Solving nite dierence schemes arising in

trivariate option pricing1

Manfred Gilli
Department of Econometrics, University of Geneva, 1211 Geneva 4, Switzerland

[email protected]
Evis Kellezi
Department of Econometrics and FAME, University of Geneva, 1211 Geneva 4,
Switzerland

[email protected]
Giorgio Pauletto
Department of Econometrics, University of Geneva, 1211 Geneva 4, Switzerland

[email protected]

Abstract
We investigate computational and implementation issues for the valuation of options on three underlying assets, focusing on the use of the nite
dierence methods. We demonstrate that implicit methods, which have
good convergence and stability properties, can now be implemented eciently due to the recent development of techniques that allow the ecient
solution of large and sparse linear systems. In the trivariate option valuation problem, we use nonstationary iterative methods (also called Krylov
methods) for the solution of the large and sparse linear systems arising
while using implicit methods. Krylov methods are investigated both in
serial and in parallel implementations. Computational results show that
the parallel implementation is particularly ecient if a ne spatial grid is
needed.

Keywords: Multivariate option pricing, nite dierence methods, implicit

methods, Krylov methods, parallel computing.
JEL codes: C63, C88, G13

Introduction

In this paper we focus on computational and implementation issues of nite

dierence methods for the valuation of multivariate contigent claims. Examples
1 This research has its roots in a discussion with Hans Amman who provided valuable
suggestions. We would like to thank Nick Webber and two referees for helpful comments
on an earlier version of this paper. Special thanks are due to Daniel Pfenniger from the
Geneva Observatory who kindly provided the support necessary to run our experiments on
the Beowulf cluster Gravitor. Financial support from the Swiss National Science Foundation
(project 125248.97) is gratefully acknowledged.

of problems resulting in multivariate partial dierential equations in nance

include the pricing of foreign currency debts, compensation plans, risk-sharing
contracts, multifactor interest rate models to mention a few.
It is generally accepted that the dimensionality of the problem is a nontrivial
issue. Up to a dimension of three, methods like the nite dierences or the nite
elements can still be used. With a greater number of state variables, Monte
Carlo is thought to be the only way out.
For bivariate problems, nite dierence methods, both explicit and implicit, have been successfuly implemented. Financial applications are proposed
by (Dempster and Hutton, 1997), (Izvorski, 1998), (Chen et al., 1998), (Chen
et al., 2001) and (Kurpiel and Roncalli, 1999). Other methods like the nite
elements (Zvan et al., 1998) and the Fourier grid method (Engelmann and
Schwendner, 1998) have also been used. For a review of numerical algorithms
for bivariate option pricing problems see (Durrleman et al., 2000). Textbook
treatments of the multivariate option pricing with nite dierence methods can
be found in (Clewlow and Strickland, 1998), (Tavella and Randall, 2000) and
(James and Webber, 2000).
For higher dimensions, implicit schemes are preferred as they are stable unconditionally on the time discretisation step. As such, acceptable precision can
be obtained in reasonable computation time, provided that we can solve the
linear system of very large scale that arises at every time step. This was a
major disadvantage in the past, given that even for the trivariate case the size
of the system was prohibitive. Previous work has focused on the explicit and
the alternating direction implicit (ADI) methods as the implicit methods were
considered to be unfeasible. An example is given by (Ekvall, 1994), who implemented the explicit and the ADI methods on a massively parallel computer. The
idea of the ADI method, based on the operator splitting technique 2 , is to split
the computations at each time step in several substeps, the scheme remaining
implicit in only one of the three dimensions of the problem, reducing the size
of the linear system to solve and maintaining the good stability feature of the
implicit methods.
The point we want to make in this paper is that, for trivariate problems, implicit nite dierence methods can be implemented, allowing to take advantage
of their good stability and convergence features. The availability of ecient
methods for the solution of large and sparse linear systems, namely nonstationary iterative methods, allows the use of implicit nite dierence methods.
The paper is organized as follows: Section two introduces the valuation
problem of a European call on three underlying assets which leads to a partial
dierential equation with three state variables and time. In order to benchmark our results, we consider the case of an option on the maximum of three
assets, for which a closed form pricing formula exists. We describe in detail the
discretisation of the partial dierential equation, deriving the dierent solution
methods and giving their time and space complexity. The nonstationary iterative methods and in particular the biconjugate gradient stabilized algorithm
2 See

e.g. (Marchuk, 1990).

(BiCGSTAB) are presented in Section three. Computational results are given

in Section four and Section ve concludes.

Solving the fundamental PDE for European

style options on three assets with nite dierence methods

Consider a risk free bond B evolving according to the dierential equation

dB(t) = B(t) r dt,
with r the risk free instantaneous interest rate, and three risky assets with values
S1 , S2 and S3 satisfying the stochastic dierential equation
dS(t) = diag(S(t)) [ dt + dW (t) ].


W is a 3-dimensional standard Brownian motion, whereas S = S1 S2 S3


is the vector of prices, = 1 2 3 is the vector of drifts and is a
three by three matrix such that  dt represents the covariance matrix of the
vector of the instantaneous return on assets, with generic element ij i j dt.
The volatility coecient of the process for Si is denoted by i and ij is the
correlation coecient between dSi /Si and dSj /Sj .
In the absence of arbitrage opportunities, it can be shown, using Ito calculus3 , that all European-style derivatives on the three risky assets must verify
the following linear parabolic partial dierential equation (PDE)
3
3

C
C
1  2C
= rSi

ij i j + rC .
t
Si
2 i,j=1 Si Sj
i=1

(1)

The terminal condition is specic for each type of payo. For example, in the
case of a European call on the maximum of three assets with strike price E, the
terminal condition is the option payo at maturity T :


C(S1 , S2 , S3 , T ) = max{S1 (T ), S2 (T ), S3 (T )} E
.
(2)
+

As shown by (Johnson, 1987), the pricing problem for this particular payo
function has an analytical solution. For more general settings, the PDE (1)
has to be solved numerically. One of the most popular methods is the nite
dierence method. It consists in replacing the partial derivatives by numerical
approximations and then solving the resulting discretised problem.
Before discretising it, the PDE should be simplied. Introducing the changes
of variables x = log(S1 /E), y = log(S2 /E), z = log(S3 /E), = T t and
3 See

for example (Baxter and Rennie, 1996) or (Bj

ork, 1998).

C(S1 , S2 , S3 , t) = Eu(x, y, z, ), leads to the forward parabolic equation with

constant coecients
u

y2 u
x2 u
2 u
)
+ (r
)
+ (r z )
2 x
2 y
2 z
2
2
2
1 u 2 1 u 2 1 u 2
+
+

+
2 x2 x 2 y 2 y 2 z 2 z
2u
2u
2u
+
x y xy +
x z xz +
y z xz ru
xy
xz
yz
(r

with initial condition

u(x, y, z, 0) =

max{ex 1, ey 1, ez 1}

(3)


+

For notational convenience, we have indexed volatility coecients by x, y and

z.
Equation (3) can also be written in a more compact way in the following
general notation
u
= V u + (D) u ru,
(4)

where V and D

1 2

r 12 x2
xy
x
y
xz
x
z
2 x

1 2
1 2
yx y x
D
=
V =
r

yz y z

2 y
2 y
r 12 z2

zx z x

zy z y

1 2
2 z

are called respectively the velocity tensor and the diusion tensor.

Location and discretisation of the PDE

In the following we describe the numerical solution for the PDE (3) which requires the solution of a large block triangular linear system. We consider the
function u(x, y, z, ) at the discrete set of points
xi = Lx + i x

i = 0, 1, . . . , Nx

yj = Ly + j y

j = 0, 1, . . . , Ny

zk = Lz + k z

k = 0, 1, . . . , Nz

m = (m 1) 

m = 1, . . . , M

where Lx , Ly and Lz are the lower bounds of the spatial grid, x , y and z
dene the spatial grid steps,  the time step and Nx , Ny , Nz and M are the
number of spatial and time steps.
For a given time m , the spatial grid points (xi , yj , zk , m ) are contained in a
rectangular parallelepiped with edges parallel to the axes x, y and z. A particular grid point is then unambiguously dened by the triplet (i, j, k). Therefore the
4

value of the function u(x, y, z, ) at a grid point (xi , yj , zk , m ) can be denoted

by um
ijk .
Let be the set of all (Nx + 1)(Ny + 1)(Nz + 1) grid points, i.e. the set of
all triplets (i, j, k). We consider the partition
=

and

= ,

where is the set of (Nx 1)(Ny 1)(Nz 1) interior points obtained for
i = 1, . . . , Nx 1, j = 1, . . . , Ny 1 and k = 1, . . . , Nz 1 and is the set of
grid points on the boundary.
In the appendix A we explain in detail how we obtain the discretised version
of the PDE (3) in matrix notation:
m+1
um
+ (1 )P um
um+1
= P u

(5)

m
where um
, m = 1, . . . , M is the vector of interior grid points at time m, u is
the vector of all grid points and 0 1. Figure 1 illustrates the structure
of matrix P corresponding to a grid with ve points in each direction. The
columns of matrix P which correspond to grid points uijk are indicated
with a circle and those which correspond to grid point uijk are indicated
with a bullet.

...
...
...
...
...
...
k=1
k=2
k=3
k=4
.
.
.
.
.
.........................................k=0
......................................................................................................................................................................................................................................................................................................................................................................................................................................
...
... j=0 j=1 j=2 j=3 j=4 .... j=0 j=1 j=2 j=3 j=4 .... j=0 j=1 j=2 j=3 j=4 .... j=0 j=1 j=2 j=3 j=4 .... j=0 j=1 j=2 j=3 j=4 ....
...
...
...
...
...
...
.
.
.
.
.
.
...
.................................................................................................................
..................
..................
.......................................................
..................
..................
.......................................................
..................
..................
............................................................................................................................
..
..
..
..
..
..
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
....
....
....
....
....
..
................................................................................................................................................................................................................................................................................................................................................................................................................................................................

Figure 1: Pattern of matrix P for Nx = Ny = Nz = 4. Columns marked with

bullets form set and columns marked with circles form set .
The columns of matrix P are then partitioned into two matrices A and B,
such that the columns of A correspond to the indices of set and the columns
of B to the indices of set , i.e.
m
= Aum
P um
+ Bu .

Finally we can rewrite the set of equations (11) as

m+1
m
(I A)um+1
= (I + (1 )A)um
+ Bu + (1 )Bu .

(6)

The matrix on the left in Figure 2 corresponds to matrix A extracted from

matrix P given in Figure 1. The matrix on the right in the same gure shows
the pattern for a larger matrix A with Nx = Ny = Nz = 6.
..
..
..
..
k=1
k=2
k=3
...
..
..
..
................................................................................................................................................................................
...
...
...
..
... j=1 j=2 j=3 ... j=1 j=2 j=3 ... j=1 j=2 j=3 ....
.............................................................................................................................................................................
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
..
..........................................................................................................................................................................
....
....
....
....
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
..
...
..............................................................................................................................................................................
..
...
...
...
...
...
...
...
...
...
...
...
...
...
..
...
....
...
...
...
...
...
...
...
...
...
...
...
...
...
..
...
....
...
...
...
..
..
..
...
................................................................................................................................................................................

Figure 2: Pattern of matrix A for Nx = Ny = Nz = 4 and Nx = Ny = Nz = 6.

Solution methods for the discretised problem

Dierent settings for the parameter transform the set of equations (6) into
dierent solution schemes. The solution for these dierent situations is discussed
in this section.
Denoting N1 , N2 , . . . , Nd the number of steps in each direction of a d dimensional problem, the number of grid points || and the number of interior points
|| is
d
d


(Ni + 1) ,
|| =
(Ni 1)
|| =
i=1

i=1

and the number of neighbours to a given grid point is 2d2 .

Explicit method ( = 0)
If is set to zero, equation (6) becomes
m
= (I + A)um
um+1
+ Bu

where all elements on the right hand side are known at time step m . This
equation can be solved explicitly for each time step m . Its numerical stability
and convergence depend on the spectral radius of matrix I + A which has to be
inferior to unity.
When solving the system it is however more ecient to consider the form
m
= um
um+1
+ P u

where we do not need to construct matrix P . The indices and elements of

the nonzero columns corresponding to a given row are dened in the appendix
(Algorithm 4 and expressions (1315)).
The explicit method for an arbitrary number d of space variables is summarized in Algorithm 1. s is the set of all neighbours of grid point s in the
d dimensional parallelepiped that occur in the nite dierence approximations
for the partial derivatives.
Algorithm 1 Pseudo-code for the explicit method.
1: Initialize u0 , c
2: for m = 0 : M 1 do
3:
for s do
4:
Compute s
m
5:
um+1
= um
s
s + c u s
6:
end for
7: end for

Complexity
Statement 4 needs 2(d 1) operations to compute the index s and 2d2 operations to compute the set of indices s . The addition and the scalar product in
Statement 5 need 1 + 2(1 + 2d2 ) operations. We then establish that the number
of elementary operations for Algorithm 1 is given by
M (6d2 + 2d + 1) ||

(7)

and as we need two arrays um+1 and um the space complexity of the algorithm
is
2 ||.
There is a possibility of reducing the operation count given in (7) by computing rst all sets s and store them in an array of dimension (1 + 2d2 ) ||.
This reduces the factor in expression (7) from 6d2 + 2d + 1 to 4d2 + 3.
Implicit Euler method ( = 1)
If is set to one, equation (6) becomes
m+1
(I A)um+1
= um
+ Bu

and um+1
is the solution of a linear system. Algorithm 2 implements the implicit

method.
The implicit method is rst order accurate and is in this respect not superior
to the explicit method. Its advantage is however that it is stable and convergent
for all choices of space and time steps.

Algorithm 2 Pseudo-code for the implicit method.

1:
2:
3:
4:
5:
6:

Initialize u0
Compute A and B
for m = 0 : M 1 do
Compute boundary values um+1

m+1
Solve (I A)um+1
= um
+ Bu

end for

-method (0 < < 1)

If , 0 < < 1 we obtain a -weighted average of the explicit and the implicit
method. In the particular case where = 1/2 we have the Crank-Nicolson
method. Equation (6) becomes in this case
m+1
m
1
(I 12 A)um+1
= (I + 12 A)um
+ 2 B(u + u ).

The Crank-Nicolson method is second order accurate and its complexity is almost identical to the complexity of the implicit method. There are a few more
operations involved for the evaluation of the right-hand side vector of the linear
system.

Krylov methods for solving the linear system

The CFL condition for the explicit method implies a prohibitively large number
of time steps.4 For implicit methods the CFL condition is satised regardless of
the number of time steps. However, the implicit method requires the solution
of a linear system of size || at every time step. As the system becomes large,
it is crucial to choose an ecient solution method.
The structure of the matrix of the linear system is shown in Figure 2. This
matrix can either be considered as a banded matrix with bandwidth (Nx 1)Ny +
1 or as a block tridiagonal matrix with blocksize (Nx 1) (Nx 1). For the
solution of the linear system, one may think to apply an LU decomposition which
exploits the particular structure of the matrix. It appears however that direct
methods are unfeasible for systems of the size considered in our application.5
Iterative methods oer a possibility to overcome the diculties of direct
methods. We experimented the nonstationary iterative methods also called
Krylov subspace methods.6
4 The CFL condition is named after Courant, Friedrich and Lewy (see Courant et al. (1928)
or Courant et al. (1967)). This is only a necessary condition for the convergence of nite
dierence methods: If it is satised, the method might converge to the exact solution.
5 Gilli and Pauletto (1997) discuss this problem in the case of the solution of large economic
models.
6 A presentation of these techniques can be found in Barrett et al. (1994), Freund et al.
(1992), Axelsson (1994), Kelley (1995) and Saad (1996). For applications in economics see
Gilli and Pauletto (1998), Pauletto and Gilli (2000) and Mrkaic (2001).

Contrary to stationary methods, such as the Jacobi or the Gauss-Seidel

methods, these techniques use information that changes from iteration to iteration. For a linear system Ax = b, Krylov methods compute the ith iterate
as
i = 1, 2, . . . .
x(i) = x(i1) + d(i)
The attractive characteristic of the nonstationary iterative methods is that the
ith update d(i) can be easily performed on sparse matrices and requires little
storage. The convergence speed is generally also better than for stationary
iterative methods.
For the application presented in the paper, we used the biconjugate gradient
stabilized method (BiCGSTAB) introduced by van der Vorst (1992). Algorithm 3 represents the pseudo-code for solving a linear system Ax = b with this
method.7
Algorithm 3 Pseudo-code for BiCGSTAB.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:

Initialize x(0)
Compute r(0) = b Ax(0) , 0 = 1, 1 = r(0) r(0) , = 1, = 1, p = 0, v = 0
for k = 1, 2, . . . until convergence do
= (k /k1 )(/)
p = r + (p v)
v = Ap
= k /(r(0) v)
s = r v
t = As
= (t s)/(t t)
x(k) = x(k1) + p + s
r = s t
k+1 = r(0) t
end for

Complexity
If BiCGSTAB is used with no preconditioning, which is the case in our application, the algorithm computes in statement 2 a matrix-vector product and a
norm, and, for each iteration, 4 inner products, 6 vector updates (also called
saxpy operations) and 2 matrix-vector products. The number of elementary
operations corresponding to these computations is 2 (2d2 + 1) || for the matrixvector product, 2 || for one saxpy as well as for one inner product, and 4 ||
for the computation of the norm.
The overall complexity of the implicit method for d = 3 and M time steps
is then
M || (42 + 96k)
7 See

also (Barrett et al., 1994).

where k is the number of iterations for BiCGSTAB to converge.8

Algorithm 3 needs memory to store matrix A which has nnz = (2d2 + 1) ||
nonzero entries and 7 vectors of size corresponding to the order of matrix A.9
Considering a storage scheme for the sparse matrix A that needs 12 nnz + 4 ||
bytes we need for d = 3 a total of
12 38 || + 4 || + 7 8 || = 516 ||
bytes of memory where integers occupy 4 bytes and reals 8 bytes.
In order to illustrate the fact that sparse direct methods do not constitute
an alternative to the nonstationary iterative methods, we compare in Figure 3
the space and time complexity of a sparse LU with BiCGSTAB as a function of
the number of grid points N in each of the three directions.

25
20
15

.
.
.
.
.
.
.
.
.
.
.
.
.
.

Gbytes

Gops / time step

LU
BiCGSTAB

LU
BiCGSTAB

10

..

0
0

. .
. .
. .
. .
.
.
. . .
. . . . . . . . .

50

100

150

.
.

.
.

.
. .
.
. .

.
.
.

0
0
50
100

. .

200

150

200

Figure 3: Comparison of computational and space complexity of a sparse LU

against BiCGSTAB as a function of grid points in each of the three directions.

It clearly appears that the sparse direct method reaches its limits for a
number of grid points below 30, whereas for BiCGSTAB we can easily reach
grid sizes beyond 200, in particular with parallel machines where memory is less
of an issue.

Computational results

In order to benchmark our results, we compute the price of an instrument

for which an analytical solution exists. We report the values of a call on the
maximum of three assets S1 , S2 and S3 with strike price E = 10, maturity
T = 1, risk free rate of return r = 0.1, volatility and correlation coecients
among assets 1 = 2 = 3 = 0.2, 12 = 23 = 13 = 0.5 and initial values
8 This operation count neglects the computation of the right-hand side in statement 5 of
Algorithm 2.
9 It is possible to implement a matrix free version of the algorithm.

10

S1 (0) = S2 (0) = S3 (0) = 10. The analytical solution of the call corresponding
to this set of parameters is 2.2672.10
In the following we present computational results obtained both with a serial
and a parallel implementation.
Serial implementation
We implemented the explicit method described in Algorithm 1, the implicit
method described in Algorithm 2 and the BiCGSTAB method as described in
Algorithm 3 in a MATLAB 5.x programming environment.11 Computations
have been executed on a 500 MHz Pentium III PC with 512 Mbytes of memory.
With 71 space grid points in each of the three directions, the implicit method
needs 516 (71 2)3 170 Mbytes of memory for data storage. This is the
largest problem instance we can solve with MATLAB on our machine.
Parallel implementation
As commented in the previous paragraph, the space complexity of the implicit
method limits the size of the problem that can be solved in a serial environment.
This is a situation where Krylov methods become attractive as they are also
well suited for parallel implementation, allowing to tackle larger grid sizes. For
our application we employed the software PETSc 2.1.0 (Portable and Extensible
Toolkit for Scientic Computing) developed and maintained at the Mathematics
and Computer Science Division of the Argonne National Laboratory, see (Balay
et al., 2001). PETSc is a set of routines designed to solve large-scale computational problems in a parallel environment. The software is written in C and can
be linked to C, C++ or Fortran programs.
We used a computing platform constituted by 20 biprocessor PC LINUX
cluster. The individual machines are Pentium III running at 500 MHz and are
equipped with 384 Mbytes of memory each. These 20 PCs are part of a larger
cluster of 64 PCs, where the other 44 are Pentium II running at 400 MHz. We
chose to use the rst twenty in order to preserve a homogenous computing
environment.
On this cluster we solved with the implicit method problems with a grid
size of N = 200 in each of the three directions corresponding to linear systems
with 7 880 599 equations. With 37 processors, we achieved a computing speed
of about 0.7 Gop/s.
10 The analytical formula for a call on the maximum of three assets is derived in (Johnson,
1987). The evaluation of this formula requires the numerical integration of the trivariate
normal density function. We used the algorithms proposed by (Genz, 2001; Genz, 1993) and
MULNOR of (Schervish, 1984).
11 The MATLAB les containing the code can be downloaded from the URL www.unige.ch/
ses/metri/gilli/.

11

Scalability analysis
To evaluate the performance of the parallel algorithm on our parallel machine,
we performed an empirical scalability analysis. Let T1 be the time of a serial
execution for a xed size of our problem. As a measure for the problem size
we take the size of the linear system to be solved which is equal to the number
of interior grid points || = (N 1)3 , when we set Nx = Ny = Nz = N . Let
Tp be the time taken by the parallel algorithm to solve the same problem on
p identical processors. Then the ratio Sp = T1 /Tp is called the speedup and
it indicates what is the relative benet of solving a problem in parallel over a
sequential implementation.
As processors spend time in communicating, they cannot dedicate all the
time to computations of the algorithm. Eciency is a measure of the fraction
of time during which processors are usefully employed; it is dened as the ratio
of speedup to the number of processors, Ep = Sp /p. Figure 4 shows the speedup
and the eciency as functions of the number of processors for dierent problem
sizes ranging from N = 100 to N = 200 and which correspond to systems of
approximatively one to eight million equations.
The speedup curves decrease after an optimum number of processors. However this optimum lies beyond the number of processors available for our application. We observe that larger instances of the problem yield higher speedup
and eciency for the same number of processors.
25

0.65
Efficiency

Speedup

20

0.7

N=100
N=120
N=140
N=160
N=180
N=200

15
10

0.6
0.55

N=100
N=120
N=140
N=160
N=180
N=200

0.5
0.45

5
0.4

0
0

10
20
30
Number of processors

40

0.35
0

10
20
30
Number of processors

40

Figure 4: Speedup and eciency versus the number of processors for solving
the implicit scheme in parallel.

To investigate whether it is possible to keep the eciency constant as we

increase the size of the problem together with the number of processors used,
we construct an isoeciency curve by taking the dierent combinations number
of processors versus problem size that yield the same eciency level. Figure 5
shows the isoeciency curve of level E = 0.6. This empirical relationship is
roughly linear suggesting that our parallel implementation is scalable.

12

Problem size (N1)

x 10

8
7
6
5
4
3
2
15

20
25
30
35
Number of processors

40

Figure 5: Isoeciency curve.

Conclusion

Based on the computational results performed in two dierent computing environments, we conclude that implicit nite dierence methods can be eciently
used for the valuation of options on three underlying assets. The availability
of nonstationary iterative methods and the fact that they are well suited for
parallelization makes the solution of very large linear systems possible. In our
experiments, the size of the system that we solve in parallel is approximatively
eight million. We think that a matrix free implementation of the algorithm
would enhance the performance and, with faster processors and more memory,
we can go even further, allowing for ner space grids.
In the serial case, the maximum grid size that we can solve in a standard
PC in Matlab environment is of about 70 in each direction.
We have used the Krylov methods without preconditioning. It would be
worthwhile studying whether preconditioning would improve the performance
of the algorithm. The inuence of the diusion and velocity tensors in the
convergence speed of the Krylov methods can also be the object of further
research.

Discretisation of the PDE

Replacing all partial derivatives in equation (3) by nite dierence approximations we can write the following approximation to PDE (3) which holds for all
grid points um
ijk :
m

= vx x + vy y + vz z + xx dxx + yy dyy + zz dzz

m

+ xy dxy + xz dxz + yz dyz rum

ijk .

13

(8)

The 9 approximations q , q Q = {x, y, z, xx, yy, zz, xy, xz, yz} for the partial
derivatives with respect to the space variables x, y and z are dened as convex
linear combinations of central nite dierence approximations at time steps
m + 1 and m
m
q = q qm+1 + (1 q )qm .
(9)
The temporal weighting factors q , q Q take their values in [0, 1] and determine
the time at which the partial derivatives with respect to the state variables are
evaluated. The central nite dierence approximations qm , q Q are:
xm

ym

zm

m
xy

m
xz

m
yz

m
m
m
um
um
i+1,j,k ui1,j,k
i+1,j,k 2uijk + ui1,j,k
m
xx
=
2
2 x
x
m
m
m
m

2u
um
u
i,j+1,k
i,j1,k
i,j+1,k
ijk + ui,j1,k
m
yy
=
2
2 y
y
m
m
m
m
ui,j,k+1 ui,j,k1
ui,j,k+1 2um
ijk + ui,j,k1
m
zz
=
2 z
2z
m
m
m
m
ui+1,j+1,k ui1,j+1,k ui+1,j1,k + ui1,j1,k
4 x y
m
m
m

u
um
i+1,j,k+1
i1,j,k+1 ui+1,j,k1 + ui1,j,k1
4 x z
m
m
m
um
i,j+1,k+1 ui,j1,k+1 ui,j+1,k1 + ui,j1,k1
4 y z

The approximation for the partial derivative with respect to time m is

dened as a linear combination between a backward approximation at time
m + 1 and a forward approximation at time m
m

= m
=

m
um+1
ijk uijk

um+1
ijk


um
ijk

+ (1 m )

m
um+1
ijk uijk


(10)

and is independent from m .

Assigning the same value to all q , q Q and substituting the denitions (9) and (10) for the approximations for the space respectively time derivatives into equation (8) and collecting terms, we obtain
m+1
m
m
um+1
ijk uijk = c uijk + (1 ) c uijk

um
ijk

(11)

where um
ijk denotes a one-dimensional array. The set of indices ijk corresponds
to the points in the grid which are neighbours of um
ijk and which occur in the
nite dierence approximations for the partial derivatives. The product c um
ijk ,

14

where c is a row vector, can be developed as

c um
ijk

m
m
m
= c1 um
i,j1,k1 + c2 ui1,j,k1 + d6 ui,j,k1 c2 ui+1,j,k1
m
m
m
c1 um
i,j+1,k1 + c3 ui1,j1,k + d4 ui,j1,k c3 ui+1,j1,k
m
m
+d2 um
(12)
i1,j,k + d1 ui,j,k + d3 ui+1,j,k
m
m
m
m
c3 ui1,j+1,k + d5 ui,j+1,k + c3 ui+1,j+1,k c1 ui,j1,k+1
m
m
m
c2 um
i1,j,k+1 + d7 ui,j,k+1 + c2 ui+1,j,k+1 + c1 ui,j+1,k+1

where the coecients are

c1 =
c4 =
c7 =

 Dxy
4x y
 Vx
2x
 Dxx
2x

c2 =
c5 =
c8 =

 Dxz
4x z
 Vy
2y
 Dyy
2y

c3 =
c6 =
c9 =

d1 = r  2c7 2c8 2c9

d4 = c8 c5
d5 = c8 + c5

d2 = c9 c6
d6 = c7 c4

 Dyz
4y z
 Vz
2z
 Dzz
2z

d3 = c9 + c6
d7 = c7 + c4

and where Vi are the elements of the velocity tensor and Dij those of the diusion
tensor in PDE (4).
Figure 6 reproduces the fragment of the grid containing point uijk (bullet)
and the set of neighbours ijk (circles).
(i,j+1,k+1)
...
. ......................................... . ......................................... .
......
...... ..
..... ...
..... ...
..... ...
.
..........
.
.
.
..
......
.
......
...
.
......
......
......
...
....
...
..
.....
.....
.....
...
....
........................................... ........................................... ......
.
.
..
...
.
.
.
.
.
... ...
.... ...
.
.... ...
.
.
.
.
.
.
.
.
.
.
.
...
...
.
.
.
.
.
...
... .. ..
.
.... .. ...
.
.
.
.
.
.
.
.
.
.
.
.
...
.
.
.
.
.
.
..
..... . .................................... . .............................. ..
.
.
.
...
.... ...........
.
.
.
(i+1,j+1,k)
.
.
........................................ ............................................... ..... ...... . ....
.
.
.
... ...
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
. .
. .
..
...
.... .... .........
....
....
... ... ..............
... .... ...........
.. .........
...
...
... .. .......
... . .......
...
................................ ..................................
...
.
...
.... .
.... . ...
...
...
...... .. ..
...... .
..... ... ...
...
...
..... .. ..
..... .. .. ..
.....
.
... .... ......... .. ... .... ........... .. ....
... ........
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... .. ......
.
.
.
.. ......
.
.
..................... ...
... ...
....
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
(i1,j1,k) .
.
.
.
.
...
.
.
.
.
.... ... .........
.... ... .........
....
.
..
..
.. .........
.. .........
... ...........
...
. ......
. ....
. ..
...
.............................. .................................. ..
...
.
... (i+1,j,k1)
.... ....
...
.
.
.
.
.
.
.
.
.
.
.
.
...
.
.
.
....
....
.....
...
....
....
.....
.....
......
... .........
.. .......
.. ......
..................................... . ........................................ . ...........................................................
.....
.
.
.
................
(i,j1,k1)
........
.....
.
.
.
.
..

Figure 6: Grid fragment with node uijk (bullet) and neighbours uijk (circles).
In order to write the equations (12) for all the grid points uijk in matrix
form, we index the space grid points uijk with a one-dimensional array s. The
index sijk corresponding to a grid point uijk is dened as
sijk = i + j(Nx + 1) + k(Nx + 1)(Ny + 1)

i=0,1,...,Nx
j=0,1,...,Ny
k=0,1,...,Nz .

(13)

We now construct a matrix P where the rows contain the coecients of vector
c. From Figure 6 and from expression (13) we see that the vector o of relative
15

Algorithm 4 Relative column indices of nonzero elements in a row of P .

1:  = 0
2: for k = 1 : 1 do
3:
for j = 1 : 1 do
4:
for i = 1 : 1 do
5:
if i = 0 j = 0 k = 0 then
6:
=+1
7:
o() = i + j (Nx + 1) + k (Nx + 1)(Ny + 1)
8:
end if
9:
end for
10:
end for
11: end for

positions of the |ijk | = 18 neighbors of a grid point uijk can be computed with
the algorithm 4.
The row indices of matrix P corresponding to the points uijk are dened
by
rijk = i + (j 1)(Nx 1) + (k 1)(Nx 1)(Ny 1) ,

i=1,...,Nx 1
j=1,...,Ny 1
k=1,...,Nz 1

(14)

and a row rijk of matrix P , corresponding to the equation dening the grid
point uijk has the form
(15)
M (rijk , sijk + o) = c
where


c = c1 c2 d6 c2 c1 c3 d4 c3 d2 d1 d3 c3 d5 c3 c1 c2 d7 c2 c1
is dened in (12). Equation (11) in matrix form becomes
m+1
um+1
um
+ (1 )P um
.
= P u

References
Axelsson, O. (1994). Iterative Solution Methods. Oxford University Press. Oxford, UK.
Balay, S., W. D. Gropp, L. Curfman McInnes and B. F. Smith (2001). PETSc
Users Manual. Technical Report ANL 95/11 - Revision 2.1.0. Argonne National Laboratory. http://www.mcs.anl.gov/petsc/petsc.html.
Barrett, R., M. Berry, T. F. Chan, J. Demmel, J. Donato, J. Dongarra, V. Eijkhout, R. Pozo, C. Romine and H. van der Vorst (1994). Templates for the
Solution of Linear Systems: Building Blocks for Iterative Methods. SIAM.
Philadelphia, PA.
Baxter, M. and A. Rennie (1996). Financial Calculus: An Introduction to
Derivative Pricing. Cambridge University Press.
16

Bjork, T. (1998). Arbitrage Theory in Continuous Time. Oxford University

Press.
Chen, A. H., J. A. Conover and J. W. Kensinger (1998). Valuing exible manufacturing facilities as options. Quarterly Jouranl of Economics and Finance
38, 651674.
Chen, A. H., J. A. Conover and J. W. Kensinger (2001). Evaluating virtual
options. Mimeo.
Clewlow, L. and C. Strickland (1998). Implementing Derivatives Models. Wiley
series in nancial engineering. Wiley.

Courant, R., K.O. Friedrichs and H. Lewy (1928). Uber

die partiellen Diferenzengleichungen der mathematischen Physic. Mathematische Annalen
100, 3274.
Courant, R., K.O. Friedrichs and H. Lewy (1967). On the partial dierence
equations of mathematical physics. IBM Journal 11, 215234.
Dempster, M. and J. Hutton (1997). Numerical valuation of cross-currency
swaps and swaptions. In: Mathematics of Derivative Securities (M. Dempster and S. R. Pliska, Ed.). Cambridge University Press. Cambridge.
pp. 473503.
Durrleman, V., A. Kurpiel, G. Riboulet and T. Roncalli (2000). Topics on twostate option pricing. Mimeo. http://gro.creditlyonnais.fr/content/
wp/pricing2d.pdf.
Ekvall, N. (1994). Experiences in the Pricing of Trivariate Contingent Claims
with Finite Dierence Methods on a Massively Parallel Computer. Computational Economics 7(2), 6372.
Engelmann, B. and P. Schwendner (1998). The pricing of multi-asset options
using a Fourier grid method. Journal of Computational Finance 1(4), 63
72.
Freund, R. W., G. H. Golub and N. M. Nachtigal (1992). Iterative solution of
linear systems. Acta Numerica pp. 57100.
Genz, A. (1993). Comparison of methods for the computation of multivariate
normal probabilities. Computing Science and Statistics 1(25), 400405.
Genz, A. (2001). Numerical computation of bivariate and trivariate normal
probabilities. Mimeo, http://www.math.wsu.edu/math/faculty/genz/
homepage.
Gilli, M. and G. Pauletto (1997). Sparse direct methods for model simulation.
Journal of Economic Dynamics and Control 21(6), 10931111.

17

Gilli, M. and G. Pauletto (1998). Krylov methods for solving models with
forward-looking variables. Journal of Economic Dynamics and Control
22(6), 12751289.
Izvorski, I. (1998). A nonuniform grid method for solving PDEs. Journal of
Economic Dynamics and Control 22(6), 14451452.
James, J. and N. Webber (2000). Interest Rate Modelling. Wiley. Chichester,
UK.
Johnson, H. (1987). Options on the maximum or the minimum of several assets.
Journal of Financial and Quantitative Analysis 22, 277283.
Kelley, C. T. (1995). Iterative Methods for Linear and Nonlinear Equations.
SIAM. Philadelphia, PA.
Kurpiel, A. and T. Roncalli (1999). Hopscotch methods for two-state nancial
models. Journal of Computational Finance 3(2), 5389.
Marchuk, G.I. (1990). Splitting and alternating direction methods. In: Handbook
of Numerical Analysis (P.G. Ciarlet and J.-L. Lions, Ed.). Vol. I. pp. 197
464. North-Holland. Amsterdam.
Mrkaic, M. (2001). Policy iteration accelerated with Krylov methods. Journal
of Economic Dynamics and Control. (Forthcoming).
Pauletto, G. and M. Gilli (2000). Parallel Krylov Methods for Econometric
Model Simulation. Computational Economics 16(12), 173186.
Saad, Y. (1996). Iterative Methods for Sparse Linear Systems. PWS Publishing
Company. MA.
Schervish, M. (1984). Multivariate Normal Probabilities with Error Bound. Applied Statistics 33, 8187.
Tavella, D. and C. Randall (2000). Pricing Financial Instruments : The Finite
Dierence Method. Wiley series in Financial Engineering. Wiley. New York.
van der Vorst, H. A. (1992). Bi-CGSTAB: A Fast and Smoothly Converging
Variant of BI-CG for the Solution of Nonsymmetric Linear Systems. SIAM
Journal for Scientic and Statistical Computing 13(2), 631644.
Zvan, R., P.A. Forsyth and K.R. Vetzal (1998). A general nite element approach for PDE option pricing models. Mimeo.

18