com/abstract=1759668
Applications of Periodic and QuasiPeriodic Expansions to
Options Pricing
Dominique Bang
1
1
Bank of America Merrill Lynch, London
Dominique.Bang@baml.com
Abstract. We present a new formulation of the call payo as a convergent
trigonometric series, where the corresponding frequencies are selected accord
ing to a minimum variance algorithm. When the characteristic function of
the logmoneyness is available, this provides an ecient alternative to trans
form methods for option valuation (such as Fourier Integral), reducing the
frequency spectrum from a continuum to a discrete set.
1 Introduction
A standard approach to pricing vanilla options is based on integral representations
(such as the Fourier integral) of a call payo, requiring the characteristic function
of the logmoneyness to be known in a closed form, as is the case for many models
of interest (e.g. Black&Scholes, Heston, Stein&Stein, Cox process, Levy processes
etc..). These methods generally rely on a numerical integration of the (inverse)
Fourier Integral, which can be computationally intensive. Moreover, the related
integral being highly oscillatory and thus prone to inaccuracy, care needs to be
taken in the way the integral is truncated and discretized. In particular, in the
case of Heston dynamics, a number of tricks can be used to improve the accuracy
of the numerical integration while keeping reasonable the number of evaluations of
the characteristic function as described comprehensively in [2]. Alternatively, an
interesting idea has been developped in Fang and Oosterlee [4]
1
where the call price
is approximately expanded in a Fourier Series using an approximate density based
on integral truncation; the authors demonstrate the eciency of their approach
compared with Carr and Madan [3] brute force integration in the case of Heston
Dynamics. In this note, we develop a dierent approach, seeking for an exact series
representation of a call payo in term of trigonometric functions. The idea is to nd
(recursively) the combination of trigonometric functions that is closest(in a sense to
be specied later) to the call price and then, as for the Fourier Integral, to apply the
expectation operator. In a sense, this approach amounts to performing a Principal
Components Analysis of the Call option payo itself, using trigonometric functions
(for which the expectation is known in a closed form via the characteristic function
of the logmoneyness). We demonstrate constructively the existence of such a series,
show its convergence toward the call payo in a strong sense (for the chosen norm)
1
Thanks to Leif Andersen for pointing out this reference and for constructive conversations.
1
Electronic copy available at: http://ssrn.com/abstract=1759668
and give an error estimate when the call price is computed using a truncation of
the series. Finally, as an illustration, we apply this method to option pricing under
Heston dynamics. Additionnally, we provide in passing a simple (approximate but
arbitrarily accurate) representation of a call option in terms of Fourier series, similar
to Fang and Oosterlees cosine expansion, but with an improved rate of convergence.
2 Notations and General Framework
The payo of a call struck at K maturing at T on the underlying S
T
is given by:
(S
T
K)
+
= K(e
x
e
x
2
e
x
2
)
x
= ln(
S
T
K
)
From contour integral theory, we have that:
(x)
= e
x
2
=
1
_
+
0
f
(x)
2
+
1
4
d, f
(x)
= cos(x)
yielding
(S
T
K)
+
= S
T
Ke
x
2
(x) (1)
This decomposition formula shows how to recover the call payo using a con
tinuum of frequencies in the Fourier space. Applying the expectation operator to
this equality yields a valuation formula for the call option, provided that the char
acteristic function of the logmoneyness at maturity, (s) = E
T
(e
(
1
2
+is)x
), is known
in closed form. As pointed out above, this integral, being highly oscillatory, needs
particular care in the way it is discretized and/or truncated. More over, as a Fourier
transform of an unsmooth fonction, and per Heisenberg principle, the integrand de
cays slowly to zero, or, in other terms high frequencies are necessary to describe well
the underlying payo function. The aim of this note is to develop an alternative to
this approach by providing a trigonometric series representation of (x) rather than
an integral, that is to project the function on a discrete set of frequencies rather
than a continuum. More precisely, we aim to write
=
n=1
n
e
n
where
n
is a scalar and e
n
is a function such that e
n
span
_
f
(p
j
)
j
_
1jn
, where
the dierentiation is performed w.r.t (the element
p
p
f
(x) corresponds to a
polynomial of degree p times a sin() or a cos() depending on the parity of p). The
frequencies
j
are selected through a minimum variance algorithm (which, in our
case, is formulated as a maximization of a normalized scalar product
2
to be dened
2
This approach is comparable to the one developped by Laskar [5] in the context of KAM theory.
The latter proved to be very successful in areas as diverse as Celestial Mechanics and High Energy
Particles Dynamics.
2
Electronic copy available at: http://ssrn.com/abstract=1759668
later on) and can be seen as a PCA. Next, we introduce the framework in which this
analysis takes place, which is naturally Hilbert space. More precisely, consider the
weighted space H
=
_
+
0
g(x)h(x)(x)dx
g = g
=
_
g, g
(x) > 0
_
+
0
x
p
(x)dx < , p 0
where the last technical condition is designed to ensure that for any 0, the
elements f
remains in H
)
0
, satisfying some optimal constraints and complete w.r.t . To gain further
intuition about the forthcoming construction, we emphasize the following statement:
Lemma 1. let g
and
denote a xed element in H
,
_
= arg min  g

2
.
Find
= arg max
_
,
g
g

_
2
, =
,g

g

2
.
Proof 1. From  g

2
= 0 we must have =
,g
g

2
, so that
 g

2
= 
2
2 , g
+
2
g

2
= 
2
_
,
g
g

_
2
which shows that minimizing  g

2
is equivalent to maximizing
_
,
g
g

_
(the
lemma is also clear from a Lagrange Multipliers perspective).
The rst formulation amounts to nding the element g
1
which maximizes
_
,
f
f

_
, set e
1
=
f
1

f
1

and
1
=
_
,
f
1

f
1

_
and
3
restart the same procedure with
1
=
1
e
1
and so forth. In section 4, we provide
a rigorous formalization of this algorithm, stating the main result of this note, that
is the existence of (e
n
)
nN
along with the convergence of the series
i=1
i
e
i
to
wards ., but prior to that, in section 3 we give a selfcontained result which allows
for (approximate) call pricing in a very ecient and simple manner. This result is
based on Fourier Series expansion of a periodization of .
3 Periodic decomposition of a Quasicall payo
The following result
3
gives a proxy for the call price using the characteristic function
(s) = E
T
(e
(
1
2
+is)x
) of the logmoneyness.
Theorem 3.1 Let A be a real positive number. Consider the 2A periodic
function
A
(x) such that
A
(x) = (x) on x A, and C
A
(F, K) the (approximate)
call price by replacing with
A
in eq.1. Then we have:
C
A
(F, K) = F K
_
l=0
c
l
_
l
A
_
_
c
l
= 4A
1 (1)
l
e
A
2
A
2
+ (2l)
2
, c
0
= 2
1 e
A
2
A
and the following error estimates hold:
Error in prices:
C(F, K) C
A
(F, K)
FK
_
N (K exp
A
) + 1 N (K exp
A
)
Error in normal (
N
) and Black (
B
) vol for an atthemoney option:
N
A
_
2
T
F
_
N (K exp
A
) + 1 N (K exp
A
)
B
A
_
2
T
_
N (K exp
A
) + 1 N (K exp
A
) ()
where N() is the cumulative distribution function of S
T
. Proof: see Appendix A.
Remark 3.i The idea of using Fourier Series rather than Integral has been rst
exploited by Fang and Oosterlee in [4]. Though, the approach is slightly dierent in
their paper, as it is the integral representation of the density itself which is truncated,
leading eventually to a cosine series. We note that their approach involves a total of
three levels of approximations whereas ours only involves two(in the choice for the
threshold A and the truncation of the Fourier Series) with an improved decay rate
of the Fourier coecients. The latter being rescaled by (.) which usually vanishes
exponentially fast, this shouldn t make too much of a dierence in practice, except
3
The idea to directly periodize originates from a conversation with Alex Lipton. We wish to
thank him in passing for his interest in this work.
4
when the distribution of the underlying is close to singular (e.g. when the option
maturity is very small) and where the call is close to its intrinsic value. Also, our
Formula is simpler, and the error estimate for the choice of A quite transparent
(N(x) can be easily estimated for small/large x). The previous formula has been
mentioned in passing, as the core subject of this paper is to provide an exact series
expansion, rather than a proxy. However, its great simplicity compared to the
forthcoming quasiperiodic expansion sometimes outweights the fact that it is only
an imperfect representation of the call.
4 QuasiPeriodic decomposition of a call payo
The next theorem is the main result of this note: it proves constructively the exis
tence of a countable trigonometric family (e
n
)
nN
satisfying some optimality criteria
and on which can be decomposed, in a Hilbert sense.
Theorem 4.1 There exists an orthonormal family (e
n
)
nN
H
N
with asso
ciated frequencies (
n
0)
nN
, indices (p
n
0)
n1
and scalars (
n
0)
nN
such
that:
=
i=1
i
e
i
= , e
i
e
i
span
_
f
(p
j
)
j
_
1ji
where the convergence of the series should be understood as a strong convergence in
H
= arg max
0
0
() where
0
()
=
_
,
f
f

_
2
(2)
set and store:
e
1
=
f
1
_
_
+
0
cos
2
(
1
x)(x)dx
,
1
= , e
1
, p
1
= 0 (3)
2. nd
n+1
for n > 1 such that:
n+1
= arg max
0
n
() where
n
()
=
_
,
f
n
i=1
f
, e
i
e
i
f
n
i=1
f
, e
i
e
i

_
2
5
set and store:
e
n+1
= lim
+
n+1
f
n
i=1
f
, e
i
e
i
f
n
i=1
f
, e
i
e
i

n+1
= , e
n+1
p
n+1
= min
_
p such that f
(p)
n+1
i=1
_
f
(p)
n+1
, e
i
_
e
i
= 0
_
Moreover (for completeness and practical interest) we have for any n 1 :
lim
n
() = 0
Proof: see Appendix B. As for the existence of the family (e
n
)
nN
, see B.1, for
the convergence of the associated series toward , see B.2 and for the last statement
n
(
n
) = 0, it comes directly from a second order Taylor expansion associated with
a maximality argument.
Remark 4.i In (b) the limit is taken from above lim
+
n+1
. This choice is made
to resolve the intrinsic ambiguity on e
n+1
(which can be replaced by e
n+1
without
any disruption).
Remark 4.ii The fact that
n
(
n
) = 0 is useful in practice, providing informa
tion about the zeros of
n
(
n
) when solving for
n+1
(in particular,
n+1
=
n
, and
moreover
n+1
should be fairly far from
n
). In practice, one can observe a persis
tence of the zeros of
n
() that is
n
(
i
) = 0 not only for i = n but also for i n
(or at least it remains small). However, this is not necessarily the case for all n.
Thus, a frequency
j<n1
may reappear when solving for arg max
n
(). When this
occurs, there is an indeterminacy in lim
+
n+1
f
n
i=1
f
,e
i
e
i

f
n
i=1
f
,e
i
e
i
, as both numerator
and denominator collapse, involving naturally the derivatives of the elements f
(see
LHopitals rule). The integer p
n
can be seen as a counter of the visits of the fre
quency
n
before time n and also corresponds to the order of dierentiation in f
n
when
n
= 0. For the special case
n
= 0 the degree of dierentiation is twice the
number of visits n.
Remark 4.iii We insist that, by construction, once e
n
has been computed and
stored, only one additional frequency is necessary to generate e
n+1
. This means
that the algorithm can be used on an adaptative basis with a low computational
cost (the numbers used to compute e
n
can be stored and reused for e
n+1
). This is of
prime importance in option pricing, where the expectation operator will be applied
to the elements e
n
and thus, will lead to evaluations of the characteristic function,
which is generally computationnally intensive. In our case, the evaluations of the
characteristic function will be stored to be possibly reused when/if incrementing n.
5 Option Pricing
5.1 Valuation Formula and Error Estimate
We apply the previous approach to the pricing of a call option in a model for which
the characteristic function of the logmoneyness () = E
T
(e
(
1
2
+i)x
) is known, as
6
suming that (
j
)
j1
, (p
j
)
j1
, (e
j
)
j1
and (
j
)
j1
have been found following the
procedure in 4 such that:
e
x
2
=
i=1
i
e
i
, e
i
=
i
j=1
C
ij
p
j
f
p
j

j
F
= E(S
T
)
for some constants C
i,j
. Also we introduce (.), the density function of the log
moneyness x = log
_
F
T
K
_
, along with its symmetrization (s)
= (s) + (s). We
dene the exact and approximate (truncated) price of the call option by
C(F, K)
= E (S
T
K)
+
C
n
(F, K)
= F KE
_
e
x
2
n
i=1
i
e
i
_
and assume that the norm 
> 0
such that f
< C
f
for any f H
H
.
Theorem 5.1 Under the previous assumptions, the following results hold:
Option Valuation
For the exact price we have
C(F, K) = F K
_
i=1
i
i
j=1
C
ij
p
j
()
p
j

=
j
_
and for the proxy:
C
n
(F, K) = F K
_
n
j=1
n
j
p
j
()
p
j

=
j
_
n
j
=
n
i=j
C
ij
i
Error Estimate
We have the following bound for the truncation error:
C(F, K) C
n
(F, K) C
FK
_
,
n
i=1
2
i
and for the normal implied vol error atm we have:
N
n
_
,
n
i=1
2
i
and for the lognormal vol atm approximately
B
n
_
,
n
i=1
2
i
7
Proof: See appendix C
Remark 5.i As for the standard Fourier approach, this formula allows a strip
of options for dierent strikes to be priced at once since
E
_
exp
_
(
1
2
+ iz)x
__
= exp
_
(
1
2
+ iz) ln(K)
_
E
_
exp
_
(
1
2
+ iz) ln(F
T
)
__
and the evaluations of the characteristic functions can be stored and reused for dif
ferent strikes.
Remark 5.ii We emphasize that the computation of C
n
(F, K) requires exactly n
evaluations of the characteristic function (or its derivatives); moreover, as already
stated above, one increment of n will result in a single new evaluation of , pro
vided the past evaluations have been stored, which makes an adaptative procedure
possible.
Remark 5.iii The error estimate shows clearly the impact of the norm 
on the
deviation of the proxy from the exact price. Obviously C
should be reasonable so
that the series converge suciently rapidly.
5.2 How to choose the truncation level n?
From the last section, it is clear that the error depends directly on the product
C
_
,
n
i=1
2
i
. Once has been chosen, the number of terms necessary to
achieve a given accuracy can be easily inferred by setting
C
FK
_
,
n
i=1
2
i
<
when C
TZ, Z N(0, 1)
(s) = (s) =
1
2T
exp(
s
2
2
2
T
)
where N(0, 1) denotes the standard normal distribution. For , we use an exponen
tial weight
(s) = exp(s)
A few calculations show that f
< C
f
T
yielding C
=
_
_
2e
2
T = 1, a good choice for our proxy distribution will be (s) = exp(s). In that
case, our experiments suggests that values n 15 20 produce very good results,
not only when the terminal distribution of the underlying is lognormal, but also
for richer dynamics such that stochastic volatility, as long as the variance of the
8
process is reasonnably in line with the one implied by . When C
is not known
a priori, another possible strategy is to evaluate the elements (e
i
)
1i100
once and
for all, and then increment the number of elements used within this family until
convergence is achieved (e.g the results do not change by more than some given
quantity). This adaptative procedure is made possible as the evaluations of the
characteristic functions are stored and reused. In the next section we examine the
choice for the weight .
5.3 Example of Weights
In this section we provides useful quantities related to the practical implementation
when the weights are either gaussian or exponential. We recall that the scalar
product in H
is dened by:
g, h
=
_
+
0
g(x)h(x)(x)dx
where, without loss of generality, we impose
_
+
0
(x)dx = 1.
5.3.1 Gaussian Weight
This case corresponds to Black&Scholes dynamics for the underlying, thus, for an at
themoney option, the logmoneyness can be assumed to be gaussian (and centered
for simplicity). This results in the following table:
distribution Gaussian
(x)
2
2
exp(
x
2
2
2
T
)
f
, f
exp
_
2
+
2
2
2
T
_
cosh (
2
T)
f

2
1
2
[1 + exp (2
2
2
T)]
, f
[exp(m
2
)erfc(m)] , m =
_
1
2
i
_
_
T
2
, 2N(
T) exp(
2
T
2
)
Remark 5.ivThe Gaussian distribution is not necessary appropriate when the un
derlyings distribution has fat tails, which is the case in stochastic volatility models.
5.3.2 Exponential Weight
The last remark suggests that we might consider a more slowly decaying density,
and a good tractable candidate is the exponential density for which we will analyze
9
in more details.
distribution Exponential
(x) exp(x)
f
, f
2
2
+
2
+
2
(
2
+(+)
2
)(
2
+()
2
)
f

2
2
+2
2
2
+4
2
, f
(
+
1
2
)
(
+
1
2
)
2
+
2
,
1+
In this case, we can work out the rst term in the expansion. A short calculation
leads to:
0
(, ) =
, f
2
f
, f
=
_
(
+
1
2
)
(
+
1
2
)
2
+
2
_
2
2
+2
2
2
+4
2
5
4
_
_
4
+
8
25
2
2
+
32
25
2
cos (
1
)
1
2
_
6+2
(25
2
+8+32)
(25
2
+8+32)
1
8
2(1+2)
(25
2
+8+32)
4+16+11
2
+
(25
2
+8+32)
3+
(25
2
+8+32)
for = 1 this becomes:
655
4
. 4374
1
24
8959+737
65
1777+303
65
. 6942
e
1
(x)
cos
655
4
x
31+
65
651
1. 1299 cos (0. 4374 x)
The rst element e
1
capture
2
1
. 48198 of the total variance , = 0.5. That is, in
relative terms approximatively 96.4% of the total variance. The 20 rst frequencies
have been computed numerically:
=
_
0.4374, 1.6489, 0.0000, 0.8941, 2.9970, 4.6829, 2.3117, 6.5990, 3.7550, 1.2114,
8.7081, 5.5708, 10.997, 7.5851, 13.455, 9.7852, 16.074, 18.836, 12.160, 0.0000
_
=
_
69.42%, 10.61%, 5.202%, 4.034%, 3.963%, 1.910%, 1.077%, 1.056%, 0.800%, 0.663%
0.645%, 0.474%, 0.423%, 0.306%, 0.293%, 0.212%, 0.211%, 0.158%, 0.153%, 0.138%
_
Note that the twentieth term is the rst occurrence of a frequency revisit. Using the
rst three terms covers 99.2% of the variance etc.. Next, we apply this methodology
to Heston Dynamics.
10
5.4 An application to Heston Dynamics
In this section we assume that the underlying follows Heston dynamics, that is,
using similar notations than in Lipton[1]:
dS
t
S
t
=
v
t
dW
(S)
t
dv
t
= (1 v
t
) dt +
v
t
dW
(v)
t
, v
0
= 1
dt = dW
(S)
t
dW
(v)
t
From now on we consider an initial spot S
0
= 5% and the base case: = 40%,
= 100%, = 0, = 10% and T = 5y. The standard deviation in the base case is
80% justifying the use of an exponential weight with a caracteristic constant = 1.
In the tables in Appendix D, we vary individually the dierent model parameters
and compare a 20 terms quasiperiodic expansion with a benchmark brute force
integration, reporting the errors between the two. The results suggest that a very
good accuracy is reached with 20 terms, for a large spectrum of model parameters.
6 Conclusion
The contribution of this paper is two folds: rst we have slightly simplied and im
proved an original idea developped in [4] and second we have developped a new exact
decomposition of the call payo in a polynomialtrigonometric series. Existence and
convergence results have been established in a rigorous framework, yielding a val
uation formula along with error estimates when truncation is applied for practical
purpose. The expansion has been tested under Heston dynamics, demonstrating
the robustness and excellent accuracy of the approach for a number of terms as
low as 20. The method could, in principle, be applied to other dynamics whenever
the caracteristic function of the logspot is known in a closed form. We plan, in a
near future, to apply this technique to close to singular distribution (such as those
appearing for very short term maturity) known to pose challenging problems when
used in a Fourier Transform framework.
Aknowledgements
I am indebtful to Cyril Grunspan for initial motivating conversations, to Yann Ticot
for his insights through out this work and to Professor Peter Hawkes for useful
suggestions. Also I want to thank my colleagues in Bank Of America Merrill Lynch,
and more specially Leif Andersen, Alex Lipton and Henrik Rasmussen for useful
comments and support. Last, i am grateful to prof. Cornelius W. Oosterlee and
prof. Fang Fang for interesting feed back and discussion.
11
A Proof of Theorem 3.1
A
(x) being 2A periodic, even, piecewise C
1
and C
0
, can be decomposed into its
Fourier (cosinus) Series:
A
(x) =
l=0
c
l
cos
_
l
A
x
_
c
l>0
= 4A
1 (1)
l
e
A
2
A
2
+ (2l)
2
, c
0
= 2
1 e
A
2
A
Thus, we have that:
C
A
(F, K) = E
T
_
S
T
Ke
x
2
A
(x)
_
= F K
l=0
c
l
E
T
_
e
x
2
cos
_
l
A
x
__
yielding
C
A
(F, K) = F K
_
l=0
c
l
_
l
A
_
_
By CauchySchwartz,
C(F, K) C
A
(F, K) = K
E
T
_
e
x
2
((x)
A
(x))
K
_
E
T
[e
x
]
_
E
T
_
((x)
A
(x))
2
ds
K
_
E
T
[e
x
]
_
E
T
_
1
x>A
ds
=
FK
_
N (K exp
A
) + 1 N (K exp
A
)
At the money K = F, we have, in term of normal volatilities:
N
A
_
2
T
F
_
N (K exp
A
) + 1 N (K exp
A
)
and in term of lognormal volatility, approximately:
B
A
_
2
T
_
N (K exp
A
) + 1 N (K exp
A
)
B Proof of Theorem 4.1
B.1 Existence result
In this section we show that the recursive construction described in 4 is valid at any
order
12
Proof: We prove the result by induction. For n = 1, we have that
f
 =
_
+
0
cos
2
(x)(x)dx > 0
, f
=
_
+
0
cos(x)(x)(x)dx
0
()
=
_
_
+
0
cos(x)(x)(x)dx
_
2
_
+
0
cos
2
(x)(x)dx
, 0
which shows that
1
is a positive continous function of . Moreover, using Lebesgues
Lemma, we see that:
lim
_
+
0
cos
2
(x)(x)dx =
1
2
_
+
0
(x)dx
lim
_
+
0
cos(x)(x)(x)dx = 0
so that:
lim
0
() = 0
Relying now on the continuity of
1
associated with a compacity argument, we can
ensure that there exist
0 such that
= arg max
0
_
_
,

_
2
_
we now let
1
=
e
1
=
f
1
_
_
+
0
cos
2
(
1
x)(x)dx
and assume now a valid construction of (e
i
)
1in
and (
i
)
1in
up to the rank n.
We proceed as for the rst term, considering
e
=
f
n
i=1
f
, e
i
e
i
f
n
i=1
f
, e
i
e
i

=
i
, i = 1..n
which is well dened since f

2
>
n
i=1
f
, e
i
2
as f
/ Span(e
1
, .., e
n
) when =
i
, i = 1, .., n.
Let us consider now
j
0, j n. We have trivially f
n
i=1
f
j
, e
i
e
i
= 0
so that we can consider
p = max
_
l 0/ 0 k l,
k
f
k

j
=
n
i=1
_
k
f
k

j
, e
i
_
e
i
_
13
then, using Taylor expansion up to order p, we have that:
f
i=1
f
, e
i
e
i
=
(
j
)
p+1
(p + 1)!
_
p+1
f
p+1

j
n
i=1
_
p+1
f
p+1

j
, e
i
_
e
i
+ o(1)
_
so that
e
=
(
j
)
p+1

j

p+1
p+1
f
p+1

n
i=1
_
p+1
f
p+1

j
, e
i
_
e
i
p+1
f
p+1

n
i=1
_
p+1
f
p+1

j
, e
i
_
e
i
= arg max
0
n
() < . Also,
in virtue of what preceed,
e
n+1
= lim
+
n+1
f
n
i=1
f
,e
i
e
i

f
n
i=1
f
,e
i
e
i
is well dened. Thus, we can perform this
construction at any order.
B.2 Convergence result
n
=
n
i=1
i
e
i
(e
i
)
1i
being orthonormal, we have:
,
n
,
n
=
n
i=1
2
i
As an increasing bounded sequence,
n
,
n
is convergent with limit l 0 (this
also yields
= 0) . Per

n+p
n

2
=
n+p
n
,
n+p
=
n+p
i=n+1
2
i
i=n+1
2
i
we see that
n
is a Cauchy sequence in H
]
2
= > 0
e
=
f
i=1
, e
i
_
e
i
_
i=1
, e
i
_
2
by orthogonality, we have that:
, e
= , e
14
and from
= 0, there exists N
1
such that n N
1
we have
2
n
<
2
. Also, we
have that:
lim
n
e
n
= e
e
n
=
f
n1
i=1
, e
i
_
e
i
_
n1
i=1
, e
i
_
2
such that there exists N
2
> 0 verifying n N
2
, , e
N
2
>
2
. Considering now
N = max(N
1,
N
2
), we see that we must have simultaneously , e
N
2
>
2
and
2
N
<
2
which clearly violates the optimality of e
N
. Thus we must have = = 0 which
completes the proof.
C Proof of Theorem 5.1
For the valuation formulae, we have that:
E
_
e
x
2
e
i
_
=
i
j=1
C
ij
E
_
e
x
2
p
j
f
p
j

j
_
=
i
j=1
C
ij
_
p
j
p
j
E
_
e
x
2
e
ix
_

j
_
=
i
j=1
C
ij
p
j
()
p
j

=
j
_
yielding:
C(F, K) = F K
i=1
i
i
j=1
C
ij
p
j
()
p
j

=
j
_
C
n
(F, K) = F K
_
n
j=1
n
j
p
j
()
p
j

=
j
_
,
n
j
=
n
i=j
C
ij
i
Error estimates are a simple consequence of CauchySchwartz inequality.
D Numerical results and tables
In the next tables we detail the deviations from the exact values(computed via
Fourier Integral and an adequate number of discretizations points) in term of implied
volatilities for dierent set of parameters.
15
Table 1: Varying atm Volatility
strike 10% 20% 40% 50% 70% 100%
1% 1.46E02 2.15E04 6.55E06 5.17E07 3.03E05 2.36E04
3% 5.62E04 1.09E04 3.30E05 2.31E05 2.39E05 1.07E04
5% 3.87E04 3.19E05 1.36E05 1.47E05 2.10E05 9.64E05
7% 4.63E05 3.45E05 2.46E05 1.96E05 2.25E05 1.01E04
9% 2.21E03 1.58E04 3.54E05 2.39E05 2.44E05 1.11E04
Table 2: Varying correlation
strike 0.9 0.5 0 0.5 0.9
1% 1.97E05 1.46E04 6.55E06 3.32E05 1.58E03
3% 6.89E05 3.79E05 3.30E05 2.05E05 3.35E05
5% 2.91E05 1.99E05 1.36E05 2.05E05 6.59E05
7% 8.22E06 1.10E05 2.46E05 3.38E05 2.22E05
9% 4.72E05 3.48E05 3.54E05 2.80E05 2.43E05
Table 3: Varying vol of var
strike 1% 10% 50% 100% 300% 500%
1% 8.74E05 8.22E05 1.37E05 6.55E06 4.74E06 7.56E05
3% 2.80E05 2.78E05 2.64E05 3.30E05 1.10E04 1.44E04
5% 2.45E05 2.44E05 2.23E05 1.36E05 1.18E04 5.29E04
7% 2.60E05 2.59E05 2.44E05 2.46E05 1.72E05 1.38E05
9% 2.90E05 2.88E05 2.73E05 3.54E05 2.01E04 5.33E04
Table 4: Varying time to maturity
strikeT 6M 1Y 2Y 5Y 10Y 20Y
1% 1.34E01 3.54E03 2.88E04 6.55E06 1.59E05 5.33E05
3% 9.53E04 2.29E04 8.98E05 3.30E05 1.82E05 2.74E05
5% 8.20E05 1.81E06 1.55E05 1.36E05 1.27E05 2.45E05
7% 2.16E04 9.57E05 5.31E05 2.46E05 1.58E05 2.59E05
9% 1.61E03 3.06E04 1.06E04 3.54E05 1.88E05 2.83E05
16
References
[1] Andersen L. and Andreasen J.(2002), Volatile Volatilities, Risk Magazine.
[2] Andersen L. and Piterbarg V.(2010), Interest Rates Modeling, Vol. 1, Chapt. 9,
Atlantic Financial Press.
[3] Carr P.P. and Madan D.B.(1999), Option valuation using the fast Fourier trans
form. J. Comp. Finance, 2:6173.
[4] Fang F. and Oosterlee C.W.(2008), A novel pricing method for european options
based on FourierCosine series expansions, SIAM SISC.
[5] J. Laskar(2001), Frequency Map Analysis and quasiperiodic decompositions, pro
ceedings of the school Hamiltonian systems and Fourier analysis, Porquerolles.
[6] Lipton A.(2002), The vol Smile Probem, Risk Magazine.
17