INTERNATIONAL JOURNAL OF c 2010 Institute for Scientific

NUMERICAL ANALYSIS AND MODELING Computing and Information

Volume 7, Number 2, Pages 303–320

ADI FINITE DIFFERENCE SCHEMES FOR OPTION PRICING

IN THE HESTON MODEL WITH CORRELATION

K. J. IN ’T HOUT AND S. FOULON

Abstract. This paper deals with the numerical solution of the Heston par-
tial differential equation (PDE) that plays an important role in financial op-
tion pricing theory, Heston (1993). A feature of this time-dependent, two-
dimensional convection-diffusion-reaction equation is the presence of a mixed
spatial-derivative term, which stems from the correlation between the two un-
derlying stochastic processes for the asset price and its variance.
Semi-discretization of the Heston PDE, using finite difference schemes on
non-uniform grids, gives rise to large systems of stiff ordinary differential equa-
tions. For the effective numerical solution of these systems, standard implicit
time-stepping methods are often not suitable anymore, and tailored time-
discretization methods are required. In the present paper, we investigate four
splitting schemes of the Alternating Direction Implicit (ADI) type: the Douglas
scheme, the Craig–Sneyd scheme, the Modified Craig–Sneyd scheme, and the
Hundsdorfer–Verwer scheme, each of which contains a free parameter.
ADI schemes were not originally developed to deal with mixed spatial-
derivative terms. Accordingly, we first discuss the adaptation of the above
four ADI schemes to the Heston PDE. Subsequently, we present various nu-
merical examples with realistic data sets from the literature, where we consider
European call options as well as down-and-out barrier options. Combined with
ample theoretical stability results for ADI schemes that have recently been ob-
tained in In ’t Hout & Welfert (2007, 2009) we arrive at three ADI schemes
that all prove to be very effective in the numerical solution of the Heston PDE
with a mixed derivative term. It is expected that these schemes will be useful
also for general two-dimensional convection-diffusion-reaction equations with
mixed derivative terms.

Key Words. Initial-boundary value problems, convection-diffusion equations,

mixed derivatives, Heston model, option pricing, method-of-lines, finite differ-
ence methods, ADI splitting schemes.

1. Introduction
In the Heston model, values of options are given by a time-dependent partial
differential equation (PDE) that is supplemented with initial and boundary condi-
tions [7, 14, 22, 24]. The Heston PDE constitutes an important two-dimensional
extension to the celebrated, one-dimensional, Black–Scholes PDE. Contrary to the

Received by the editors March 8, 2009 and, in revised form, September 4, 2009.
2000 Mathematics Subject Classification. 65L05, 65L20, 65M06, 65M12, 65M20.
This work is supported financially by the Research Foundation–Flanders, FWO contract no.
G.0125.08.
303
304 K. J. IN ’T HOUT AND S. FOULON

Black–Scholes model, however, to date in the Heston model no closed-form ana-

lytical formulas have been found for any but the simplest options, and therefore
numerical techniques are applied.
A well-known and versatile strategy for the numerical solution of initial-boundary
value problems for multi-dimensional PDEs is the method-of-lines approach. In this
approach, the PDE is first discretized in the spatial variables, yielding large systems
of stiff ordinary differential equations. These, so-called, semi-discrete systems are
subsequently solved by applying a suitable numerical time-stepping method. Due to
the large size and the multi-dimensional structure of the obtained semi-discrete sys-
tems, standard time-stepping methods, such as the popular Crank–Nicolson scheme
(trapezoidal rule), are often not effective anymore, and tailored time discretization
methods are required.
For the numerical solution of the semi-discrete Heston PDE we shall study in
this paper splitting schemes of the Alternating Direction Implicit (ADI) type. In
the past decades, ADI schemes have been successful already in many application
areas. A main and distinctive feature of the Heston PDE, however, is the presence
of a mixed spatial-derivative term, stemming from the correlation between the two
underlying stochastic processes for the asset price and its variance. It is well-
known that ADI schemes were not originally developed to deal with such terms.
In the present paper, we will investigate the adaptation of several important ADI
schemes to the numerical solution of the Heston PDE with arbitrary correlation
factor ρ ∈ [−1, 1]. As test cases we will consider European call options and down-
and-out barrier options. Through various numerical examples with realistic data
sets from the literature, combined with ample theoretical stability results that have
recently been obtained, we arrive at three ADI schemes that all prove to be very
effective in the numerical solution of the Heston PDE with a mixed derivative term.
It is expected that these schemes will be useful also for general two-dimensional
convection-diffusion-reaction equations with mixed derivative terms.
An outline of our paper is as follows.
Section 2 discusses the Heston PDE and its numerical discretization. In Section
2.1 we formulate the Heston PDE together with initial and boundary conditions for
European call options. In Section 2.2 we describe a finite difference discretization
of the Heston PDE. A non-uniform spatial grid is used to capture the important
region around the strike. In Section 2.3 we formulate the ADI type schemes under
consideration in this paper for the semi-discrete Heston PDE with a mixed deriva-
tive term: the Douglas scheme, the Craig–Sneyd scheme, the Modified Craig–Sneyd
scheme, and the Hundsdorfer–Verwer scheme. Each of these contains a free param-
eter θ. We discuss the different origins of the four schemes and review theoretical
stability results that were recently obtained in [9, 10] concerning their application
to multi-dimensional convection-diffusion equations with mixed derivative terms.
Section 3 contains extensive numerical experiments. In Section 3.1 we study
the accuracy of our finite difference discretization in various examples of parameter
sets for the Heston model obtained from the literature. Here the availability of
Heston’s analytical pricing formula for European call options makes an actual com-
putation of the global spatial errors possible. In Section 3.2 we perform numerical
experiments with all the ADI schemes above, where we analyze the behavior of the
global temporal errors for each example introduced in Section 3.1. As an alterna-
tive method, we also consider a Runge–Kutta–Chebyshev scheme. In Section 3.3
we discuss numerical experiments for down-and-out call options. For these exotic
 ADI FINITE DIFFERENCE SCHEMES FOR OPTION PRICING 305

options a closed-form analytical pricing formula has only been obtained [14] in the
literature if the correlation ρ = 0.
Section 4 summarizes our conclusions concerning the four ADI schemes in the
numerical solution of the Heston PDE with a mixed derivative term. Subsequently,
several issues for future research are discussed.

2. Numerical discretization of the Heston PDE

2.1. The Heston PDE. Let u(s, v, t) denote the price of a European option if at
time T − t the underlying asset price equals s and its variance equals v, where T is
the given maturity time of the option. Heston’s stochastic volatility model implies
[7, 14, 22] that u satisfies1 the parabolic PDE

∂u ∂2u ∂2u ∂2u ∂u ∂u

(2.1) = 12 s2 v 2 + ρσsv + 12 σ 2 v 2 + (rd − rf )s + κ(η − v) − rd u
∂t ∂s ∂s∂v ∂v ∂s ∂v
for 0 ≤ t ≤ T , s > 0, v > 0. The Heston PDE (2.1) can be viewed as a time-
dependent convection-diffusion-reaction equation, on an unbounded two-dimensional
spatial domain. The parameter κ > 0 is the mean-reversion rate, η > 0 is the long-
term mean, σ > 0 is the volatility-of-variance, ρ ∈ [−1, 1] is the correlation between
the two underlying Brownian motions, and rd , rf denote the domestic and foreign
interest rates, respectively. In this paper we always assume that 2κη > σ 2 , which
is known as the Feller condition.
For a European call option, the payoff yields the initial condition

(2.2) u(s, v, 0) = max(0, s − K)

where K > 0 is the given strike price of the option. Further, a boundary condition
at s = 0 holds,

(2.3) u(0, v, t) = 0 (0 ≤ t ≤ T ).

At the boundary v = 0 no condition is specified. From the assumption 2κη > σ 2 it

follows that this is an outflow boundary.
As a preliminary step towards the numerical solution of the initial-boundary
value problem for the Heston PDE, the spatial domain is restricted to a bounded
set [0, S] × [0, V ] with fixed values S, V chosen sufficiently large. The following
additional conditions at s = S and v = V are imposed for a European call option,
cf. [7, 27]:

∂u
(2.4) (S, v, t) = e−rf t (0 ≤ t ≤ T ),
∂s

(2.5) u(s, V, t) = se−rf t (0 ≤ t ≤ T ).

Clearly, the boundary conditions (2.3), (2.5) are of Dirichlet type, whereas (2.4)
is of Neumann type. We take S = 8K and V = 5. This yields a negligible modeling
error with respect to (2.1)–(2.3) on the unbounded domain for a wide range of
parameter values.

1We assume w.l.o.g. that the market price of volatility risk is equal to zero.
306 K. J. IN ’T HOUT AND S. FOULON

2.2. Space discretization: finite difference schemes. For the initial-boundary

value problem (2.1)–(2.5) we perform a spatial discretization on a Cartesian grid by
finite difference (FD) schemes. Here we apply non-uniform meshes in both the s-
and v-directions such that relatively many mesh points lie in the neighborhood of
s = K and v = 0, respectively. The application of such non-uniform meshes greatly
improves the accuracy of the FD discretization compared to using uniform meshes.
This is related to the facts that the initial function (2.2) possesses a discontinuity
in its first derivative at s = K and that for v ≈ 0 the Heston PDE is convection-
dominated. It is also natural to have many grid points near the point (s, v) = (K, 0)
as in practice this is the region in the (s, v)-domain where one wishes to obtain op-
tion prices. The type of non-uniform meshes that we employ has recently been
considered e.g. by Tavella & Randall [24] and Kluge [13].
We first define the mesh in the s-direction. Let integer m1 ≥ 1 and constant
c > 0. Let equidistant points ξ0 < ξ1 < . . . < ξm1 be given by
ξi = sinh−1 (−K/c) + i · ∆ξ (0 ≤ i ≤ m1 )
with
1
∆ξ = [sinh−1 ((S − K)/c) − sinh−1 (−K/c)].
m1
Then a non-uniform mesh 0 = s0 < s1 < . . . < sm1 = S is defined through the
transformation
(2.6) si = K + c sinh(ξi ) (0 ≤ i ≤ m1 ).
This mesh is smooth in the sense that there exist real constants C0 , C1 , C2 > 0 such
that the mesh widths ∆si = si − si−1 satisfy
2
(2.7) C0 ∆ξ ≤ ∆si ≤ C1 ∆ξ and |∆si+1 − ∆si | ≤ C2 (∆ξ)
uniformly in i and m1 . The parameter c controls the fraction of mesh points si
that lie in the neighborhood of the strike K. In particular,
∆si ≈ c ∆ξ whenever si ≈ K.
In our numerical experiments we have taken c = K/5.
We define a non-uniform mesh in the v-direction analogously. Let integer m2 ≥ 1
and constant d > 0. Consider equidistant points given by ηj = j · ∆η for j =
0, 1, . . . , m2 with
1
∆η = sinh−1 (V /d).
m2
Then we define a mesh 0 = v0 < v1 < . . . < vm2 = V through
(2.8) vj = d sinh(ηj ) (0 ≤ j ≤ m2 )
and write ∆vj = vj − vj−1 . Also the mesh (2.8) is smooth. The parameter d
controls the fraction of mesh points vj that lie near v = 0. It holds that
∆vj ≈ d ∆η whenever vj ≈ 0.
In our experiments we have taken d = V /500.
As an illustration, Figure 1 displays the spatial grid defined by (2.6), (2.8) for
the (small) sample values m1 = 30, m2 = 15 where K = 100. Clearly, there are
many grid lines near s = K and v = 0.
We subsequently formulate the FD schemes that we use. Let f : R → R be
any given function, let x0 < x1 < x2 < . . . < xm be any given mesh points and
∆xi = xi − xi−1 . To approximate the first derivative f ′ (xi ), we consider three FD
schemes:
 ADI FINITE DIFFERENCE SCHEMES FOR OPTION PRICING 307

3
V

0
0 100 200 300 400 500 600 700 800
S

Figure 1. Sample grid defined by (2.6), (2.8) for m1 = 30, m2 =

15, K = 100.

(2.9a) f ′ (xi ) ≈ αi,−2 f (xi−2 ) + αi,−1 f (xi−1 ) + αi,0 f (xi ) ,

(2.9b) f ′ (xi ) ≈ βi,−1 f (xi−1 ) + βi,0 f (xi ) + βi,1 f (xi+1 ) ,
(2.9c) f ′ (xi ) ≈ γi,0 f (xi ) + γi,1 f (xi+1 ) + γi,2 f (xi+2 ) ,
with coefficients given by
αi,−2 = ∆xi −∆xi−1 −∆xi
∆xi−1 (∆xi−1 +∆xi ) , αi,−1 = ∆xi−1 ∆xi , αi,0 = ∆x∆x i−1 +2∆xi
i (∆xi−1 +∆xi )
,
−∆xi+1 ∆xi+1 −∆xi ∆xi
βi,−1 = ∆xi (∆xi +∆xi+1 ) , βi,0 = ∆xi ∆xi+1 , βi,1 = ∆xi+1 (∆xi +∆xi+1 ) ,
γi,0 = ∆x−2∆x i+1 −∆xi+2
i+1 (∆xi+1 +∆xi+2 )
, γi,1 = ∆xi+1 +∆xi+2 −∆xi+1
∆xi+1 ∆xi+2 , γi,2 = ∆xi+2 (∆xi+1 +∆xi+2 ) .

To approximate the second derivative f ′′ (xi ), we deal with the FD scheme

(2.10) f ′′ (xi ) ≈ δi,−1 f (xi−1 ) + δi,0 f (xi ) + δi,1 f (xi+1 ) ,
where
2 −2 2
δi,−1 = ∆xi (∆xi +∆xi+1 ) , δi,0 = ∆xi ∆xi+1 , δi,1 = ∆xi+1 (∆xi +∆xi+1 ) .

Next, assume f : R2 → R is a function of two variables (x, y). We consider

approximating the mixed derivative fxy (x, y). Let xi , ∆xi be as above, let mesh
points y0 < y1 < y2 < . . . < yn in the y-direction be given, and write ∆y j =
yj − yj−1 . Denote by β̂i,k the coefficients analogous to βi,k in (2.9b), but then
relevant to the y-direction. For the discretization of the mixed derivative we use
the FD scheme
X1
∂2f
(2.11) (xi , yj ) ≈ βi,k β̂j,l f (xi+k , yj+l ) .
∂x∂y
k, l =−1

Clearly, the approximation (2.11) can be viewed as obtained by application of (2.9b)

successively in the x- and y-directions.
308 K. J. IN ’T HOUT AND S. FOULON

The FD schemes above are all well-known. Formulas (2.9b), (2.10), (2.11) are
central schemes, whereas (2.9a), (2.9c) are upwind schemes. We note that these
schemes have previously been applied by Kluge [13] in the numerical solution of
the Heston PDE. Through Taylor expansion it can be verified that each of the
formulas (2.9), (2.10), (2.11) has a second-order truncation error, provided that the
function f is sufficiently often continuously differentiable and the meshes in the x-
and y-directions are smooth, cf. (2.7).
The actual FD discretization of the initial-boundary value problem for the Heston
PDE is performed as follows.
In view of the Dirichlet boundary conditions (2.3) and (2.5), the grid in [0, S] ×
[0, V ] is given by
G = {(si , vj ) : 1 ≤ i ≤ m1 , 0 ≤ j ≤ m2 − 1}.
At this grid, each spatial derivative appearing in (2.1) is replaced by its correspond-
ing central FD scheme (2.9b), (2.10), or (2.11) – except in the region v > 1 and at
the boundaries v = 0 and s = S.
In the region v > 1 we apply the upwind scheme (2.9a) for ∂u/∂v whenever
the flow in the v-direction is towards v = V . This is done so as to avoid spurious
oscillations in the FD solution when the volatility-of-variance σ is close to zero.
At the outflow boundary v = 0 the derivative ∂u/∂v is approximated using the
upwind scheme (2.9c). All other derivative terms in the v-direction vanish at v = 0,
due to the factor v occurring in (2.1), and hence, these terms do not require further
treatment.
At the boundary s = S the spatial derivatives in the s-direction need to be
considered. First, the Neumann condition (2.4) at s = S implies that the mixed
derivative ∂ 2 u/∂s∂v vanishes there. Next, the derivative ∂u/∂s is directly given by
(2.4). Finally, we approximate ∂ 2 u/∂s2 at s = S = sm1 using the central scheme
(2.10) with the virtual point S + ∆sm1 , where the value at this point is defined by
extrapolation using (2.4).
The FD discretization described above of the initial-boundary value problem
(2.1)–(2.5) for the Heston PDE yields an initial value problem for a large system
of stiff ordinary differential equations (ODEs),

(2.12) U ′ (t) = AU (t) + b(t) (0 ≤ t ≤ T ), U (0) = U0 .

Here A is a given m × m–matrix and b(t) (t ≥ 0), U0 are given m–vectors with m =
m1 m2 . The vector U0 is directly obtained from the initial condition (2.2) and the
vector function b depends on the boundary conditions (2.3)–(2.5). For each given
t > 0, the entries of the solution vector U (t) to (2.12) constitute approximations to
the exact solution values u(s, v, t) of (2.1)–(2.5) at the spatial grid points (s, v) ∈ G,
ordered in a convenient way.

2.3. Time discretization: ADI schemes. Acquiring an effective numerical

time-discretization method for the spatially discretized Heston problem (2.12) is
a key step in arriving at a full numerical solution scheme for the Heston PDE that
is both efficient and robust.
Let ∆t > 0 be a given time step and let temporal grid points be given by
tn = n∆t for n = 0, 1, 2, . . .. A well-known method for the numerical solution of
stiff initial value problems for systems of ODEs
(2.13) U ′ (t) = F (t, U (t)) (0 ≤ t ≤ T ), U (0) = U0
 ADI FINITE DIFFERENCE SCHEMES FOR OPTION PRICING 309

is the Crank–Nicolson scheme or trapezoidal rule. This method defines approxima-

tions Un to the exact solution values U (tn ) of (2.13) successively for n = 1, 2, 3, . . .
by

(2.14) Un = Un−1 + 21 ∆tF (tn−1 , Un−1 ) + 21 ∆tF (tn , Un ).

In our case of (2.12) we have

(2.15) F (t, w) = Aw + b(t) for 0 ≤ t ≤ T, w ∈ Rm .

Thus, each step (2.14) requires the solution of a system of linear equations involving
the matrix (I − 12 ∆tA) where I denotes the m×m identity matrix. Since (I − 12 ∆tA)
does not depend on the step index n, one can compute a LU factorization of this
matrix once, beforehand, and next apply it in all steps (2.14) to obtain Un (n ≥ 1).
The Crank–Nicolson scheme can be practical when the number of spatial grid
points m = m1 m2 is moderate. In our application to the two-dimensional Heston
PDE, however, m usually gets large and the Crank–Nicolson scheme becomes inef-
fective. The reason for this is that (I − 21 ∆tA), and hence the matrices in its LU
factorization, possess a bandwidth that is directly proportional to min{m1 , m2 }.
For the numerical solution of the semi-discretized Heston problem (2.12) we shall
consider in this paper splitting schemes of the ADI type. We decompose the matrix
A into three submatrices,
A = A0 + A1 + A2 .

We choose the matrix A0 as the part of A that stems from the FD discretization
of the mixed derivative term in (2.1). Next, in line with the classical ADI idea, we
choose A1 and A2 as the two parts of A that correspond to all spatial derivatives
in the s- and v-directions, respectively. The rd u term in (2.1) is distributed evenly
over A1 , A2 . The FD discretization described in Sect. 2.2 implies that A1 , A2 are
essentially tridiagonal and pentadiagonal, respectively.
Write b(t) from (2.12) as b(t) = b0 (t) + b1 (t) + b2 (t) where the decomposition is
analogous to that of A. Next, define functions Fj (j = 0, 1, 2) by

(2.16) Fj (t, w) = Aj w + bj (t) for 0 ≤ t ≤ T, w ∈ Rm .

Then for F given by (2.15) we have the splitting F = F0 + F1 + F2 . Clearly, F0 6= 0

whenever the correlation factor ρ 6= 0.
Let θ be a given real parameter. In the following we formulate four splitting
schemes for the initial value problem (2.13). Here we assume that F stems from a
FD discretization of a general 2D convection-diffusion equation with a mixed de-
rivative term that is decomposed similarly as above for the semi-discrete Heston
PDE. All four schemes generate, in a one-step manner, approximations Un to the
exact solution values U (tn ) of (2.13) successively for n = 1, 2, 3, . . .:

Douglas (Do) scheme :



 Y0 = Un−1 + ∆tF (tn−1 , Un−1 ),



(2.17) Yj = Yj−1 + θ∆t (Fj (tn , Yj ) − Fj (tn−1 , Un−1 )) (j = 1, 2),





Un = Y2
310 K. J. IN ’T HOUT AND S. FOULON

Craig–Sneyd (CS) scheme :


 Y0 = Un−1 + ∆tF (tn−1 , Un−1 ),





 Yj = Yj−1 + θ∆t (Fj (tn , Yj ) − Fj (tn−1 , Un−1 )) (j = 1, 2),






(2.18) Ye0 = Y0 + 12 ∆t (F0 (tn , Y2 ) − F0 (tn−1 , Un−1 )) ,







 Yej = Yej−1 + θ∆t (Fj (tn , Yej ) − Fj (tn−1 , Un−1 )) (j = 1, 2),





Un = Ye2

Modified Craig–Sneyd (MCS) scheme :



 Y0 = Un−1 + ∆tF (tn−1 , Un−1 ),







 Yj = Yj−1 + θ∆t (Fj (tn , Yj ) − Fj (tn−1 , Un−1 )) (j = 1, 2),






 Yb0 = Y0 + θ∆t (F0 (tn , Y2 ) − F0 (tn−1 , Un−1 )) ,

(2.19)



 Ye0 = Yb0 + ( 12 − θ)∆t (F (tn , Y2 ) − F (tn−1 , Un−1 )) ,







 Yej = Yej−1 + θ∆t (Fj (tn , Yej ) − Fj (tn−1 , Un−1 )) (j = 1, 2),






Un = Ye2

Hundsdorfer–Verwer (HV) scheme :


 Y0 = Un−1 + ∆tF (tn−1 , Un−1 ),







 Yj = Yj−1 + θ∆t (Fj (tn , Yj ) − Fj (tn−1 , Un−1 )) (j = 1, 2),




(2.20) Ye0 = Y0 + 12 ∆t (F (tn , Y2 ) − F (tn−1 , Un−1 )) ,





 Yej = Yej−1 + θ∆t (Fj (tn , Yej ) − Fj (tn , Y2 )) (j = 1, 2),







Un = Ye2 .
In the Do scheme (2.17), a forward Euler predictor step is followed by two implicit
but unidirectional corrector steps, whose purpose is to stabilize the predictor step.
The CS scheme (2.18), the MCS scheme (2.19) and the HV scheme (2.20) can
be viewed as different extensions to the Do scheme. They all perform a second
predictor step, followed by two unidirectional corrector steps.
Each of the above splitting schemes treats the mixed derivative part F0 in a fully
explicit way. Note that in the special case where F0 = 0, the CS scheme reduces to
the Do scheme, but the MCS scheme (with θ 6= 12 ) and the HV scheme do not.
The F1 and F2 parts are treated implicitly in all four schemes. In every step
of each scheme, systems of linear equations need to be solved involving the two
matrices (I − θ∆tAj ) for j = 1, 2. Like for the Crank–Nicolson scheme, these
matrices do not depend on the step index n, and thus one can determine their LU
factorizations once, beforehand, and next apply them in all time steps to compute
Un (n ≥ 1). However, the bandwidths of the matrices given by LU factorization
 ADI FINITE DIFFERENCE SCHEMES FOR OPTION PRICING 311

of (I − θ∆tAj ) (j = 1, 2) are now fixed, i.e., independent of m1 and m2 . As a

consequence, in each of the four splitting schemes, the number of floating point
operations per time step is directly proportional to m, which yields a big reduction
compared to the Crank–Nicolson scheme.
For any parameter θ, the classical order of consistency of the Do scheme – for
general F0 , F1 , F2 – is just one. An advantage of the CS, MCS and HV schemes is
that they can attain order of consistency two for general F0 , F1 , F2 . Taylor expan-
sion shows that the CS scheme has order two if and only if θ = 21 . Subsequently,
the MCS and HV schemes have order two for any given θ. With the latter two
schemes, the parameter θ can thus be chosen to meet additional requirements. A
virtue of all four splitting schemes is that all internal vectors Yj , Yej form consistent
approximations to U (tn ).
The Do scheme can be regarded as a direct generalization of the classical ADI
schemes for 2D diffusion equations by Douglas & Rachford [5] and Peaceman &
Rachford [17] to the situation where a mixed spatial derivative term is present
in the equation. This generalization was considered by McKee & Mitchell [15] for
diffusion equations and then by McKee et. al. [16] for convection-diffusion equations.
The CS scheme was developed by Craig & Sneyd [4] with the aim to arrive at
a stable second-order ADI scheme for diffusion equations with mixed derivative
terms.
The MCS scheme has recently been introduced by In ’t Hout & Welfert [10] to
obtain more freedom in the choice of θ as compared to the second-order CS scheme.
The HV scheme was designed by Hundsdorfer [11] and Verwer et. al. [26] for the
numerical solution of convection-diffusion-reaction equations arising in atmospheric
chemistry, cf. also [12]. The application of the HV scheme to equations containing
mixed derivative terms has recently been studied by In ’t Hout & Welfert [9, 10].
Our formulation of the ADI schemes (2.17)–(2.20) is similar to the type of formu-
lation used in [11]. In the literature, these schemes are also sometimes referred to
as Stabilizing Correction schemes, and are further closely related to Approximate
Matrix Factorization methods and IMEX methods, cf. e.g. [12].
ADI schemes have already been applied by various researchers for the numerical
solution of PDE models in finance. Lipton [14] describes an ADI scheme for 2D
convection-diffusion equations with a mixed derivative term that is closely related
to the Do scheme with θ = 21 . Andreasen [2] uses the Do scheme with θ = 12 for
certain interest rate models without a mixed derivative term. Randall [19] applies
the CS scheme with θ = 12 to the Heston PDE with a mixed derivative term. We
note that actual numerical results were not reported in loc. cit. The HV scheme
has recently been considered for the application to PDE models in finance by In
’t Hout [8], where an initial experiment in the case of the Heston PDE with a
mixed derivative term is discussed. As a corollary of our present paper we will find
that the choice of θ = 0.3 for the HV scheme [8] is not optimal with respect to
robustness.
Theoretical stability results for all four ADI schemes – relevant to FD discretiza-
tions of 2D convection-diffusion equations with a mixed derivative term – have been
derived in [4, 9, 10, 15, 16]. These results concern unconditional stability, i.e., with-
out any restriction on the time step ∆t. The analysis in loc. cit. has been performed
following the classical von Neumann method (Fourier transformation), where the
usual assumptions are made that the coefficients are constant, the boundary con-
dition is periodic, the spatial grid is uniform, and stability is considered in the
l2 -norm.
312 K. J. IN ’T HOUT AND S. FOULON

Currently, the most comprehensive stability results for the Do, CS, MCS and
HV schemes relevant to multi-dimensional PDEs with mixed derivative terms have
been obtained in [9, 10]. We review the main conclusions from loc. cit. relevant to
the situation of our paper.
The Do and CS schemes are both unconditionally stable when applied to 2D
convection-diffusion equations with a mixed derivative term whenever θ ≥ 12 . In
particular, the second-order CS scheme is unconditionally stable.
The MCS and HV schemes are unconditionally stable when applied to 2D pure √
diffusion equations with a mixed derivative term whenever θ ≥ 31 and θ ≥ 1 − 12 2,
respectively. Recall that the MCS and HV schemes are of order two for any given
θ. At this moment, unconditional stability results for the MCS and HV schemes
in the general situation of 2D equations, with convection, are lacking. It was
conjectured [9] however that the HV scheme is unconditionally stable when applied
to 2D√convection-diffusion equations with a mixed derivative term whenever θ ≥
1 1
2 + 6 3.

3. Numerical experiments
3.1. Spatial discretization error. In this section we consider four numerical
examples and assess the actual convergence behavior of the FD discretization (2.12)
of the Heston PDE defined in Sect. 2.2. For any given numbers of mesh points m1 ,
m2 in the s- and v-directions, we define the global spatial discretization error at
time t = T by

e(m1 , m2 ) = max |u(si , vj , T ) − Uk (T )| : 12 K < si < 23 K, 0 < vj < 1 .
Here u denotes the exact European call option price function, satisfying the initial-
boundary value problem (2.1)–(2.3) for the Heston PDE on the unbounded domain.
Next, Uk designates the component of the exact solution U to (2.12) that corre-
sponds to the grid point (si , vj ).
Clearly, the global spatial error is defined via a maximum norm. The set of
asset prices s ∈ ( 21 K, 32 K) and variances v ∈ (0, 1) in our definition encompasses
most situations of practical interest. We note that the modeling error, that was
introduced by restricting the domain of the Heston PDE to a bounded set, is
also contained in e(m1 , m2 ). In our experiments this contribution turns out to be
negligible.
For the actual computation √ of the global spatial errors e(m1 , m2 ) we apply the
HV scheme with θ = 12 + 61 3 to (2.12) with the small time step ∆t = T /1000
to obtain a sufficiently accurate approximation of U (T ). Subsequently, we employ
an implementation of Heston’s semi-analytical formula [7] to acquire values of u.
For calculating the single integrals occurring in this formula we use a numerical
quadrature rule. Numerical difficulties one can encounter in the implementation,
due to the presence of multi-valued complex functions, have recently been discussed
in [1]. We adopt the algorithm proposed in loc. cit. where branch cuts in the
complex plane are correctly taken into account.
We perform numerical experiments in the four cases of parameter sets given by
Table 1. Observe that in three of the four cases there is a substantial correlation
factor ρ. Only in case 4 the correlation factor is relatively small.
Case 1 has been taken from Albrecher et. al. [1], where we have chosen T = 1.
Case 2 comes from Bloomberg [3]. A special feature of this parameter set is that
σ is close to zero, which implies that the Heston PDE is convection-dominated in
the v-direction. Values for rd , rf and T were not specified in [3] and have been
 ADI FINITE DIFFERENCE SCHEMES FOR OPTION PRICING 313

Case 1 Case 2 Case 3 Case 4

κ 1.5 3 0.6067 2.5
η 0.04 0.12 0.0707 0.06
σ 0.3 0.04 0.2928 0.5
ρ -0.9 0.6 -0.7571 -0.1
rd 0.025 0.01 0.03 0.0507
rf 0 0.04 0 0.0469
T 1 1 3 0.25
K 100 100 100 100

Table 1. Parameters for the Heston model and European call options.

Case 1 Case 2

150 150
100 100
U U
50 50
0 0
1 1
200 200
0.5 100 0.5 100
V 0 0 S V 0 0 S

Case 3 Case 4

150 150
100 100
U U
50 50
0 0
1 1
200 200
0.5 100 0.5 100
V 0 0 S V 0 0 S

Figure 2. European call option price functions u in the four cases

given by Table 1.

chosen separately. Case 3 has been taken from Schoutens et. al. [21]. Here the
Feller condition is only just met. Finally, case 4 stems from Winkler et. al. [27].
Figure 2 displays the exact option price functions u corresponding to the four
cases of Table 1 on the domain (s, v) ∈ [0, 200] × [0, 1].
Figure 3 subsequently shows for each case from Table 1 the global spatial errors
e(2m2 , m2 ) vs. 1/m2 for m2 = 10, 20, . . . , 100. We have taken m1 = 2m2 as it turns
out that, for efficiency reasons, one can use much less points in the v-direction than
in the s-direction. To determine the numerical order of convergence p of the spatial
discretization, we have fitted in each case a straight line to the outcomes for the
global spatial errors. Accordingly, in the cases 1, 2, 3, 4 we found the respective
314 K. J. IN ’T HOUT AND S. FOULON

Case 1 0
Case 2
0
10 10

−1 −1

Spatial error
Spatial error
10 10

−2 −2
10 10

−3 −3
10 10 −2 −1
−2 −1
10 1/m2 10 10 1/m2 10

0
Case 3 Case 4
0
10 10

−1 −1
Spatial error

Spatial error
10 10

−2 −2
10 10

−3 −3
10 −2 −1
10 −2 −1
10 1/m2 10 10 1/m 10
2

Figure 3. Global spatial errors e(2m2 , m2 ) vs. 1/m2 for m2 =

10, 20, . . . , 100 in the four cases given by Table 1 with least-squares
fit superimposed.

orders p = 1.9, 2.0, 2.1, 2.4. This clearly suggests that the FD discretization of the
Heston PDE described in Sect. 2.2 is convergent of appr. order two.
We remark that for the global spatial errors in relative sense, we obtained in
each case that it is close to 1.0 percent for m2 = 30 and decreases to approximately
0.1 percent for m2 = 100. Here asset-variance pairs (si , vj ) are considered such
that the option value u(si , vj , T ) is always at least 1.
Finally, we repeated the numerical experiments above for uniform grids. Then
the global spatial errors are always found to be much larger, compared to the non-
uniform grid under consideration with the same number of points. Furthermore,
the error behavior is erratic in this case. The latter can be seen to be related to the
relative position of the point s = K to the mesh in the s-direction, and concerns
a known phenomenon, see e.g. Tavella & Randall [24]. This behavior is much less
noticeable for (suitable) non-uniform grids.
3.2. Temporal discretization error. In this section we present numerical ex-
periments for the ADI schemes (2.17), (2.18), (2.19), (2.20) which yields important
insight into their actual stability and convergence behavior in the application to
semi-discretized Heston problems (2.12) with non-zero correlation.
We define the global temporal discretization error at time t = T = N ∆t by

eb (N ; m1 , m2 ) = max |Uk (T ) − UN, k | : 12 K < si < 23 K, 0 < vj < 1 ,
 ADI FINITE DIFFERENCE SCHEMES FOR OPTION PRICING 315

Case 1 Case 2
1 1
10 10

−1 −1
10 10

Temporal error
Temporal error −3
10
−3
10

−5 −5
10 10

−7 −7
10 10 −3 −2 −1 0
−3 −2 −1 0
10 10 10 10 10 10 10 10
1/N 1/N

1
Case 3 1
Case 4
10 10

−1 −1
10 10
Temporal error

Temporal error
−3 −3
10 10

−5 −5
10 10

−7 −7
10 −3 −2 −1 0
10 −3 −2 −1 0
10 10 10 10 10 10 10 10
1/N 1/N

Figure 4. Global temporal errors eb (N ; 100, 50) vs. 1/N in the

four cases given by Table 1. Schemes: RKC with ε = 10 (dotted
line), Do with θ = 12 (diamond), CS with θ = 21 (circle), MCS with
√
θ = 13 (grey circle), HV with θ = 1 − 12 2 (star) and HV with
√
θ = 12 + 16 3 (square).

where the index k is such that Uk (T ) and UN, k correspond to the spatial grid point
(si , vj ). Clearly, the global temporal error is defined for the same (s, v)-domain as
the global spatial error and we also deal again with the maximum norm, cf. Sect. 3.1.
Motivated by the theoretical stability and accuracy results discussed in Sect. 2.3,
we shall consider the Do and CS schemes with θ = 12 and the MCS scheme with
√
θ = 13 . Next, we consider the HV scheme for the two values θ = 1 − 12 2 ≈ 0.293
√
and θ = 12 + 16 3 ≈ 0.789, to which we refer in the following as HV1 and HV2,
respectively. We apply all these ADI schemes in each of the four cases of parameter
sets for European call options in the Heston model listed in Table 1.
In addition, we also apply the Runge–Kutta–Chebyshev (RKC) scheme. This
is an explicit second-order Runge–Kutta scheme which has been constructed such
that its stability region includes a large interval [−β, 0] along the negative real axis.
For a complete discussion of this method see e.g. [12, 23]. The RKC scheme has a
free parameter ε. Since the Heston PDE contains a convective part, we have chosen
ε = 10, cf. [25]. In this case β ≈ 0.34(ν 2 − 1), where ν denotes the number of stages
of the scheme. Accordingly, in a given application to a semi-discretized Heston
316 K. J. IN ’T HOUT AND S. FOULON

1
Case 1 Case 2
1
10 10

−1 −1
10 10

Temporal error

Temporal error
−3 −3
10 10

−5 −5
10 10

−7 −7
10 −3 −2 −1 0
10 −3 −2 −1 0
10 10 10 10 10 10 10 10
1/N 1/N

1
Case 3 1
Case 4
10 10

−1 −1
10 10
Temporal error

Temporal error
−3 −3
10 10

−5 −5
10 10

−7 −7
10 −3 −2 −1 0
10 −3 −2 −1 0
10 10 10 10 10 10 10 10
1/N 1/N

Figure 5. Global temporal errors eb (N ; 200, 100) vs. 1/N in the

four cases given by Table 1. Schemes: RKC with ε = 10 (dotted
line), Do with θ = 21 (diamond), CS with θ = 21 (circle), MCS with
√
θ = 13 (grey circle), HV with θ = 1 − 12 2 (star) and HV with
√
θ = 12 + 16 3 (square).

problem (2.12)p with time step ∆t, the number of stages is taken as the smallest
integer ν ≥ 1 + 3∆t r[A] where r[A] denotes the spectral radius of A.
Figures 4, 5 display for m1 = 2m2 = 100 and m1 = 2m2 = 200, respectively,
the results for the global temporal errors eb (N ; m1 , m2 ) vs. 1/N for a range of step
numbers N between N = 1 and N = 1000. We applied the HV2 scheme with
N = 5000 to obtain a reference value for U (T ).
A first observation from Figures 4, 5 is that in the cases 1, 3, 4 the RKC scheme
has global temporal errors that are often comparable to those of the MCS scheme.
For relatively large ∆t, however, the RKC scheme requires a large number of stages
ν since in all cases r[A] ≈ 5.1 ·104 (if m1 = 2m2 = 100) and r[A] ≈ 2.2 ·105 (if m1 =
2m2 = 200). In our experiments, RKC always turned out to be considerably less
efficient than MCS. Note that because the RKC scheme does produce fair temporal
errors for all N in each of the cases 1, 3, 4, this suggests that the eigenvalues of
the corresponding matrices A all lie close to the negative real axis in these three
cases. In case 2, the RKC scheme clearly shows instability for, at least, values
N < 100. We explain this from the corresponding matrices A to have eigenvalues
 ADI FINITE DIFFERENCE SCHEMES FOR OPTION PRICING 317

in the left half of the complex plane with substantial imaginary parts. This is
also conceivable, since in case 2 the Heston PDE is convection-dominated in the
v-direction, cf. Sect. 3.1.
For the Do, CS, MCS and HV2 schemes, the Figures 4, 5 clearly reveal that in
all cases 1–4 the global temporal errors always stay below a moderate bound and
decrease monotonically with N . This is a very favorable result and indicates an
unconditionally stable behavior of these ADI schemes in all cases. Note that it is a
non-trivial result, as it does not directly follow e.g. from the von Neumann analysis
discussed in Sect. 2.3.
For the HV1 scheme, case 2 shows a peak in the global temporal errors which is
higher if m1 = 2m2 = 200 than if m1 = 2m2 = 100. We conjecture that the HV1
scheme is just conditionally stable, under a CFL condition. We mention in passing
that the same is found for the HV scheme with value θ = 0.3, cf. [8].
Subsequent analysis of the results in Figures 4, 5 yields for the MCS and HV2
schemes in each case 1–4 a convergence order equal to 2.0. Moreover, the conver-
gence behavior is of the form C(∆t)2 (0 < ∆t ≤ τ ) with constants C, τ > 0 that
are only weakly dependent on the number of spatial grid points m. Hence, the MCS
and HV2 schemes show a stiff order of convergence equal to two. This clearly agrees
with their orders of consistency, cf. Sect. 2.3. Remark that this is not obvious, as
the order of consistency is a priori only relevant to fixed, non-stiff ODEs.
For the HV1 scheme, we obtain a stiff order of convergence equal to two in the
cases 1, 3, 4.
The Do and CS schemes exhibit in all cases an undesirable convergence behavior,
with temporal errors that are relatively large for modest time steps ∆t. This
atypical behavior becomes more pronounced when m1 , m2 get larger. Though
the results are not included in the figures, we remark that the (time-consuming)
Crank–Nicolson scheme shows a similar behavior. The cause for this phenomenon
is related to the fact that at s = K the payoff function (2.2) is non-smooth and the
Do, CS and Crank–Nicolson schemes do not sufficiently damp local (high-frequency)
errors incited by this. A remedy for this situation is to first apply, at t = 0, two
backward Euler steps with step size ∆t/2, and then to proceed onwards from t = ∆t
with the time-stepping schemes under consideration, cf. Rannacher [20] and also
e.g. [6, 12, 18]. By adopting this damping procedure, we recover in each case 1–4 a
stiff order of convergence equal to one for the Do scheme and equal to two for the
CS scheme.
Concerning implementation, we mention that all codes have been written in
Matlab, version 7.8, where all matrices have been defined as sparse. As an indication
for the computing times, our implementation of the CS, MCS, HV schemes each
took per time step about 0.005 cpu-sec (if m1 = 2m2 = 100) and 0.02 cpu-sec
(if m1 = 2m2 = 200) on an Intel Duo Core T5500 1.6 GHz processor with 1 GB
memory. For the Do scheme these times are appr. halved. Note that the times are
directly proportional to the number of spatial grid points m.

3.3. Down-and-out call options. The FD discretization of the Heston PDE

described in Sect. 2.2 is readily adapted to more exotic options such as barrier
options. We briefly discuss here down-and-out call options. If B ∈ (0, K) denotes
the down-and-out barrier, then the spatial domain becomes [B, S] × [0, V ]. Next,
the boundary conditions (2.3), (2.5) change to, respectively,

u(B, v, t) = 0 (0 ≤ t ≤ T )
318 K. J. IN ’T HOUT AND S. FOULON

and
u(s, V, t) = (s − B)e−rf t (0 ≤ t ≤ T ).

Finally, mesh points in the s-direction are given by (2.6) with

ξi = sinh−1 ((B − K)/c) + i · ∆ξ (0 ≤ i ≤ m1 )
and
1
∆ξ = [sinh−1 ((S − K)/c) − sinh−1 ((B − K)/c)].
m1
The above modifications clearly concern only a few lines in the implementation.
Lipton [14] derived a semi-analytical formula for the prices of double barrier
options in the Heston model provided the correlation ρ = 0 and rd = rf . By using
Lipton’s formula with a large value for the upper barrier, the semi-discrete Heston
PDE (2.12) for down-and-out barrier options has been validated in all four cases
from Table 1, where we chose the barrier B = 95 < 100 = K and set ρ = 0,
rd = rf = 0.03. Moving the upper boundary for s to S = 14K so as to reduce the
modeling error, again an appr. second-order convergence behavior for the global
spatial errors is obtained in each case, cf. Sect. 3.1.
We subsequently applied each of the ADI schemes from Sect. 3.2 together with
the RKC scheme to the semi-discrete Heston PDE for down-and-out barrier op-
tions where B, S were taken as above and we considered the four original cases
from Table 1. For all schemes it turned out to be advantageous, to a larger or
smaller extent, to employ the damping procedure at t = 0 to obtain a regular
behavior of the global temporal errors when the time step ∆t is relatively large.
The conclusions concerning the observed stability and convergence behavior of the
schemes are similar to those of Sect. 3.2. In particular, we find that in all cases
1–4 the Do scheme has a stiff order of convergence equal to one and the CS, MCS,
HV2 schemes possess a stiff order of convergence equal to two. Next, the global
temporal errors for the latter three schemes are always close to each other in the
cases 2 and 4, but in the cases 1 and 3 the HV2 scheme has somewhat larger errors
than the CS, MCS schemes. Finally, for the HV1 scheme we observe again a peak
in the global temporal errors in case 2. In the other three cases from Table 1 the
HV1 scheme shows a stiff order of convergence equal to two.
4. Conclusions and future research
Among the schemes discussed in this paper, the MCS scheme with θ = 13 , used
with damping at t = 0, seems preferable for the fast, accurate and robust numerical
solution of the semi-discrete Heston PDE with arbitrary correlation
√ ρ ∈ [−1, 1]. The
CS scheme with θ = 21 and the HV scheme with θ = 12 + 16 3, both applied with
damping, form good alternatives. All three ADI schemes show an unconditionally
stable behavior combined with a stiff order of convergence equal to two. For the
CS scheme it is essential to apply damping at t = 0, whereas for the above HV
scheme the error constant can be somewhat larger than for the MCS scheme, √ when
damping is used. The RKC scheme as well as the HV scheme with θ = 1 − 12 2 lack
robustness, as they can have an unstable behavior if the volatility-of-variance σ is
close to zero. Also, the RKC scheme appears to be relatively inefficient. Finally,
the Do scheme shows an unconditionally stable behavior but has only a stiff order
of convergence equal to one whenever the correlation ρ is non-zero.
We conclude by mentioning some issues for future research. A main issue is a
theoretical analysis of the stability and convergence properties of the ADI schemes
that have been observed in the experiments, cf. Sect. 3.2. Next, a study of the
 ADI FINITE DIFFERENCE SCHEMES FOR OPTION PRICING 319

performance of ADI schemes relevant to other exotic options in the Heston model
is of much interest. The use of variable time steps, especially near the initial
time t = 0, is likely to increase the efficiency, cf. e.g. [13]. Finally, ADI schemes
can be attractive in the numerical solution of other multi-dimensional PDEs from
finance with mixed derivative terms, e.g. the three-dimensional hybrid Heston–
Hull–White model. The extension of the ADI schemes (2.17)–(2.20) to such PDEs
is straightforward. Positive results on unconditional stability of these schemes in
arbitrary spatial dimensions, for pure diffusion equations with mixed derivative
terms, have recently been proved in [10].

Acknowledgments
The first author wishes to thank Willem Hundsdorfer for suggesting to consider
the RKC scheme and for providing a Matlab implementation of this scheme. He is
also grateful to Marc Spijker for his comments, which have enhanced the presenta-
tion of this paper. Further, the authors would like to thank an anonymous referee
for constructive remarks.

References
[1] H. Albrecher, P. Mayer, W. Schoutens & J. Tistaert, The little Heston trap, Wilmott Mag.,
January 2007, 83–92.
[2] J. Andreasen, Back to the future, Risk 18 (2005) 104–109.
[3] Bloomberg Quant. Finan. Devel. Group, Barrier options pricing under the Heston model,
2005.
[4] I. J. D. Craig & A. D. Sneyd, An alternating-direction implicit scheme for parabolic equations
with mixed derivatives, Comp. Math. Appl. 16 (1988) 341–350.
[5] J. Douglas & H. H. Rachford, On the numerical solution of heat conduction problems in two
and three space variables, Trans. Amer. Math. Soc. 82 (1956) 421–439.
[6] M. B. Giles & R. Carter, Convergence analysis of Crank–Nicolson and Rannacher time-
marching, J. Comp. Finan. 9 (2006) 89–112.
[7] S. L. Heston, A closed-form solution for options with stochastic volatility with applications
to bond and currency options, Rev. Finan. Stud. 6 (1993) 327–343.
[8] K. J. in ’t Hout, ADI schemes in the numerical solution of the Heston PDE, in: Numerical
Analysis and Applied Mathematics, eds. T. E. Simos et. al., AIP Conf. Proc. 936 (2007)
10–14.
[9] K. J. in ’t Hout & B. D. Welfert, Stability of ADI schemes applied to convection-diffusion
equations with mixed derivative terms, Appl. Num. Math. 57 (2007) 19–35.
[10] K. J. in ’t Hout & B. D. Welfert, Unconditional stability of second-order ADI schemes
applied to multi-dimensional diffusion equations with mixed derivative terms, Appl. Num.
Math. 59 (2009) 677–692.
[11] W. Hundsdorfer, Accuracy and stability of splitting with Stabilizing Corrections, Appl. Num.
Math. 42 (2002) 213–233.
[12] W. Hundsdorfer & J. G. Verwer, Numerical Solution of Time-Dependent Advection-Diffusion-
Reaction Equations, Springer, Berlin, 2003.
[13] T. Kluge, Pricing derivatives in stochastic volatility models using the finite difference method,
Dipl. thesis, TU Chemnitz, 2002.
[14] A. Lipton, Mathematical Methods for Foreign Exchange, World Scientific, Singapore, 2001.
[15] S. McKee & A. R. Mitchell, Alternating direction methods for parabolic equations in two
space dimensions with a mixed derivative, Computer J. 13 (1970) 81–86.
[16] S. McKee, D. P. Wall & S. K. Wilson, An alternating direction implicit scheme for parabolic
equations with mixed derivative and convective terms, J. Comp. Phys. 126 (1996) 64–76.
[17] D. W. Peaceman & H. H. Rachford, The numerical solution of parabolic and elliptic differ-
ential equations, J. Soc. Ind. Appl. Math. 3 (1955) 28–41.
[18] D. M. Pooley, K. R. Vetzal & P. A. Forsyth, Convergence remedies for non-smooth payoffs
in option pricing, J. Comp. Finan. 6 (2003) 25–40.
[19] C. Randall, PDE Techniques for Pricing Derivatives with Exotic Path Dependencies or Exotic
Processes, Lecture notes, Workshop CANdiensten, Amsterdam, 2002.
320 K. J. IN ’T HOUT AND S. FOULON

[20] R. Rannacher, Finite element solution of diffusion problems with irregular data, Numer.
Math. 43 (1984) 309–327.
[21] W. Schoutens, E. Simons & J. Tistaert, A perfect calibration ! Now what ?, Wilmott Mag.,
March 2004, 66–78.
[22] S. E. Shreve, Stochastic Calculus for Finance II, Springer, New York, 2004.
[23] B. P. Sommeijer, L. F. Shampine & J. G. Verwer, RKC: An explicit solver for parabolic
PDEs, J. Comp. Appl. Math. 88 (1997) 315–326.
[24] D. Tavella & C. Randall, Pricing Financial Instruments, Wiley, New York, 2000.
[25] J. G. Verwer, B. P. Sommeijer & W. Hundsdorfer, RKC time-stepping for advection-diffusion-
reaction problems, J. Comp. Phys. 201 (2004) 61–79.
[26] J. G. Verwer, E. J. Spee, J. G. Blom & W. Hundsdorfer, A second-order Rosenbrock method
applied to photochemical dispersion problems, SIAM J. Sci. Comp. 20 (1999) 1456–1480.
[27] G. Winkler, T. Apel & U. Wystup, Valuation of options in Heston’s stochastic volatility
model using finite element methods, in: Foreign Exchange Risk, eds. J. Hakala & U. Wystup,
Risk Publ., 2002.

Department of Mathematics and Computer Science, University of Antwerp, Middelheimlaan 1,

B-2020 Antwerp, Belgium
E-mail : [email protected]
URL: http://win.ua.ac.be/~kihout/

KBC, Havenlaan 12, B-1080 Brussels, Belgium

E-mail : [email protected]