6) plus

δl (z) = 0, as z → 0, for l = 0, 1, ..., s − 1

and a moment condition


Z1
w(τ )τ j dτ = 0, for j = 0, 1, ..., q − s − 1
0
Qs
on the quadrature points, with ω(τ ) ≡ i=1 (τ − τi ). The formula to obtain the rational
functions {Pi (z)}si=1 in [20] is
s l
!
X l! X (−z)j
τil Pi (z) = r(z) − , l = 0, 1, ..., s − 1. (10)
i=1
(−z)l+1 j=0
j!

3. Padé Schemes for Inhomogeneous Problems


Using Padé schemes as rational function approximations of the matrix exponential functions
in (4, 5) is an important and common tool. Let Pn,m (z) and Qn,m (z) be two polynomials of
degree n and m, respectively, defined as follows [10]:
n
X (m + n − j)!n!
Pn,m (z) = (−z)j ,
j=0
(m + n)!j!(n − j)!
m
X (m + n − j)!m!
and Qn,m (z) = (z)j .
j=0
(m + n)!j!(m − j)!

The following useful properties of (n, m)−Padé are given, for example, in Lambert [10]: The
(n, m)−Padé approximant Rn,m to e−z is: (i) A-acceptable if n = m, (ii) A0 -acceptable if
n 6 m, and (iii) L-acceptable if n = m − 1 or n = m − 2.
For example, the Backward Euler scheme is R0,1 (z) = (1 + z)−1 , the Crank-Nicolson
scheme is R1,1 (z) = (1 − 21 z)(1 + 12 z)−1 , and other higher order schemes of present interest
include the following:
1 1 1 2 1
R0,3 (z) = (1 + z + z 2 + z 3 )−1 , R1,2 (z) = (1 − z)(1 + z + z 2 )−1 ,
2 6 3 3 6
1 1 2 1 1 2 −1
R2,2 (z) = (1 − z + z )(1 + z + z ) ,
2 12 2 12
1 2 1 3 1 4 1 5 −1
R0,5 (z) = (1 + z + z + z + z + z ) ,
2 6 24 120
2 1 3 3 1
R2,3 (z) = (1 − z + z 2 )(1 + z + z 2 + z 3 )−1 ,
5 20 5 20 60
1 1 2 1 3 1 1 1 3 −1
R3,3 (z) = (1 − z + z − z )(1 + z + z 2 + z ) .
2 10 120 2 10 120
Smoothing Schemes for Inhomogeneous Parabolic PDEs 7

We shall design our schemes for inhomogeneous problems based on the positivity pre-
serving (0, 2m − 1)–Padé schemes as damping devices. These practical Padé schemes do
well in approximating the exponential function in the sense that their symbols capture the
properties of positivity and monotone convergence to zero at infinity of e−x , x > 0— thus
avoiding the amplification or introduction of oscillations in high frequency components of
the error. Although the first subdiagonal (m-1, m)–Padé schemes are also L-stable and
sometimes have a smaller constant in the local truncation error, their symbols change sign
or their derivatives do, which can cause unphysical oscillations due to amplification in the
error, depending on the relative sizes of the diffusion rate, the spatial mesh size and time
steps.
As an example, let us begin by mentioning the backward Euler scheme for inhomogeneous
problems. Using the rational approximation r(z) = R0,1 (z) and τ1 = 21 , this first order scheme
is vn+1 = R0,1 (kA) vn + kP1 (kA)f tn + 21 k , for which P1 (z) = 1+z
1
. Next, we shall give the
more complicated versions, involving higher order schemes. These standard schemes form
the basis for the new family of damping schemes of this.

3.1. A fourth Order Scheme for Inhomogeneous Problems


√ √
Let r(z) = R2,2 (z), and define τ1 = 3−6 3 and τ2 = 3+ 3
6
, the Gaussian quadrature points of
order 2. Then the system (10) reduces to:

1
P1 (z) + P2 (z) = − (r (z) − 1) , (11)
z
1
τ1 P1 (z) + τ2 P2 (z) = 2 (r (z) − 1 + z) ,
z
which results in a fourth order scheme as follows:

vn+1 = R2,2 (kA) vn + kP1 (kA) f (tn + τ1 k) + kP2 (kA)f (tn + τ2 k)

where √ √
1 1 − 63 z 1 1 + 63 z
P1 (z) = 1 2 , P2 (z) = 1 2.
2 1 + 12 z + 12 z 2 1 + 12 z + 12 z
The above mentioned scheme performs with fourth order accuracy only if the initial data
has sufficient regularity.
To achieve good convergence and robust performance with diagonal Padé schemes some
damping treatment is necessary. The idea is to use positivity preserving (0, 2m − 1)-Padé
schemes for initial damping and then use diagonal (m, m)-Padé schemes after that. That
the global accuracy of the (0, 2m − 1)-Padé schemes is one less does not matter because
the damping scheme is used only two times, and the local truncation error, one higher than
the global, then truly affects the accuracy. Next, we develop the details for an important
damping scheme in our proposed fourth order numerical method.

3.2. A Third Order Scheme for Inhomogeneous Problems


Taking the rational approximation r(z) = R0,3 (z), a good choice for its positivity properties,
and using the Gaussian quadrature points τ1 and τ2 as above, we construct the following
8 Wade, Khaliq, Yousuf, Vigo-Aguiar

third order scheme:

vn+1 = R0,3 (kA) vn + kP1 (kA) f (tn + τ1 k) + kP2 (kA)f (tn + τ2 k)

1
√  √ 
11+ 6
3 − 3 z + 61 1 − 3 z 2
P1 (z) = ,
2 1 + z + 21 z 2 + 16 z 3
1
√  √ 
11+ 6
3 + 3 z + 61 1 + 3 z 2
P2 (z) = .
2 1 + z + 21 z 2 + 16 z 3

4. The New Smoothing Schemes IPSP(m)


In this section we will present the new family of smoothing/damping schemes in general.
Similar to the homogeneous case, we use the idea of combining diagonal Padé schemes Rm,m
with the positivity preserving Padé schemes R0,2m−1 . Because of the positivity preserving
property of these starting schemes, we shall use the notation IPSP(m) for our inhomogeneous
positively smoothed Padé schemes. Each IPSP(m) scheme is of order 2m. The proposed
family of smoothing schemes is as follows:
 Pm
 R0,2m−1 (kA)vn + k
 Pi (kA)f (tn + τi k), 0 6 n < p;
i=1
(IPSP(m)) vn+1 = Pm (12)
 Rm, m (kA)vn + k
 Pi (kA)f (tn + τi k), n > p,
i=1

where the Pi are solutions of equation (10) when r = R0,2m−1 or Rm,m , respectively. Here,
p > 2 is the number of initial damping steps, where p can always be taken as 2 independent
of m. Taking p > 2, however, can sometimes result in a bit more damping, for example,
p = 4 occasionally works better. Much larger values of p do not give significant improvement.
For √example, the
√ fourth order scheme, using second order Gaussian quadrature points
3− 3 3+ 3
τ1 = 6 , τ2 = 6 , has the following form:
 1 1 −1
v1 = 1 + kA + (kA)2 + (kA)3 v0 + kP1 (kA)f (t0 + τ1 k) + kP2 (kA)f (t0 + τ2 k),
2 6
 1 1 −1
v2 = 1 + kA + (kA)2 + (kA)3 v1 + kP1 (kA)f (t1 + τ1 k) + kP2 (kA)f (t1 + τ2 k),
2 6
with P1 (kA) and P2 (kA) as
√  √ 
3− 3 1− 3
 
1 2 1 1 −1
P1 (kA) = 1+ kA + (kA) 1 + kA + (kA)2 + (kA)3
2 6 6 2 6
√  √ 
3+ 3 1+ 3
 
1 1 1 −1
P2 (kA) = 1+ kA + (kA)2 1 + kA + (kA)2 + (kA)3 ,
2 6 6 2 6
and for n > 2
1 1 1 1 −1
1 − kA + (kA)2 1 + kA + (kA)2 vn + kP1 (kA)f (tn + τ1 k)

vn+1 =
2 12 2 12
+kP2 (kA)f (tn + τ2 k),
Smoothing Schemes for Inhomogeneous Parabolic PDEs 9

where P1 (kA) and P2 (kA) are


 √ 
1 3 1 1 −1
P1 (kA) = 1− kA 1 + (kA) + (kA)2
2 6 2 12
 √ 
1 3 1 1 −1
P2 (kA) = 1+ kA 1 + (kA) + (kA)2 .
2 6 2 12

4.1. Partial Fraction Form of the IPSP(m) Schemes


A problem now arises, how to accurately and efficiently compute these solutions despite
the presence of higher order matrix polynomials that must be inverted. These schemes are
designed to share the same denominator, and we may take advantage of recent advances in
parallelization of Padé schemes, albeit for a somewhat different application, developed by a
number of authors, particularly, by Gallopoulos and Saad [6], Serbin [19], Khaliq et al. [8],
and Ruesch et al. [18]. Allowing the use of complex arithmetic, it has become possible to
implement more efficiently the higher order Padé schemes in parallel or in serial by using
the partial fraction technique. Even in serial, it is best to employ these splittings for the
calculation of the inverse in order to avoid disproportionate roundoff error.
The partial fraction form of the rational functions r(z) = Rn,m (z) and {Pi (z)}m
i=1 requires
us to consider two cases, n < m for subdiagonal Padé schemes and n = m for diagonal Padé
schemes. If n < m, then we utilize
q1 q1 +q2  
X wj X wj
Rn,m (z) = +2 <
j=1
z − cj j=q +1
z − cj
1

and the corresponding {Pi (z)}m


i=1 take the form

q1 q1 +q2  
X wij X wij
Pi (z) = +2 < , i = 1, 2, ..., m.
j=1
z − c j j=q +1
z − c j
1

Here, Rn,m as well as the Pi have q1 real and 2q2 nonreal poles {cj } , with q1 + 2q2 = n, and
0 0
where wj = Pn,m (cj ) /Qn,m (cj ) and wij = Ni (cj ) /D (cj ) . For the case when m = n, we
utilize
q1 q1 +q2  
m
X wj X wj
Rm,m (z) = (−1) + +2 <
j=1
z − cj j=q +1
z − cj
1

and
q1 q1 +q2  
X wij X wij
Pi (z) = +2 < , i = 1, 2, ..., m.
j=1
z − c j j=q +1
z − c j
1

The following algorithm results:


Parallel Algorithm
m
P
1. For i = 1, ..., q1 + q2 , solve (kA − ci I) yi = wi vs + kwij f (ts + τj k) ;
j=1

q1
P q1P
+q2
2a. (If n < m) Compute vs+1 = yi + 2 < (yi ) ;
i=1 i=q1 +1
10 Wade, Khaliq, Yousuf, Vigo-Aguiar

q1 q1P
+q2
2b. (If n = m) Compute vs+1 = (−1)m vs +
P
yi + 2 < (yi ) .
i=1 i=q1 +1

In order to construct some complete examples of smoothing schemes in their partial


fraction form, we have computed the poles and weights of the schemes (0, 3)-Padé, and (2,
2) -Padé approximations of e−z .
I. For the (0, 3)−Padé scheme, we have q1 = q2 = 1 and

c1 = −1.5960716379833, c2 = −0.7019641810083 − i1.807339494452,


w1 = 1.475686517795720, w2 = −0.7378432588979 + i0.365017840801,
w11 = 0.25964745169791, w12 = −0.3128364277412 + i0.472314917248,
w21 = 0.66492666056455, w22 = 0.3505493716099 − i0.04941905457189.

The algorithm works out to be


vs+1 = y1 + 2<(y2 ),
where
(kA − c1 I) y1 = w1 vs + kw11 f (ts + τ1 k) + kw12 f (ts + τ2 k) ,
and
(kA − c2 I) y2 = w2 vs + kw21 f (ts + τ1 k) + kw22 f (ts + τ2 k) ,
which can be solved in parallel on two machines for a speedup, or on a serial machine, to
preserve accuracy.

II. The fourth order scheme, (2, 2) −Padé, works out as follows: q1 = 0, q2 = 1 and

c = −3. − i1.732050807568877,
w = −6. + i10.39230484541327,
w11 = −0.86602540378 + i3.232050807569,
w12 = 0.866025403784 + i0.23205080757.

The algorithm looks like this:


vs+1 = vs + 2<(y),
where
(kA − cI) y = wvs + kw11 f (ts + τ1 k) + kw12 f (ts + τ2 k) .
Now we must put these schemes together, using (0, 3)-Padé twice as a damping scheme
at the start, thereafter with (2, 2)-Padé.

5. Convergence in Hilbert Space


In this section we will prove convergence results for the smoothing scheme in the Hilbert
space case, assuming A is a self-adjoint operator. The approach we wish to follow for the in-
homogeneous case is based upon the analysis of Brenner, Crouziex, and Thomée described in
[3] as well as summarized in Thomée [20, Ch. 9]. We are showing the proof as a modification
for the case when the time stepping scheme is a combination of two rational approximations
Smoothing Schemes for Inhomogeneous Parabolic PDEs 11

instead of that given in [20]. Also we consider the case when initial data is nonsmooth. In
this analysis we use the spaces Ḣ s = D(As/2 ) as defined in [20], with norm
N
!1/2
X
|v|s = (As v, v)1/2 = kAs/2 vk = λsj (v, φj )2 ,
j=1

where {φj }N N
j=1 are orthonormal eigenfunctions of A and {λj }j=1 the eigenvalues, which are
positive.
We assume f ∈ Ḣ s to have sufficient regularity. For the starting scheme we use the shorter
notation rs (z) to indicate the (0, 2m − 1)-Padé initial damping scheme and correspondingly
we use just Rk f (tn ) for Rks f (tn ) because context is clear. For the main scheme we utilize
the notation rm (z) for Rm,m (z) and correspondingly Rk f (tn ) = Rkm f (tn ). Also we shall
use the idea that the operator Ek = r(kA) is said to stable in H if kEkn k 6 C for n > 1,
0 < k 6 k, nk 6 t, cf. [20].

Theorem 5.1. [20, Thm. 9.1] Assume A is a self-adjoint operator on a Hilbert space
H and the time discretization scheme (12) is accurate of order q = 2m, where m is a positive
integer. Suppose f (l) (t) ∈ Ḣ 2q−2l for l < q, t > 0. Then there exists a constant C = C(t)
such that
 q−1
X Ztn 
q (q)
kvn − u(tn )k 6 Ck t−q
n kvk + tn Sl + kf kds , 0 6 n 6 n, 0 < k 6 k. (13)
l=0 0

where Sl = sups6tn |f (l) (s)|2q−2l


Proof. Note, by [20] the solution operator Ek = r(kA) is stable in H, where r is either
of R0,2m−1 or Rm,m . The starting scheme is
n−1
X
vn = rsn (kA)v + k rsn−1−j (kA)Rks f (tj ), for n = 1, 2, (14)
j=0

and with rs = R0,2m−1 . The main scheme, with rm = Rm,m , is as follows:


1
X
n−2
vn = r m (kA)rs2 (kA)v + k n−2
rm (kA)rs1−j (kA)Rks f (tj )
j=0
n−1
X
n−1−j
+k rm (kA)Rkm f (tj ), n > 3. (15)
j=2

Also, using E(t) = e−tA , the exact solution of equation (2) can be written as:
1
X n−1
X
u(tn ) = E(tn )v + k E(tn−1−j )Ik f (tj ) + k E(tn−1−j )Ik f (tj ), (16)
j=0 j=2

where
Z1
Ik f (tj ) = E(k − sk)f (tj + sk)ds.
0
12 Wade, Khaliq, Yousuf, Vigo-Aguiar

The error E n = vn − u(tn ), for n > 2, can be written as:

E n = rm
n−2
(kA)rs2 (kA)v − E(tn )v
X 1
n−2
(kA)rs1−j (kA)Rks f (tj ) − E(tn−1−j )Ik f (tj )

+k rm
j=0
n−1
X
n−1−j

+k rm (kA)Rks f (tj ) − E(tn−1−j )Ik f (tj )
j=2
= E0n + Esn + Em
n
, (respectively)

where E0n is the error for the corresponding homogeneous equation, Esn is the error for the
n
inhomogeneous part of the starting scheme, and the error Em is due to the inhomogeneous
part of the main scheme.
The error term E0n can be approximated by the established result [22, Theorem 3.1] as
follows:

kE0n k = k rm
n−2
(kA)rs2 (kA) − E(tn ) k 6 Ck q t−q

n kvk (17)
n−1−j n
After adding and subtracting rm (kA)Ik f (tj ) in the error term Em and rearranging its
terms, we arrive at
n−1
X n−1
X
n n−1−j n−1−j

Em = k rm (kA) − E(tn−1−j ) Ik f (tj ) + k rm (kA)(Rkm − Ik )f (tj )
j=2 j=2
n n
= Em1 + Em2 . (respectively) (18)

Following the approach given by Thomée [20, Theorem 9.1] we have the following estimate
n n
for Em1 and Em2 ,
n
Rt
kEm1 k 6 Ck q t2n |f |2q ds. (19)

which is bounded by the right hand side of (13). Also,


t
n−1
X q−1
X n−1 Zj+1
X
n
kEm2 k 6 Ck q+1 |f (l) (tj )|2q−2l + Ck q kf (q) kds. (20)
j=2 l=0 j=2 t
j

Incorporating (19) into the right hand side of (20), we therefore arrive at the following
estimate for the main scheme:
t
n−1
X q−1
X n−1 Zj+1
X
n
kEm k 6 Ck q+1 |f (l) (tj )|2q−2l + Ck q kf (q) kds. (21)
j=2 l=0 j=2 t
j

Using a similar approach, we can also develop the estimate for the starting scheme as:
t
1
X q−1
X 1 Zj+1
X
kEsn k 6 Ck q+1 |f (l) (tj )|2q−2l + Ck q kf (q) kds, 0 6 n 6 2. (22)
j=0 l=0 j=0 t
j
Smoothing Schemes for Inhomogeneous Parabolic PDEs 13

Combining (21) and (22),


t
n−1
X q−1
X n−1 Zj+1
X
kEsn k + n
kEm k 6 Ck q+1 (l)
|f (tj )|2q−2l + Ck q
kf (q) kds,
j=0 l=0 j=0 t
j

q−1
X Ztn
6 Ck q tn Sl + Ck q kf (q) kds. (23)
l=0 0

where Sl = sups6tn |f (l) (s)|2q−2l , and (17) together with (23) complete the proof.

6. Numerical Experiments
In this section we demonstrate the performance of the new algorithm on some important
examples form financial mathematics, using a Black-Scholes PDE model. Consider the
following nonself-adjoint advection-diffusion type equation [23] with spatially variable coef-
ficients:
∂V 1 ∂2V ∂V
+ σ 2 S 2 2 + rS − rV = 0. (24)
∂t 2 ∂S ∂S
Here S represents the price of the underlying asset (serving as a space variable), and V (S, t)
the value of the option at time t before the expiry time T ; parameters σ, r and E are the
volatility of the underlying asset, the interest rate, and the exercise/strike price of the option,
respectively. One can find the derivation, background, and technical details of this model
in [7, 13, 23]. To determine V (S, t) uniquely we must specify other conditions that involve
information about the particular option, e.g., initial and boundary conditions.
Our first example is of a so-called butterfly option which has three strike prices. The
payoff for this option has three corners at strike prices. The second example is of a binary
call option whose payoff has a jump discontinuity at the strike price.
Convergence results for Examples 6.1 and 6.2 are obtained using the fourth order PSP(2)
smoothing scheme. In our first experiment, we employ a spatial discretization using a fourth
order central difference scheme and compute combined fourth order convergence in space
and time. In our second experiment, we compute with a very small spatial mesh so that
there is relatively no error in the x variable to show fourth order convergence in time. All
the numerical schemes are implemented in serial, except we use the split versions described
earlier. Without these splittings the higher polynomial functions of the matrices that must
be inverted would cause numerical difficulties.
The Delta of an option is an important parameter in pricing and hedging of that option.
It is the rate of change of the option value with respect to the asset price, [7, 23]. In
our examples, we compute the Delta of an option using the second order central difference
formula:
∂V V k − Vi−1
k
(Si , tk ) ≈ i+1
∂S Si+1 − Si−1
Another important parameter used in hedging strategies is the Gamma of an option. It is the
rate of change of Delta with respect to asset price. If Gamma is small, Delta changes slowly,
14 Wade, Khaliq, Yousuf, Vigo-Aguiar

and adjustments to keep a portfolio Delta neutral need to be made relatively infrequently.
We compute the Gamma of an option using a second order central difference formula:
k
∂2V Vi+1 − 2Vik + Vi−1
k
(S i , t k ) ≈ , (h = Si+1 − Si ).
∂S 2 h2

6.1. A Butterfly Spread


A Butterfly Spread is a combination of three options with three strike prices, in which one
contract is purchased with two outside strike prices and two contracts are sold at the middle
strike price. We will use a call option in this example, solving the following PDE model:

∂V 1 ∂2V ∂V
+ σ 2 S 2 2 + rS − rV = 0, 0 6 S 6 100, t ∈ [0, 0.5], (25)
∂t 2 ∂S ∂S
with payoff at expiry being
V (S, T ) = max(S − E1 , 0) − 2 max(S − E2 , 0) + max(S − E3 , 0)
where E1 , E2 , and E3 are the three strike prices with E1 < E2 < E3 and E2 = (E1 + E3 )/2.
This is a homogeneous problem with corners in the initial data, i.e. the payoff function, at
E1 , E2 , and E3 ; its Delta has three jump discontinuities. The behavior of Delta at expiry
can be summarized with the following conditions:



 0, for 0 6 S < E1 ;
∂V 
1, for E1 6 S < E2 ;
lim− = (26)
t→T ∂S 
 −1, for E2 6 S < E3 ;
0, for S > E3 .

Tables 1 and 2 show the order of convergence for the fourth order undamped scheme
(2,2)-Padé (# 1) and the fourth order smoothing scheme PSP(2) (# 2). The convergence
order is evaluated at the strike price E2 = 0.5. It is clear from Table 1 that the scheme is
not showing fourth order of convergence, while Table 2 shows the expected order of conver-
gence. The column ”Order” represents the exponent on k in the ratio of the error sequence,
computed in the standard fashion as the ratio of the logarithms of the ratios between the
errors and step sizes, respectively. As there is no analytic solution available, the error has
been developed relative to a reference solution computed separately on an extremely fine
grid with the backward Euler scheme. The reference solution at the strike price E2 = 0.5 is
≈ 0.02102705683555.
Figure 1 is the graph of the payoff function for the Butterfly Spread. Time evolution
graphs shown in Figures 2 and 3 are obtained using (2,2)-Padé and PSP(2), respectively.
Figure 2 shows oscillations at the three strike prices E1 , E2 , and E3 since there is irregularity
at these points. These oscillations are recovered using our smoothing scheme as shown in
Figures 3. Figures 4 and 5 show the time evolution plots of Delta for the Butterfly Spread
using (2,2)-Padé and PSP(2) respectively. Figure 4 shows large oscillations at the three strike
prices E1 , E2 , and E3 due to a jump discontinuity at these points. Again these solutions
are accurately recovered using our smoothing scheme as shown in Figure 5. The parameters
used for the Figures 2–5 are as follows: σ = 0.5, r = 0.1, ∆t = 0.2, ∆S = 1.0 and three
strike prices E1 = 40, E2 = 50 E3 = 60.
Smoothing Schemes for Inhomogeneous Parabolic PDEs 15

Table 1. Convergence results for the Butterfly Spread using the undamped (2,2)-Padé scheme. Parameter
values are: x0 = 0, xn = 1, T = 0.5, r = 0.1, and σ = 0.5.

∆S ∆t Strike Price (E2 ) Error (E2 ) Order


0.0010000 0.2500 0.03111779 1.009e-002 —
0.0005000 0.1250 0.02612218 5.095e-003 0.98584
0.0002500 0.0625 0.02356926 2.542e-003 1.00304
0.0001250 0.0313 0.02229881 1.272e-003 0.99926
0.0000625 0.0156 0.02166293 6.359e-004 1.00001

Table 2. Convergence results for the Butterfly Spread using the smoothing scheme PSP(2). We have used
the same parameters as in Table 1.

∆S ∆t Strike Price (E2 ) Error (E2 ) Order


0.0010000 0.2500 0.02122575 1.987e-004 —
0.0005000 0.1250 0.02104830 2.124e-005 3.22565
0.0002500 0.0625 0.02102885 1.789e-006 3.56937
0.0001250 0.0313 0.02102719 1.325e-007 3.75493
0.0000625 0.0156 0.02102707 8.707e-009 3.92811

10

6
Payoff

0
0 10 20 30 40 50 60 70 80 90 100
Asset Price

Figure 1. The Payoff function of the Butterfly Spread cf. [23, p. 259].
16 Wade, Khaliq, Yousuf, Vigo-Aguiar

4.5

3.5

3
Option Value

2.5

1.5

0.5

0
1

0.5

80 90 100
50 60 70
0 20 30 40
Time 0 10
Asset Price

Figure 2. The Butterfly Spread using the undamped (2,2)-Padé scheme.

4.5

3.5

3
Option Value

2.5

1.5

0.5

0
1

0.5

80 90 100
50 60 70
0 20 30 40
Time 0 10
Asset Price

Figure 3. The Butterfly Spread using the smoothing scheme PSP(2).


Smoothing Schemes for Inhomogeneous Parabolic PDEs 17

0.8

0.6

0.4

0.2

0
Delta

−0.2

−0.4

−0.6

−0.8

−1
1

0.5

0 60 70 80 90 100
Time 20 30 40 50
0 10
Asset Price

Figure 4. The Delta of the Butterfly Spread using the undamped (2,2)-Padé scheme.

0.8

0.6

0.4

0.2

0
Delta

−0.2

−0.4

−0.6

−0.8

−1
1

0.5

0 60 70 80 90 100
Time 20 30 40 50
0 10
Asset Price

Figure 5. The Delta of the Butterfly Spread using the smoothing scheme PSP(2).
18 Wade, Khaliq, Yousuf, Vigo-Aguiar

6.2. A Digital Call Option


The preceding example was of homogeneous type (f = 0); we now move to consider a
nonhomogeneous problem in financial mathematics. Here, one boundary condition (31) is
time-dependent, which can be equivalently reformulated as a nonhomogeneous problem. In
this example we solve the Black-Scholes PDE model

∂C 1 2 2 ∂ 2 C ∂C
+ σ S + rS − rC = 0, 0 6 S 6 80, t ∈ [0, 0.5], (27)
∂t 2 ∂S 2 ∂S
for the case when the payoff is not continuous. The payoff function for this call option, also
called ‘cash-or-nothing,’ is given as:

A, for S > E;
lim− C(S, t) = (28)
t→T 0, for S < E.

where the constant A > 0 represents the payoff amount at expiry. Discontinuities in the
payoff reduce the order of convergence, which is restored by averaging the payoff as:

 A, for S > E;
lim− C(S, t) = A/2, for S = E; (29)
t→T
0, for S < E.

When S = 0, the asset remains at zero for all later times and hence the payoff is zero. This
gives the boundary condition

C(0, t) = 0, for all 0 6 t 6 T. (30)

When S is large, the option is almost certain to pay off the amount A. Therefore after
discounting for interest, we assume

C(S, t) ≈ Ae−r(T −t) , S → ∞. (31)

The analytic solution for this digital option, as given in [7, 23], is:

C(S, t) = Ae−r(T −t) N (d2 ), (32)

where
log(S/E) + (r − 12 σ 2 )(T − t)
d2 = √ (33)
σ T −t
and N (·) is the cumulative distribution function for a standardized normal random variable:
Zx
1 1 2
N (x) = e− 2 s ds. (34)

−∞

The Delta of the digital call option, partial derivative of (32) with respect to S, has form:

∂C Ae−r(T −t) N 0 (d2 )


= √ (35)
∂S σS T − t
Smoothing Schemes for Inhomogeneous Parabolic PDEs 19

Table 3. Convergence results for the Digital Call option using the (2,2)-Padé. Parameters are: x 0 = 0, xn =
1, T = 0.5, r = 0.05, and σ = 0.2.

∆S ∆t Strike Price (E2 ) Error (E2 ) Order


0.0001 0.5000 0.5279693254 4.355e-003 —
0.0001 0.2500 0.5301633433 2.161e-003 1.01082
0.0001 0.1250 0.5312449746 1.080e-003 1.00120
0.0001 0.0625 0.5317849008 5.399e-004 1.00002
0.0001 0.0313 0.5320547600 2.701e-004 0.99948

Table 4. Convergence results for the Digital Call option using PSP(2).

∆S ∆t Strike Price (E2 ) Error (E2 ) Order


0.0001 0.5000 0.5322132399 1.116e-004 —
0.0001 0.2500 0.5323137134 1.110e-005 3.32913
0.0001 0.1250 0.5323239089 9.066e-007 3.61423
0.0001 0.0625 0.5323247500 6.549e-008 3.79113
0.0001 0.0313 0.5323248117 3.796e-009 4.10875

and the behavior of Delta at expiry can be summarized as [7]:



0, for S > E;
∂C 
lim = ∞, for S = E; (36)
t→T − ∂S
0, for S < E.

Tables 3 and 4 show the order of convergence for the fourth order undamped scheme (2,2)-
Padé the smoothing scheme IPSP(2), respectively, at T = 1. It is evident that the ordinary
(2,2)-Padé scheme is not showing the expected fourth order, while IPSP(2) does. The con-
vergence is computed at the strike price E where the payoff has a jump discontinuity. The
analytic solution at the strike price is approximately 0.53232481545376
Figure 6 is the graph of the payoff function for the digital call option. The time evolution
graphs of the digital call option, shown in Figures 7 and 8, are obtained using the (2,2)-Padé
and PSP(2) schemes, respectively. A discontinuity in the payoff function results in spurious
oscillation in the (2,2)-Padé scheme, which are eliminated by the smoothing scheme PSP(2).
Figures 9 and 10 are the time evolution graphs of the Delta function using (2,2)-Padé and
PSP(2) schemes respectively. The parameter used to obtained these graphs are as follows:
σ = 0.3, r = 0.05, ∆t = 0.1, and ∆S = 0.2667. Figure 11 shows the graph of the Gamma
function using the (2,2)-Padé scheme combined with the smoothing scheme PSP(2).

6.3. European Call Option (Dimensionless Form)


Considering the Black-Scholes PDE model (24) with

V (0, t) = 0, V (S, t) ∼ S, as S → ∞, and V(S, T) = max(S − E, 0)


20 Wade, Khaliq, Yousuf, Vigo-Aguiar

1.1

0.9

0.8

0.7

0.6
Payoff

0.5

0.4

0.3

0.2

0.1

−0.1
0 10 20 30 40 50 60 70 80
Asset Price

Figure 6. The Payoff function of the Digital Call option.

0.8
Option Value

0.6

0.4

0.2

0
0.5
0.4 80
0.3 60
0.2 40
0.1 20
0 0
Time
Asset Price

Figure 7. The Digital Call Option using the undamped scheme (2,2)-Padé.
Smoothing Schemes for Inhomogeneous Parabolic PDEs 21

0.8
Option Value

0.6

0.4

0.2

0
0.5
0.4 80
0.3 60
0.2 40
0.1 20
0 0
Time
Asset Price

Figure 8. The Digital Call Option using the smoothing scheme PSP(2).

0.3

0.25

0.2
Option Value

0.15

0.1

0.05

−0.05
0.5
0.4 80
0.3 60
0.2 40
0.1 20
0 0
Time
Asset Price

Figure 9. The Delta of the Digital Call Option using the undamped scheme (2,2)-Padé.
22 Wade, Khaliq, Yousuf, Vigo-Aguiar

0.2

Option Value 0.15

0.1

0.05

0
0.5
0.4 80
0.3 60
0.2 40
0.1 20
0 0
Time
Asset Price

Figure 10. The Delta of the Digital Call Option using the smoothing scheme IPSP(2).

−3
x 10
5
PSP(2)
(2,2)−Pade
4

1
Gamma

−1

−2

−3

−4

−5
25 30 35 40 45 50 55
Asset Value

Figure 11. Comparison of the Gamma values for the Digital Call option using the undamped (2,2)-Padé
scheme and the smoothing scheme IPSP(2). The parameters used for this graph are σ = 0.3, r = 0.05, ∆t =
0.02, and ∆S = 0.5.
Smoothing Schemes for Inhomogeneous Parabolic PDEs 23

0.8

0.6
u

0.4

0.2

0
5
4
0.2
3
0
2 −0.2
−0.4
1 −0.6
0 −0.8
t
x

Figure 12. Time evolution graph of a European call option using (2,2)-Padé.

and applying the following transformation


τ
S = Eex , t=T− 1 2, C = Eu(x, τ ).
2
σ

a dimensionless and forward in time form of (24) is obtained, cf. [23]. We will solve the
transformed equation using parameter values from [13], which is o the form

∂u ∂2u ∂u
= 2 + (c − 1) − cu, t0 6 t 6 tf , a6x6b (37)
∂t ∂x ∂x

where c= 1 rσ2 , r = 0.065, σ = 0.8, a = ln 25 , b = ln 75 , t0 = 0, tf = 5, and with initial


 
2
condition
u(x, 0) = max(exp(x) − 1, 0)
and boundary conditions

7 − 5 exp(−kt)
u(a, t) = 0, u(b, t) = .
5
Figures 12 and 13 are the surface plots of European Call Option in dimensionless form
using (2,2)-Padé and PSP(2), respectively. The parameter values used are σ = 0.8, r = 0.01,
T = 5.0, ∆t = 0.25, and ∆S = 0.02. At x = 0 the payoff has a corner and therefore
its derivative is discontinuous there. The graph for the (2,2)-Padé shows some unwanted
oscillations at x = 0 where as the graph of PSP(2) shows no oscillations. The oscillations at
x = 0 become worse for the delta function shown in Figure 14. Again, these oscillations are
recovered by PSP(2), as shown in Figure 15.
24 Wade, Khaliq, Yousuf, Vigo-Aguiar

0.8

0.6
u

0.4

0.2

0
5
4
0.2
3
0
2 −0.2
−0.4
1 −0.6
0 −0.8
t
x

Figure 13. Time evolution graph of a European call option using PSP(2).

1.4

1.2

0.8
du

0.6

0.4

0.2

0
5
4
0.2
3
0
2 −0.2
−0.4
1 −0.6
0 −0.8
t
x

Figure 14. Time evolution graph of the Delta of a European call option using (2,2)-Padé.
Smoothing Schemes for Inhomogeneous Parabolic PDEs 25

1.4

1.2

0.8
du

0.6

0.4

0.2

0
5
4
0.2
3
0
2 −0.2
−0.4
1 −0.6
0 −0.8
t
x

Figure 15. Time evolution graph of Delta of a European call option using PSP(2).

In this experiment we consider a well known problem given in a number of papers, for
example Voss and Khaliq [21] and Serbin [19]:

∂u ∂u2
= + tν (νx(1 − x) + 2t), (38)
∂t ∂t2
u(0, t) = 0, u(1, t) = 0, t > 0,
u(x, 0) = 0, 0 < x < 1.

This problem has a nontrivial forcing term, although smooth initial data. We present it at
this time to demonstrate that the advantages of the scheme presented in this article are not
only for problems with nonsmooth initial data, but also include more efficient computational
procedures and parallelizability. In addition, to show flexibility in the companion spatial
discretization, we here employ a spatial discretization with the Chebyshev spectral method.
Since the initial data is smooth, the (2,2)-Padé scheme shows fourth order convergence,
yet the IPSP(2) scheme performs about the same and also is more robust to be used with
confidence for more difficult PDE problems. Thomeé [20] has also developed a fourth order
scheme for inhomogeneous problems with smooth data using a (2,2)-Padé approximation
of the matrix exponential function. That scheme in its original form, however, requires
inverting higher order matrix polynomial functions, which present computational difficulty.
Our modified scheme using the partial fraction technique not only does not require inverting
higher order matrix polynomials but also performs computationally more efficiently. For
fourth order convergence, it is only required to solve simple Backward Euler like problems,
albeit with complex arithmetic.
26 Wade, Khaliq, Yousuf, Vigo-Aguiar

Table 5. Convergence results for the Example [19] using the (2,2)-Padé scheme. Parameter values are:
x0 = 0, xn = 1, T = 1, and ν = 3.

∆x ∆t Err. (2, 2) Ord. (2,2) Err. IPSP(2) Ord. IPSP(2)


0.010000 0.05000 2.984e-006 — 2.932e-006 —
0.005000 0.02500 1.840e-007 4.01916 1.833e-007 3.99935
0.002500 0.01250 1.147e-008 4.00379 1.146e-008 4.00037
0.001250 0.00625 7.164e-010 4.00104 7.162e-010 3.99964
0.000625 0.00313 4.478e-011 3.99975 4.478e-011 3.99930

References
[1] K. E. Atkinson, An Introduction to Numerical Analysis, Second ed., John Wiley & Sons, Inc. 1989.
[2] P. Brenner and V. Thomée, On Rational Approximations of Semigroups, SIAM J. Numer. Anal. 16
(1979), 683–694.
[3] P. Brenner, M. Crouzeix and V. Thomée, Single Step Methods for Inhomogeneous Linear Differential
Equations in Banach Space, RAIRO Anal. Numér. 16 (1982), 5–26.
[4] M. Crouzeix and V. Thomée, On the Discretization in Time of Semilinear Parabolic Equations with
Nonsmooth Initial Data, Math. Comp. 49 (1987), 359–377.
[5] M. Crouzeix, V. Thomée, and L.B. Wahlbin, Error Estimates for Spatially Discrete Approximations of
Semilinear Parabolic Equations with Initial Data of Low Regularity, Math. Comp. 53 (1989), 25–41.
[6] E. Gallopoulos and Y. Saad, On the Parallel Solution of Parabolic Equations, Preprint, CSRD Report
854 (1988), Univ. of Illinois, Urbana-Champaigne.
[7] D. J. Higham,An Introduction to Financial Option Valuation, Cambridge University Press, 2004.
[8] A.Q.M. Khaliq, E.H. Twizell and D.A. Voss, On Parallel Algorithms for Semidiscretized Parabolic Partial
Differential Equations Based on Subdiagonal Padé Approximations, Numer. Meth. PDE 9 (1993), 107–
116.
[9] A.Q.M. Khaliq and B.A. Wade, On Smoothing of the Crank-Nicolson Scheme for Nonhomogeneous
Parabolic Problems, J. Comput. Meth. Sci. & Eng. (JCMSE) 1(1) (2001), 107–124.
[10] J.D. Lambert, Numerical Methods for Ordinary Differential Systems, John-Wiley & Sons, Chichester,
2000.
[11] M. Luskin and R. Rannacher, On the Smoothing Property of the Crank-Nicolson Scheme, Appl. Anal.
14 (1982), 117–135.
[12] M. Luskin and R. Rannacher, On the Smoothing Property of the Galerkin Method for Parabolic Equa-
tions, SIAM J. Numer. Anal. 19 (1982), 93–113.
[13] S. N. Neftci, An Introduction to the Mathematics of Financial Derivatives, Academic Press, 2000.
[14] A. Pazy, Semigroups of Linear Operators and Applications to Partial Differential Equations, Appl.
Math. Ser. vol. 44, Springer-Verlag, Berlin. 1983.
[15] D. M. Pooley, K.R. Vetzal and P.A. Forsyth, Convergence Remedies for Non-Smooth Payoffs in Option
Pricing, J. Comput. Finance 6(4) (2003), 25–40.
[16] R. Rannacher, Finite Element Solution of Diffusion Problems with Irregular Data, Numer. Math. 43
(1984), 309–327.
[17] R. Rannacher, Discretization of the Heat Equation with Singular Initial Data, Zeit. Ang. Math. Meth.
(ZAMM) 62 (1982), 346– 348.
Smoothing Schemes for Inhomogeneous Parabolic PDEs 27

[18] M.F. Reusch, L. Ratzan, N. Pomphrey, and W. Park, Diagonal Padé Approximations for Initial Value
Problems, SIAM J. Sci. & Stat. Comp. 9 (1988), 829–838.
[19] S. Serbin, A Scheme for Parallelizing Certain Algorithms for the Linear Inhomogeneous Heat Equation,
SIAM J. Sci. & Stat. Comp. 13 (1992), 449–458.
[20] V. Thomée, Galerkin Finite Element Methods for Parabolic Problems, Ser. Comp. Math. 25, Springer-
Verlag, Berlin, 1997.
[21] D. Voss and A.Q.M. Khaliq, Time–Stepping Algorithms for Semidiscretized Linear Parabolic PDEss
Based on Rational Approximations with Distinct Real Poles, Adv. Comput. Math. 6 (1996), 353–363.
[22] B.A. Wade, A.Q.M. Khaliq, M. siddique, and M.Yousuf, Smoothing with Positivety–Preserving Padé
Schemes for Parabolic Problems with Nonsmooth data, Numerical Methods for Partial Differential Equa-
tions, (2005) 21(3), 553-573.
[23] P. Wilmott, S. Howison, and J. Dewynne,The Mathematics of Financial Derivatives Vol. A and B,
Cambridge University Press, 1995.
[24] Y. Yan, Smoothing Properties and Approximation of Time Derivatives for Parabolic Equations: Con-
stant Time Steps, IMA J. Numer. Anal. 23 (2003), 465–487.

You might also like