Option Pricing Formulae using Fourier

Transform: Theory and Application

Martin Schmelzle*
April 

Abstract
Fourier transform techniques are playing an increasingly important role in
Mathematical Finance. For arbitrary stochastic price processes for which the
characteristic functions are tractable either analytically or numerically, prices for a
wide range of derivatives contracts are readily available by means of Fourier inversion
methods. In this paper we first review the convenient mathematical properties of
Fourier transforms and characteristic functions, survey the most popular pricing
algorithms and finally compare numerical quadratures for the evaluation of density
functions and option prices. At the end, we discuss practical implementation details
and possible refinements with respect to computational efficiency.

Keywords: option pricing, characteristic functions, Fourier inversion methods, Lévy

processes, stochastic volatility, density approximation, numerical quadratures.

*
Contact information:

martin.schmelzle@pfadintegral.com
1 Introduction
The first important contribution to quantitative modeling of derivatives securities
dates back to Black and Scholes (1973) who deduce a precise formula for the value of
a European option on an underlying whose price on maturity follows a log-normal
diffusion process. Despite the success of the Black–Scholes model on pricing and
hedging derivatives, Merton (1976) noted early that options quoted on the markets
differ systematically from their predicted values, which led up to questioning the
distributional assumptions based on geometric Brownian motion.
By adding jumps to the archetypal price process with Gaussian innovations
Merton (1976) is able to partly explain the observed deviations from the benchmark
model which are characterized by fat tails and excess kurtosis in the returns
distribution [for an overview of ‘stylized facts’ on asset returns see Cont (2001),
statistical properties of implied volatilities are summarized in Cont & al. (2002)]. In
the sequel also other authors develop more realistic models based on jump processes
[e. g. Eberlein and Keller (1995), Madan & al. (1998) and Kou (2002)]. They derive
option values from an analytical form of the conditional density function, for the
value of the underlying on maturity given the initial state. Many of these original
results are quite complicated requiring special functions or infinite summations.
As an alternative to model the option payoff directly by an analytical stochastic
process, it has been recognized that by mapping the characteristic function of the
density function to the payoff in Fourier space, option values can be usually
computed much easier for these sophisticated processes. The characteristic function
developed as a tool for the solution of problems in Probability Theory is the Fourier
transform of the density function and the main idea using the transform methods is
then to take an integral of the payoff function over the probability distribution
obtained by inverting the corresponding Fourier transform. There is growing interest
in applying these methods using characteristic functions and Fourier transforms
which stems from the need to apply more complex pricing models than the Gaussian,
which are more conveniently characterized through a characteristic function
primarily rather than a probability distribution.
Transform methods turn out to be a very effective tool for the solution of many
technical problems, since calculations in Image space are often much easier than in
the spatial domain. The solution to the problem in Image space is then described
through an Image function. Examples for these Image functions are calculations in
Laplace or Fourier Space which are widely used in financial applications. To finally
obtain the solution in original Space domain the Image function has to be inverted
via inversion methods.
The Fourier transform is a widely used and a well understood mathematical tool
from Physics and Engineering disciplines applicable to numerous tasks, for example
signal processing [Allen and Mills (2004)], or as a method for solving partial
differential equations [Duffy (2004)]. It is interesting to note that already Merton
(1973) noticed the possibility of using Fourier transforms to solve the Black–Scholes
partial differential equation.
Inside the field of Finance the Fourier inversion method was first proposed in the
Stein and Stein (1991) stochastic volatility model that uses the transform method in
order to find the distribution of the underlying. Settling on the characteristic function
approach, Heston (1993) obtains an analytical pricing formula for the valuation of
European options with time varying volatility of the underlying. By using a variant of
Lévy’s Inversion Theorem for the probability functions the options prices result as the
differences from the numerical evaluation of two Fourier integrals.
Since then Fourier inversion methods became a very active field of research in
finance theory. Bakshi and Madan (2000) provide an economic foundation for
characteristic functions, generalize the approach of Heston (1993) and Stein and Stein
(1991) in many significant ways, and develop valuation formulae for a wide variety of
contingent claims. Duffie & al. (2000) offer a comprehensive survey that the Fourier
methods are applicable to a wide range of stochastic processes, the class of
exponential affine jump diffusions.
A numerically very efficient methodology is introduced in Carr and Madan
(1999) who pioneer the use of fast Fourier transform algorithms by mapping the
Fourier transform directly to call option prices via the characteristic function of an
arbitrary price process. Lee (2004) generalizes their approach to other payoff
functions and unifies it with other known Fourier pricing elements. In Carr and Wu
(2004) the authors extend the Carr and Madan (1999) methodology for general
claims and apply these to time changed Lévy processes, the class of generalized affine
models [Filipović (2001)] and Quadratic activity rate models [Leippold and Wu
(2002)].
Given the characteristic function for some price dynamics and a Fourier
transformed payoff function determining the contract type, Lewis (2001) develops a
highly modular pricing framework. In the Carr and Madan (1999) approach the
whole option price is Fourier transformed including the particular payoff function,
whereas Lewis (2001) entirely separates the underlying stochastic process from the
derivative payoff by the aid of the Plancherel–Parseval Theorem and obtains a variety
of valuation formulae by the application of Residual Calculus.
More techniques have been developed to compute option values as an integral in
Fourier space, using Fourier transform methods. Throughout the paper we will
mention some of them and give more details. Density calculations and option pricing
are then, as we will see, a matter of numerical integration for the Fourier inversion,
usually employing direct integration methods or the Discrete Fourier Transform,
where the Fast Fourier Transform is an efficient way to compute it in practice.
Methodological aspects concerning numerical implementation should be considered
sensibly since the semi-infinite domain Fourier integrals might exhibit oscillations in
the integrand, which is complicating numerical analysis.

| 3
 Computational performance also becomes a critical factor when calibrating the
models to observed market prices which can afterwards be used to value exotic
products or devise hedging strategies. The complexity of these new models require
sophisticated numerical algorithms which ensure fast and reliable option pricing,
calculations of hedge parameters, volatility surface calculations and density
approximations.

The structure of the paper is organized as described in the following. In the next
section, we briefly outline the general valuation framework for European option
pricing. We start by considering the standard martingale pricing approach and then
point to the inherent connections to Arrow–Debreu securities and state price
densities. In Section 3, we give an overview of the most fundamental ideas and
mathematical tools needed for characteristic function methods, we also present the
most relevant properties of Fourier Transforms and their inversion. Section 4
provides an exhibition of the most popular Fourier pricing algorithms. We begin with
the Black–Scholes style valuation formula and proceed with the more flexible single
integral solutions. With a focus on an intuitive understanding, we emphasize on a
concise presentation of the main results of the various methodologies without too
much technical details. Thereafter, we accentuate some of the similarities between
these approaches and point out recent developments which are based on these
procedures or which apply them to other fields in mathematical finance. In Section 5,
we finally show how to apply Fourier inversion methods to density approximations
and option pricing. To do so we first present and discuss some qualitative features of
popular numerical quadratures and fast Fourier transform methods. Then, we give
detailed numerical examples, show convergence properties of the Fourier integrands,
compare accuracy and run times. At the end, in Section 6, we review various
mathematical and numerical methods regarding implementational issues and
computational performance. A thorough quantitative and numerical analysis is a
important factor and relevant for the stability of the Fourier transform methods, the
choice of adequate truncation levels for the numerical evaluation and integration
sample spacing of the Fourier integrands. Section 7 concludes.

2 Valuation Problem for European Options

Asset prices are modeled by stochastic processes whose evolutions are governed by an
underlying probability measure. In the well established theory of arbitrage pricing,
there exists a risk neutral probability measure under which asset prices are arbitrage
free. Moreover, in terms of probability, the absence of arbitrage is essentially
equivalent to the existence of a risk neutral equivalent martingale measure for the
stochastic process making the underlying process become a martingale.
Consequently, as Ross (1976) and Cox and Ross (1976) have shown, the price of a
European option can be reduced to taking their discounted expected values of a

| 4
future payoff with respect to a risk neutral measure
. This is known as martingale
pricing or pricing by expectation
  , , 
   
 | . (2.1)
For a European call option with strike , risk-free rate  and time to maturity , the
value of the arbitrage free option with payoff 
max  , 0
   
is then given by
 , , 
   
    | , (2.2)
#

       ! | d ,
$
#

      ! | d .
%

! | . It is the conditional transition probability density of the underlying  at

The expectation is calculated on the basis of the risk neutral density function

time  conditional upon the information flow  (filtration) of the asset price
available up to time & [Harrison and Pliska (1981)]. From Harrison and Kreps (1979)
we know that under technical conditions the existence of an equivalent martingale

equivalent martingale measure makes the discounted price process '

    a
measure and the absence of arbitrage are essentially equivalent properties. This

martingale and satisfies the martingale property '$


(' ).
Under the risk neutral pricing measure, all assets have the same expected rate of
return which is the risk free rate. In other words, under no-arbitrage conditions the
risk preferences of investors do not enter into valuation decisions, thus, we can
assume risk neutrality and price any security in a preference free world. For recent
treatments and more details on the approach to arbitrage pricing see e. g. Delbaen and
Schachermayer (2006) and the references therein.
Commonly, the risk neutral probability measure is used for arbitrage pricing. A
market place free from arbitrage opportunities is usually characterized by a unique set
of state contingent claims. Arrow–Debreu prices, or similarly Arrow–Debreu

pays off one unit of the numéraire at time , if, and only if, a specific state occurs. For
securities, are the most elementary contingent claims. An Arrow–Debreu security

this reason Arrow–Debreu prices are also known as state prices. The continuous state
equivalent of Arrow–Debreu securities constitutes a state price density.
Any possible payoff can be replicated with a linear combination of state prices
whereas this particular replication portfolio implies a unique arbitrage free price.
These elementary relations are illustrated in Figure 1.

| 5
 5 5
4 4
3 3
Profit / Loss

Profit / Loss
2 2
1 1
0 0
... K-2 K-1 K K+1 K+2 ... ... K-2 K-1 K K+1 K+2 ...
-1 -1
-2 -2
Spot ST Spot ST

Figure 1: Arrow–Debreu securities. Left: Payoff diagram of a state price. Right: Payoff diagram for a
European call option.

This observation suggests that risk neutral probabilities can be expressed in terms of

density is related to the risk neutral density by    ! and can be derived by twice
state prices and vice versa. Following Breeden and Litzenberger (1978), the state price

differentiating either a call or put option with respect to its strike price
+,
!
  * -
+ , %./
(2.3)
0.

pays off one unit at time  exactly if 

 . Suppose  is the final payoff to a
This result shows that it is possible to construct an Arrow–Debreu security which

derivative security maturing at 

#
 
 1   d . (2.4)
#

Following Carr (2003) this is a spectral decomposition of the payoff  into the
payoffs or delta claims 1   respectively from an infinite collection of Arrow–
Debreu securities
+,
   
1    .
+ , 
(2.5)

The risk neutral probability function and the state price density are completely
interchangeable concepts, only distinguished by a discount factor.
Under the martingale pricing approach, an option value can be represented as a
convolution of a payoff function with a discounted probability density function or
equivalently the state price density of the state variables. It is well known that if the
density of the underlying is known in closed form, option prices can be obtained by a
single integration of their payoff against this density function.
For example if we assume Brownian motion for the price dynamic, the density

variable having a log-normal distribution. Under
its density is given by

function is known in closed form. The stock price on maturity then reveal a random

< >
1 789 /0 :89 /; :  = > ??@
,
! 
 ,

,= > 
 3√26
(2.6)

| 6
i. e. the density of a Normal Distribution with mean ln $ C :  3 , ?  and
<
,
variance 3 , . By a single integration the value for the Black–Scholes call of today is
obtained
< >
#
1 789 /0 :89 /; :  = > ??@
,
$ , , 
       d .

,= > 
 3√26
(2.7)
%

Unfortunately, more sophisticated price dynamics like affine jump diffusions and
Lévy processes often do not possess density functions in closed form or have quite
complicated analytical expressions involving special functions and infinite
summations. However, for many of the more advanced asset price models
characteristic functions are available in closed form. By analogy, if the characteristic
function of the underlying is tractable, option prices can also be obtained by a single
integration.

3 Fourier Transform, Inversion Theorem and Characteristic

Functions
The power and versatility of Fourier Analysis and the Calculus of Characteristic
Functions have been used in many recent publications relating to probability
distributions. Generally this is gaining strong momentum in the field of financial
modeling, too. The successful application of these techniques require a considerable
degree of mathematical sophistication. The purpose of this section is to give an
overview of this theory, and to show how to apply characteristic functions and
Fourier transforms in mathematical finance.

3.1 Characteristic Functions

If not in all statistical analyses, at least in most of them, the calculation of the
distribution function of random variables is required. A very interesting fact is that
even if the random variable of interest does not have an analytical expression, the
characteristic function of this random variable always exists. There is a one to one
relationship between the probability density and a characteristic function. If the
characteristic function is known in closed form, is tractable numerically, or given by
empirical data, then we can compute the distribution function by using the Inversion

The characteristic function DE F

( GHE ) of a real valued random variable I
theorem.

is defined for arbitrary real numbers F as the expectation of the complex valued
transformation  GHE , where i
√1 is the imaginary unit. If KE L is the probability
function (PDF) of the random variable then the integral

| 7
 #
DE F
( GHE )
 GHM KE L dL, (3.1)
#

function KE L denoted by O KE L . At a given F, DE F is a single random variable

defines the expected value and is by definition the Fourier transform of the density

and for ∞ Q F Q ∞ we have a stochastic process. An intuitive understanding of

characteristic functions is offered by Epps (1993) who give clear geometric

statistical applications. From Euler’s Identity the random variable I can be

interpretations of the characteristic functions to elucidate their properties and use in

represented by sine and cosine function as

 GHE
cosFI C i sinFI . (3.2)
From this operation we can see that exponential  GHE for a given F represents a point

can be considered as a contour in the complex plane defined by |U|

1, where |U| is
on the unit circle in the complex plane. The unit circle is a circle with unit radius. It

the complex modulus which can be interpreted as a distance measure.

u=1.0

 GHE 

u=0.5

u=0.0

function. Right: Wrapped around distribution of FI, I~Φ0, 1 at F

0, 0.5, 1 and characteristic
Figure 2: Geometrical interpretations of characteristic functions. Left: Operation of a characteristic

function of Φ0, 1 on 0, 1. Source: Epps (1993), p 34.

Since DE F
( GHE ), the characteristic function evaluated at any F can be
interpreted as the center of mass of the distribution of FI wrapped around the unit
circle in the complex plane. If the domain of DE F is given by the whole real line, we
have the geometrical representation of the characteristic function as a curve confined

we can see that the norm of the characteristic function |DE F | is always within the
to an infinitely long circular ‘cylinder’ or ‘tube’ which is illustrated in Figure 2. Further

unit circle. At F
0 we have the value one since in this case FI is degenerated.
Comparing the wrapped around distribution of FI and FI for F Y 0, it
becomes apparent that one is the mirror image of the other about the real axis. The

| 8
opposite sign, hence D ZZZZZZZZ
E F
DE F is the complex conjugate of DE F . Due to
real parts of their centers are equal and the imaginary parts are equal, too, but of

this symmetry around F

0, it is sufficient to concentrate on F [ 0 to fully describe

If a characteristic function is absolutely integrable over the real line ∞, C∞

the distribution function.

then I has an absolutely continuous probability distribution. This is said to be

integrable in the Lebesgue sense and belongs to \< ] .

Below follows a summary of elementary properties of characteristic functions

1. DE F always exists since |iFI | is a continuous and bounded function for all finite
real F and L; further ^_∞ K L dL^ ` _∞|KL | dL so that the defining integral
∞ ∞

2. DE 0
1 for any distribution
converges absolutely

3. ^DI F ^ ` 1
4. ZZZZZZZZ
DE F
DE F
5. If a
b C cI, then Dd F
iHe DE cF
6. If I< and I, are stochastically independent, then the characteristic function Dd F of
the new random variable a
I< C I, is the product of the characteristic function of
each random variable DEf F DE> F , which are allowed to be drawn from different
distributions

moments and cumulants of a distribution function. By gth differentiation of the

Another important property of characteristic functions is the possibility to derive the

characteristic function at the origin of F, the gth moment I h  of the distribution
(if they exist) is obtained
1d* DE F -
h
 I h 

ih dFh H.$.
(3.3)

ih I are computed by taking the logarithm of DE F which is then called the

In the same way that the characteristic function generates the moments, cumulants

cumulant characteristic function lnDE F

1d*
h
lnDE F
ih I
h -
i dFh
(3.4)
H.$.

Thus, important qualitative statistical properties such as variance, skewness and

kurtosis have analogons in Image space and are readily available if the characteristic
function is known
I
i< I , (3.5)

jklI
i, I , (3.6)

| 9
 in I
m I
,
i, I n/ ,
(3.7)

ip I
g I
.
i, I ,
(3.8)

These fundamental properties make characteristic functions a viable tool in

Probability Theory and to applications in stochastic inference.

3.2 The Inversion Theorem

An essential property of characteristic functions is their one to one relationship with
distribution functions. Every random variable possesses a unique characteristic
function and the characteristic function indeed characterizes the distribution
uniquely [Waller (1995)]. The Inversion Theorem is the Fundamental Theorem of the
Theory of Characteristic Functions since it links the characteristic function back to its
probability distribution via an inverse Fourier transform.
Based on the fundamental results from Lévy (1925) who gave a general inversion
formula, Gurland (1948) and Gil-Pelaez (1951) develop useful representations of the
Inversion Theorem. In the following we will follow the form given by Gil-Pelaez
(1951) [see Waller & al. (1995) for a review on inversion methods].

Pelaez inversion integral for cumulative distribution function (CDF) _# KE L qL

M
The inversion algorithms are based on the following particular form of the Gil-

1 1 #
 GHM DE F
rE L
sI ` L
 dF.
2 26 iF
(3.9)
#

the characteristic function. Taking the derivate of rE L yields the probability density
We see that the recovered distribution function is expressed as an integral in terms of

function KE L
1 #
KE L
O < DE F 
 GHM DE F dF.
26
(3.10)
#

between KE L and DE F .
The comparison of (3.1) with (3.10) exhibits the reciprocal relationship which exists

The Fourier transform of a real valued function can be written as a two

U
b C ic, where tU
b is the real part and uU
c is the imaginary part of the
dimensional vector in the complex plane. These complex values are expressed as

complex number U with b and c being real numbers. The complex conjugate is given
by Uv
b  ic, the modulus is |U|
√b, C c, and the real part is tU
U C Uv / 2
and imaginary part uU
U  Uv / 2i. Further we have
DE F C DE F
tDE F 
,
2
(3.11)

| 10
 DE F  DE F
uDE F 
,
2i
(3.12)

which implies that the function DE F is even in its real part and odd in its imaginary
part for all F. An even function is said to be symmetric with respect to F
0.
Therefore, the integral of the positive and the negative halves of an even function are
the same.

integrals. Using these results the expression for KE L reduces to

This special symmetry property allows us to further simplify the Fourier

1 $
1 #
KE L
tw  GHM DE F dFx C tw  GHM DE F dFx,
26 # 26 $

1 #
ZZZZZZZZZZZZZZZZZZ 1 #

tw  yHM DE F dFx C tw  GHM DE F dFx,
26 $ 26 $

1 #

t w2  GHM DE F dFx,
26 $

1 #

t( GHM DE F )dF.
6
(3.13)
$

Similar calculations for the CDF yield

1 1 #
 GHM DE F   GHM DE F
rE L
C dF,
2 26 iF
(3.14)
$
# ZZZZZZZZZZZZZZZZZ
1 1  yHM DE F   GHM DE F

C dF,
2 26 $ iF
1 1 #
 GHM DE F

 tw x dF,
2 6 iF
(3.15)
$

1 1 #
 GHM DE F

 uw x dF.
2 6 F
(3.16)
$

For the probability sI [ L

_M KE L qL we can use the simple relation
#

rEz L
sI [ L
1  rE L to obtain the complementary CDF (cCDF)
1 1 #
 GHM DE F
rEz L
C tw x dF,
2 6 iF
(3.17)
$

1 1 #
 {HM DE F

C uw x dF.
2 6 F
(3.18)
$

In certain cases it will be convenient to employ Euler’s identity and express the
complex exponential as two separate cosine and sine functions. For (3.1) we have the
equivalent representation

| 11
 # #
DE F
cosFL KE L dL C i sinFL KE L dL. (3.19)
# #

Using this transformation the density function in (3.10) results in a real valued
integral of a real variable [see Abate and Whitt (1992)]
1 #
KE L
cosFL tDE F  C sinFL uDE F  dF.
6
(3.20)
$

And finally, for the CDF we get

1 1 #
dF
rE L
 cosFL uDE F   sinFL tDE F  .
2 6 F
(3.21)
$

3.3 Elementary Properties of the Fourier Transform

the Fourier transform from KL as K|F

O KF or just O K.
Next, we mention some more relevant properties of Fourier transforms. We denote

Linearity For arbitrary b, c the transform is a linear operator

O bKL C c}L F
bK|F C c}~F . (3.22)
Translation The Fourier transform turns a multiplication by a variable L$ into
translation
O K L  L$ F
 GHM; K|F . (3.23)

Differentiation If  [ 0 is an integer, K € is piecewise continuously differentiable,

and each derivative is absolutely integrable on the entire real line, then
O(K € )F
iF € K|F . (3.24)
Thus, a differentiation converts to multiplication in Fourier space.

Convolution Theorem One of the most fundamental and important properties of

Fourier transforms is convolution (Faltung). In Fourier space a convolution is
mapped into multiplication. Define the convolution by
#
K  } L
KL  ‚ }‚ d‚, (3.25)
#

where  denotes the convolution operator. Then we have

O K  }F
K|F }~F , (3.26)
and
O K}F
K|F  }~F . (3.27)

| 12
Thus, the Fourier transform of the convolution of two functions equals the product of
the Fourier transforms of each of the functions.

Absolutely and Square Integrable Functions Many concepts in the field of Fourier

absolutely integrable (or simply integrable) if the integral of its absolute value over ]
Transform Theory draw heavily on the notion of absolute integrability. A function is

is finite
#
|KL | dL Q ∞. (3.28)
#

This is a very important relation since for a Fourier transform and its inverse to exist,

functions is called \< or \< ] . Functions that are integrable on an interval b, c are
then the condition in equation (3.28) must hold. The space of absolutely integrable

said to be on \< b, c .

If K L is absolutely integrable, which is true for K L ƒ \< ] , then the Fourier
transform exists. Given O KF is absolutely integrable, the inverse Fourier
transform O < (OK)F exists. However, it is important to note that the Fourier
transform of an \< ] function is not necessarily integrable even though it often is.
This is why we cannot simply assume to take its inverse transform.
Fourier transforms are extendable to square integrable functions as well. A
function is square integrable if
#
|K L |, dL Q ∞. (3.29)
#

The space of square integrable functions is denoted by \, or \, ] . The space \,

delineates an inner product space. Functions \< b, c that are integrable on an
interval b, c also satisfy
„
|K L |, dL Q ∞. (3.30)
e

on an interval b, c is defined on the inner product space as well, i. e. a vector space
The last inequality is of great importance since it implies that a function well defined

equipped with an inner product relation …†,†‡. The notion of inner products is
introduced in the next part about the Plancherel–Parseval Theorem.

The Plancherel Theorem and Parseval’s Theorem Another very important property
is that under certain conditions inner or scalar products are preserved under Fourier
transforms. In the context of Probability Theory, these results are especially
significant for the reconstruction of a distribution from its characteristic function.

| 13
 The scalar product or inner product of two functions KL and }L on \, is
defined as
#
…K, }‡
K L ZZZZZZ
}L dL. (3.31)
#

And equally for the Fourier transforms O K and O }

#
…O K, O }‡
OKZZZZZZ
O }dL . (3.32)
#

Since KL
_  GHM O KqF,
< #
,ˆ #
then we have
# #
1 #
ZZZZZZqL

KL }L  GHM O K}F dFdL,

# # 26 #

1 # #

O K  GHM }F dLdF,
26 # #

1 #

O KZZZZZZ
O}dF.
26
(3.33)
#

There are subtle aspects involved in the interchange of integration carried out in the
preceding equations. By applying Fubini’s Theorem, which states in principle that if a
function of two variables is absolutely integrable over a region, then its iterated
integrals and its double integral over the region are all equal. Hence, we may freely

If K
} than the integral of the squared modulus of K is equal to the integral of
interchange the order of integration.

the squared modulus of its Fourier transform O K

# #
|K dL

|, |OK|, dF. (3.34)

# #

Via a Fourier inversion the scalar or inner product can be transformed from Fourier
domain to spatial domain and vice versa. This identity is sometimes also referred to as

series. The inner product on \, ] , restricted to those functions that are also
the Parseval’s Identity, who discovered a discrete version in the context of Fourier

absolutely integrable, furnishes a inner product operation …†,†‡ for \< ‰ \, , where the
inner product induces a norm ŠL Š
‹…L, L‡ on this space. One of the main
consequences of the Plancherel Theorem or Parseval’s Idendity for later purposes is
that Fourier transforms preserve the norms of functions.
There seems to be no unique naming convention about the theorem, i. e. in
different fields the theorem might be referred to as the Plancherel Identity or the
Parseval’s Identity. This is why we will refer to it as the Plancherel–Parseval theorem
within this paper.

| 14
function in terms of real valued transform variables F. In certain cases it is convenient
Generalized Fourier Transform Up to this point we have defined the characteristic

axis. For this case we can extend the domain of the transform variable F to the
or even necessary to integrate a characteristic function along a line parallel to the real

complex plane F Œ U ƒ , where the characteristic function is well behaved [Lukacs

(1970), Theorem 7.1.1]. The set of values for U
U C iU{ , for which the expectation
in Eq. (3.1) is well defined, is within some strip of regularity ŽE with  Q uU Q 
parallel to the real U-axis. In these U-plane strips the extended characteristic functions
D  U are typically regular functions, also known as analytic characteristic functions.
Following Lukacs (1970), an analytic charactistic function is a characteristic function

in the complex U-plane. Under this extended definition D  U is called the generalized
which coincides with a regular analytic function in some neighborhood of the origin

Fourier transform [see also Titchmarsh (1975)]. Since the transform variable is now
extended to the complex plane the transform is also sometimes referred to as the
complex Fourier transform. The inverse of this generalized Fourier transform is given
by
1 G‘’ #
K L
 G‘M DU dU.
26
(3.35)
G‘’ #

transform if uU [ 0 and the cumulant generating function if uU Q 0 as special

It is interesting to note that the generalized Fourier transform contains the Laplace

cases, provided that they are well defined.

Generally, the properties for the ordinary Fourier transform also apply with little
or no modification to the generalized Fourier transform. For example, we mention

numbers U by an integration along a straight line parallel to the real axis

the important fact that the Plancherel–Parseval identity also work for complex

#
1 #
K}v dL
O KU C iU{ ZZZZZZ
O}U C iU{ dU ,
26
(3.36)
# #

if the functions K L and }L are well behaved at U{ . This will be an important
property for later purposes.

Further informations on the topic and a comprehensive reference for the theory of
characteristic functions can be found in Lukacs (1970). More detailed discussions in
the context of Fourier transforms can be found in Rudin (1987) and Titchmarsh
(1975). Broad treatments of Fourier Theory including other scientific fields are e. g.
Allen and Mills (2004) and Duffy (2004).

4 Pricing Formulae using Characteristic Functions

With the knowledge of the characteristic function of the stochastic process
determining the price dynamic of the state variables option prices are available by

| 15
Fourier inversion methods. In general there are two approaches in literature
considering inverse Fourier transform. The first approach obtains option prices with
respect to the Fourier inversion of cumulative distribution functions with an
appearance like the classical Black–Scholes formula. This route was pioneered by the
famous work of Heston (1993) and was refined and extended in many ways. Most
noticeable is the work by Bakshi and Madan (2000) who give a clear economical
interpretation of characteristic functions with respect to market completeness and
Arrow–Debreu securities in a state space framework. Ross (1976) showed that
options are completing or spanning markets by expanding the asset space. Bakshi and
Madan (2000) demonstrate that markets can be equivalently spanned by

It is convenient to specify the characteristic function D  as the expected value of

characteristic functions.

the complex exponential of the logarithmic price L

ln 
#
D  F
( GHM )
 GHM ! L dL, (4.1)
#

where ! is the risk neutral density of L relative to the martingale measure
. By

Euler identity we can express D  F as 
cosFL C i sinFL . As we have seen the
payoff of a European option can be replicated by a portfolio of Arrow–Debreu
securities and similarly the payoff on characteristic functions can be composed of
trigonometric sine and cosine functions. Therefore the characteristic function can be
viewed as an Arrow–Debreu security in Fourier space
1 #
1  M  
 GH“ dF.
” %

26
(4.2)
#

The characteristic function of the state price density is defined as follows

D  F
(    GHM ), (4.3)
#

 GHM    ! L dL,
#

which corresponds to a hypothetical claim that pays  GHM at .

The second approach considers the pricing of options analogue to the Fourier
inversion of the probability density function. Carr and Madan (1999) show how this
analogy can be used by analytically relating the Fourier transform of an option to its
characteristic function. Along the same idea Lewis (2001) develops an option pricing
framework where the payoff structure is modeled explicitly and takes the Fourier
transform with respect to the state variable, whereas Carr and Madan (1999) apply
the transform to the strike value. In these cases the argument of the characteristic
function is evaluated in a particular domain of the complex plane depending on the
specific option payoff to ensure the existence of the Fourier transform.
Both Fourier inversion methodologies are applicable to a wide range of European
payoff structures. In the following we will focus on European options to illustrate the

| 16
various transform methods prevailing in the option pricing literature. Starting from
the quasi Black–Scholes formula first introduced in Heston (1993), we discuss the
similar developments by Attari (2004) and Bates (2006) and move on to the more
flexible Carr and Madan (1999), Lewis (2001) and Lipton (2002) option price
characterizations.

4.1 The Black–Scholes Style Formula

Beginning with Heston (1993), many authors use the Fourier inversion methods to
solve advanced valuation problems. Generalizing previous work Bakshi and Madan
(2000) demonstrate the power and versatility of using characteristic functions for the
pricing of contingent claims. They illustrate the use of Fourier transformed state price
densities or Arrow–Debreu securities respectively, and how generic payoffs can be
spanned by positions in characteristic functions. In fact Bakshi and Madan (2000)
reduce the valuation problem to the estimation of Arrow–Debreu securities under

The price of a European call with spot  and strike  is

appropriately modified equivalent probability measures.

 $ , , 
   
    , (4.4)
#

       !  d .
$

Decomposing the expression yields a portfolio of Arrow–Debreu securities

# #
 $ , , 
  
 ! d    
 !  d , (4.5)
% %
# #

    M !L dL      !L dL,
h h

Π<     Π, ,
where g
ln . The quantities Π< and Π, are both conditional probabilities of
finishing in-the-money at maturity. Π< is computed with the stock as numéraire asset
whereas Π, is computed with a zero coupon bond as underlying numéraire asset. For
Π, we use _h ! L dL which is the probability sL – g . The characteristic function
#

for Π, is D, F
D  F
_#  GHM ! L dL, hence
#

#
1 #
Π,
—  GHM D  F dF˜ dL.
26
(4.6)
h #

Changing the order of integration yields to

1 # #
Π,
D  F —  GHM dL ˜ dF,
26
(4.7)
# h

which simplifies to

| 17
 1 1 #
 GHh D  F
Π,
C t w x dF.
2 6 iF
(4.8)
$

The situation for Π<

   _h  M !L dL is a bit more subtle, since in this case the
#

stock serves as numéraire. Introducing a change of measure from
to
™ by a Radon-

Nikodym derivative [compare Geman & al. (1995)] we get
™
d
 M0

.
d
 M0 
(4.9)

™ the Fourier transform of Π< is defined as

With this new measure

( M0  GHM ) D  F  i

™ ( GHM )

.
  M 0  D  i
(4.10)

Because of the no arbitrage condition 
 

$   we get D  i as its
characteristic function and for ( M0  GHM ) we get D  F  i . Further note that
 M !L is nonnegative for all L and the integral over 0, ∞ equals one. Hence, we
can treat 
™ ( GHM ) as a probability function and invert it accordingly
1 1 #
 GHh D  F  i
Π<
C tw x dF.
2 6 iFD  i
(4.11)
$

characteristic function, i. e. whether it is defined for ln /0   or ln  [see Section

The specific form of the formulae depend on the chosen definition of the
/

5.2]. By using the second definition and depending on the definition of g, the above
;

expression can be rewritten in a slightly different form. For example, by changing g to

a ‘dimensionless’ moneyness g
ln %; C  we obtain [as described in Lewis (2001)]
/

1 1 #
 GHh D  F  i
Π<
C tw x dF,
2 6 iF
(4.12)
$

1 1 #
 GHh D  F
Π,
C tw x dF.
2 6 iF
(4.13)
$

The two integrals Π< and Π, can be combined into one integral. By rearranging the

1
involved equations the Black–Scholes style formula for a call option reduces to
$ , , 
     
2
1 #  GHh D  F  i  GHh D  F
C  tw x   t w x dF.
(4.14)
 
6 $ $ iF iF
While most authors use the real part of the complex valued integrand, it is also
possible to use equation (3.18) instead of (3.17) [see, e. g. Duffie & al. (2000) and
Bates (1996)]. This leads to

| 18
 1 1 #
 GHh D› F
К
C uw x dF.
2 6 F
(4.15)
$

Applying Euler identity to the probabilities Π< and Π, it yields to

1 1 #
u(D› F ) cosFg  t(D› F ) sinFg
К
C dF.
2 6 F
(4.16)
$

4.2 Attari’s Approach

inversion formula exploits the inherent relationship between the two probabilities Π<
Attari (2004) proposes a modification to the Black–Scholes Style solution. His

and Π, of a European option. Merging the two integrals Π< and Π, into one
expression the calculation requires only one integral. Pointing out that !L

_#  GHM D  F dF and Π<

_h  M !L dL, substituting the expression for ! L
< # #
,ˆ
in Π< and defining g
ln
%“ œ0
/;
yields
#
1 #
Π<
 —
M
 GHM D, F dF˜ dL .
26
(4.17)
h #

Changing the order of integration by Fubini’s theorem gives

1 # #
Π<
D  F —  M  GHM dL ˜ dF.
26
(4.18)
# h

Using the fact that *Π< |h.#

1, since a zero strike option equals , we have
1 1 # #
Π<
C D  F —  M  GHM dL ˜ dF,
2 46
(4.19)
# h

1 # h
 D  F —  M  GHM dL ˜ dF.
46 # #

Integrating the expression yields

1 1 #
 GHG h
Π<
C D  F dF
2 26 iF C i
(4.20)
#

1 #
 GHG Ÿ C  GHG Ÿ
 lim D  F — ˜ dF.
46 ŸŒ# # iF C i
Expanding the last term and taking F ΠF in the resulting second term gives
1 #
 GHG Ÿ 1 #
 GHG Ÿ
lim D  F — ˜ dF  lim D  F — ˜ dF.
46 ŸŒ# iF C i 46 ŸŒ# iF  i
(4.21)
# #

| 19
In the first integrand we have a pole at F
i while for the integrand we have a pole
at F
i. By using Residue Theorem the second integral for Π< and 26i times the
residues at its poles we get 26i Resi C Resi £
26i : ?

G¤0 G ¤0 G
pˆ ,

1 1 #
GHG h
D  i
Π<
C D  F dF C .
2 26 iF C i 2
(4.22)
#

The term D  i equals 1 (since *Π< |h.#

1), hence we have
h #
 GHh D  F
Π<
1 C dF.
26 # iF C i
(4.23)

Bringing it together with the expression of the call value this results in
h #
 GHh D  F
$ , , 
$ —1 C dF˜
26 # iF C i
(4.24)

1 1 #
 GHh D  F
    — C dF˜.
2 26 # iF
Rearranging terms, applying Euler identity, and symmetry for real valued functions
finally turns to
1
 $ , , 
$     
2
(4.25)

uDF  tDF 
# ¥tDF  C ¦ cosFg C ¥uDF   ¦ sinFg
1 F F
 
 
dF.
6 $ 1 C F,
This formulation offers a pricing formula involving a single one dimensional
integration. We note that in comparison to the Black–Scholes like solution the
integrand now has a quadratic term in the denominator which ensures a faster
convergence of the integrand.

4.3 Bates’ Approach

A similar approach to that of Attari (2004) is outlined in Bates (2006). Here, the value
§¨
§%
 _% ! d
of a European call option is evaluated by using the discounted CDF
  #

inversion formula for an arbitrary stochastic process governing ! , and finally
[see Breeden and Litzenberger (1978)], substituting the

integrating with respect to  results in

%
1 1   F
/; D
# GH 89
 $ , , 
$      © C dFª.
2 26 iF1  iF
(4.26)
#

| 20
$ is an integration constant determined by the value of a zero strike call. Using the
fact that option prices are real valued yields to
%
1 1   F
/; D
# GH 89
 $ , , 
$      © C t« ¬ dFª.
2 6 iF1  iF
(4.27)
$

Compared to the Black–Scholes style solution an advantage against is that this

formulation requires only one integration and has an integrand which will converge
faster due to the quadratic term in the denominator.

4.4 Carr and Madan Approach

Carr and Madan (1999) develop a different method designed to use the fast Fourier
transform (FFT) to value options. Unfortunately, the FFT cannot be applied directly

evaluation point F
0. Therefore, instead of solving for the risk neutral exercise
to evaluate the integrals we mentioned above, since the integrands are singular at the

probabilities of finishing in-the-money they introduce a new technique with the key

the logarithmic strike price g. With this specification and a FFT routine a whole
idea to calculate the Fourier transform of a modified call option price with respect to

range of option prices can be obtained within a single Fourier inversion.

variables L
ln  and g
ln  we get for a European call option
Again, beginning from the risk neutral valuation formula and a change of

 g
   
    , (4.28)
#

    M   h !L dL.
h

Unfortunately, expressing the call in terms of the log strike in (4.28),  g tends to $
as g goes to ∞
#
 g
    M   # !L dL, (4.29)
#
#

  
 M !L dL,
#

   
 M .
From martingale property 
 
$   we see that limhŒ#  g
$ is not
converging to zero. Hence,  g is not \< and a Fourier transform does not exist [see

factor  ­h with  [ 0 it is possible to make  g an integrable function

integrability condition (3.28)]. However, by introducing an exponential damping

i g
 ­h  g , (4.30)
where the modified call price i g is an integrable function, since

| 21
 #
| ­h  g |dg Q ∞, (4.31)
#

for a suitable . The effect of an exponential damping factor on the integrability of

call price functions is shown in Figure 3.

100% 35%
alpha=0.75
30%
80% alpha=1.25
25%
Price/Spot

Price/Spot
60% alpha=1.75
20%
alpha=2.25
40% 15%
10%
20%
5%
0% 0%
-6 -4 -2 0 2 -6 -4 -2 0 2
log(K/S) log(K/S)

the call price tends to $ as g approaches minus infinity. Right: Damping the call price function with
Figure 3: Effect of damping on the square integrability of call price functions. Left: Without damping

different values for the damping factor . Heston model parameters: 


100, 
1, 
0.05,
®
2, ¯
0.01, °
0.5, 3±
0.25 and ²$
0.02.

The Fourier transform of i g is then given by

#
³ F
 GHh i g dg, (4.32)
#
# #

 GHh  ­h     M   h  ! L dL dg,
# #
# #

 GHh
 ­h     M   h ! L dL dg,
# h
# M

   !L —  M   h  GHh  ­h dg˜ dL.
# #

For the inner integral, representing the call payoff, we get

M M M
 M   h  GHh  ­h dg
 M  ­GH h dg   ­GH< h dg, (4.33)
# # #

M 1

( ­GH h )#  ( ­GH< h )# .
M M
 C iF  C iF C 1
Taking the limit for limhŒ#  ­GH h
0 with  [ 0 we get
 ­<GH M  ­<GH M
—  ˜.
 C iF  C 1 C iF
(4.34)

Hence leading to

| 22
 #
 ­<GH M  ­<GH M
³ F
  
! L —  ˜ dL,
 C iF  C 1 C iF
(4.35)
#
#
 ­<GH M

  
! L dL.
#  C iF  C 1 C iF

Taking the Fourier transform for _# !L  ­<GH M

_# !L  GH­< M we get
# #

the characteristic function for the risk neutral price process D  F   C 1 i .

Finally we have
   D  F   C 1 i
³ F
,
 , C   F, C i2 C 1 F
(4.36)

where ³F is expressed in terms of the characteristic function D  .

In a last step, given ³F , an inverse Fourier transform multiplied by the
reciprocal of the exponential factor yields to the undamped call prices
 ­h #
 g
 GHh ³F dF,
26
(4.37)
#

 ­h #

t( GHh ³F ) dF.
6 $

This method is viable when  is chosen in a way that the damped option price is well
behaved. Damping the option price with  ­h makes it integrable for the negative axis
g Q 0. On the other hand for g [ 0 the option prices increase by the exponential  ­h
which influences the integrability for the positive axis. A sufficient condition of i g
to be integrable for both sides (square integrability) is given by ³0 being finite

Plancherel–Parseval idendity]. With F

0
[compare Section 3.3 on the Fourier transform to square integrable functions via the

   D   C 1 i
³ 0
.
, C 
(4.38)

Thus we need D   C 1 i Q ∞ which is equivalent to

D   C 1 i

( <­ ) Q ∞ (4.39)
Therefore, i g is well behaved when the moments of order 1 C  of the underlying
exist and are finite. If not all moments of  exist, this will impose an upper bound on
.
The corresponding put values can be obtained by defining
´ g
 ­h µ g , (4.40)

   D  F   C 1 i
with Fourier transform

³¶ F
, .
    F, C i2 C 1 F
(4.41)

| 23
And finally the put prices can be written as
 ­h #
µ g
t( GHh ³¶ F ) dF.
6
(4.42)
$

For the put formula to be well defined, is suffices to choose an appropriate  [ 0 in a

way that 
 ­  Q ∞ [see Lee (2004)].
The Carr and Madan (1999) formulation is different to the previous described
ones in a sense that not only the risk neutral density function but the whole option
price is Fourier transformed. This is possible through the introduction of an
additional degree of freedom which makes the integral in (4.37) in fact a generalized
Fourier transform, since
1 G‘’ #
1 #G‘’
 G‘h
³U dU
t( G‘h ³U ) dU ,
26 6
(4.43)
G‘’ # $G‘’

 ‘’ h #

t( G‘ h ³U C iU{ ) dU .
6 $

Therefore, the modification technique removes the non-smooth-behavior of the pole

at the origin by shifting the pole to a less numerically sensitive position on the
imaginary axis. The interpretation of the integration of a damped function as a
contour integral in the complex plane is first considered by Lewis (2001) and is
discussed in detail in Lee (2004).
Carr and Madan (1999) notice that for very short maturities, the call value
approaches its non analytic intrinsic value causing the integrand in the Fourier
inversion to be highly oscillatory and therefore difficult to integrate numerically. To
mitigate this numerical inconvenience, the authors propose an alternative approach
which they call the Time Value method. The reasoning is quite similar to the previous
approach. But in this case the call price is obtained via the Fourier transform of a
modified time value, where the modification involves a hyperbolic sine function

Let U g denote the time value of an out-of-the-money option, i. e. for g Q L we

instead of an exponential function.

have the put price for U g and for g [ L we have the call price. Scaling $
1 for
simplicity, U g is defined by
#
U  g
   ( M   h ·M¸h,h¸$ C  M   h ·M¹h,h¹$ )!L dL , (4.44)
#

where · denotes the indicator function. Furthermore let º F be the Fourier

transform of U g
#
º F
 GHh U g dg. (4.45)
#

The prices of out-of-the-money options can be obtained by an inverse Fourier

transform

| 24
 1 #
U  g
 GHh º F dF.
26
(4.46)
#

expression for º F follows as

By substituting (4.44) into (4.45) and writing in terms of characteristic functions1 the

1 D  i D  F  i
º F
   —   ˜.
iF C 1 iF iFiF C 1
(4.47)

There are no issues regarding the integrability of this function as g tends to ∞ or ∞,
but the time value at g
0 can get rather steep as  Π0 which is making the Fourier

function sinhg
transform wide and oscillatory. This effect can be alleviated by considering a damping

#
¼ F
 GHh sinhg U g dg,
#
#
 ­h   ­h

 GHh
U g dg,
# 2
º F  i  º F C i

.
2
(4.48)

Hence, the time value of an option follows a Fourier inversion

1 1 #
U  g
 GHh ¼ F dF,
sinhg 26 #

1 1 #

 GHh ¼ F dF.
sinhg 6
(4.49)
$

Another refinement to the sharp peak at g

0 is due to the simple observation that
g Q L for puts and g [ L for calls unnecessarily creates a small discontinuity in
U 0 . By using ln r with r
  instead, U g is ensured to be continuous [see
McCulloch (2003)]. The Time Value modification is also designed for the use of the
FFT and claims more regular results for short maturities.

4.5 Lewis’ Approach

The Fourier pricing setup introduced in Lewis (2001) generalizes previous work on
Fourier transform methods2. Key to his contribution is the idea to express the option
value as convolution of generalized Fourier transforms and then apply the

uniformity for the normalization of D i . If the martingale property requires D i
  the
1
The formulation slightly differs from the one in Carr and Madan (1999), there seems to be no

expressions are all equal.

2
The pricing method was first developed in Lewis (2000) as ‘The Fundamental Transform’, however,
due to the detailed discussion and thorough derivations in Lewis (2001), the pricing framework is
commonly referred to as the Lewis (2001) methodology.

| 25
Plancherel–Parseval identity instead of integrating over a discounted transition
density times a payoff function. The transform representations of option prices may
be interpreted as contour integrals in the complex plane. By shifting the contours and

For each derivative there exists a known payoff function  at maturity. The
applying Residual Calculus it is possible to generate alternative pricing formulae.

payoff for a call option is L

 M    with L
ln  . Instead of transforming
the whole option price including the payoff function as described in Carr and Madan

representations in Fourier space  ½ U

_#  G‘M L dL. Unfortunately,  M   
#
(1999), Lewis (2001) emphasizes the fact that payoff functions also have

is an unbounded function and does not belong to \< . Thus, the defining integral is
½ U
OL  Œ ∞ and the Fourier transform does not exist. This
not finite as 
issue can be circumvented by considering an exponential damping factor, just like in

Fourier transform, since  M     ‘’M Œ 0 as L Œ ∞ for some appropriate U{ . This

the Carr and Madan (1999) approach. For a modified payoff we have a regular

corresponds to a Fourier transform along a line in the complex plane where the path
of integration is parallel to the real axis. A direct calculation for the call option payoff
yields to
#
½ U

  G‘M  M    dL ,
#
#

 G‘M  M   dL ,
89 %

 
G‘< M G‘M M.#

*—  ˜-
iU C 1 iU 89 %,

 G‘<  G‘

0—  ˜,
iU C 1 iU
 G‘<

 ,
U ,  iU
(4.50)

with U being a complex valued number. The upper limit L

∞ only exist under the
condition uU [ 1 which implies that the Fourier transform is well behaved only
within a certain strip of regularity Ž¾ in the complex domain ( G‘< has a branch
point at U
i). It is interesting to note that the transformed payoff for the put has the

U{ Q 0.
same functional form, but it is defined in a different strip in the complex plane with

| 26
 6

4
Re(CallPayoff(z))

1
0

-1
-2 1,8
-4
-3,7
-3,4

1,45
-3,1
-2,8
-2,5
-2,2
-1,9

Im(z)
-1,6
-1,3
-1
-0,7
-0,4
-0,1
0,2
0,5
0,8
1,1
1,1

1,4
1,7
2
2,3
2,6
2,9
3,2
3,5
3,8
Re(z)

5000
4000
3000
Re(PutPayoff(z))

2000
1000
0
-1000
-2000
-3000
-4000
-4
-3,6
-3,2
-2,8
-2,4
-2
-1,6
-1,2
-0,8

-0,1
-0,4
0

-0,2
0,4

-0,3
0,8

-0,4
1,2

-0,5
1,6
2

-0,6
2,4

Re(z)
-0,7
2,8

-0,8
3,2

-0,9
3,6
-14

Im(z)

options with uU [ 1. Lower figure: Strip for put options with uU Q 0.
Figure 4: Strip of regularity for transformed call and put payoff functions. Upper figure: Strip for call

In general, the strip of regularity is restricted to  Q uU Q  for typical derivatives

payoffs with ,  ƒ ] [see Table 1]. Given the generalized payoff transform the
corresponding inverse transform is then
1 G‘’ #
 L
½ U dU.
 G‘M 
26
(4.51)
G‘’ #

| 27
Table 1: Generalized Fourier Transforms for a variety of financial claims. The expression 1L  L¿
denotes the Dirac delta function and · is the indicator function.

L ½ U
OL 
 Ž¾
Financial Claim Payoff Function Payoff Transform Strip of Regularity

 G‘<
 M     uU [ 1
U ,  iU
Call

 G‘<
   M   uU Q 0
U ,  iU
Put

 G‘<
min M ,  0 Q uU Q 1
Covered Call;
Cash-Secured Put U ,  iU
 G‘
·“ ” À%  uU [ 0
iU
Cash-or-Nothing Call

 G‘
·“ ” Á% uU Q 0
iU
Cash-or-Nothing Put

 G‘<
 M ·“ ”À%  uU [ 1
iU C 1
Asset-or-Nothing Call

 G‘<
 M ·“ ”Á% uU Q 0
iU C 1
Asset-or-Nothing Put

1L  ln   G‘ entire U-plane

Delta function;
Arrow–Debreu
Money market 1 261U uU
0

Assuming that we have a well defined characteristic function D  U with U ƒ ŽE for

an arbitrary price dynamic and a transformed payoff  ½ U with U ƒ Ž¾ , we can
proceed by applying martingale pricing to obtain the option value
 $ , , 
   
L , (4.52)

  
G‘’ #

 w ½ U dUx,
 G‘M 
26 G‘’ #

   G‘’ #

½ U dU,

( G‘M )
26 G‘’ #

   G‘’ #

D  U 
½ U dU.
26 G‘’ #

Above we have stated that the extended characteristic function D  U with U ƒ  is

well defined in ŽE . Due to the reflection symmetry property for a real valued function
D  U
D  Uv , D  U is well behaved in the conjugate strip of regularity ZZZ
ZZZZ ŽE as
well, where the real U-axis is the line of symmetry. For the whole integral being well

| 28
behaved we need a U ƒ ŽÂ in a way that there is a intersection between ZZZ ŽE and Ž¾
ZZZ
(ŽÂ
ŽE ξ ). If the inversion contour is taken along this strip ŽÂ then the integral
converges absolutely and we may change the order of integration by Fubini’s
Theorem. The change of integration is required to take the expectation inside the
integral.

Strip of Regularity Ž¾ uU

Contour of Integration

U{
1

U
0 tU

Strip of Regularity ŽE

from the real axis to U ƒ ŽÂ in order to make both the option payoff (here the call) and the price
Figure 5: Contour of integration for the European call option. The contour of integration is shifted

process well behaved.

The valuation formula is obtained by considering the integration of a state price

density times the payoff function as a convolution representation in Fourier space by
a direct application of the Plancherel–Parseval identity
#
   G‘’ #
  
L !L qL
D  U 
½ U dU,
26
(4.53)
# G‘’ #

where D  U is the conjugate Fourier transform of the risk neutral density of the log
spot.
Applying the payoff transform of a call option (4.50) the call price is given by
   G‘’ #
dU
 $ , , 
  G‘h D  U ,
26 U,
 iU
(4.54)
G‘’ #

with g
ln %; C  in the phase factor  G‘h and U{ ƒ ŽÂ . With the definition of g,
/

Incidentally the same formula works for the put option value by imposing U{ Q 0.
the forward stock price is equal to the strike price for the option at-the-money.

The integrand in (4.54) is regular in the strip ZZZ

ŽE , except in the case where the
denominator U  iU is zero. This will happen at two poles, at U
0 and U
i. The
,

residue for the pole at U

i is

| 29
    G‘h D  U
Resi
lim wU  i —  ˜x,
‘ŒG 26 U ,  iU
(4.55)

D  i i i


.
26 26
By moving the integration contour to U{ ƒ 0,1 and according to the Residue
Theorem the call option value must equal the integral along uU
U{ minus 26i
times the residue at U
i which is . This is leading to a slightly different formula
   G‘’ #
dU
 $ , , 
$   G‘h D  U .
26 U ,  ÄU
(4.56)
G‘’ #

A convenient choice is to shift the contour by U{

,, because then the path of
<

U
F C , and employing the symmetry property for real valued functions gives
integration is symmetrically located between the two poles. The change of variables
G

#
√  / , i dF
 $ , , 
$  t Å GHh D  ¥F  ¦Æ ,
6 2 F, C 1
$ 4
#
√  / , i dF

$  t Å GHh D  ¥F  ¦Æ .
6 2 F, C 1
4
(4.57)
$

For this particular case the change of sign in the characteristic function does not
change the value of the integral as long as the sign in the corresponding phase factor
is changed as well. The first representation can be found in Itkin (2005) while in
Lewis (2001) the lower expression is deduced.

then defined by uU Q 0. By shifting the contour crossing both poles at U

0 and
For the put option a similar contour as in (4.54) can be used, however, the strip is

U
i we additionally pick up the residue for U
0 and find that
   G‘h D  U
Resi
lim wU —  ˜x,
‘Œ$ 26 U ,  iU
(4.58)

   D  0 i i  


 .
26 26
By Residue Theorem we then find an alternative formulation for the put option
   G‘’ #
dU
µ$ , , 
     G‘h D  U .
26 U,  ÄU
(4.59)
G‘’ #

Putting together the resulting formulae for calls and puts we obtain

      C µ, (4.60)
and receive the put call parity relation.

| 30
 By moving the contours to exactly uU
1 and uU
0 the integrals become
principal value integrals and by Residue Calculus one half of the residues are picked
up. Calculating the resulting two integrals at the two involved poles recover the Black-
Scholes like solution [see Lewis (2000), Lewis (2001) and Sepp (2003) for details].

4.6 Lipton’s approach

A similar approach to Lewis (2001) is described by Lipton (2002) in the context of FX
option pricing. Lipton presents his approach tailored towards some selected models.
First we state a general formulation
   # < i dF
 $ , , 
$   , D ¥F  ¦ ,
:GH ?h
26 2 F, C 1
4
#

   # < i dF

$  t Å :GH,?h D  ¥F  ¦Æ .
6 2 F, C 1
4
$
(4.61)

The derivation of the Lipton (2002) approach is similar to the reasoning of Lewis
(2001) by considering a Plancherel–Parseval style relation [see Lipton (2001)].
As an example, we derive the formulation for Black–Scholes model given in

have DÇ/
exp ÉiUÊ  , U , 3 , Ë with Ê
 , 3 , . Substituting U
F C , into DÇ/
Lipton (2002). Recalling the characteristic function for the Gaussian model (5.10) we
< < G

gives exp É  :F, C p?Ë and finally leads to

=> <
,
< < =>
   #
 :GH ?h:H > ? 
, p ,
Ç/ $ , , 
$  dF,
26 1
F, C4
(4.62)
#

   # < ><?=
>
dF

$  t w :GH,?h:H x .

1
p ,
6 $ F, C 4

We note that the Lipton formulation looks somewhat different to the Lewis approach,
and they are in fact equivalent. The connection lies in between the term √  > in
0

Lewis’ formula and the phase factor 

f
:GH ?h
> in Lipton’s formula. Expanding the last
term yields to  GHh  , then bringing 
f f
h h
> >

definition of g, and we get the results of Lewis’ representation.

outside the integrand and using the

4.7 Concluding Remarks and Recent Developments

In this section we revisited numerous Fourier inversion methods for the pricing of
European options. There are a number of similarities and parallelisms between those
prevailing concepts. The method developed at first in Heston (1993) offers the same

| 31
intuitive interpretation of the valuation problem as the classical Black–Scholes
formula and it was remarked the method of choice for Fourier pricing applications
until more efficient methods came up. The two methodologies from Attari (2004) and
Bates (2006) exploit the tight relation of the risk neutral exercise probabilities offering
a single integral solution with advantageous convergence properties. An algorithm
similar to Attari (2004) and Bates (2006) is developed in Wu (2008) by treating an
option value analogous to a cumulative density function.
In Lewis (2001), the direct relationship of the log spot characteristic function in
Image space is used from a more general point of view by representing the valuation
formulae as contour integrals in the complex plane. The impact of an additional
damping parameter is hereby characterized as the effect on the option prices
depending on the choice of a particular path of integration in the complex plane. In
this sense, the Carr and Madan (1999) methodology appears as a special case of Lewis
option pricing formulae. Another aspect is that the entire separation of Fourier
transforms for the underlying price dynamics and the payoff function allows for a
modular pricing framework which is facilitating particular valuation problems for
more advanced payoff structures.
A great effort is made by Lee (2004) in unifying and generalizing the previous
work of Duffie & al. (2000), Bakshi and Madan (2000), Carr and Madan (1999) and
Lewis (2001). In Duffie & al. (2000) four different payoff classes are considered and
are extended to the Carr and Madan (1999) framework. From Bakshi and Madan
(2000) the concept of a discounted characteristic function is adapted to allow for
stochastic interest rates. By recognizing the Carr and Madan (1999) formulation as a
contour integral [see Lewis (2001)] complementary valuation formulae are provided
by shifting the contour integrals in the complex plane and applying Residue Theory.
Moreover Lee (2004) provides error bounds for the discretized Fourier inversion and
discusses error minimization strategies.
Other authors have derived similar formulae for the valuation problem of
European options. For example, the approach chosen by Raible (2000) is nearly
identical to the Lewis (2001) methods by using a convolution representation for the
expectation of the payoff at maturity. The only difference is considering a two sided
(or bilateral) Laplace transform of a modified payoff function instead of a generalized
Fourier transform. His discussion also includes other European payoffs like power
and self-quanto options. The inversion integrals are then evaluated by means of the
FFT algorithm. Besides, methods for simple vanilla products using similar
assumptions like Raible (2000) or Lewis (2001), Borovkov and Novikov (2002)
develop an approach for the valuation of barrier type options based on the integration
of moment generating functions for general Lévy processes.
One purpose of this section is to give a detailed overview of the underlying
theory to value options with Fourier inversion methods. These methods are refined in
many ways for more elaborate valuation problems, or serve as building blocks to

| 32
other areas in the field of mathematical finance. In the following we concisely present
more methods and techniques which are tightly related to the discussed approaches.
As mentioned previously, the Fourier transform based approach works for a lot
of European payoff structures. Agliardi (2009) provides a comprehensive
characterization for Compound Options, Multi-period Power Binaries, Power
Digitals, and Supershare Options including relevant strips of regularity for arbitrary
Lévy processes. While the mathematical foundations in Section 3 suffice for the
payoff functions considered in this paper, more elaborate payoffs or multidimensional
applications need a much more detailed consideration. A systematic analysis is
offered by Eberlein & al. (2009) of the conditions required for the existence of Fourier
transform formulae in a general framework, i. e. when options have arbitrary payoff
functions and possibly depend on the path of the stochastic process.
Dempster and Hong (2000) introduce an integration methodology based on
double fast Fourier transforms for spread option pricing that is efficient for geometric
Brownian motion and more sophisticated price processes. A recent contribution for
the use of Fourier transform methods for spread options comes from Hurd and Zhou
(2009) by expressing the spread option payoff in terms of the gamma function and
applying the FFT technology from Carr and Madan (1999).
The introduced Fourier techniques have been extended in many different ways,
including the pricing of path dependent derivatives. For instance Cardi (2005)
develops a valuation formula to discrete Barrier options based on the Lewis (2001)
methodology. Use of FFT methods to price discrete Asian options is considered in
Benhamou (2000).
A versatile numerical method to the valuation problem of early exercise options
and certain path dependent options by means of a quadrature method (QUAD) is
introduced in Andricopoulos & al. (2003). By recognizing that the payoff of an option
can be segmented, the integration of the payoff is only carried out over continuous
segments of the payoff and evaluated with numerical quadratures. However, the
method requires analytical transition densities like the Gaussian. To relax this
requirement O’Sullivan (2005) introduces the QUAD-FFT method, by considering
the one to one connection from density functions with their characteristic functions
makes the QUAD method applicable for a wide range of stochastic processes. Exotic
features can be incorporated by applying the option pricing formula recursively. The
Convolution method (CONV) developed in Lord & al. (2007) follows a similar idea
by considering option values as a convolution of a payoff function with a probability
distribution and taking the QUAD and QUAD-FFT methods into Fourier domain.
In an empirical study about Lévy option pricing models by Ji and Zapatero
(2008) the authors extend the quadratic approximation method from Barone-Adesi
and Whaley (1987) which approximates the early exercise premium of American
options under Black–Scholes assumptions to exponential Lévy models. American
options are then the sum of the European Fourier prices and the approximated early
exercise premium.

| 33
 It must not be the case that the characteristic functions are known in closed
form, for instance the Heston (1993) characteristic function is an analytic solution to
a system of ordinary differential equations (ODE). In fact, these ODE’s can be solved
numerically, too, which is often the only choice for similar advanced price dynamics
[see e. g. Dobránszky (2009)]. Fourier pricing is even applicable in non parametric
models by replacing the unknown characteristic function with its empirical version,
e. g. Binkowski (2007) infers the empirical characteristic functions from quoted
option prices and compares the results to some parametric cases.
Kruse and Nögel (2005) extend the Fourier valuation method to European
forward starting options assuming the stochastic volatility framework initially
introduced in Heston (1993) by a change of measure technique. Nunes and Alcaria
(2009) amend the two dimensional Kruse and Nögel (2005) approach to a single
integral solution based on the Attari (2004) formulation.
Fourier inversion methods are applicable for the efficient calculations of
derivatives in other frameworks as well. By noting that the call and put like payoffs
also occur frequently in insurance industry Dufresne & al. (2009) consider the
Plancherel–Parseval identity to calculate stop-loss or excess-of-loss premiums. In
addition to the exponential damping ensuring the integrability of the payoffs, they

Compound Poisson, (Generalized) Pareto and -stable distributions.

additionally examine polynomial damping factors and illustrate their findings with

Another important area is the valuation of interest rate derivatives. Attari (2004)
is also considering his modification to interest rate products. Andreasen (2006)
develops a full yield curve model with stochastic volatility following the work of
Andersen and Andreasen (2002) and using results from Heston (1993), Lewis (2000)
and Lipton (2002). The resulting model resembles a shifted Heston model which
allows the use of the introduced Fourier inversion techniques and gives good fit to
observed cap and swaption prices.
In Bouziane (2008) a general pricing framework is established for the pricing of
interest rate derivatives based on the methodology introduced in Lewis (2001). In this
flexible framework a wide range of interest rate derivatives are priced with either
conditional or unconditional exercise rights. He further develops an efficient method
to price derivatives for a whole range of strike prices similar to the FFT methodology
from Carr and Madan (1999) named the IFFT algorithm.
Fourier based methods also gain increasingly popularity in the field of
defaultable assets. For example, Grundke (2007) presents an integrated market and
credit portfolio model and extends the CreditMetrics model by correlated interest rate
and credit spread risk. Sepp (2006) generalizes the CreditGrades model for stochastic
volatility and jump diffusion specifications, estimates default probabilities using
equity options, and applies these models for the modeling of credit default swap and
equity default swap spreads. Stochastic volatility frameworks for credit and interest
rate derivatives using numerical quadratures can be found in Tahani (2004) and
Tahani and Li (2007).

| 34
 In Gatheral (2006) an overview of applications to volatility derivatives is given.
Further, he shows how to infer at-the-money Black–Scholes implied volatilities or the
at-the-money Volatility Skew directly from characteristic functions without the need
of first calculating the price and then using a root finding algorithm in conjunction
with a Black–Scholes pricing engine. The exposition in this section resembles only a
small subset of possible applications, but gives an idea of the versatility of Fourier
methods in financial modeling.

5 Applications of Inverse Fourier Methods to Distribution

Functions and Option Pricing
In this section we will illustrate some of the most widely used concepts of numerical
evaluation of inverse Fourier integrals. Two methods will be compared. On one hand
the direct integration methods, on the other hand the fast Fourier transform
algorithms are considered. Each of these procedures has their own advantageous
properties compared to the other routines. In any case, the application of the
procedures requires considering numerical and technical issues with precautions.

5.1 Numerical Evaluation of Inverse Fourier Integrals

To obtain probability densities or option prices, the inversion of the Fourier integrals
can be evaluated by means of standard numerical integration schemes. Numerical
integration of functions is a broad subject and rich in terms of possible techniques. In
this section we will briefly review some of the most frequently used numerical
quadrature schemes to evaluate the semi-infinite domain of oscillatory functions
encountered in Fourier analysis.

Direct Integration Methods

It is well known that an integral can be represented as the area under the curve of a
function in between a specified interval. A powerful procedure is to approximate the
searched interval by simple geometrical objects like rectangles. More sophisticated
rules are the Newton–Côtes formulae where the interval is approximated by some
interpolating polynomials usually in Lagrange form which in turn are easily
computable. The Trapezoidal and Simpson’s rules are probably the most popular and
commonly used Newton–Côtes integration schemes available. While having a clear
geometrical and intuitively obvious interpretation, there are alternatives which offer
higher computational efficiency, namely the family of Gaussian quadratures. A
Gaussian quadrature tries to give an accurate estimate for an integral by optimally
choosing abscissas where to evaluate a particular function. Following the
Fundamental Theorem of Gaussian Quadrature the optimal abscissas are exactly the
roots of orthogonal polynomials for the same interval and weighting function. In the
sequel we implement and compare a number of widely used algorithms.

| 35
quadrature formula usually specified for an integration domain over 1,1. A Ì-
Gauss–Legendre quadrature formula This is perhaps the best known Gaussian

point quadrature is then given by

< Ï<

K L dL Í Î { KL{ . (5.1)
< {.$

function L
1. Thus the nodes are the Ì zeroes of the Legendre polynomial of
The resulting orthogonal polynomials are the Legendre polynomials for the weighting

degree Ì.
The integration interval may be changed to arbitrary interval b, c by following
change of variables

cb cCb cb

„ Ï<

KL dL Í Î { K ¥ C L { ¦.
2 2 2
(5.2)
e {.$

Proceeding this way the quadrature is applied to the transformed integrand with
modified integration limits.
Depending on the actual problem it may be an appropriate procedure to apply
the Gauss–Legendre formula not on the whole interval but decompose the integration
region into subintervals and sum up the particular approximations
# Ï< ›< Ñ
Ð L
K L dL
lim Î K L dL.
ό#
(5.3)
$ ›.$ ›Ñ

The fixed step size Ò denotes the discretisation length and Ì is the truncation point.
As Ì Œ ∞ and Ò Œ 0 the numerical quadrature will approach the theoretical value.

Gauss–Laguerre quadrature formula The Gauss–Laguerre quadrature evaluates an

exponentially weighted integral from zero to infinity as
# # #
K L dL
 M  M KL dL
{  M K L dL, (5.4)
$ $ $
Ï< Ï<

Í Î {  KL{
Î { KL{ .
M’

{.$ {.$

function L
 M . This quadrature is particular interesting since all of our
The resulting orthogonal polynomials are the Lagrange polynomials with weighting

applications have semi-infinite domains of integration ranges.

Adaptive Simpson quadrature The idea of using adaptive algorithms for the
numerical calculation of integrals is dating back to at least McKeeman (1962). Since
then many new and sophisticated algorithms have been developed, among these

| 36
Gander and Gautschi (2000) propose an efficient and reliable implementation of the
adaptive Simpson rule.
The adaptive control strategy divides the area of integration in subintervals,
evaluates the integral at each region and uses an error estimate of the integral to check
if a specified error tolerance is met. At regions where the function is well
approximated by a quadratic function, only a few function evaluations are needed, in
other areas the adaptive strategy evaluates the subintervals in a recursive manner.

Adaptive Gauss–Lobatto quadrature The second quadrature implemented in the

paper from Gander and Gautschi (2000) builds upon a Gauss–Lobatto quadrature
modified by a Kronrod extension to add an effective error control procedure. The
Lobatto formula has preassigned abscissas at the end points of the interval. The
remaining nodes and all weights are then determined in a way to obtain the highest
exactness possible. The Kronrod extension is used to provide an estimate of the
approximation error. If the error exceeds a specified tolerance, regions where the
function is not so well behaved will be divided over again and again.

Adaptive Gauss–Kronrod quadrature The Gauss–Kronrod quadrature is an

extension of the Gauss–Legendre algorithm. For well behaved regions the error
estimates will be small and the numerical approximations will be accepted. For
integration areas where the function is not smooth, the algorithm subdivides the
interval until the desired accuracy will be reached.

For detailed discussion on numerical quadratures we refer to the extensive literature

available on this subject. Key reference on numerical integration is Davis and
Rabinowitz (1984). Detailed descriptions in a financial context can be found in Fusai
and Roncoroni (2008) and Judd (1998).

Fast Fourier Transform and Fractional Fourier Transform

Fourier transform (DFT) which maps a vector Ó

қ Ï<›.$ onto some vector Ôh Ó
The Fast Fourier Transform (FFT) is a highly efficient implementation of the discrete

Ï<
,ˆ
Ôh Ó
Î  G Ï ›h қ , g
0, … , Ì  1 (5.5)
›.$

By choosing Ì as a power of two, the complexity of the algorithm reduces from an

order of Ì , for the direct numerical integration methods to that of Ì log , Ì
operations to compute Ì values. The Fourier transform of the vector Ó will be
denoted with Ôh Ó ƒ Ï . The vector Ó ƒ Ï corresponds to Ì values of the
function Ò evaluated at the points ×
F› Ï<
›.$ .
A result of the FFT procedure will be a new vector Ø
Kh Ï<
h.$ , where each of
the elements Kh correspond to a specified value of L
Lh . For example, for the

| 37
purpose of density evaluations Lh will typically be chosen in a way that the values are

demonstrate that the vector Ø will contain option values corresponding to carefully
located around the mode. In the context of option pricing, Carr and Madan (1999)

chosen Ù
Lh Ï< h.$ in a sense that the resulting values are around the at-the-money
level, i. e. the values of L are typically kept close to zero.
Following Carr and Madan (1999) we denote the grid size of the vector × with Ú
and Û is the spacing of Ù. In the case of conventional FFT procedures Ú and Û are
restricted by imposing an inverse relation
26
ÚÛ
.
Ì
(5.6)

The finer the summation grid spacing Ú, the coarser the output spacing returned
from the FFT procedure will be and vice versa.

The fundamentally inflexible nature imposed by the restriction (5.6) was addressed by
Bailey and Swarztrauber (1991, 1994) who overcome this constraint by developing a
generalization of the conventional DFT, the Fractional Fourier Transform (FRFT).
It is defined as
Ï<

Üh Ó, 
Î  G,ˆ›h­ қ , g
0, … , Ì  1 (5.7)
›.$

The fractional parameter  may be any arbitrary real or complex valued number. In
this framework both Ú and Û may now be chosen freely which will together determine
the FRFT parameter 
ÚÛ. The ordinary FFT and its inverses are special cases of
the FRFT for 
. A Ì-point FRFT can be implemented by invoking two forward
<
Ï
and one inverse 2Ì-point FFT calculations. The algorithms works as follows, define
two 2Ì-point sequences Ý and Þ as
‚›
қ  Gˆ› 0`ßQ
>­

‚›
0 à ` ß Q 2

U›
 0`ßQà
(5.8)
Gˆ› > ­

U›
 Gˆ›,€ à ` ß Q 2
>­

The FRFT is then given by

Üh Ó, 
 Gˆh ­ Ôh< á(ԛ Ý áԛ Þ ), 0 ` g Q 
>
(5.9)

where Ôh< denotes an inverse FFT calculation and á denotes element-by-element

integrated and may be precomputed. The computational cost of a Ì-point FRFT

vector multiplication. The exponential factors do not depend on the function to be

procedure is about a 4Ì-point FFT in terms of the number of elementary operations.

The additional benefit by employing a FRFT instead of running a single FFT is due to

| 38
the fact that a FRFT with smaller Ì may achieve the same accuracy as using a FFT
with much larger Ì.

5.2 Examples of Characteristic Functions for Particular Price

Dynamics
In order to illustrate the basic concepts and investigate the standard numerical
procedures we will briefly introduce three candidates of exponential Lévy processes
and exponential affine models. We just state the corresponding characteristic
functions and refer to the respective references for detailed discussion and
derivations. Comprehensive treatments on Lévy processes and affine models with
Fourier pricing applications can be found in Cont and Tankov (2004), Boyarchenko
and Levendorskiĭ (2002) and Schoutens (2003).

Black–Scholes Our benchmark model and one of the simplest examples of Lévy
processes and affine (jump) diffusion is standard Brownian motion. This is the only
continuous process from the wide class of Lévy processes. In this sense the classical
Black–Scholes model can be categorized as a continuous exponential Lévy model.

variance 3 , is given by
The characteristic function for an exponential Brownian motion with instantaneous

3 , F,
D  F
( GHE0 )

exp wiÊF  x.
2
(5.10)

To prevent arbitrage the drift correction parameter Ê must be chosen in a way that
the discounted price process remains a martingale. By imposing D  i
 M0 

1 we find that Ê
 , 3 , . The analytic continuation of the characteristic function is
<

defined by the whole complex plane.

Finite Moment Log Stable This Lévy process with infinite activity was proposed by
Carr and Wu (2003) to address the observation for S&P 500 index options that the
volatility skew does not flatten as time to maturity increases. The characteristic

6
function is described by

D  F
( GHE0 )
exp ÉiFÊ  iF3 ­ sec Ë,
2
(5.11)

where Ê
3 ­ sec , is the convexity adjustment term to sustain the martingale
ˆ­

property. The tail index  ƒ 1, 2 controls the tail behavior and 3 describes the
width of the risk neutral density. For 
2 the Finite Moment Log Stable model

3Ç/ is related to the dispersion measure for the FMLS model 3â/ via 3Ç/
3â/ √2.
(FMLS) coincides with the Black–Scholes model, where the Black–Scholes volatility

The domain for the strip of regularity ŽE is restricted to uU Q 0.

| 39
 1,06
1,04
1,02
0,10
1,00
Re(phiFMLS(z))

Im(phiFMLS(z))
0,05
0,98
0,00 -2,0
0,96 -1,4
0,94 -0,05 -0,8
1,2 -0,2
0,92 -0,10
0,1 0,4

-2,0
0,90 1,0 Im(z)

-1,4
-1,0 Im(z)

-0,8
1,6

-0,2
-2,0
-1,6
-1,1

-2,0

0,4
-0,7
-0,2
0,3

1,0
0,7
1,1

1,6
1,6
2,1
Re(z)
Re(z)

Model parameters: 
1.6 and 3
0.1.
Figure 6: Finite Moment Log Stable model in the complex plane. Left: Real part. Right: Imaginary part.

An intuitive way to illustrate the strip of regularity ŽE for a given model is to plot the
characteristic function by the movements of its real and imaginary parts on the plane
of complex numbers. By varying the state variables and model parameters it is then
possible to gain some insights on the model inherent dynamics and properties. In
Figure 6 and Figure 7 this is shown for the FMLS and Heston (1993) model
respectively for some arbitrarily chosen parameters.

Heston (1993) The Heston dynamics are one of the most prominent stochastic

square root process. More precisely the instantaneous variance ²

3 , is allowed to
volatility specifications where the instantaneous variance follows a mean reverting

be stochastic and time dependent and as such relaxing the assumption of stationary
increments. The characteristic function for the Heston (1993) is
D  F
( GHE0 )
expäF C åF ²$ , (5.12)
where äF and åF
®¯ } ç  1
ä F

æ   q   2ln — ˜è,
3±, }1
(5.13)

  q 1   ç
å F
— ˜,
3±, 1  } ç
(5.14)

with

q
‹ ,  4~¼, (5.15)
q
}
,
Cq
(5.16)

and auxiliary variables

1 1
~
 FF C i , 
®  iF3± °, ¼
3±, .
2 2
(5.17)

| 40
instead it is contained in the definition of ~. The Heston model belongs to the class of
There is no explicit convexity correction for the diffusion process in this specification,

characteristic function for L

ln  and ², conditional upon the instantaneous value
affine diffusions. These are defined in a way that the logarithm of the conditional

The domain for the extended characteristic function is the strip ŽE given by the
of the stochastic volatility state variable, is a linear function of these state variables.

interval b , b , where b Q 0 and b [ 1 solve [see Lee (2004)]

}ib  çGe 
1. (5.18)

time-homogenous process and hence the analytical strip ŽE does not depend on the
For the FMLS model the price process follows a Lévy process which is by definition a

number of finite moments. For the Heston model, however, without stationary
increments the number of moments is not invariant to the time horizon. The time
evolution of the characteristic function is plotted in Figure 7 with two different time
horizons on the Gauss plane. As the time increases the number of moments decrease.
In other words the singularities move along the imaginary axis toward the real axis
which in turn narrow increasingly the strip of regularity. This issue is known as the
moment stability of affine stochastic volatility models and is especially important for
long dated options. For certain circumstances under which the moments of the price
process can explode (become infinite) in finite time, compare e. g. Lord and Kahl
(2007), Andersen and Piterbarg (2007), Keller-Ressel (2008).
Although known for some time now, there are still substantial research efforts
made to gain a better understanding of this model. For a recent review and some
lesser known features of the Heston (1993) model see Zeliade Systems (2009).

| 41
 4
3
Abs(phiHeston) 2
1
0

-10
-8

9
-5

8
6
-3

4
3
0

1
0
2

-2
Re(z) Im(z)

-4
-5
7

-7
-8
9
-10
8
6
Abs(phiHeston)

4
2
0
-10
-8

9
-5

8
6
-3

4
3
0

1
0
2

-2
Re(z)
4

-4

Im(z)
-5
7

-7
-8
9
-10

maturity. Model parameters: ®

1, ¯
0.02, °
0, 3±
0.4 and ²$
0.02.
Figure 7: Heston (1993) model in the complex plane. Top: One year to maturity. Bottom: Eight years to

Depending on the valuation formula [e. g. the Carr and Madan (1999) formulation]
the above specified characteristic functions might have to be rewritten into the form
of
expln $ C D  F . (5.19)
Having specified three different model specifications we are now ready to explore the
pricing formulae and density functions using Fourier inversion methods.

5.3 Distribution Functions

Before considering option pricing applications we will investigate how to recover
distribution functions from tractable characteristic functions. On one hand this is per
se an interesting and important topic for applications to statistical inference and on
the other hand option prices are obtained by integrating a payoff function according
to a particular probability density.

Direct Integration
In this part we will first have a glance of some qualitative properties of the different
quadrature schemes and chosen models. Then the different integration algorithms are
compared with respect to accuracy and computational efficiency.

| 42
 f(x)
0,4

0,3

0,2

0,1

0,0

-0,1

-0,2
-6,0
-5,6
-5,2
-4,8
-4,4
-4,0
-3,6
-3,2
-2,8
-2,4

38
-2,0
-1,6
-1,2
-0,8
-0,4 32
0,0
0,4
0,8
1,2
26

1,6
2,0
2,4
2,8
3,2
20

3,6
4,0
4,4
4,8
x 14 Legendre points

5,2
5,6
6,0
8

Figure 8: Standard Normal Density function as a function of Legendre points.

In Figure 8 the density of a Standard Normal random variable is plotted as a function

of Gauss–Legendre integration points. It can be seen that the center of the density is
captured quite well by only a few points, however, by simple observation we can see
that for the tails of the probability function a lot more points are necessary to achieve
a minimum degree of accuracy. The oscillatory behavior is typical for Gaussian
quadratures where each quadrature has its own characteristics and purposes, a
Gauss–Laguerre integration mimics qualitatively the same picture only distinguished
by a slightly different curvature.
In order to assess the accuracy of the above mentioned numerical integration
schemes we begin with the ubiquitous Black–Scholes model as a benchmark to test
against. Before being interested in high precision results we show some qualitative
properties of two basic Gaussian quadratures, namely the Legendre and Laguerre
integration.

| 43
 10
BS PDF
8 Legendre
Laguerre
6

0
-60% -40% -20% 0% 20% 40% 60%
-2

1,5E-01
64 - Legendre
18 - Laguerre
1,0E-01

5,0E-02

0,0E+00
-60% -40% -20% 0% 20% 40% 60%
-5,0E-02

-1,0E-01

-1,5E-01

Figure 9: Black–Scholes density function. Top: Density with time to maturity 0.05 and volatility of
20 %. Bottom: Absolute error of 64 point Gauss–Legendre and 18 point Gauss–Laguerre integration.

In the exhibit above the density function for the Black–Scholes model is shown. As
one can see, the Legendre quadrature has increasingly ‘good’ fit in the tails whereas
the Laguerre integration shows a better fit in the center of the density. Note the
difference in this example, the Laguerre quadrature with only 18 points shows similar
quantitative properties in magnitude as the Legendre quadrature with 64 points. This
indicates that the Gauss–Laguerre scheme seems to cope quite well with this kind of
problems. Choosing 64 points for the Laguerre nodes result in an absolute error well
below 1e–12 in the range of [–0.5, 0.5]. To receive the same precision for a simple
Legendre quadrature more than 170 nodes are required.

| 44
 10 1,2
T=0.5 T=0.5
8 1
T=0.05 T=0.05
0,8
6
0,6
4
0,4
2 0,2
0 0
-60% -40% -20% 0% 20% 40% 60% 0 20 40 60 80 100 120

Figure 10: Black–Scholes probability density function for two different times to maturity and volatility
of 20 %. Left: Probability density functions. Right: Fourier integrands.

As previously noted, there is an inverse functional relation between space domain and
image domain. A steep function in space usually becomes wide in the characteristic

integrands. For increasing  the return densities become smoother and have fatter
function, a behavior which becomes apparent by the inspection of typical Fourier

tails, which lead to thinner tails of the characteristic function making it less well
behaved. This effect will differ for various price processes depending on their
decaying behavior towards infinity. This can lead to serious consequences for
numerical procedures, which will be discussed in detail later.

1,0 0

0,8 -0,002 0 20 40 60 80 100 120

0,6 -0,004
0,4 -0,006
T=0.5
0,2 -0,008
T=0.05
T=0.5
0,0 -0,01
T=0.05
-60% -40% -20% 0% 20% 40% 60%
-0,012

Figure 11: Black–Scholes cumulative density function for two different times to maturity and volatility
of 20 %. Left: Cumulative density functions. Right: Fourier integrands.

corresponding Black–Scholes density and Fourier integrands. For choosing  Q 2 the

In Figure 12 the density function for the FMLS model is illustrated together with the

model inherent skewness to the left property is clearly visible and as one can see the
tail on the left of the density decays slower than the log normal commonly referred to
as a fat tail.

| 45
 20 1,2
FMLS PDF real
1,0
15 BS PDF imag
0,8
10 0,6
0,4
5
0,2

0 0,0

-20% -10% 0% 10% 20% -0,2 0 20 40 60 80 100 120 140 160 180 200

Figure 12: FMLS density and integrand. Left: Density with 

0.05, 
1.6 and 3
0.1 together
with the corresponding Black–Scholes density. Right: Real and imaginary parts of the FMLS integrand
for the density approximation.

In Section 3.2 four different formulae are presented which allow to approximate the
cumulative density function. These are now compared with respect to convergence
properties and efficiency with the help of the Heston (1993) model specification.

0,008 2 0,010
real imag real
0,006 0
0,005 imag
0,004 -2
0,002 -4 0,000

0,000 -6 0 20 40 60 80
-0,005
-0,002 0 20 40 60 80 -8
-0,010
-0,004 -10
-0,006 -12 -0,015

Figure 13: Integrands in the Heston model with different inversion formulae. Left: Inversion formula
using the real part [Eq. (3.15)]. Right: Inversion formula using Eq. (3.14).

From Figure 13 we see the different curves of the integrands employing different
Inversion formulae. The exhibit on the left shows the integrand using the real part of
the function evaluation [Eq. (3.15)], using the imaginary part shows exactly the same
picture [Eq. (3.16)] with the exception that the real and imaginary parts are simply
interchanged. Using (3.14) the resulting integrands for the real parts pretty much
resemble the same behavior, only mirrored and slightly different in magnitude. Using
the composite Gauss–Legendre integration the CPU times with respect to equations
(3.15) and (3.16) are virtually identical (averaging over 100 density evaluations from –
0.5 up to 0.5 needs 0.035 seconds in CPU time), for (3.14), however, the running time
nearly doubles which seems logical since with this representation the quadrature has
to evaluate the characteristic functions two times (0.067 seconds on average). The
representation as in (3.21) behaves similar to (3.14), but needs only the timings like
equations (3.15) and (3.16) since the integrand needs to be evaluated only once.
The number of evaluations of the characteristic function is clearly the
determining factor and is especially pronounced in the case of the Heston (1993)
model involving two complex exponentiations (without counting repeating terms),
one complex logarithm and one complex square root.

| 46
 Now we turn to the issue of getting precise values for the desired Fourier
integrals. The analytic Black–Scholes model serves as a benchmark for the different
integration methods. As is well known, even for the Black–Scholes model which is
generally considered analytic or closed form, we need a numerical approximation for
the required Cumulative Standard Normal Distribution. Following the exposition in
West (2005) we rely on the therein proposed Hart (1968) algorithm which is accurate
to double precision throughout the real line. In Figure 14 we summarize the main
results in terms of the achieved accuracy versus CPU time in seconds. The most
efficient algorithms will be located in the left down corner. As a measure for the error
we take the root mean squared errors (RMSE) resulting from averaging the PDF
calculations from –0.6 up to 0.6 involving 121 values. For the generation of the Gauss
nodes and the adaptive Gauss–Kronrod quadrature we use the routines implemented
in the ALGLIB library3. The adaptive Simpson and Gauss–Lobatto procedures are
direct implementations of the algorithms described in Gander and Gautschi (2000).

1,00E-07

1,00E-08

1,00E-09 10E-10

comp10 ^15
1,00E-10 h=1 h=0.5
comp10 ^12
h=2 h=1
h=2 h=0.5
1,00E-11 10E-10 comp20 ^12
Leg256
RMSE

Leg128
Lag100 Lag128 Lag160 comp20 ^15
1,00E-12 Lag185
Lag64 Legendre
Leg160 Leg192

1,00E-13 Laguerre
h=2 h=2 h=0.5 h=0.5
adSimpson
h=0.1
1,00E-14 10E-10 h=1 h=1 adLobatto
10E-14 adKronrod
1,00E-15 10E-15

1,00E-16
0,001 0,01 0,1

CPU time (seconds)

Figure 14: Accuracy vs. CPU time. Accuracy measured as root mean squared error (RMSE) in relation
to the computation time in seconds (CPU time) on a double log-scale.

For the Gauss–Laguerre and Gauss–Legendre quadratures the number of integration

points is included, for the composite Gauss–Legendre the number of points per
subinterval is indicated in conjunction with the corresponding local stopping criteria.
The adaptive schemes are indicated with the desired tolerance level for the routines
beginning with 10e–10 upwards.
Having expected a more regular pattern with downward sloping curves from the
upper left to the right down corner we see that this behavior is not given in all cases.

3
Available at http://www.alglib.net/

| 47
approach zero as the step size Ò tends to zero. One potential reason for this contrary
For instance the error arising from the composite schemes should theoretically

and seemingly irregular behavior may be traced back to the increasing number of
numerical operations which can accumulate to a substantial amount of rounding
errors working within a finite precision computation environment. The number of
significant digits for the nodes and weights for the Gauss quadratures in double
precision floating point arithmetic are restricted to 16 digits. Decreasing the step size
any further will increase the rounding errors dramatically leading to completely
meaningless results. Using composite Newton–Côtes procedures this effect is less
pronounced (not shown here), since the subintervals are approximated by analytic
functions.
Simply increasing the number of integration nodes for the plain Gauss
quadratures also does not lead automatically to more precise values. This will depend
heavily on the properties of the used polynomials, i. e. whether the orthogonal
polynomials are capable to approximate the evaluated function well enough [compare
Figure 9] and as we can see choosing higher order schemes need not necessarily
imply higher accuracy. In addition, there are numerical impediments as well, e. g. for
double precision the highest number of nodes for the Gauss–Laguerre scheme that
can be generated is 365, otherwise the algorithm will produce an overflow [see e. g.
Dobránszky (2008)].
All considered adaptive procedures perform quite well, as expected the Gaussian
schemes are far superior to the Simpson rule albeit this routine works remarkably
well. In terms of computational efficiency the best overall performance is achieved by
the adaptive Gauss–Kronrod algorithm, whereas the most precise values are obtained
with the adaptive Gauss–Lobatto routine.

FFT and FRFT

The FFT and FRFT are computationally efficient algorithms for the calculation of
density functions. In this section we address some implementational issues and assess
the accuracy with respect to analytical solutions.
For the PDF inversion we need to discretize the Fourier integral
1 #
KE L
t( GHM DE F )dF,
6 $

1
Ï<

Í Î t( GHé M DE  F› £)Ú.

6
(5.20)
›.$

where Ú denotes the step size for the summation grid.

summands for the integral approximation Ì which should be a power of 2 to use the
In order to calculate the integral in (5.20) we need to specify the number of

computational efficiency of the FFT and the size of Ú determining the grid spacing in
Fourier domain. Both values will determine the effective upper bound for the

| 48
integration b
ÌÚ. The FFT returns Ì values of L, where Lh
c C Ûg with grid
spacing of Û
26/ ÚÌ yields the return grid. This implies that the returned values
are within the range c, c centered around zero with c
ÌÛ/ 2. Further setting
F›
Úß the PDF finally becomes

1
Ï<

KE Lh Í Î t( Gêë›h  GHé „ DE  F› £Ú› ),

6
(5.21)
›.$

where › are some weightings implementing the chosen integration rule. Common
choices are the Trapezoidal rule with integration weights ›
for ß
0 or
<
,
ß
Ì  1 and 1 otherwise, and the Simpson’s rule weightings ›
(3 C
<
n
1ßC11ß where 1 is the Kronecker delta function which is 1 for 
0 and zero
otherwise. Running a FFT then delivers an approximation of the discretized Fourier
integral.
As we know from Section 3.2, the probability density function is an absolutely
integrable function and does not impose any major impediments in conjunction with
FFT methods. For the cumulative density functions, however, this is not the case. The

the required evaluation point F

0. In order to circumvent this problem it is possible
FFT cannot directly be applied to recover the CDF, since the integrand is singular at

by some carefully chosen parameter  in a way that the integrand is well behaved and
to proceed in exactly the same way as shown in Section 4.4 by damping the function

If there is a damping constant  [ 0 such that the characteristic function under

directly applicable for the FFT [see e. g. Kim & al. (2009)].

consideration is absolutely integrable for all complex U with uU

, then we are
able to define a modified CDF
r­ L
 ­M sI ` L . (5.22)
Just like the modified option price we can define a modified characteristic function
1
D­ F
D iF C  .
  iF E
(5.23)

The cumulative density is then given by applying the Fourier inversion to the
dampened transform and finally undamping the inverted function
 ­M #
r L
 GHM D­ F dF,
26 #

 ­M #

t( GHM D­ F )dF.
6
(5.24)
$

Using this representation in conjunction with a discretization scheme as in (5.20) and

(5.21) we are now able to use the FFT to evaluate the cumulative density function for
a whole range of values in one single stroke.

| 49
 1,6 0,40
real real
1,4 0,35 imag
imag
1,2 0,30
1,0 0,25
0,8 0,20
0,6 0,15
0,4 0,10
0,2 0,05
0,0 0,00
-0,2 0,0 5,0 10,0 15,0 20,0 25,0 30,0 -0,05 0,0 5,0 10,0 15,0 20,0 25,0 30,0

Damping factor 
0.75. Right: Damping factor 
2.75.
Figure 15: Damping of the Fourier transform for the cumulative density function. Heston model. Left:

and is therefore not suited for the FFT. With increasing  the integrand becomes less
Without the damping constant the integrand exhibit divergent behavior around zero

[see Figure 15]. However, choosing  too big will render the numerical integration
steep and smoother, aiding the numerical quadrature to evaluate the integral precisely

unstable as well. This relationship is illustrated in Figure 16 by contrasting the FFT

there is a range of values for  delivering quite accurate results and will then slowly
calculations with the analytical solutions in the Black–Scholes model. As we can see,

worsen for  getting bigger. The minimum error in this example is around 
1.5.

1,0E+00

1,0E-02

1,0E-04

1,0E-06
RMSE

1,0E-08

1,0E-10

1,0E-12

1,0E-14
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
alpha

of damping factor  in the Black–Scholes model Ì

4096 and Ú
0.25.
Figure 16: Root mean squared error (RMSE) of FFT cumulative density approximations as a function

Having specified how to distretize the Fourier integrals and how to overcome the
impediments for the FFT with respect to the cumulative densities, we will now give
some practical examples. In the following we adapt the example in Chourdakis (2005)

volatility. For instance using Ì

4096 and Ú
0.25 for the FFT we get an output
for densities in the Black–Scholes model with time to maturity 0.25 and 30 %

spacing of Û
0.00614 and an upper integration bound of 1023.75. For the FRFT we

Ì
128 and Ú
0.25 implies an output grid with Û
2 ð <,ò
0.009375 and an
set the desired output range for the density values to [–0.6, 0.6], for example using
$.ñ

upper bound of 31.75 (with the restriction imposed by the FFT the output spacing in

| 50
this case would be Û
0.19635). The fractional parameter is determined by 

ëê
,ˆ
and results in 0.0078125 for aforementioned parameters. In Table 2 we list the

parameters and other combinations of Ì and Ú.

resulting errors and corresponding running times for the above mentioned example

The error measure is calculated only at grid points belonging to the FFT and
FRFT sampling grid. Preceding in this way we can avoid interpolation and hence no
interpolation error will affect the results. The results show that the two methods are
capable of returning very accurate values for the density function compared to
analytical values. Further, as expected, the runtimes of the FRFT are usually a lot
faster than the corresponding FFT (using the FFT implementation in the ALGLIB
library).
Table 2: Approximation error and timing of FFT vs. FRFT.

Ì Ú RMSE CPU time (sec)

FFT 2048 0.25 1.667E–14 0.057
2048 0.50 1.166E–14 0.055
4096 0.25 1.667E–14 0.108
4096 0.50 1.167E–14 0.108
8192 0.25 1.821E–14 0.260
8192 0.50 1.276E–14 0.255
FRFT 64 0.50 4.070E–06 0.014
64 1.00 9.208E–16 0.014
128 0.25 3.359E–06 0.029
128 0.50 9.813E–16 0.029
128 1.00 1.579E–15 0.029
256 0.25 9.122E–16 0.074
256 0.50 1.551E–15 0.074
256 1.00 2.765E–15 0.073

As we have seen it is impossible with the conventional FFT to evaluate the integral in

For example, the FFT for the Black–Scholes density with Ì

4096 and Ú
0.25
(5.24) on a fine grid and get a fine resolution for the density values at the same time.

returns only 107 out of the 4096 values within [–0.6, 0.6]. With respect to the range of
interest for our illustration, more than 97 % are outside the region of interest.

| 51
 2,0E-13 6,0E-06
abs error abs error
1,5E-13
4,0E-06
1,0E-13
2,0E-06
5,0E-14
0,0E+00 0,0E+00
-5,0E-14 -0,6 -0,4 -0,2 0,0 0,2 0,4 0,6 -0,6 -0,4 -0,2 0,0 0,2 0,4 0,6
-2,0E-06
-1,0E-13
-4,0E-06
-1,5E-13
-2,0E-13 -6,0E-06

analytical solution in the Black–Scholes case. Left: FFT Ì

4096 and Ú
0.25. Right: FRFT
Figure 17: Pattern of absolute error of FFT and FRFT probability density approximations versus

Ì
128, Ú
0.25.

With the FRFT we have an appealing possibility of breaking the relation between Ú
and Û and explicitly specifying both the summation grid of the Fourier integral and
the desired output range. However, both the FFT and FRFT have the limitation that
the grid points must be chosen equidistantly thus precluding the use of more
sophisticated integration algorithms like Gaussian quadrature.
Conducting the same experiments for the cumulative density reveal quite similar

damping parameter  is critical for either of the two algorithms.

implications with respect to the choice of FFT or FRFT, though the choice of the

Findings for the Heston (1993) and FMLS model against direct integration
methods reveal similar qualitative characteristics, nevertheless an exact quantification
of the errors is not an easy task since in this case approximations in one form are just
compared to approximations in another form.

5.4 Option pricing

In the previous section we implemented a number of numerical integration schemes
for probability function calculations. Now we proceed and apply the same techniques
to the described Fourier pricing algorithms. Again we illustrate typical shapes of the
Fourier integrands, compare their convergence properties, and investigate the impact
of changing the numerical quadratures with respect to accuracy and running times.

Black–Scholes Style Formula

Heston (1993) pioneers the use of Fourier inversion methods in option pricing by
assuming a Black–Scholes like functional form as solution to the valuation problem.
The difference of two probabilities under appropriate changes of measure then yield
to the theoretical value. In the sequel, we compare this concept with the analytical
solution of the Black–Scholes formula. In Figure 18 we exemplify some typical shapes
for the two integrands involved. In the left exhibit, an at-the-money option with zero
interest rates is shown. We can see a perfect symmetry of the two integrands. The two

the Φq< and Φq, values in the analytic case with q<
0.075 and q,
0.075
resulting integral values employing inversion formula (3.17) correspond perfectly to

| 52
yielding Φ0.075
0.529893 and Φ0.075
0.470107 finally giving
5.978528811 for both the call and put.

0,015 0,050
Integrand P1
0,010 Integrand P2 0,000

0,005 -0,050 0 2 4 6 8 10 12 14 16 18 20 22 24 26

0,000 -0,100
0 2 4 6 8 10 12 14 16 18 20 22 24 26
-0,005 -0,150

-0,010 -0,200 Integrand P1

Integrand P2
-0,015 -0,250

Figure 18: Integrands of the Black–Scholes model in the Black–Scholes style formula. Left: 
100,

100, 
0.25, 
0, 3
0.3. Right: 
80.

neutral probabilities of finishing in-the-money with Φq<

0.940929 and
On the right the integrands are displayed with strike 80 corresponding to the risk

Φq,
0.921117 resulting in 20.403599348 for the call and 0.403599348 for the
put respectively.
The functional form expresses the option value as the difference of two binaries,
also referred to as digital options. To be more precise the call value resembles the
result of subtracting a Cash-or-Nothing call from an Asset-or-Nothing call. We
exemplify this relation for the FMLS model using the entries in Table 1. The results
are summarized in Table 3, note that the Cash-or-Nothing call has to be scaled by the
current strike value.

FMLS model. Model parameters are 


100, 
1, 
0.05, 
1.6 and 3
0.1 where the
Table 3: Call price as the difference of an Asset-or-Nothing call and a Cash-or-Nothing call in the

contour of integration is chosen along 1.75.

Asset-or-Nothing 73.085400047
Cash-or-Nothing ð Strike 63.443665532
Call 9.641734515

While having the same appealing intuitive properties like in the Black–Scholes case,
this approach has several deficiencies with respect to the numerical evaluations of the
Fourier integrals. The numerical approximation of a Fourier integral will introduce a
discretization error and a truncation error [compare Section 6] which will double by
evaluating two integrals. This might be alleviated by considering equation (4.14),
however, it remains to evaluate the characteristic function for the price dynamics
twice which is a time critical factor. Another crucial factor are the convergence
properties of the integrands. From (4.14) and (4.15) likewise it becomes apparent that
the integrands decay approximately in a linear fashion, whereas all other approaches
under consideration have a quadratic rate of decay, usually leading to substantial
savings in computation time.

| 53
Attari (2004) and Bates (2006) Formulae
As we know from the approaches developed by Attari (2004) and Bates (2006) it is
possible to obtain single integration solutions and sharing the cumulative density
function analogy of the Black–Scholes style formula for finding option prices. In this
framework we do not have to specify a specific damping factor or contour of
integration in the complex plane to make the Fourier integrals well behaved which is
definitely simplifying implementational matters. Further, another similarity between
both formulae is that they share a quadratic term in the denominator exhibiting
exactly the same advantageous convergence properties compared to the Black–
Scholes like solution.

1,4 0,05
Attari-Bates Integrand
1,2 0,00
1,0
-0,05
0,8
-0,10
0,6
-0,15
0,4
Integrand P1
0,2 -0,20
Integrand P2
0,0 -0,25
0,0 2,0 4,0 6,0 8,0 10,0 0 2 4 6 8 10 12 14 16 18 20 22 24 26

Figure 19: Rate of decay in the Attari (2004) and Bates (2006) valuation formulae in comparison to the
Black–Scholes like solution. Left: Attari-Bates Integrand. Right: Black–Scholes style integrands. Model
parameters are chosen as in the right panel of Figure 18.

The curvatures of the integrands in Figure 19 demonstrate the much faster rate of
decay of the models with quadratic decay as the frequency variable approaches
infinity.

Carr and Madan (1999) and Time Value Formulae

The Fourier inversion method developed by Carr and Madan (1999) attracted a lot of
attention mainly due to two reasons. First, it allows to price options for a whole range
of strikes in one go and secondly it introduces an additional degree of freedom which
has a deep impact on the curvature of the involved Fourier integrands. This
additional parameter is not only crucial for the application of the FFT to option
pricing but as well serves as a scaling factor affecting the smoothness of the
integrands. A well behaved integrand helps the numerical integration schemes
significantly and hence is indispensable for highly accurate results in this framework.

the choice of the damping factor  is an important one. In the left panel the divergent
Figure 20 gives an example that is based on the Heston (1993) model. As one can see,

behavior of the integrands with reasonable choices for  that allow for a precise
behavior around zero is clearly visible and in contrast, the right panel shows a smooth

evaluation of the pricing formula. Many authors have discussed this pricing approach
in rigorous detail, most notably the works by Lee (2004) extending the approach

| 54
detailed instructions on how to chose  to receive ‘very’ high precision values.
significantly and providing an error analysis and Lord and Kahl (2007) who give

1200 30
real real
1000 25
imag imag
800 20
real
600 15
imag
400 10
200 5
0 0
-200 0,0 2,0 4,0 6,0 8,0 10,0 -5 0,0 5,0 10,0 15,0 20,0
-400 -10
-600 -15
-800 -20

Figure 20: Damping factor  as scaling parameter. Left: Real and imaginary parts of the Heston call
value integrand with 
0.1. Right: Real and imaginary parts of the Heston call value integrand
with 
1.75 (solid lines) and 
2.75 (dashed lines). Model parameters: 

100, 
1,

0.05, ®
2, ¯
0.01, °
0.5, 3±
0.25 and ²$
0.02.

a deep analysis we refer to the mentioned literature. The interplay between  and the
In the sequel we focus on some typical observations made within this framework, for

resulting function values is visualized in Figure 21. An  near the origin of the call
formula and an  around one for the put formula show complete meaningless results,
with increasing  the prices approach the true values (obtained with the adaptive

bound on  as well which depends on the number of moments of the particular price
Gauss–Kronrod quadrature). As was mentioned in Section 4.4, there is an upper

process.

12,50 5,00
10,00
7,50
2,50
5,00
Heston Call

Heston Put

2,50
0,00 0,00
-2,50 0,00 0,05 0,10 0,15 0,20 1,00 1,05 1,10 1,15 1,20 1,25
-5,00
-2,50
-7,50
-10,00
-12,50 -5,00
alpha alpha

Figure 21: Heston values as a function of damping factor . Left: Call options formula, true price:
7.504536548. Right: Put options formula, true price: 2.627478999. Model parameters: 

100,

1, 
0.05, ®
2, ¯
0.01, °
0.5, 3±
0.25 and ²$
0.02.

The examples illustrated in Figure 22 based on FFT and the fractional Fourier
transform show typical error patterns emerging from the use of these methods. They
are only for illustration purposes, a detailed overview comparing these methods and
assessing their accuracy is given in Chourdakis (2005).

| 55
 0,0E+00 1,5E-13
FFT FRFT
-1,0E-13
1E-13
-2,0E-13
-3,0E-13 5E-14

-4,0E-13 0
-5,0E-13
-5E-14
-6,0E-13
-7,0E-13 -1E-13
80 90 100 110 120 80 90 100 110 120

Figure 22: Comparison of Black–Scholes call option prices. Left: FFT with Ì
4096 and Ú
0.25.
Right: FRFT with Ì
128 and Ú
0.2 for the strike range 80 – 120. Model parameters: 


100, 
0.5, 
0.04, 3
0.2 and damping factor 
1.75.

The second approach considered by Carr and Madan (1999) substracts the intrinsic
value from the option and thus deals with time values of the option only. This simple
procedure ensures the square integrability of the valuation function and does not

g
0, in the neighborhood of this peak the FFT has serious problems due to the fact
require choosing a damping factor. Unfortunately, the function has a sharp peak at

that in this area very high frequencies are required to accurately evaluate the
integrand which are omitted by a Fourier inversion over a finite range of integration.
The source of divergence results from the non differentiability of the intrinsic value

Madan (1999) creates a discontinuity at g

0 which troubles the numerical inversion
around the at-the-money level. Additionally, the original formulation in Carr and

even more. These findings can be seen in Figure 23 where the time value function is
plotted and contrasted to the analytical solution of the Black–Scholes model. A simple
solution to the discontinuity problem is proposed in McCulloch (2003) [as described
in Section 4.4].

7,00 2,00
6,00 1,00
5,00
0,00
4,00
3,00 -1,00 80 90 100 110 120
2,00 -2,00
1,00 -3,00
0,00
-4,00
70 80 90 100 110 120 130 140
-5,00

FFT with Ì
4096 and Ú
0.25. Model parameters: 

100, 
0.5, 
0.04, 3
0.2 and
Figure 23: Out-of-the-Money Time Value function and absolute error in the Black–Scholes model.

damping factor 
1.75. The blue points indicate where exactly 
 or g
0, at this point our
FFT algorithm did not return a value.

The divergent behavior of the Time Value method can be alleviated by damping the
very high frequencies by a hyperbolic sine function. An example is given in Figure 24.
Without damping the extreme frequencies one can observe the highly oscillatory
curvature of the integrand in the left exhibit.

| 56
 400 25
without sinh sinh 1.75
300 20
sinh 4.75
200 15
100 10
0 5
-100 0
-200 -5
0,0 5,0 10,0 15,0 20,0 0,0 5,0 10,0 15,0 20,0 25,0

Figure 24: Integrand for the Time Value method. FMLS model with 

100, 
0.5, 
0.05,

1.6 and 3
0.14.

By introducing a sinh damping function the integrand becomes progressively better

behaved and hence more suitable for numerical integration schemes. However, for
choosing the damping factor too big the integrand will get to close to zero and
inhibits precise evaluation.
The Time Value method is usually considered as a robust and stable method, but,
as Borak & al. (2005) note, due to the sharp cusp around the at-the-money level it
may lead to some undesired numerical artefacts for the resulting volatility smiles.

Lipton (2002) and Lewis (2001) Formulae

With the introduction of a damping factor Carr and Madan (1999) make a big step to
numerically highly efficient pricing algorithms. By multiplying the option price
function by an exponential function to ensure the integrability of the Fourier
transform the pricing problem is effectively shifted into the complex plane which
requires the characteristic function to be evaluated in a specific domain of the
complex plane depending on the payoff function and the particular stochastic
process.
From Lewis (2001) we have a clear exposition of the interplay between the
conditions on the payoff functions and the price processes made explicit by directly
applying the Plancherel–Parseval identity to the convolution representation of the
pricing kernel with the payoff in Fourier space. By carefully choosing the contour of
integration along the complex plane and checking for singularities a flexible pricing
framework for arbitrary payoff functions and price processes is established.
Throughout the paper we illustrate several applications of this framework, for
instance the binary options in the Black–Scholes section above or the calculations of
Greeks and the state price density in the following section. Here we examine the
algorithms for plain European call and put options.

| 57
 0,10 3,00
0,00 2,00

-0,10 0,0 2,0 4,0 6,0 8,0 10,0 12,0 14,0 16,0 1,00

-0,20 0,00

-0,30 -1,00 0,0 1,0 2,0 3,0 4,0

real real
-0,40 -2,00
imag imag
-0,50 real -3,00 real
imag imag
-0,60 -4,00

value integrand with U{

2.5 (solid lines) and U{
4.5 (dashed lines). Right: Real and imaginary parts
Figure 25: Effect of choosing a contour of integration. Left: Real and imaginary parts of the Heston call

of the Heston put value integrand with U{

0.25 (solid lines) and U{
0.75 (dashed lines). Model
parameters: 
100, 
80, 
1, 
0.05, ®
2, ¯
0.01, °
0.5, 3±
0.25 and
²$
0.02.

panel two examples of possible lines of integration along U{ [ 1 are shown, for this
The graphics shown in Figure 25 depict the valuation formula in Eq. (4.54). In the left

region the call price function is well defined [Call

24.119720814]. For the right
panel two arbitrary values U{ Q 0 are chosen which define the put value function
[Put
0.218074775]. The effect of a particular contour of integration with respect to
the curvature of the integrands is considerable.
By moving the contour and applying Residual Calculus two alternative
formulations for the call and put formulae are obtainable. As illustrated in the entry of

0 Q uU Q 1. This is true for both the call and put formula only distinguished by a
the covered call and cash-secured put in Table 1 the strip of regularity is defined for

different resulting residue. The effect of specifying a particular line of integration in

resulting option values as a function of the contour parameter U{ are contrasted to the
this narrow strip seems to be more sensible than in the previous case. In Figure 26 the

‘true’ value calculated by the aid of the adaptive Gauss–Kronrod quadrature.

9,0E-10
8,0E-10
7,0E-10
6,0E-10
5,0E-10
4,0E-10
3,0E-10
2,0E-10
1,0E-10
0,0E+00
0,40 0,45 0,50 0,55 0,60

Figure 26: Absolute error of the Heston call value as a function of the contour parameter U{ .

As was pointed out in Section 4.5 the choice of U{

, is not only facilitating
<

implementations but ensures accurate results for this particular valuation formula.
The same observations hold for put value formulation and the Lipton (2002)
formulae as well as they are equivalent formulations.

| 58
5.5 Greeks and State Price Densities
The accurate valuation of financial claims is not only the key in financial modeling,
the hedging of these derivative instruments is at least as important. In order to do so
we need the various sensitivities of the specific models with respect to the state
variables and model parameters commonly referred to as Greeks. Expositions of
common Greeks in the Black–Scholes style Fourier valuation framework can be
found e. g. in Schöbel and Zhu (1999), Minenna and Verzella (2008) and including
some higher order Greeks like Vomma and Charm, Dual Delta or Dual Gamma we
refer to Reiß and Wystup (2001). A detailed description, possible uses and pitfalls of
these higher order Greeks in a Black–Scholes world may be found in a series of papers
from Haug (2003a, b). All these Greeks may be obtained by using simple and
straightforward finite differences schemes, however, in case that the characteristic
function is known these are directly available in an analytical form just like the
derivatives prices or density functions. An exposition of the derivation of Greeks in
the single integral solutions is relatively rare yet, therefore we present some arbitrarily

In the Carr and Madan (1999) approach the hedge ratio Delta Δ is given as
chosen hedge parameters in these frameworks.

follows
+
Δ
,
+
(5.25)

+  ­h #
   D  F   C 1 i

æ t w GHh x dFè,
+ 6 $  , C   F, C i2 C 1 F
+D  F   C 1 i
 ­h #    +

t « GHh , ¬ dF.
6 $  C   F C i2 C 1 F
,

Calculating the partial derivative we get

+D  F   C 1 i D  F   C 1 i  C 1 C iF

.
+ 
(5.26)

Hence yielding
 ­h #
   D  F   C 1 i  C 1 C iF
Δ
t w GHh x dF,
6  , C   F, C i2 C 1 F
(5.27)
$

which may be further simplified to [compare Eq. (4.36)]

 ­h #
   D  F   C 1 i
Δ
t w GHh x dF.
6  C iF
(5.28)
$

The Gamma Γ of an option is given by the second derivative of the option with
respect to the underlying

| 59
 +Δ +Δ +L
Γ

,
+ +L +
(5.29)

 ­h #

t( GHh    D  F   C 1 i ) dF.
,6
(5.30)
$

The Delta and Gamma Greeks shown above in addition to the sensitivity Rho are

If we are interested in calculating the Vega õ for example, we have to differentiate the
model free in the sense that they do not rely on the particular characteristic function.

characteristic function with respect to the model parameter describing the volatility
or variance which will be evidently model dependent. Even more involved is the
situation for Theta, i. e. the sensitivity with respect to the time to maturity, the time is
a relevant factor for both the general valuation formula and the particular
characteristic functions conditional upon the current time state. For the Heston
(1993) model Greeks are illustrated in detail e. g. in Nagel (2001) and Reiß and
Wystup (2001) and will not be repeated here.
In the following we will illustrate the concepts within the Lipton–Lewis valuation
framework and work out a number of Greeks for the FMLS model. Recalling (4.61)
and differentiating yields for the Delta
   #
1 < i dF
Δ
1 t Å¥iF C ¦  :GH,?h D  ¥F  ¦Æ ,
6 2 2 F, C 1
4
$
(5.31)

and
   # < i
Γ
t Å :GH,?h D  ¥F  ¦Æ dF,
,6 2
(5.32)
$

for the Gamma. The general form for Vega is given by

i
+    # < +D  :F  2? dF
õ

 t « :GH,?h ¬ ,
+²$ 6 +²$ 1
F, C4
(5.33)
$

where ²$ denotes the variance or volatility depending on the particular model.

Considering the FMLS model we do not have either of the two for  Q 2, but we have
the scale parameter 3 governing the width of the distribution. Differentiating the
characteristic function of the FMLS process with respect to 3 results in
6 6
+exp É :iF3 ­ sec 2  iF3 ­ sec 2 ?Ë
+3
(5.34)

6 6
iF3 ­ α sec 2 iF3 ­ α sec 2

©  ª
3 3

| 60
 6 6
ð exp É :iF3 ­ sec  iF3 ­ sec ?Ë.
2 2
For Rho the calculation yields
+    #
1 < i dF

  t Å¥iF C ¦  :GH,?h D  ¥F  ¦Æ
+ 6 2 2 F, C 1
4
$

# < i dF
(5.35)
 t Å , D  ¥F  ¦Æ .
:GH ?h
2 F, C 1
4
$

To complete the first order Greeks Theta reads

i
+    #
1 D  :F  2?

 t÷¥iF C ¦  C ø
+ 6 $ 2 

< i dF
ð  :GH,?h D  ¥F  ¦
2 F, C 1
(5.36)
4
# < i dF
 t Å , D  ¥F  ¦Æ .
:GH ?h
2 F, C 1
4
$

Higher order Greeks can be derived in exactly the same manner by further
differentiating the first order Greeks. For illustration purposes we arbitrarily pick up
three higher order Greeks.
Firstly we consider the DdeltaDtime, Charm or Delta Bleed which is the Delta’s
sensitivity to changes in time. The DdeltaDtime Greek indicates of what happens with
the option’s Delta when maturity comes closer. It is defined as
i
+∆z    #
1 1 D  :F  2?

  t¥iF C ¦ ÷¥iF C ¦  C ø
+ 6 $ 2 2 

< i dF
ð  :GH,?h D  ¥F  ¦
2 F, C 1
(5.37)
4
# < i dF
 t Å :GH,?h D  ¥F  ¦Æ .
2 F, C 1
4
$

Secondly, the DgammaDvol, also known as Zomma, is the sensitivity of Gamma with
respect to changes in volatility or variance

| 61
 i
+Γ    # < +D  :F  2?

, t « ¬ dF.
:GH ?h
,
+²$  6 +²$
(5.38)
$

Thirdly, the DvegaDvol, Vomma or also known as Vega convexity or Volga, is the
sensitivity of the Vega to changes in volatility or variance
+, +õ

.
+²$ , +²$
(5.39)

6 6
For the FMLS model taking the second derivative of the option gives

+õ    # iF3 ­  , sec 2 iF3 ­  sec 2

 t 
+²$ 6 $ 3, 3,
6 6
iF3 ­  , sec  ­
 2 C iF3  sec 2
3, 3,
< ú secˆ­ GH= ú secˆ­ ?
ð  :GH,?h:GH= , ,

6 6
(5.40)
,
iF3 ­ α sec iF3 ­ α sec
C , © 2  2ª
3 3

< ú secˆ­ GH= ú secˆ­ ? dF

ð  :GH,?h:GH=  .
1
, ,
F, C 4
For expository purposes, we give a numerical example for the FMLS model and
summarize the resulting Greeks in Table 4.

Other Greeks can be derived in a similar way, however, depending on the particular
characteristic function these may result in quite lengthy expressions. We can see that
many terms appear repeatedly, so we can temporarily store intermediate results which
will simplify the implementation and can reduce computation time significantly. The
implications in terms of computational efficiency and accuracy are practically
identical to the observations made throughout the paper.

| 62
Table 4: Greeks for the Finite Moment Log Stable model. FMLS model with 

100, 
0.5,

0.05, 
1.8 and 3
0.11. The first entry for each item is for the call while the second is for the
put respectively.

Call 5.952366338
Put 3.483357541

Delta 0.653499430 Rho 29.698788334

–0.346500570 –19.066707268

Gamma 0.033587476 Charm 0.092109339

0.033587476 DdeltaDtime –0.092109339

Vega 38.456732518 Zomma –0.265054082

38.456732518 DgammaDvol –0.265054082

Theta –7.670146141 Vomma / Volga –0.715079076

–2.793596581 DvegaDvol –0.715079076

On several occasions we already pointed out the great importance of the concept and
usefulness of Arrow–Debreu prices, or state price densities respectively, with respect
to the fundamental risk neutral valuation principle and Fourier pricing alike. At this
point we present a graphical illustration to emphasize these important theoretical
underpinnings.
For this purpose we employ the Fourier transform of an Arrow–Debreu security
as is shown in Table 1. The state price density or discounted risk neutral density
[compare Eq. (2.3)] then results from an infinite collection of those Arrow–Debreu
securities as in Figure 27.
Formally this reads
   G‘’ #
1  M  
 G‘ 89 /;   D  U  G‘ dU,
6
(5.41)
$#

   G‘’ # /;

 G‘ 89 %   D  U dU.
6
(5.42)
G‘’ $

| 63
 4,5
Arrow-Debreu RND
4,0

3,5

3,0

2,5

2,0

1,5

1,0

0,5

0,0
60,00 70,00 80,00 90,00 100,00 110,00 120,00 130,00 140,00

model. Model parameters: 

100, 
1, 
0.05, ®
2, ¯
0.01, °
0.5, 3±
0.25 and
Figure 27: Continuum of Arrow–Debreu securities and risk neutral density (RND) in the Heston

²$
0.02.

As we can see summing up these ‘Dirac spikes’ representing the prices of a claim with
a delta function payoff at infinitlely small increments depict a well formed density.

5.6 Concluding Remarks and Further Results from the Literature

The major computational workload for all discussed methods is determined by the
number of evaluations of the characteristic function. Driving factor for the number of
evaluations needed to obtain a given level of accuracy are the convergence properties
of the particular methodology. For option pricing applications the Black–Scholes like
valuation formula and variants hereof fail the requirement of sufficiently fast decaying
integrands in comparison to the other methods which all decay much faster due to a
quadratic term in the denominator. The models evaluating the Fourier integrals on a
particular contour of integration in the complex plane additionally offer a fine tuning
parameter affecting convergence rates which, if chosen carefully, is able to provide a
good compromise between speed and accuracy.
All conducted numerical experiments are based on straightforward
implementations of the algorithms without any (intended) optimizations. The CPU-
time requirements for the calculations depend on many factors like CPU clock or
used programming language and thus the displayed times should be viewed more in a
relative sense. In order to increase the computational efficiency of the Gauss
quadratures the nodes and weights can be pre-calculated and stored for later use
(which is done for the composite Gauss quadratures). Furthermore, we strongly
recommend to always use the highest precision available for the particular quadrature
nodes, in the literature abscissas and weights are often tabulated for less than double
precision which can deteriorate numerical results significantly4.

4
High-precision abscissas and weights tabulated up to 100 nodes and C/C++ source code for the
Gauss–Legendre quadrature are available at http://www.holoborodko.com/pavel/?page_id=679

| 64
 An important topic to pay attention to is that the implementation of
characteristic functions might require the evaluation of complex logarithms as is the
case in Heston’s formula [see e. g. Mikhailov and Nögel (2003)]. These are multivalued
functions which can introduce discontinuities while evaluating the specific integrands
and lead to meaningless option prices. This phenomena was first pointed out by
Schöbel and Zhu (1999) and was tackled since then by several authors, amongst
others, Kahl and Jäckel (2006), Lord and Kahl (2008), Fahrner (2007) or Guo and
Hung (2007). For our studies we used the ‘formulation 2’ of the characteristic function
described in e. g. Albrecher & al. (2006) or Gatheral (2006).
The calculation of option prices and density approximations by the various
numerical algorithms lead to different types of errors. By imposing an upper
integration limit with a finite number may result in a truncation error [see Section
6.4]. Evaluating the integrand only at a limited number of points introduces sampling
errors and for FFT based methods interpolation errors arise from the fact that these
algorithms return values only on an equidistant grid [compare Section 6.2]. Another
source of error are roundoff errors introduced by inexact computer arithmetic.
In the literature on characteristic function calculus a number of possibilities are
developed to control these errors. Providing error bounds for the numerical inversion
of characteristic functions is a crucial ingredient for this task [see e. g. Hughett
(1998)]. The results from Probability Theory have been amended to option pricing
purposes by considering option prices as normalized probability functions. For
example, Pan (2002) builds upon the results from Davies (1973) to provide an error
analysis to target pricing errors. Another thorough error analysis framework and
error minimizing algorithms are presented in Lee (2004).
Numerical integration schemes and FFT algorithms are subject to ongoing active
research efforts. These improvements will surely affect applications in financial
modeling as well. One example is the work from Minenna and Verzella (2007) who
apply non uniform FFT algorithms which allow for non uniform sampling of the
characteristic function and employ Gaussian quadratures by interpolating an
oversampled FFT to the pricing of options.

6 Numerical Issues and Possible Refinements to the Inverse

Fourier Methods
The focus of this section lies on implementational issues and practical details.
Computational performance and complexity are extremely important in density
calculations and pricing. We investigate different approaches that have positive effects
on computational speed and can help to stabilize Fourier inversion methods.

| 65
6.1 Inversion of the Option minus the Black–Scholes Approximation
In order to compute option prices Tankov (2009) presents a slightly different
formulation of the Time Value method proposed in Carr and Madan (1999) and Cont
and Tankov (2004). The pricing problem is defined by the Fourier transform of the
time value of the option
U g
 M   h     M   h  . (6.1)
Then the Fourier transform of the time value of a call option, assuming an adjusted
log strike g
ln
%“ œ0
/;
, is given by
#
D   F  i  1
û F
 GHh U g dg
.
iFiF C 1
(6.2)
#

Finally, the option prices result by inverting the Fourier transform

$ #
 $ , , 
$      C  GHh û F dF.
6
(6.3)
$

As was already noted by Carr and Madan (1999) the time value method may result in
a very wide and potential oscillatory Fourier transform with a poor convergence rate.
Cont and Tankov (2004) argue this is due to the non smooth behavior of the time
value, which is a non differentiable function, causing the Fourier transform to decay
too slowly at infinity. This numerically challenging problem can be reduced severely
by subtracting an analytically integrable complementary function from the original
integrand. Proceeding this way, a methodology described in e. g. Manno (1988), will
reduce the underlying curvature of the function to be integrated. The integration of
the now better behaved composite function can then be evaluated with standard
numerical quadratures and will most likely improve resulting accuracy.
In the context of option pricing this simple method was first considered by
Andersen and Andreasen (2000), who use the Black–Scholes model as a control
variate to stabilize the numerical Fourier inversion. The Black–Scholes model is a
smooth function of the strike and makes the modified option prices differentiable.
The characteristic function from the Gaussian will be subtracted from the original
integrand and the Black–Scholes prices will be added afterwards [see Tankov (2009)]
1 =
Ũ  g
 M   h    iÇ/ g ,

(6.4)

leading to
#
D  F  i  D = F  i
û' F
 GHh Ũ  g dg
,
iFiF C 1
(6.5)
#

where D = F
exp É F, C iF Ë and eventually inverting yields to
>
=ýþ
,

| 66
 $ #
 $ , , 
Ç/
= 
$ , ,  C  GHh û' F dF.
6
(6.6)
$

For this approach to work well we need to know an appropriate volatility level of the
Black–Scholes model. One way to approximate the standard deviation of the
particular price process in question is proposed in O’Sullivan (2005)

+ , DE F +
* DE F  ¦ ,
,
3Ç/

* - C ¥
+F, +F H.$
(6.7)
H.$

which can be simplified to

3Ç/
‹tD 0   uD 0  , . (6.8)
The standard deviation can now be calculated either analytically or by a simple finite
difference scheme, for example Dobránszky (2009) opts to numerically approximate

Another possibility to approximate the wanted 3Ç/ is to determine the

the derivatives with eps=1e–5.

corresponding cumulants ih of the distribution as is shown in Fang and Oosterlee

(2008). While i, corresponds to the second moment Fang and Oosterlee (2008)
recommend to include the fourth cumulant ip (or even iñ as well since especially for
short time to maturities the Levy processes might exhibit sharp peaks and fat tails,
which may be accounted for by inclusion of higher order cumulants

3Ç/
i, C ‹ip . (6.9)

Cumulants for the Heston model and some popular Lévy processes can be found in

does not exist a second moment for  Q 2, hence the second (real valued) cumulant
the appendix, Table 11 from Fang and Oosterlee (2008). For the FMLS model there

does not exist either. In this case the numerical approximation seems to be the only
reasonable choice (twice differentiation results in a complex valued cumulant as a
function of the frequency variable).
All methods work well and do not impose any numerical impediments, however,
depending on the characteristic function, the analytical expressions may become
quite large, which in turn would favor the numerical approximation. While accurate

an ad hoc choice for 3Ç/ like 0.1 or 0.2 will improve convergence significantly. On a
approximations of the standard deviations lead to optimized Fourier inversion, even

side note we emphasize the fact that the obtained values for 3Ç/ describe the standard
deviation of the stochastic process and do not necessarily have to coincide with
concepts like implied volatility.

| 67
 4,5 0,030
FMLS real
4,0
BS real 0,025
3,5
FMLS imag
3,0 BS imag 0,020

2,5
0,015
2,0

1,5 0,010

1,0
0,005
0,5

0,0 0,000
0,0 2,0 4,0 6,0 8,0 10,0 12,0 14,0 16,0 18,0 20,0

parameter: 

100, 
0.5, 
0.05, 
1.8 and 3
0.1. The solid lines depict the FMLS
Figure 28: Control variate for the FMLS model in the Lipton–Lewis framework. FMLS model

integrand and the dashed lines result for the Black–Scholes integrand based on the standard deviation
given by 31.62 % for the FMLS process (imaginary parts are denoted on the right scale). The implied
volatility for the model parameter with call price 5.567831374 is given by 15.15 %.

The use of Black–Scholes control variates is applicable to all the mentioned valuation
formulae presented in this text [see e. g. Dobránszky (2009) for the Lipton–Lewis
formula and Figure 28]. The use of the Black–Scholes model not only improves
numerical evaluations of the Fourier integrals, it further allows a thorough error
control in a valuation framework [see Tankov (2004)]. Other control variates are
conceivable if the characteristic functions of both models were close, which, however,
will unlikely be the case since ‘analytical’ solutions to other sophisticated price
dynamics typically involve infinite summation or special functions.
Without going into detail the same method works for density approximations as
well. An example in given in Jaschke and Jiang (2002) for the approximation for the
cumulative density function
1 1 #
 GHM => H>
rL  ΦL; , 3
 —DE F   GH
, ˜ dF .
2 26 iF
(6.10)
#

Applying the inverse Fourier transform to rL  ΦL; , 3 mitigates the problem
that the integrand has a pole at zero. Alternatives to the normal distribution may be
more appropriate to the problem at hand provided they have analytical densities and
corresponding Fourier transforms.

6.2 Interpolation Issues

One fundamental property of conventional FFT and FRFT methods is that the
resulting values are located on an equally spaced output grid. Quite probably not all
desired values will exactly agree on this equidistant grid, calling for some

density calculations where the desired L-values are irregularly spaced and for option
interpolation methods to find the intermediate values. This is vitally important for

| 68
prices which will be equally spaced on a log strike grid whereas options listed on the
markets are usually denoted on a strike level, not the log strike level.
In the literature the authors often regard simple linear interpolation as an
adequate algorithm yielding sufficiently accurate results. The main reasons are the
additional computational workload or the argument that choosing the discretization
sufficiently small, the interpolation error will be negligible. Exceptions are e. g.
McCulloch (2003) or Chourdakis (2005) in the context of option pricing and Menn
and Rachev (2006) for density approximations. McCulloch (2003) argues that simple
linear interpolation may give an interpolation error in excess of the Fourier inversion
computational error due to the convexity of the pricing function. Instead the use of
cubic spline interpolation is favored for the call and put formulae which are capable to
capture the curvature of the functions. These implications are exhibited in Figure 29
where the respective pricing errors of the two methods are displayed for a range of
option values.

0,003
linear spline
0,002

0,001

0,000
80 85 90 95 100 105 110 115 120
-0,001

-0,002

-0,003

pricing in the Black–Scholes model with Ì

256, Ú
0.5, 
1.75. Model parameters: 


Figure 29: Linear vs. cubic spline interpolation for option pricing. Relative errors for FFT call option

100, 
0.25, 
0.04, 3
0.3.

In comparison to the linear interpolation the cubic spline method is much more

reduce the number of summation points Ì and instead optimize the number of
accurate in error terms. If computation time is critical one might use this result to

discretization points in Fourier space to obtain a characteristic function evaluated on

a fine grid.
In numerical experiments Menn and Rachev (2006) confirm that cubic splines
offer much better overall performance than linear interpolation with respect to
density calculations. Whereas linear interpolation error is negligible in the tails where
the curvature almost vanishes it will have a significant impact in the center of the
distribution around the mode [see Figure 30]. Further Menn and Rachev (2006) are

for -stable distributions.

able to approximate an error estimate of the cubic spline interpolation quantitatively

| 69
 0,003
linear spline
0,002

0,001

0,000
-60% -40% -20% 0% 20% 40% 60%
-0,001

-0,002

-0,003

density approximations with in the Black–Scholes model with Ì

256, Ú
0.5.
Figure 30: Linear vs. cubic spline interpolation for density approximations. Absolute error for FFT

Without saying other interpolation schemes will work perfectly well depending on
the particular problem. One choice would be e. g. to consider a convex interpolation
for option prices since this is a convex function of the strike [compare e. g. Hagan and
West (2006)].
To summarize we can state that the potential increase in runtime using more
sophisticated interpolation algorithms is in many cases justified by a potentially
substantial improvement in accuracy.

6.3 Caching Technique

The FFT valuation framework along the lines of Carr and Madan (1999) is usually
considered as the de facto standard for the fast pricing for a series of option prices.
This is because the FFT, like the FRFT, is able to deliver outputs for a whole range of
different strike levels in one go. Though, the simultaneous calculation for a given set
of strikes is not restricted to the FFT or FRFT methods. For instance, in Bates (1996)
the simple observation was made, that the values of the characteristic function for
given model parameters do not depend on a particular strike level. Thus, it is possible
to compute these values once, store them in a cache and reuse as desired for any
strike. For a detailed discussion of this caching technique we refer to Kilin (2007)
where this algorithm is contrasted to the FFT and FRFT methods using a number of
affine (jump) diffusions and time changed Lévy processes. Using the Attari (2004)
pricing algorithm [compare Section 4.2] for the direct integration method and
conducting some calibration exercises with all three methods the author is able to
show that the caching technique is orders of magnitude faster than FFT and FRFT.
An important aspect to note is that the caching technique can be used in
conjunction with a whole range of sophisticated integration schemes like Gaussian
quadratures. Therefore, the integration algorithm is not restricted to equally spaced
grid points as is the case for the FFT and FRFT.

arbitrary L-values are not related to the evaluations of the characteristic function for
The same technique works for density approximations as well. The values for

specific model parameters. This makes the method a viable tool for computationally

| 70
intensive density calculations without the need for FFT methods and regularly spaced
grid points.
In conclusion, the caching technique is a perfect candidate for tasks where the
computational workload is high like calibration to market prices, calculations of
corresponding hedge parameters or volatility surface and portfolio simulations.

6.4 Transforming the Domain of Integration

In theory all the considered Fourier integrals so far are continuous Fourier integrals,
i. e. are defined continuously from minus infinity or zero to plus infinity. In practice
though, we usually have to evaluate these integrals numerically. This implies that we
have to discretize the integral in some form for summation and, further, have to
truncate the infinite or semi infinite integral to a finite one (apart from the Gauss–
Laguerre quadrature). For the discretization we have discussed a number of
possibilities, now we consider potential ways of how to truncate the integration
domain.
A very elegant way is to determine the asymptotics for the characteristic
functions letting the frequency variable approach infinity. This is done for some
popular price dynamics like Heston (1993), Schöbel and Zhu (1999) and the Black
and Scholes (1973) model in Lord and Kahl (2007). They then evaluate the
transformed integration domain by applying an adaptive Gauss–Lobatto quadrature.
This way highly accurate derivative prices are achievable, but not in all cases it is
possible to transform the domain of integration analytically.

that the absolute real and imaginary parts of the characteristic function DE F are
Alternatively Chourdakis (2005) outlines a numerical method based on the fact

less or equal than its modulus

|tDE F | ` |DE F |,
|uDE F | ` |DE F |.
(6.11)

we may regard the characteristic function as zero in a numerical sense for large F.
As we know from Section 3.1 the characteristic function equals zero at infinity, that is

critical FeM such that

Using these facts we may then use the modulus of the characteristic function to find a

|DE FeM | ` 10 , (6.12)

where à denotes the desired precision level. Usually simple integer search methods
will suffice for practical tasks avoiding the need for computationally intensive solver
methods. Due to the oscillatory nature of Fourier integrals it may be theoretically
possible that there is a contribution of the characteristic function beyond the
determined truncation point, which nevertheless imposed no problems during made
numerical experiments.

| 71
 Instead of determining an upper integration limit a priori an alternative is to let
the numerical quadrature chose an upper truncation point adaptively. For instance,
the composite Gauss–Legendre quadrature [Eq. (5.3)] is predestined for this

which we use intensively throughout the paper. The number of summations Ì of the
procedure which is proposed in e. g. Sepp (2003) or Yan and Hanson (2006) and

sub-intervals each evaluated by a Gauss–Legendre quadrature is determined by a

contribution of the last strip becomes less than 10e–12 with a step size Ò
1 for the
local stopping criterion. Sepp (2003) recommends to stop the integration if the

single integral solutions and Ò

10 for the Black–Scholes style valuation formula,

the last strip to the total integration becomes smaller than 0.5e–7 and consider Ò
5
whereas Yan and Hanson (2006) stop the integration if the ratio of the contribution of

as a good compromise between speed and accuracy for the Black–Scholes like

In Lewis (2000) an approximation for an upper truncation level FeM is provided

formula.

in the example code on p 68 for the Heston (1993) model

Ð
FeM
max w1000, x,
‹²$ 
(6.13)

where Ð is set to 10 in this particular case. This ‘dimensionless’ relation reflects the
inverse dependency of the integrand on the time to maturity, a property we have seen

6.5]. For other models and pricing algorithms other choices for Ð and 1000 may be
in the manifold graphical illustrations of various integrands [compare also to Section

more appropriate and should be tested thoroughly. Ð can be roughly interpreted as

means, the heavier the tails of a distribution the bigger Ð should be. Whereas for a
the number of standard deviations for the particular price process in view, which

thin tailed Black–Scholes model a value of say 5 is usually more than sufficient, 10 or
even more seems appropriate for stochastic volatility and jump processes.
A similar procedure again for the Heston (1993) model together with the Attari
(2004) algorithm is presented in Staunton (2006)
ln 0.0001
FeM
Ð ,
√
(6.14)

where Ð is set to 10 as well. This formulation also respects the fact that the integrand
depends inversely on the time to maturity, but does the scaling without a specific
volatility or variance amendment.
There are a number of possibilities to find an upper integration point, which is
definitely an important implementational issue to ensure accurate results. Simply
making some ad hoc choices about the truncation levels should be avoided at all costs.
This might work just fine for some circumstances, but inherits the risk that a
significant part of the integrand is missed leading to meaningless results or that there
is virtually no contribution to the final result rendering the procedure quite inefficient
and random.

| 72
 Since we deal with integrals on a semi-infinite domain, another possibility would
be the use of Gauss–Laguerre quadratures not caring about truncation altogether. A

say like Ì
128, store the nodes and weights and stop integration when the
reasonable procedure would be to choose a Laguerre quadrature with a high degree,

contribution of the last strip becomes smaller than a predefined tolerance level. This
would reduce the numerical complexity for relatively well behaved functions, but
inherits the risk of not being sufficiently determined especially for short expiration
dates and is not appropriate for highly oscillatory integrands and thus should be
employed only with precaution.

6.5 Transforming the Fourier Integrands

The relevant domain of integration may vary quite significantly depending on the
model parameters and more importantly the remaining time to maturity. In Nagel
(2001) the author is outlining a standardizing procedure to transform the Fourier
integrals which will narrow the differences with respect to the meaningful integration
regions. The integrands are normalized by the standard deviation of the return
process on maturity. The author illustrates the procedure for the Black–Scholes like
valuation formula; here we will adapt this method to the single integral solutions.

F with the scaled version F Œ F¿ç

± , where ²$ denotes the instantaneous
H
Considering for instance the Lewis (2001) approach we substitute the Fourier variable

‹ ;

version with F¿ç then reads

variance of the price dynamics. Applying the transformation to (4.57) the modified

4√  / ,
 $ , , 
$ 
6
i dF
_$ t É GH
 h D  :F¿ç  2?Ë
#
1 C F¿ç ,
(6.15)

ð ,
‹²$ 
where we finally have to rescale the resulting integral to obtain the desired option
prices. The concept works equally fine for the other presented pricing algorithms as
well. The standard deviation of the distribution function can be computed via explicit
calculation of the second moment or corresponding cumulant respectively or by
approximating the standard deviation with finite differences as in Section 6.1.
The impact from this moment scaling procedure results in a drastically reduced
length of the needed integration range and a much less pronounced sensitivity to the
maturity. These implications are illustrated below for the Heston (1993) model.

| 73
 4,5 0,030 70 0,25
real real
4,0 60
0,025 imag 0,20
3,5 imag
50
3,0 0,020 0,15
2,5 40
0,015 0,10
2,0 30
1,5 0,010 0,05
20
1,0
0,005 10 0,00
0,5
0,0 0,000 0 -0,05
0,0 1,0 2,0 3,0 4,0 5,0 0,0 0,5 1,0 1,5 2,0 2,5

Figure 31: Impact of transforming the Fourier integrands. Left: Integrand with a standard numerical
integration. Right: Slope with normalized integrands.

Using the transformation procedure allows applying the numerical quadrature on a

predefined integration interval and therefore is potentially able to ease the numerical
implementation. However, in this case a much finer resolution of the quadrature is
needed to obtain sufficiently accurate results in comparison to the standard algorithm
without the modification.

6.6 Log Spacing

To minimize computation one would like to have the lowest number of sampling
points possible. In the context of Fourier inversion from image domain to space
domain we can exploit the knowledge that the function to be integrated usually will
exhibit a general decline in amplitude with increasing transform variable. Instead of
sampling as often at low frequencies as at high frequencies, it seems reasonable to
evaluate the integrand less frequently at higher frequencies.
In the following we adopt an idea from Meisel (1968) and implement an
integration scheme in which the sampling intervals increase on a logarithmic scale.
The characteristic function is then evaluated using unequally sampled intervals,
where the intervals are equally spaced on a logarithmic scale. For this purpose we
apply a simple procedure similar to the logspace function found in many standard
numerical packages5 and amend it to ensure it starts at zero. At least we have to
specify a starting and final value of the sequence and the number of samples to
generate. Further possibilities are to decide whether or not to include the last sample
or to define the base of the log space (which usually is base 10 by default).

5
See e. g. http://docs.scipy.org/doc/numpy/reference/generated/numpy.logspace.html

| 74
 4,5
Heston Integrand
4,0 adaptive Gauss-Lobatto sampling

3,5 log spacing

3,0

2,5

2,0

1,5

1,0

0,5

0,0
0,0 5,0 10,0 15,0 20,0 25,0 30,0

spacing parameter with starting value b

0, final value c
10 and number of steps
100. Heston
Figure 32: Sampling intervals on a log scale in comparison to adaptively sampled intervals. Log

model parameter: 

100, 
0.5, 
0.05, ®
2, ¯
0.01, °
0.5, 3±
0.1 and
²$
0.02.

In conjunction with the composite Gauss–Legendre quadrature using an adaptive

upper integration limit we are able to compute the Heston call price accurate up to at
least 13 digits over the full strike range from 50 to 150 in comparison to the adaptive
Gauss–Kronrod quadrature with a chosen tolerance level of 1e–15 for the given
example in Figure 32.
Table 5: Timings for different quadrature schemes. CPU time in seconds averaging over 100 call price
calculations with strikes ranging from 51 to 150 (Heston calls with parameters as in Figure 32). For all
routines the tolerance level is set to 1e–15.

Quadrature CPU time (sec)

adaptive Simpson 0.09266
adaptive Gauss–Lobatto 0.08451
adaptive Gauss–Kronrod 0.01797
composite Gauss–Legendre 0.04123
log spacing 0.00849

The time needed for this operation is only a negligible fraction of the time the
adaptive procedure takes. In Table 5 the log spacing procedure is compared to other
numerical quadratures as well. The time differences are quite remarkable and suggest
that this approach is a very powerful method to help improve the numerical
evaluation of Fourier integrals in terms of accuracy and in minimizing computation
time.
Combining the log spacing technique with the caching technique gave the fastest
density approximations and pricing for a set of options along the strike and time

| 75
dimension in comparison to all the other mentioned methods during numerical
experiments.

7 Conclusion
Financial modeling in the area of option pricing involves detailed knowledge about
stochastic processes describing the asset payoffs. For sophisticated price dynamics
these are most conveniently characterized through functions in Image space. By a
mapping of the probability function from spatial domain to the unit circle in the
complex plane, expected values of a future payoff are then available in the form of an
integral representation. With Fourier inversion methods these integral transforms
allow the pricing for a whole range of possible payoff structures.
In this work we outline general features of Fourier transform techniques
applicable to both the modeling of density functions and option pricing. Further we
describe a number of modeling alternatives and accentuate on the similarities and
subtleties of the different frameworks. Then we present several commonly used
numerical quadratures to evaluate the integral representations and exemplify their
use to density approximations and derivatives pricing. Finally, we present ‘best
practices’ on how the quantitative modeling of the Fourier techniques may be
improved by considering implementational issues, methodological aspects and
computational performance.
The Fourier pricing techniques and Fourier inversion methods for density
calculations add a versatile tool to the set of advanced techniques for pricing and
management of financial derivatives.

References
Abate, S. and W. Whitt (1992) The Fourier-series method for inverting transforms of
probability distributions, Queueing Systems: Theory and Applications 10(1–2),
5–88.

Agliardi, R. (2009) The quintessential option pricing formula under Lévy processes,
Applied Mathematics Letters 22(10), 1626–1631.

Albrecher, H., P. Mayer, W. Schoutens and J. Tistaert (2006) The Little Heston Trap,
Working Paper, Linz and Graz University of Technology, K.U. Leuven, ING
Financial Markets.

Allen, R. L. and D. W. Mills (2004) Signal Analysis: Time, Frequency, Scale, and
Structure, IEEE Press (Wiley–Interscience), New York.

| 76
Andersen, L. and J. Andreasen (2000) Jump-diffusion Processes: Volatility Smile
Fitting and Numerical Methods for Pricing, Review of Derivatives Research 4,
231–262.

Andersen, L. and J. Andreasen (2002) Volatile Volatilities, Risk, December, 163–168.

Andersen, L. and V. Piterbarg (2007) Moment explosions in stochastic volatility

models, Finance and Stochastics 11, 29–50.

Andreasen, J. (2006) Stochastic Volatility for Real, Working Paper, Bank of America.

Andricopoulos, A., M. Widdicks, P. Duck and D. Newton (2003) Universal Option

Valuation using Quadrature Methods, Journal of Financial Economics 67, 447–
471.

Attari, M. (2004) Option Pricing using Fourier Transforms: A Numerically Efficient

Simplification, Working Paper, Charles River Associates. Available at SSRN:
http://ssrn.com/abstract=520042

Bailey, D. and P. Swarztrauber (1991) The fractional Fourier transform and

applications, SIAM Review 33, 389–404.

Bailey, D. and P. Swarztrauber (1994) A fast method for the numerical evaluation of
continuous Fourier and Laplace transforms, SIAM Journal on Scientific
Computing 15, 1105–1110.

Bakshi, G. and D. Madan (2000) Spanning and derivative-security valuation, Journal

of Financial Economics 55, 205–238.

Barone-Adesi, G. and R. E. Whaley (1987) Efficient analytic approximation of

american option values, Journal of Finance 42, 301–320.

Bates, D. S. (1996) Jumps and Stochastic Volatility: Exchange Rate Processes Implicit
in Deutsche Mark Options, Review of Financial Studies 9(1), 69–107.

Bates, D. S. (2006) Maximum Likelihood Estimation of Latent Affine Processes,

Review of Financial Studies 19, 909–965.

Benhamou, E. (2000) Fast Fourier Transform for Discrete Asian Options, EFMA 2001
Lugano Meetings. Available at SSRN: http://ssrn.com/abstract=269491

| 77
Binkowski, K. (2007) Pricing of European Options using Empirical Characteristic
Functions, Macquarie University, Sydney.

Black, F. and M. S. Scholes (1973) The Pricing of Options and Corporate Liabilities,
Journal of Political Economy 81, 637–654.

Borak, S., K. Detlefsen and W. Härdle (2005) FFT Based Option Pricing, Discussion
Paper, Sonderforschungsbereich (SFB) 649, Humboldt Universität Berlin.

Borovkov, K. and A. Novikov (2002) On a new approach to calculating expectations

for option pricing, Journal of Applied Probability 39, 889–895.

Bouziane, M. (2008) Pricing Interest-Rate Derivatives – A Fourier-Transform Based

Approach, Lecture Notes in Economics and Mathematical Systems 607,
Springer, Berlin.

Boyarchenko, S. I. and S. Z. Levendorskiĭ (2002) Non-Gaussian Merton-Black-Scholes

Theory, World Scientific.

Breeden, D. and R. Litzenberger (1978) Prices of state-contingent claims implicit in

option prices, Journal of Business 51, 621–651.

Cardi, G. (2005) Exotic Options under Lévy Processes, Ph.D. Dissertation, University
of Bergamo.

Carr, P. P. (2003) Option Pricing using Integral Transforms, Presentation. Available at

http://www.math.nyu.edu/research/carrp/papers/pdf/integtransform.pdf

Carr, P. P. and D. B. Madan (1999) Option valuation using the fast Fourier transform,
Journal of Computational Finance 2(4), 61–73.

Carr, P. P. and L. Wu (2003) The Finite Moment Log Stable Process and Option
Pricing, Journal of Finance 58, 753–777.

Carr, P. P. and L. Wu (2004) Time-changed Lévy processes and option pricing, Journal
of Financial Economics 71, 113–141.

Chourdakis, K. (2005) Option pricing using fractional FFT, Journal of Computational

Finance 8(2), 1–18.

Cont, R. (2001) Empirical properties of asset returns: stylized facts and statistical
issues, Quantitative Finance 1, 223–236.

| 78
Cont, R., J. da Fonseca and V. Durrleman (2002) Stochastic Models of Implied
Volatility Surfaces, Economic Notes 31(2), 361–377.

Cont, R. and P. Tankov (2004) Financial Modelling with Jump Processes (Financial
Mathematics Series), Chapman & Hall/CRC, New York.

Cooley, J. and J. Tukey (1965) An algorithm for the Machine Calculation of Complex
Fourier Series, Mathematics of Computation 19, 297–301.

Cox, J. C. and S. A. Ross (1976) The valuation of options for alternative stochastic
processes. Journal of Financial Economics 3, 145–166.

Davies, R. B. (1973) Numerical inversion of a characteristic function, Biometrika

60(2), 415–417.

Davis, P. J. and P. Rabinowitz (1984) Methods of Numerical Integration (Computer

Science and Applied Mathematics), 2nd ed., Academic Press, NewYork.

Delbaen, F. and W. Schachermayer (2006) The Mathematics of Arbitrage, Springer,

Berlin.

Dempster, M. A. and S. S. Hong (2002) Spread Option Valuation and the Fast Fourier
Transform, Technical Report WP26/2000, University of Cambridge.

Dobránszky, P. (2008) Numerical Quadratures to Calculate Lévy Base Correlation,

Technical Report, Finalyse SA, FORTIS Merchant Banking and KU Leuven.

Dobránszky, P. (2009) Option Pricing Using Numerically Evaluated Characteristic

Functions, Presentation on the 3rd Conference on Numerical Methods in
Finance, Paris. Available at
http://peter.dobranszky.com/files/NumericallyEvaluatedCFs.pdf

Duffie, D., J. Pan and K. Singleton (2000) Transform analysis and asset pricing for
affine jump–diffusions, Econometrica 68, 1343–1376.

Duffy, D. G. (2004) Transform Methods for Solving Partial Differential Equations,

2nd ed., Chapman & Hall/CRC, New York.

Dufresne, D., J. Garrido and M. Morales (2009) Fourier Inversion Formulas in Option
Pricing and Insurance, Methodology and Computing in Applied Probability 11,
359–383.

| 79
Eberlein, E., K. Glau and A. Papapontoleon (2009) Analysis of Fourier Transform
Valuation Formulas and Applications, to appear in Applied Mathematical
Finance, arXiv:0809.3405v4

Eberlein, E. and U. Keller (1995) Hyperbolic Distributions in Finance, Bernoulli 1(3),

281–299.

Epps, T. W. (1993) Characteristic Functions and Their Empirical Counterparts:

Geometrical Interpretations and Applications to Statistical Inference, American
Statistician 47, 33–38.

Fahrner, I. (2007) Modern Logarithms for the Heston Model, International Journal of
Theoretical and Applied Finance 10, 23–30.

Fang, F. and C. Oosterlee (2008) A novel pricing method for European options based
on Fourier-cosine series expansions, SIAM Journal on Scientific Computing
31(2), 826–84.

Fang, F., R. Lord and C. W. Oosterlee (2006) Fast and accurate Methods in Pricing
Early Exercise Options under Lévy Processes, Working Paper, Rabobank
International, Delft University of Technology and Center for Mathematics and
Computer Science (CWI), Amsterdam. Available at http://mpra.ub.uni-
muenchen.de/1952/

Filipović, D. (2001) A General Characterization of one Factor Affine Term Structure

Models, Finance and Stochastics 5(3), 389–412.

Fusai, G. and A. Roncoroni (2008) Implementing Models in Quantitative Finance:

Methods and Cases (Springer Finance), Springer, Berlin.

Gander, W. and W. Gautschi (2000) Adaptive Quadrature—Revisited, BIT 40(1),

March, 84–101.

Gatheral, J. (2006) The Volatility Surface: A Practitioner’s Guide (Wiley Finance

Series), John Wiley & Sons, New York.

Geman, H., N. El Karoui and J.-C. Rochet (1995) Changes of Numéraire, Changes of
Probability Measure and Option Pricing, Journal of Applied Probability 32(2),
443–458.

Gil-Pelaez, J. (1951) Note on the inversion theorem, Biometrika 38(3–4), 481–482.

| 80
Guo, J. and M. Hung (2007) A Note on the Discontinuity Problem in Heston’s
Stochastic Volatility Model, Applied Mathematical Finance 14, 339–345.

Grundke, P. (2007) Computational aspects of integrated market and credit portfolio

models, OR Spectrum 29, 259–294.

Gurland, J. (1948) Inversion formulae for the distribution of ratios, Annals of

Mathematical Statistics 19, 228–237.

Hagan, P. S. and G. West (2006) Interpolation methods for curve construction,

Applied Mathematical Finance 13(2), 89–129.

Hart, J. F. & al. (1968) Computer Approximations (SIAM Series In Applied

Mathematics), John Wiley & Sons, New York. Algorithm 5666 for the error
function.

Haug, E. (2003a) Know Your Weapon Part 1, Wilmott Magazine, May, 49–57.

Haug, E. (2003b) Know Your Weapon Part 2, Wilmott Magazine, July, 50–56.

Harrison, J. and D. Kreps (1979) Martingales and arbitrage in multiperiod securities

markets, Journal of Economic Theory 20(3), 381–408.

Harrison, M. and S. Pliska (1981) Martingales and stochastic integrals in the theory of
continuous trading, Stochastic Processes Applications 11(3), 215–260.

Heston, S. L. (1993) A Closed-form Solution for Options with Stochastic Volatility

with Application to Bond and Currency Options, Review of Financial Studies
6(2), 327–343.

Heston, S. L. and S. Nandi (2000) A Closed-Form GARCH Option Pricing Model,

Review of Financial Studies 13(3), 585–625.

Huang, J. and L. Wu (2004) Specification Analysis of Option Pricing Models Based on

Time-Changed Lévy Processes, Journal of Finance 59, 1405–1440.

Hughett, P. (1998) Error Bounds for Numerical Inversion of a Probability

Characteristic Function, SIAM Journal on Numerical Analysis 35(4), 1368–
1392.

Hurd, T. R. and Z. Zhou (2009) A Fourier transform method for spread option
pricing, Preprint, arXiv:0902.3643v1

| 81
Itkin, A. (2005) Pricing options with VG model using FFT, arXiv:physics/0503137v1

Jaschke, S. R. and Y. Jiang (2002) Approximating Value at Risk in Conditional

Gaussian Models, In: Applied Quantitative Finance, Eds. W. Härdle, T. Kleinow
and G. Stahl, Springer, Berlin.

Ji, M. and F. Zapatero (2008) Empirical Performance of Lévy Option Pricing Models.
Available at SSRN: http://ssrn.com/abstract=1266380

Judd, K. L. (1998) Numerical Methods in Economics, The MIT Press, Cambridge.

Kahl, C. and P. Jäckel (2006) Not-so-complex logarithms in the Heston model,

Working Paper, University of Wuppertal and ABN AMRO.

Keller-Ressel, M. (2008) Affine Processes – Theory and Applications to Finance,

Ph.D. Dissertation, Vienna University of Technology.

Kilin, F. (2007) Accelerating the Calibration of Stochastic Volatility Models, Working

Paper, Frankfurt School of Finance & Management and Quanteam AG.

Kim, Y., S. Rachev, M. Bianchi and F. Fabozzi (2009) Computing VaR and AVaR in
infinitely divisible distributions, Technical Report, University of Karlsruhe.

Kou, S. G. (2002) A jump-diffusion model for option pricing, Management Science 48,
1086–1101.

Kruse, S. and U. Nögel (2005) On the pricing of forward starting options in Heston’s
model on stochastic volatility, Finance and Stochastics 9, 223–250.

Lee, R. W. (2004) Option Pricing by Transform Methods: Extensions, Unification, and

Error Control, Journal of Computational Finance 7(3), 51–86.

Leippold, M. and L. Wu (2002) Asset pricing under the Quadratic Class, Journal of
Financial and Quantitative Analysis 37(2), 271–295.

Lévy, P. (1925) Calcul des Probabilités, Gauthier-Villars et Cie, Paris.

Lewis, A. (2000) Option Valuation under Stochastic Volatility: With Mathematica

Code, Finance Press, Newport Beach.

| 82
Lewis, A. (2001) A Simple Option Formula for General Jump-Diffusion and other
Exponential Lévy Processes, Envision Financial Systems and OptionCity.net,
California. Available at http://optioncity.net/pubs/ExpLevy.pdf

Lipton, A. (2001) Mathematical methods for foreign exchange, World Scientific.

Lipton, A. (2002) The vol smile problem, Risk, February, 61–65.

Lord, R. and C. Kahl (2007) Optimal Fourier Inversion in Semi-Analytical Option

Pricing, Discussion Paper TI No. 2006–066/2, Tinbergen Institute. Available at
SSRN: http://ssrn.com/abstract=921336

Lord, R. and C. Kahl (2008) Complex logarithms in Heston-like models. Rabobank

International and ABN AMRO, Available at SSRN:
http://ssrn.com/abstract=1105998

Lord, R., F. Fang, F. Bervoets and C. Oosterlee (2007) A Fast Method for Pricing
Early-Exercise Options with the FFT, Lecture Notes In Computer Science 4488,
415–422.

Lukacs, E. (1970) Characteristic Functions, 2ed ed., Charles Griffin & Co, London.

Manno, V. P. (1988) A simple procedure for reducing numerical integration errors

near singularities, Communications In Applied Numerical Methods 4, 713–716.

Madan, D. B., P. P. Carr and E. C. Chang (1998) The Variance Gamma Process and
Option Pricing, European Finance Review 2, 79–105.

McCulloch, J. H. (2003) The Risk-Neutral Measure and Option Pricing under Log-
Stable Uncertainty, Working Paper, Ohio State University.

McKeeman, W. M. (1962) Algorithm 145: Adaptive numerical integration by

Simpson’s rule, Communications of the ACM 5(12), p 604.

Meisel, W. (1968) A Numerical Integration Formula Useful in Fourier Analysis,

Communications of the ACM 11, p 51.

-stable distributions, Computational Statistics & Data Analysis 50, 1891–1904.

Menn, C. and S. T. Rachev (2006) Calibrated FFT-based density approximations for

Merton, R. C. (1973) Theory of rational option pricing, Bell Journal of Economics and
Management Science 4, 141–183.

| 83
Merton, R. C. (1976) Option pricing when the underlying stock returns are
discontinuous, Journal of Financial Economics 3, 125–144.

Mikhailov, S. and U. Nögel (2003) Heston’s Stochastic Volatility Model –

Implementation, Calibration and some Extensions, Wilmott Magazine, July, 74–
79.

Minenna, M. and P. Verzella (2007) Fast Option Pricing using Non Uniform Discrete
Fourier Transform on Gaussian Discretization Grids, CONSOB, University of
Milano Bicocca.

Minenna, M. and P. Verzella (2008) A revisited and stable Fourier transform method
for affine jump diffusion models, Journal of Banking & Finance 32, 2064–2075.

Nagel, H. (2001) Optionsbewertung bei stochastischer Volatilität, Deutscher

Universitäts Verlag, Gabler, Wiesbaden.

Nunes, J. P. and T. R. Alcaria (2009) Two Extensions to Forward Start Options

Valuation, ISCTE Business School, Caixa Económica Montepio Geral. Available
at SSRN: http://ssrn.com/abstract=1329370

O’Sullivan, C. (2005) Path Dependent Option Pricing under Lévy Processes, EFA
2005 Moscow Meetings Paper. Available at SSRN:
http://ssrn.com/abstract=673424

Pan, J. (2002) The jump-risk premia implicit in options: Evidence from an integrated
time-series study, Journal of Financial Economics 63, 3–50.

Raible, S. (2000) Lévy Processes in Finance: Theory, Numerics and Empirical Facts,
Ph.D. Dissertation, University of Freiburg.

Reiß, O. and U. Wystup (2001) Efficient Computation of Option Price Sensitivities

using Homogeneity and other Tricks, Journal of Derivatives 9, 41–53.

Ross, S. A. (1976) Options and Efficiency, The Quarterly Journal of Economics 90(1),
75–89.

Rudin, W. (1987) Real and Complex Analysis (International Series in Pure and
Applied Mathematics), 3rd ed., McGraw-Hill, New York.

Schöbel, R. and J. Zhu (1999) Stochastic Volatility with an Ornstein Uhlenbeck

Process: An Extension, European Finance Review 3, 23–46.

| 84
Schoutens, W. (2003) Lévy Processes in Finance: Pricing Financial Derivatives (Wiley
Series in Probability and Statistics), John Wiley & Sons, Chichester.

Sepp, A. (2003) Pricing European-Style Options under Jump Diffusion Processes with
Stochastic Volatility: Applications of Fourier Transform, Working Paper,
University of Tartu.

Sepp, A. (2006) Extended CreditGrades Model with Stochastic Volatility and Jumps,
Wilmott Magazine, September, 50–62.

Staunton, M. (2006) Stochastic Volatility without Complex Numbers, Wilmott

Magazine, May, 46–48.

Stein, E. and J. Stein (1991) Stock Price Distributions with Stochastic Volatility: An
Analytic Approach, Review of Financial Studies 4, 727–752.

Tahani, N. (2004) Valuing Credit Derivatives Using Gaussian Quadrature: A

Stochastic Volatility Framework, Journal of Futures Markets 24(1), 3–35.

Tahani, N. and X. Li (2007) Pricing Interest Rate Derivatives under Stochastic

Volatility, Working Paper, York University.

Tankov, P. (2004) Lévy Processes in Finance: Inverse Problems and Dependence

Modelling, Ph.D. Dissertation, École Polytechnique, Paris.

Tankov, P. (2009) Calibration de Modèles et Couverture de Produits Dérivés, Lecture

Notes, École Polytechnique et Université Paris VII, Edition 2009. Available at
http://people.math.jussieu.fr/~tankov/MA/

Titchmarsh, E. C. (1975) Introduction to the Theory of Fourier Integrals, Reprint of

2nd ed., Oxford University Press, London.

Waller, L. A. (1995) Does the Characteristic Function Numerically Distinguish

Distributions, The American Statistician 49(2), 150–152.

Waller, L. A., B. W. Turnbull and J. M. Hardin (1995) Obtaining Distribution

Functions by Numerical Inversion of Characteristic Functions with
Applications, The American Statistician 49(4), 346–350.

West, G. (2005) Better Approximations to Cumulative Normal Functions, Wilmott

Magazine, May, 70–76.

| 85
Wu, L. (2008) Modeling Financial Security Returns using Lévy Processes, In:
Handbooks in Operations Research and Management Science: Financial
Engineering, Volume 15, Eds. J. Birge and V. Linetsky, Elsevier, North-Holland.
Available at http://faculty.baruch.cuny.edu/lwu/papers/handbooklevy.pdf

Yan, G. and F. B. Hanson (2006) Option Pricing for a Stochastic-Volatility Jump-

Diffusion Model with Log-Uniform Jump-Amplitudes, Proceedings of the 2006
American Control Conference, Minneapolis, Minnesota, June 14–16.

Zeliade Systems (2009) Heston 2009, Zeliade Systems White Paper, September.
Available at http://www.zeliade.com/whitepapers/zwp-0004.pdf

| 86