Quantitative Finance

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/rquf20

Closed-form option pricing for exponential Lévy

models: a residue approach

Jean-Philippe Aguilar & Justin Lars Kirkby

To cite this article: Jean-Philippe Aguilar & Justin Lars Kirkby (2023) Closed-form option pricing
for exponential Lévy models: a residue approach, Quantitative Finance, 23:2, 251-278, DOI:
10.1080/14697688.2022.2152365

To link to this article: https://doi.org/10.1080/14697688.2022.2152365

Published online: 15 Dec 2022.

Submit your article to this journal

Article views: 45

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=rquf20
Quantitative Finance, 2023
Vol. 23, No. 2, 251–278, https://doi.org/10.1080/14697688.2022.2152365

Closed-form option pricing for exponential Lévy

models: a residue approach
JEAN-PHILIPPE AGUILAR *† and JUSTIN LARS KIRKBY‡
†Société Générale, 17 Cours Valmy, Paris - La Défense Cedex, 92987, France
‡School of Industrial and Systems Engineering, Georgia Institute of Technology, Atlanta, GA, 30318, USA

(Received 3 March 2022; accepted 22 November 2022; published online 15 December 2022 )

Exponential Lévy processes provide a natural and tractable generalization of the classic Black–
Scholes–Merton model which account for several stylized features of financial markets, including
jumps and kurtosis. In the existing literature, closed-form option pricing formulas are sparse for
exponential Lévy models, outside of special cases such as Merton’s jump diffusion, and complex
numerical techniques are required even to price European options. To bridge the gap, this work
provides a comprehensive and unified pricing framework for vanilla and exotic path independent
options under the Variance Gamma (VG), Finite Moment Log Stable (FMLS), one-sided Tem-
pered Stable (TS), and Normal Inverse Gaussian (NIG) models. We utilize the Mellin Transform
and residue calculus to obtain closed-form series representations for the price of several options,
including vanillas (European), digitals, power, and log options. These formulas provide nice theoret-
ical representations, but are also efficient to evaluate in practice, as numerous numerical experiments
demonstrate. The closed-form nature of these option pricing formulas makes them ideal for adoption
in practical settings, as they do not require complicated pricing methods to achieve high-accuracy
prices, and the resulting pricing error is reliably controllable.

Keywords: Lévy process; Stable process; Variance Gamma process; Normal inverse Gaussian
process; Stochastic volatility; Option pricing; Mellin Transform
AMS subject classifications (MSC 2020): 60E07, 60E10, 60H35, 65C30, 65T50, 91G20, 91G30
JEL Classifications: C00, C02, G10, G12, G13

1. Introduction As a prominent example, exponential Lévy models are a

direct generalization of the Black–Scholes process which
1.1. Lévy markets and existing pricing techniques retain the mathematical tractability afforded by stationary
and independent increments yet are able to parsimoniously
The problem of financial derivatives valuation is a classic one
capture the heavy (asymmetric) tails needed to generate
in quantitative finance, which draws perpetual interest due
implied volatility skews. These models have seen widespread
to its practical importance and mathematical intrigue. Var-
application in areas such as option pricing (Boyarchenko
ious models have been proposed to capture the risky asset
and Levendorskii 2000, 2002a, Kou 2002, Boyarchenko
which underlies financial derivatives, starting with the clas-
sic Black–Scholes–Merton (BSM) model. While BSM is a and Levendorskiǐ 2002, Schoutens 2003, Fang and Oost-
useful starting point for modeling a risky asset, its inabil- erlee 2008, Boyarchenko and Levendorskii 2011, Fusai
ity to capture market-implied volatilities limits its practical et al. 2011, Kirkby 2015, 2017, 2018, Kirkby and
use for derivatives valuation. The empirical inconsistencies Deng 2019), risk management (Asmussen and Klüppel-
of the Black–Scholes model compared to market reality are berg 1996, Barndorff-Nielsen et al. 2001, Cui et al. 2021,
also well documented, see for example, Bates (1996), Carr Kirkby et al. 2020), and insurance (Embrechts 2000,
and Wu (2003b), and Lee and Hannig (2010). To account Chan 2004, Wang and Zhang 2019, Zhang et al. 2020, Xie
for its deficiencies, in terms of both the empirical prop- and Zhang 2020, Kirkby and Nguyen 2021).
erties of risky assets as well as observed market implied Given the importance of Lévy models in financial model-
volatility smiles, several extensions have been proposed. ing, a variety of competing for numerical approaches have
been proposed for option valuation. Techniques such as Par-
tial Differential Integro Equations (PIDE) extend the classic
∗ Corresponding author. Email: [email protected] PDE approaches, see for example (Matache et al. 2004, Cont
© 2022 Informa UK Limited, trading as Taylor & Francis Group
252 J.-P. Aguilar and J. L. Kirkby

and Voltchkova 2005, Li and Zhang 2016). The related Roper (2009), Muhle-Karbe and Nutz (2011), Figueroa-López
radial basis technique has also proven successful (Brum- and Forde (2012), and Andersen and Lipton (2013).
melhuis and Chan 2014, Chan 2016). For models with a
representation as a subordinated Brownian motion, includ-
ing Variance Gamma (VG) and Normal Inverse Gaussian 1.2. Contributions
(NIG), Monte Carlo methods are also feasible (Avramidis and
This work considers a recent development in option pric-
L’Ecuyer 2006, Benth et al. 2006, Kalemanova et al. 2007).
ing under exponential Lévy models which are based on the
An especially efficient strain of Monte Carlo simulation
Mellin transform. Early numerical applications of the Mellin
techniques based on the Fourier transform/Wiener–Hopf
transform in finance targeted the PIDE, such as Panini and
factorization were developed by Kuznetsov et al. (2011),
Srivastav (2004) and Frontczak and Schobel (2010). More
Boyarchenko (2012), Boyarchenko and Levendorskiǐ (2019), recently, Li and Rodrigo (2017), Guardasoni et al. (2020), and
and Boyarchenko and Levendorskii (2020a). Let us also Rodrigo (2021) utilize Mellin transforms to obtain integral
mention truncated asymptotic expansion techniques such as representations of options prices, such as barrier options. By
the Gram–Charlier or Edgeworth series that involve a span- contrast, we apply the theory directly to the convolution oper-
ning over a basis of orthogonal (Hermite) polynomials (Jar- ator as it appears in the risk-neutral option pricing formula and
row and Rudd 1982, Jondeau and Rockinger 2001, Heston derive closed-form analytical pricing formulas using residue
and Rossi 2016); in the VG case, applications of quadra- calculus.
ture (Gauss-Laguerre) methods have also been introduced The Mellin transform (see appendix 2 for definition and
(Loregian et al. 2012). basic properties, and (Flajolet et al. 1995) for a comprehen-
Due to the availability of a closed-form characteris- sive introduction) is closely related to the bilateral Laplace
tic function in most cases, Fourier transform-based tech- transform via the change of variables x → e−x , but its inver-
niques are among the most prevalent for option pricing sion kernel x−s (instead of epx in the Laplace case) generates
under Lévy-driven models and are generally efficient. Since a fundamental correspondence which, schematically, reads
the seminal works of Heston (1993) and Levendorskii and
Boyarchenko (1998), which introduced the Fourier trans-
pole of f ∗ (s) in s = sn −→ term in x−sn in the series
form to derivatives pricing, and (Carr and Madan 1999),
which first considered the use of the Fast Fourier Transform expansion of f (x), (1)
(FFT) in this context, extensions such as Lewis (2001), Lord
et al. (2008), and Jackson et al. (2008) have sought to improve for f a function defined over the positive reals and f ∗ its
the convergence and efficiency of Fourier techniques. Mellin transform. As we will see, in all our applications, the
As explained in Boyarchenko and Levendorskiǐ (2002), variable x will represent the log forward moneyness k and,
Boyarchenko and Levendorskii (2011), and Boyarchenko and thanks to (1), we will obtain representations of option prices
Levendorskiǐ (2015), the use of the FFT for pricing European as fast converging series of positive powers of k (and other
options can decrease both the accuracy and speed of the pro- parameters, depending on the model). Of course, the existence
cedure. Very efficient realizations of the inverse Fourier trans- of the mapping (1) is conditioned to the possibility of express-
form are introduced in Boyarchenko and Levendorskii (2011), ing the inverse transform as a sum of residues only, that is, that
Boyarchenko and Levendorskiǐ (2015) and Boyarchenko and a Jordan style lemma holds for the integrand; in our cases,
Levendorskiǐ (2019), which overcome this deficiency of the only gamma functions of linear arguments will be involved,
FFT and lead to accuracy that is often several orders of which enables the Jordan condition to be directly expressible
magnitude greater than conventional methods. For example, as a simple difference of some coefficients, except in the NIG
the recent method of Boyarchenko and Levendorskiǐ (2019) model where a supplementary hypothesis involving the log
and Boyarchenko and Levendorskii (2020b) can achieve forward moneyness will be needed.
machine precision accuracy in microseconds for European In details, this work extends and generalizes the results
options, and 10−5 in dozens of microsecond for barrier derived in Aguilar (2019, 2020b), providing a comprehensive
options, respectively. The results in Boyarchenko and Leven- and unified pricing framework for the Variance Gamma (VG),
dorskii (2022a, 2022b) are even more efficient and represent Finite Moment Log Stable (FMLS), one-sided Tempered Sta-
some of the most promising research in this area. ble (TS), and Normal Inverse Gaussian (NIG) models. The
The work of Fang and Oosterlee (2008) introduced contributions of our work are as follows:
cosine basis expansion, which has been extended to
• We provide a comprehensive pricing framework for
wavelet and local basis methods in Ortiz-Gracia and
vanilla and exotic path-independent options under a
Oosterlee (2013), Kirkby (2015), Leitao et al. (2018),
large class of exponential Lévy models, based on
and Boyarchenko et al. (2022). Other techniques such
the machinery of Mellin transforms. This frame-
as Wiener–Hopf factorization (Boyarchenko and Lev-
work unifies some existing literature and extends
endorskii 2002b, 2002a, Kuznetsov 2010, Kudryavtsev
its generality.
and Levendorskiǐ 2009, Kudryavtsev and Zanette 2013, • Using residue calculus, we derive a number of
Boyarchenko and Levendorskii 2013, Phelan et al. 2018) new closed-form pricing formulas under the general
are also quite powerful and are used to price contracts (asymmetric) VG and NIG dynamics, which are
with lookback/barrier features. Finally, the use of small-time two of the most prevalent exponential Lévy models.
asymptotics has been applied for pricing Lévy models in The VG model is one of the more commonly used
 Closed-form option pricing for exponential Lévy models: a residue approach 253

Lévy models in industry, and the analytical pric- 2. Exponential Lévy motions and option pricing
ing formulas are of value in practical applications
such as calibration; some closed formulas have pre- 2.1. Lévy process model
viously been derived using the Mellin transform in
We start by recalling some basics of Lévy processes; we refer
Aguilar (2020b) in the particular case where the
to the classical monographs (Bertoin 1996, Sato 1999) for a
VG process is symmetric, while the formulas we
complete introduction to these rich mathematical objects, and
derive in the present paper are valid for the more
to Cont and Tankov (2004) for a comprehensive overview of
general case of an asymmetric process, thus allow-
their applications to financial modeling.
ing for skewed distributions that are commonly
Throughout, we assume the existing of a risk-neutral mea-
observed in practice. In the NIG case, we focus on
sure Q, and let (, F, {Ft }t≥0 , Q) be a filtered probability
extending the results obtained in Aguilar (2021a) to
space, under which we will conduct pricing. Let T > 0 and
more exotic European payoffs. We also present new
S : t ∈ [0, T] → St be the market price of some financial asset,
closed-form pricing formulas for the one-sided TS
seen as the realization of a time-dependent random vari-
model which are more general than the one derived
able {St }t∈[0,T] on R+ . We assume that St evolves under Q
for the FMLS model in Aguilar and Korbel (2019)
according to
as they allow for a non-null tempering parameter.
• Numerical experiments are provided which demon- dSt
strate the accuracy and efficiency of these analytical = (r − q)dt + dXt , t ∈ [0, T], (2)
St
results compared with six competing Fourier trans-
form techniques, and reference prices are provided where r ∈ R is the risk-free interest rate and q ∈ R is the divi-
to aid in future research.† dend yield (both assumed to be deterministic and continuously
compounded), and where {Xt }t∈[0,T] is a Lévy process. Recall
Since there exists a respectable variety of competing meth- that a Lévy process {Xt }t≥0 is a càdlag process satisfying
ods for this problem, it is important to highlight the key X0 = 0 (Q-almost surely) and whose increments are indepen-
practical advantages of our approach, besides its mathematical dent and stationary. The characteristic function X (u, t) :=
interest: EQ [eiuXt ] satisfies for t ≥ 0

(1) Closed-form expressions: this feature is very desir- X (u, t) = etψX (u) , ψX (u) := log X (u, 1), (3)
able from a practical adoption perspective, because the
formulas are simple to implement. where ψX (u) is called the Lévy symbol or characteristic expo-
(2) Error control: given the series nature of the pricing for- nent. The Lévy symbol is entirely determined by a triplet
mulas, and smooth convergence of the partial sums, (a, b, X (dx)), given by the Lévy–Khintchine formula
achieving a specified error tolerance is straightforward,
 +∞
which is ideal for model calibration. Experiments with 1  
the four models demonstrate our ability to reliably con- ψX (u) = a iu − b2 u2 + eiux − 1 − iux1{|x|<1}
2 −∞
trol pricing error at any prescribed tolerance across a
full range of strikes. × X (dx), (4)
(3) Robustness at all maturity ranges: as we will show,
many commonly applied Fourier methods deteriorate where a is specified below to ensure a risk-neutral pricing
for short times to maturity (a month or less), while our model and b is the Brownian (or diffusion) component. The
formulas actually increase in accuracy. measure X (dx), assumed to be concentrated on R\{0} and
to satisfy
 +∞
min(1, x2 )X (dx) < ∞,
1.3. Structure of the paper −∞

The remainder of this work is organized as follows: The is called the Lévy measure of the process and determines its
mathematical modeling framework is introduced in Section 2, tail behavior and the distribution of jumps.
which outlines our general pricing strategy for Exponential • When X (R) < ∞, one speaks of a process with
Lévy motions. Section 3 derives new formulas for the one- finite activity (or intensity); under this hypothesis
sided TS model, as well as the FMLS model. Section 4 only a finite number of jumps can occur on each
studies the VG model, for both the symmetric and asym- time interval. This configuration corresponds to the
metric cases, and in section 5, we consider the NIG model. class jump-diffusion processes, which have found
For each of the model classes, new closed-form pricing for- option pricing applications, for instance, via the
mulas are derived, and are supported by numerical examples Kou Model (Kou 2002) or the Merton model (Mer-
which demonstrate convergence and accuracy compared with ton 1976); applications to credit risk have also been
existing approaches. The work is concluded in section 6. proposed (Zhou 1997, Lipton 2002).
• When X (R) = ∞, one speaks of a process with
infinite activity (or intensity); this class is far richer,
† Source code is made publicly available at: because an infinite number of jumps can occur on
https://github.com/jkirkby3/PROJ_Option_Pricing_Matlab. any finite time interval. When furthermore b = 0
254 J.-P. Aguilar and J. L. Kirkby

(no Brownian component), one speaks of a pure 2.2.1. Representation in the Mellin space. We now present
jump or purely discontinuous process. Among the an alternative representation for the pricing formula in (10).
most prominent processes with infinite activity are Using definition (A1) in appendix 2, let us define the Mellin
the Finite Moment Log Stable (FMLS) process transform of the density function by
by Carr and Wu (2003a), the Variance Gamma
(VG) process by Madan et al. (1998) and the Nor-  ∞
mal inverse Gaussian (NIG) process by Barndorff- fX∗ (s, τ ) = fX (x, τ )xs−1 dx (11)
Nielsen (1997). 0
• when X (R− ) = 0 (i.e. the Lévy measure is sup-
ported by the positive real axis and the process has for x > 0, and let us assume that this transform converges in
only upward jumps), one says that the process Xt is a strip {s ∈ C, c− < Re(s) < c+ } for some real numbers c− <
spectrally positive; a process will be said to be spec- c+ . From the Mellin inversion formula (A4), we have
trally negative if −Xt is spectrally positive. Last,
let us mention that, if Xt is a subordinator, that  c+i∞
is, an almost surely non-decreasing process, then ds
fX (x, τ ) = fX∗ (s, τ )x−s ,
X (R− ) = 0. c−i∞ 2iπ
Under the dynamics (2), the terminal market price is given
by for any c ∈ (c− , c+ ). Let us also assume that the payoff
(r−q+ωX )τ +Xτ function in (10) is not zero only on R+ , that is
ST = St e , (5)
where τ := T − t is the time to maturity and ωX is the mar-
 
tingale adjustment (also called convexity adjustment, or com- supp x → P(St e(r−q+ωX )τ +x , K1 , . . . , Kn ) ⊂ R+ , (12)
pensator), which is determined by the fact that the discounted
stock price must be a Q martingale:
and denote its Mellin transform by P∗ :
Q (r−q)τ
E [ST |Ft ] = e St . (6)
 ∞
Given the form of the exponential process (5), the condi- ∗
 
P (s) = P St e(r−q+ωX )τ +x , K1 , . . . , Kn xs−1 dx, (13)
tion (6) is equivalent to: 0
 
ωX = −ψX (−i) = − log EQ eX1 . (7)
which is assumed to converge in a strip {s ∈ C, c̃− < Re(s) <
c̃+ } for some real numbers c̃− < c̃+ . If we now choose c ∈
2.2. Pricing strategy (c− , c+ ) ∩ (1 − c̃+ , 1 − c̃− ) where the intersection is assumed
In this section, we first recall the general setup for pric- to be nonempty, it follows from (10) that, under the exponen-
ing contingent claims in exponential Lévy models; more tial Lévy dynamics (2), the value at time t of a contingent
details can be found in several classical monographs, such as claim C delivering a payoff P(ST , K1 , . . . , Kn ) at its maturity
Schoutens (2003) and Boyarchenko and Levendorskiǐ (2002). T is equal to:
We then present the pricing strategy employed in this work,
using the Mellin transform.  c+i∞
−rτ ds
Given a path-independent payoff function P, i.e. a posi- C=e fX∗ (s, τ )P∗ (1 − s) . (14)
c−i∞ 2iπ
tive function depending only on the terminal value ST of the
market price and on some strike parameters K1 , . . . , KN > 0,
then the value at time t of a contingent claim delivering a Formula (14) allows to express the price of a contingent claim
payoff P at its maturity is equal to the following risk-neutral in the Mellin space as the product of two Mellin transforms
expectation: (the transform of the density function, and the transform of
  the payoff function).
C = EQ e−rτ P(ST , K1 , . . . , Kn )|Ft . (8)
Remark 2.1 The factorized formula (14) is particularly well-
In all of the following, we will denote by fX (x, t) the density suited to the pricing of path-independent instruments in
function of Xt with respect to the Lebesgue measure exponential Lévy models because:
dQ[Xt+τ < x|Xt = 0] (1) The Mellin transform of most Lévy densities is either
fX (x, τ ) = . (9)
dx known in exact and simple form (VG and NIG pro-
The conditional expectation (8) can be achieved by integrat- cesses), or straightforward to derive from equivalent
ing all possible realizations for the payoff over the density integral representations such as the Laplace transform
function; given the terminal price (5), this results in: (one-sided stable and tempered stable processes);
 +∞ (2) Although condition (12) may appear restrictive, is not
−rτ the case in practice, because option payoffs are typi-
C=e P(St e(r−q+ωX )τ +x , K1 , . . . , Kn )fX (x, τ )dx.
−∞ cally non null on the positive or negative real axis only.
(10) Introduce the forward strike F and the log-forward
 Closed-form option pricing for exponential Lévy models: a residue approach 255

moneyness k (iii) A European call option has the payoff Peur (ST , K) :=
(ST − K)1{ST >K} := [ST − K]+ . If kX < 0, its Mellin
St transform (13) exists in the strip {s ∈ C, Re(s) < cz }
F := Ke−rτ , k := log + (r − q)τ ,
K and is equal to:
as well as the martingale moneyness kX , which  cz +i∞
(z) (1 − z) (−s + z)
includes the martingale adjustment (7), p∗eur (s) =K (−1)−z
cz −i∞ (1 − s)
kX = k + ωX τ , dz
× (−kX )s−z , (19)
2iπ
and consider, for instance, the usual condition 1{ST >K} ;
comparing to condition (12), we have for any cz ∈ (−1, 0).
  (iv) A power call option (European style) has the payoff
supp x → 1{St e(r−q+ωX )τ +x >K} = (−kX , ∞) (15) Ppow (ST , K) := [STa − K]+ for some positive number
a. If kX(a) < 0, its Mellin transform (13) exists in the
which is included in R+ as soon as kX < 0, i.e. for strip {s ∈ C, Re(s) < cz } and is equal to:
out-of-the-money (OTM) options. To consider ITM
options, it suffices to transfer the calculation to R− by  cz +i∞
using p∗pow (s) = K (−1)−z a−z
cz −i∞
1{ST >K} = 1 − 1{ST <K} (16)
(z) (1 − z) (−s + z) dz
combined with the martingale property for ST and to × (−kX(a) )s−z ,
(1 − s) 2iπ
define the Mellin transform on R− instead of R+ (see (20)
further details in section 2.2.3).
Let us now establish the various Mellin transforms (13) that for any cz ∈ (−1, 0).
will be useful in the paper; to that extent, it will be useful to (v) A log call option has the payoff Plog (ST , K) :=
generalize the concepts of moneyness and martingale money- [log SKT ]+ . If kX < 0, its Mellin transform (13) exists
ness defined in Remark 2.1 in order to take power options into in the strip {s ∈ C, Re(s) < −1} and is equal to:
account: for a > 0 we define
(−kX )s+1
St p∗log (s) = . (21)
k (a) := log 1 + rτ , kX(a) = k (a) + ωX τ , s(s + 1)
K a

Proof See appendix A.5. 

so that with this notation, k := k (1) and kX := kX(1) . We will
also make extensive usage of the usual financial notations
2.2.2. Link with the characteristic exponent. It is well-
+ −
[A] := A1{A>0} , [A] := −A1{A<0} . known that the Mellin transform can be interpreted as the
‘dilatation’ analog to the Fourier transform (or to the Bilat-
The next result provides the Mellin transforms for payoffs that eral Laplace transform) via the group isomorphism exp :
are considered in this work. (R, +) → (R+ , .). More precisely, one can write
Proposition 2.1. (Mellin transforms for path-independent
fX∗ (s, τ ) = F (fX (ex , τ )) (−s/(2iπ )), (22)
payoff functions).
(i) A digital cash or nothing call option has the payoff where F denotes the Fourier operator in the spatial (x) vari-
Pc/n (ST , K) := 1{ST >K} . If kX < 0, its Mellin trans- able (see, for instance, Bertrand et al. (2000) and references
form (13) exists in the strip {s ∈ C, Re(s) < 0} and is therein). From the point of view of Lévy processes, this
equal to: relationship can be made more precise by introducing the
(−kX )s cumulant generating function, which is defined in terms of
p∗c/n (s) = − . (17)
s the characteristic exponent as
(ii) A digital asset or nothing call option has the payoff
Pa/n (ST , K) := ST 1{ST >K} . If kX < 0, its Mellin trans- MX (p) = ψX (−ip). (23)
form (13) exists in the strip {s ∈ C, Re(s) < cz } and is
equal to: Using the Laplace inversion formula allows to write the
density function as
 cz +i∞
(z) (1 − z) (−s + z)
p∗a/n (s) = K (−1)−z  cp +i∞
(1 − s) dp
cz −i∞ fX (x, τ ) = e−px eτ MX (p) , (24)
dz cp −i∞ 2iπ
× (−kX ) s−z
, (18)
2iπ
where cp belongs to the analyticity strip of MX (p). Replac-
for any cz ∈ (0, 1). ing in the Mellin transform of the density function (11) and
256 J.-P. Aguilar and J. L. Kirkby

integrating over the x-variable yields deduced by difference, as

 cp +i∞ Pa/n (ST , K) − Peur (ST , K)
dp Pc/n (ST , K) =
fX∗ (s, τ ) = (s) ps eτ MX (p) . (25) K
. (28)
cp −i∞ 2iπ
Moneyness. If kX > 0 and if, therefore, condition (12) is
Formally, using the change of variables P := MX (p), we can, not satisfied for the asset or nothing call, it suffices to use the
therefore, write the Mellin transform for the density function martingale property to write
as an inverse Laplace transform over the temporal variable:
  EQ [ST 1{ST >K} |Ft ] = St e(r−q)τ − EQ [ST 1{ST <K} |Ft ] (29)
−1
−1 s dMX (P)
fX∗ (s, τ ) = (s)L−1
(MX (P)) ,τ . (26)
dP and to evaluate the expectation in the right hand side (which
is supported by the negative real axis only) exactly like in the
Remark 2.2 Note that the interest of (26) in practice depends OTM case. Same technique holds for the European option, by
on the complexity of MX (p) or, equivalently, of the char- writing that
acteristic exponent ψX (u). For instance, in the one-sided
tempered stable model and its particular cases (FMLS, BSM), EQ [(ST − K)1{ST >K} |Ft ] = St e(r−q)τ − K
it essentially reduces to
− EQ [(ST − K)1{ST <K} |Ft ].
(30)
MX (p) = (λ + p)α (27)
The above methodology is summarized in figure 1.
where λ is the tempering parameter and α the stability param- Jordan condition and contour closing. As already observed
eter and, therefore, the inverse function MX−1 (P) is straight- from the factorized Formula 14, the price of a path indepen-
forward to obtain (see details in the proof of Lemma 3.1). On dent option can be expressed as a complex integral of the
the contrary, in the Variance Gamma or NIG cases, MX (p) product of two Mellin transforms, namely the Mellin trans-
features some more complicated terms (logarithms etc., see form of the process density and the Mellin transform of the
appendix 5) and, therefore, it will be more advantageous to option’s payoff function. In Proposition 2.1, we have shown
directly use the closed-form expressions available for the den- that these latter transforms could necessitate the introduction
sity functions, that are obtained by integrating the Gaussian of another variable, which actually turns (14) into an inte-
density conditionally to the temporal subordination, and fea- gral over two complex variables. Moreover, in the asymmetric
ture Bessel functions admitting convenient representations in VG and NIG cases, we will see that the densities themselves
the Mellin space. also call for the introduction of another Mellin variable, which
turns (14) into an integral over 3 complex variables. In a com-
plex analysis language, we will therefore deal with integrals
2.2.3. General methodology. In order to simplify the proofs
of the type
below, let us detail the general strategy of proof that apply to
all models. In particular, it will be enough to prove the pricing 
formulas for an OTM asset nothing call to deduce the price of ω, c + iRn = {z ∈ Cn , Re(zk ) = ck , k = 1 . . . n} ,
the European and cash or nothing calls, and for all moneyness c+iRn
(31)
situations (OTM, ITM, ATM).
for n ∈ N and for a complex differential n-form ω (see for
From digital to European We first note that evaluating asset
instance Griffiths and Harris (1978) for an introduction to the
or nothing options allows to deduce immediately the value of
main concepts of multidimensional complex analysis).
the European and cash or nothing options:
In the case where (31) is a so-called Mellin–Barnes integral
• One remarks from (18) and (19) that the asset or (see definition (A13) in appendix A.2), then the possibility of
nothing and European calls admit the same Mellin– expression (31) as a sum of (multidimensional) residues has
Barnes representation, except that the integration been studied in full in Passare et al. (1994) and Zhdanov and
must be performed along a vertical line in the Tsikh (1998) and reduces to a two-step procedure:
strip < 0, 1 > in the asset or nothing case and in
(1) Step 1: Choose a cone  = 1 × · · · × n ⊂ Cn
< −1, 0 > in the European case. This means that,
whose edge, or distinguished boundary, is the integra-
when left closing the contour of integration, one
tion set of (31), that is
should consider residues associated with all sin-
gularities located left of the line Re(z) = 0 in the
asset or nothing case, but only singularities located c + iRn = ∂0  = ∂1 × · · · × ∂n (32)
left of the line Re(z) = −1 in the European case.
This is why summation formulas starts at n3 = 0 for (2) Step 2: Ensure that  is located in the admissible sub-
asset or nothing options but at n3 = 1 for European space (again, see definitions and details in appendix
options (see e.g. Formula 3.1). A.2)
• Once asset or nothing and European prices are
known, cash or nothing prices can be immediately  := {s ∈ Cn , Re( .s) < Re( .c)}. (33)
 Closed-form option pricing for exponential Lévy models: a residue approach 257

Figure 1. General methodology: it suffices to establish formulas for OTM asset or nothing options to deduce the price of all digital and
European contracts.

Table 1. Convergence of the residue series for the models under already mentioned, this is the case for subordinators; most
consideration in the paper. famous examples of such processes include the Gamma (or
Model Type of convergence exponential) process and the Inverse Gaussian process. As
they have upward jumps only, they are commonly used to
TS(−) Absolute. materialize a nonlinear (stochastic) passage of time (a positive
FMLS Absolute. jump corresponding to a period of intense trading activity).
VG (symmetric and Absolute.
asymmetric)
If (R+ ) = 0, we say that the process is spectrally nega-
NIG (symmetric and Asymptotic. Strong empirical evidence tive; such processes have been introduced in the context of
asymmetric) towards convergence. option pricing in Carr and Wu (2003a), but have also been
proved to be useful in other areas of quantitative finance
such as credit risk and default modeling (see e.g. Madan and
Schoutens (2008)). In this section, we will be particularly
If Step 1 and Step 2 are fulfilled, then one can express the interested in spectrally negative tempered stable processes,
Mellin–Barnes integral as an absolutely convergent series of their particular cases and their extensions.
residues associated to the poles in :

ω= Resω. (34) 3.1. Negative tempered stable model
c+iR2 
By Tempered Stable (TS) processes, we mean Lévy processes
Remark 2.3 As already mentioned, this theory is only valid having a Lévy measure of the form:
for Mellin–Barnes integrals; from Proposition 2.1, we see that


the Mellin transforms of the payoff functions under consider- c+ e−λ+ x c− e−λ− |x|
ation involve only gamma functions, and, therefore, it follows ts (dx) := 1{x>0} + 1{x<0} dx, (35)
x1+α+ |x|1+α−
from the pricing formula (14) that the 2-step procedure can be
applied if and only if the Mellin transform of the process den-
sity involves only gamma functions as well. As we will see in where c+ , c− ∈ R, λ+ , λ− > 0 and, typically, α+ , α− ∈ (0, 2).
the rest of the paper, this will be the case for the majority of the Tempered stable processes are related to the so-called trun-
models, namely the one-sided TS model (denoted TS(−) ), the cated Lévy flights that were introduced in physics and econo-
FMLS model and the symmetric and asymmetric VG models. physics in Mantegna and Eugene Stanley (1994) in order to
For the NIG model, however, the presence of a Bessel func- capture two different behaviors: a central part that is approx-
tion will prevent us from using the theory, and, therefore, we imately Lévy-stable and tails that are approximately expo-
will be only able to prove the asymptotic convergence of the nential (in order to ensure finiteness of the second moment).
residue series; numerical evidence strongly suggests that they The processes by Mantegna and Eugene Stanley (1994), how-
actually are convergent series. The situation is summarized in ever, were not introduced in terms of the formalism (35), and
Table 1. the definition of Tempered Stable processes in terms of their
Lévy measure and characteristic function was introduced in
Koponen (1995) in the fully symmetric case (c+ = c− , λ+ =
λ− , α+ = α− = 1), and their generalization started with one-
3. One-sided processes sided versions in Boyarchenko and Levendorskii (1999), later
extended to two-sided versions. Such two-sided processes
One-sided processes are Lévy processes whose Lévy measure were called KoBoL (Koponen–Boyarchenko–Levendorskii)
is concentrated on the negative or positive real axis only. As processes in Boyarchenko and Levendorskiǐ (2002) (see also
258 J.-P. Aguilar and J. L. Kirkby

a discussion in Rosinski (2007)), and contain many sub- we will denote the probability density for the TS(−) (σ , α, λ)
families: when λ+ = λ− = 0, one simply speaks of a sta- model by fts− (x, t).
ble process, and, when α+ = α− := Y , of a CGMY (Carr–
Geman–Madan–Yor) process (Carr et al. 2002), which itself Lemma 3.1. Let c1 < 1; then for any x > 0 the following
contains the Variance Gamma process (case Y = 0). Mellin–Barnes representation holds:
 s1
λx  c1 +i∞
Remark 3.1 We note that truncated Lévy flights lead to very λα ω − t e (1 − s1 ) x ds1
fts− (x, t) = e .
inconvenient characteristic exponents for efficient numeri- αx c1 −i∞ (1 − α ) (−ω− t) α
s1 1
2iπ
cal valuation, which can limit their use in practice, see (39)
Boyarchenko and Levendorskii (2022c). In general, stable
Lévy processes are much better than the truncated Lévy Proof The moment-generating function of the negative tem-
flights, as far as the analytical properties are concerned. pered stable process satisfies
Moreover, efficient calculations are possible for stable Lévy α
−λα )
processes but not for truncated Lévy flights, see Boyarchenko EQ [epXt ] = etψts− (−ip) = e−ω− t((λ+p) ,
and Levendorskii (2020a, 2022b). and exists on the half plane {Re(p) > −λ}. From the Laplace
In the context of stable distributions with α+ = α− := α ∈ inversion formula we can deduce that the probability density
(1, 2), it is common to use the Feller parametrization, which is  cp +i∞
α dp
goes as follows: λα ω − t
fts− (x, t) = e e−px e−ω− t(λ+p) ,
cp −i∞ 2iπ
⎧ πα
α
⎪
⎨ σ := −(c+ + c− ) (−α) cos 2 for any cp > −λ. Using the change of variables p → (λ + p)α
(36) and introducing a Mellin–Barnes representation for the arising
⎩ β := c+ − c− .
⎪ 1
e−p α x term (see Table A2) for x > 0, we obtain
c+ + c−
λx  c1 +i∞
With this parametrization, σ is called the scale parameter λα ω − t e
fts− (x, t) = e (s1 )x−s1
and α the stability (or kurtosis) parameter. Values of interest α c1 −i∞
for the parameter α lie in the interval (1, 2) for most finan-  cp +i∞
1−s1 dp ds1
cial applications, and in that case one speaks of a Paretian × epω− t p α −1 , (40)
cp −i∞ 2iπ 2iπ
distribution (see details in Mittnik and Rachev (2000); ini-
tial calibrations for the cotton market in Mandelbrot (1963) for any c1 > 0. The p-integral is equal to (−ω− t)− α / (1 −
1−s1

estimated α
1.7); when α = 2, the distribution degenerates 1−s1
) (see Bateman (1954) or any table of inverse Laplace
α
into the usual normal distribution. The parameter β ∈ [−1, 1] transforms); changing the variable s1 → 1 − s1 completes the
measures the asymmetry of the distribution: when β < 0, the proof. 
distribution is skewed to the left and, when β > 0, it is skewed
to the right. Remark 3.2 Lemma 3.1 shows that the density in the
TS(−) (σ , α, λ) model is a function of the ratio x/(−ω− t) α ,
1
In this section, we will be particularly interested in the case
where the process Xt in (2) satisfies β = −1, or equivalently, which is actually a consequence of the self-similarity prop-
c+ = 0, meaning that the distribution admits only a left fat erty (Embrechts and Maejima 2000) of one-sided stable and
tail - this configuration is known under the term of maxi- tempered stable processes (any re-scaling of time is equiva-
mal negative asymmetry hypothesis. We will call this model lent to an appropriate re-scaling of space). Note also that the
the negative tempered stable model and use the abbreviation representation (39) is valid for x > 0 but can also be extended
TS(−) (σ , α, λ); the Lévy measure reads in this case to x < 0 with the result
λx  c1 +i∞ ( sα1 ) (1 − s1 )
σα e−λ|x| λα ω − t e
ts− (dx) = − 1{x<0} . (37) fts− (x, t) = e
πα
(−α) cos 2 |x|1+α αx c1 −i∞ ( α s1 ) (1 − α−1
α−1
α 1
s)
 s1
|x| ds1
It follows from the Lévy-Khintchine representation (4) that × 1 . (41)
the Lévy symbol for the TS(−) (σ , α, λ) model reads (for an (−ω− t) α 2iπ
appropriate choice of drift):
The result has been established in Mainardi et al. (2001) in
 ∞ the case λ = 0 (and in the context of fractional analysis); it is
ψts− (u) = (eiux − 1)ts− (dx) = −ω− ((λ + iu)α − λα ) , straightforward to extend it to the case of a non-null tempering
−∞
parameter λ.
where we have introduced the notation ω− := σ α / cos πα
2
. Let us introduce a supplementary Mellin–Barnes represen-
The corresponding martingale adjustment follows tation for the exponential term in (39),
immediately:
 c2 +i∞
λx ds2
ωts− = −ψts− (−i) = ω− ((λ + 1)α − λα ) . (38) e = (−λ)−s2 (s2 )x−s2 ,
c2 −i∞ 2iπ
which was derived in Boyarchenko and Levendorskii (1999). for c2 > 0, and given a contingent claim delivering a payoff
∗
In all of the following, we will assume that α ∈ (1, 2), and P(ST , K1 , . . . , Kn ) at t = T, let us denote by Pts− the following
 Closed-form option pricing for exponential Lévy models: a residue approach 259

Mellin transform (evaluated at s = s1 − s2 ): n1

kts− (−ω− τ )−
n1 −n2
α . (46)
 ∞
∗
Pts− (s1 , s2 ) := P(St e(r−q+ωts− )τ +x , K1 , . . . , Kn )xs1 −s2 −1 dx, Proof Following the methodology of section 2.2.3, we only
0 detail the proof of (i) in the OTM case. In the case of an asset
(42) or nothing payoff, then, using the notations of Proposition 2.1,
which is assumed to exist into some subset {(s1 , s2 ) ∈ the Mellin transform (42) can be rewritten as: Pts− ∗
(s1 , s2 ) =
C2 , Re(s1 , s2 ) ∈ A} where A ⊂ R2 . ∗
pa/n (s1 − s2 ) and, therefore, using (18) and Proposition 3.1,
Remark 3.3 We note the slight difference between the gen- we can write the option price as:
eral form of the Mellin transform for the payoff function (13), α   
Feλ ω− τ c1 +i∞ c2 +i∞ c3 +i∞
and its application (42) to the TS(−) (σ , α, λ) case, that is, the Ca/n = (−λ)−s2 (−1)−s3
substitution s = s1 − s2 . This is, of course, a direct conse- α c1 −i∞ c2 −i∞ c3 −i∞
quence of the introduction of a second Mellin variable for the (1 − s1 ) (s2 ) (s3 ) (1 − s3 ) (−s1 + s2 + s3 )
exponential term arising in the density (39). Similar substi- (1 − sα1 ) (−s1 + s2 + 1)
tutions will also happen in the Variance Gamma and Normal
s1 ds1 ds2 ds3
inverse Gaussian cases. × (−kts− )s1 −s2 −s3 (−ω− τ )− α , (47)
(2iπ )3
As an immediate consequence of Lemma 3.1 and of the
pricing formula (10) we have: where the integrand exists and is analytical in the
polyhedron G = {s := (s1 , s2 , s3 ) ∈ C3 , R(s1 ) < 1, Re(s2 ) >
Proposition 3.1. In the TS(−) (σ , α, λ) model, the value 0, 0 < Re(s3 ) < 1, Re(−s1 + s2 + s3 ) > 0}. We denote c =
at time t of a contingent claim C delivering a payoff (c1 , c2 , c3 ) ∈ G. Let us now apply the 2-step procedure that
P(ST , K1 , . . . , Kn ) at its maturity T and satisfying the support has been detailed in section 2.2.3:
condition (12) is:
• Step 1: We remark that, by construction, the poles
α  c1 +i∞  c2 +i∞
e−(r−λ ω− )τ
(1 − s1 ) (s2 ) occurring in (s1 , s2 , s3 ) = (n1 − n2 − n3 , −n2 , −n3 ),
C= (−λ)−s2 n1 , n2 , n3 ∈ N (induced by the singularities at neg-
α c1 −i∞ c2 −i∞ (1 − sα1 )
ative integers of the functions (s2 ), (s3 ) and
∗
s1 ds1 ds2 (−s1 + s2 + s3 )) define a cone  in C3 whose
× Pts− (s1 , s2 )(−ω− τ )− α , (43)
(2iπ )2 edge can be arbitrarily chosen to be any element
in G - hence we can choose ∂0  = c.
for any (c1 , c2 ) ∈ ((−∞, 1) × (0, ∞)) ∩ A. • Step 2:
Let us now consider the application of Proposition 3.1 to The characteristic vector (see (A14) in appendix
the case of various path-independent options. Note that Propo- 2) associated with (47) is
sition 3.1 is formulated in the case where the payoff function ⎡ ⎤ ⎡1 ⎤
is supported by R+ , i.e. if (12) is satisfied; it is straightforward −1 − 1 + α1 + 1 α
−1
to extend the result to R− thanks to Remark 3.2 and by trans- = ⎣ 1 + 1 − 1 ⎦ = ⎣ 1 ⎦.
ferring the definition of the Mellin transform (42) to R− . We 1−1+1 1
note that 1/ (−n) = 0 for n ∈ N0 in the formulas below.
As, by hypothesis, α1 − 1 < 0, it follows that  ⊂
Formula 3.1 Digital and European calls The value at time t  , where  := {s ∈ C3 , Re( .s) < Re( .c)}.
of a call option in the TS(−) (σ , α, λ) model is: Using the singular behavior of the gamma func-
(i) Asset or nothing: tion (A3), the residues associated to the singulari-
ties in  are:
α ∞

Feλ ω− τ λn2 (1 − n1 + n3 )n2 α
Feλ ω− τ (−1)n1 +n2 +n3
Ca/n =   (−λ)n2 (−1)n3
α n1 ,n2 ,n3
n !n ! 1 − n1 −nα2 −n3
=0 1 2 α n1 !n2 !n3 !
n1 −n2 −n3 (1 − n1 + n3 + n3 ) (1 + n3 )
n1
kts− (−ω− τ )− α . (44)
(1 − n1 −nα2 −n3 ) (1 − n1 + n3 )
n1 −n2 −n3
(ii) European: (−kts− )n1 (−ω− τ )− α .
∞
Feλ ω− τ λn2 (1 − n1 + n3 )n2
α
Using the definition of the Pochhammer symbol
Ceur =  
α n !n ! 1 − n1 −nα2 −n3
n ,n =0 1 2
and the particular values of the gamma func-
1 2
n3 =1 tion at positive integers (see (A21) and (A18) in
n1 −n2 −n3 appendix 3), simplifying and summing all residues
n1
kts− (−ω− τ )− α . (45) in virtue of the residue theorem (A16), we obtain
the series (44).
(iii) Cash or nothing:

−(r−λα ω− )τ ∞

e λ (1 − n1 )n2
n2
Cc/n =   Proposition 3.1 also allows to obtain closed pricing for-
α n !n
n ,n =0 1 2
! 1 − n1 −n
α
2
mulas for non-linear payoffs, such as power calls; in the
1 2
260 J.-P. Aguilar and J. L. Kirkby

following formula, we consider only the European case, but (ii) European:
digital power calls are straightforward to derive too.
∞
F 1 n1 −n3
Formula 3.2 Power call The value at time t of a European Ceur =  n1 −n3
 k−
n1
(−ω− τ )− α .
power call in the TS(−) (σ , α, λ) model is: α n =0 n1 ! 1 − α
1
n3 =1
∞
Feλ ω− τ λn2 an3 (1 − n1 + n3 )n2 (53)
α

Cpow =   (iii) Cash or nothing:

α n !n ! 1 − n1 −nα2 −n3
n ,n =0 1 2
1 2
n3 =1 ∞
e−rτ 1 n1
(a) n1
(kts−
n1 −n2 −n3
) (−ω− τ )− α . (48) Cc/n =   k−
n1
(−ω− τ )− α .
α n =0 n1 ! 1 − n1
α
1

Proof The proof is similar to proving (ii) in Formula 3.1, by (54)

remarking that the Mellin transform (42) is, in the case of a Remark 3.4 ATMF approximations Assuming that options
∗
power payoff, Pts− (s1 , s2 ) = p∗pow (s1 − s2 ), by Proposition 2.1 are at the money forward, i.e. that St = F, then keeping only
(representation (20)).  the leading term (n1 = 0, n3 = 1) in (53), we can approximate
the European call price by
3.2. Finite Moment Log Stable model  1

St (−ω− τ ) α
The Finite Moment Log Stable model FMLS(σ , α) corre- Ceur =   + O (−ω− τ ) . (55)
α 1 + α1
sponds to the negative tempered stable model with tempering
parameter λ going to 0, i.e. FMLS(σ , α) = TS(−) (σ , α, 0). It
was originally introduced in Carr and Wu (2003a) for mod- Letting α → 2 in (55),
√ recalling that in that case ω− = −σ
2

eling the maturity pattern of the implied volatility smirk (i.e. and that (3/2) = π /2, we obtain
the difference of implied volatility between out-of-the-money
1 √
calls and puts) of index options, the stability parameter α Ceur
√ St σ τ ,
being responsible for the slope of the option’s smirk. The π
model has been subsequently studied many times, for instance
from the point of view of portfolio risk in Robinson (2015), which is the well known Brenner–Subrahmanyam approxima-
√
and with applications to profit-and-loss attribution in Aguilar tion (Brenner and Subrahmanyam 1994) for the BSM(σ 2)
and Korbel (2019). model.
The Lévy measure of the FMLS(σ , α) process reads Let us mention that, recently, formulas for more exotic
instruments have been obtained in the FMLS(σ , α) model, for
σα 1
ts− (dx) = − πα 1{x<0} , (49) example in the case of log options (which are basically options
(−α) cos 2 |x|1+α on the rate of returns of the underlying asset, see details e.g.
in Wilmott (2006)) and log contracts, which were introduced
while the Lévy symbol is known to admit the representation in Neuberger (1994) for volatility hedging purposes (a vari-
 απ  ance swap being actually a delta-hedged log contract). Pricing
ψfmls (u) = σ α |u|α 1 + i tan sgnu . (50) formulas were obtained in Aguilar (2020a, 2021b) and are as
2
follows:
Note that (50) shows that the FMLS(σ , α) model actually
Formula 3.4 Log options and log contract In the FMLS
recovers the Black–Scholes–Merton model√ when α → 2,
(σ , α) model, the value of the log options are:
more precisely FMLS(σ , 2) = BSM(σ 2). Letting λ → 0
in (38) yields the FMLS convexity adjustment (i) Log call:

σα ∞
ωfmls = ω− = , (51) e−rτ 1 1−n
cos πα Clog =   k−
n
(−ω− τ ) α . (56)
2 α n=0 n! 1 + 1−n
α
and, still under the λ → 0 limit, it is clear that only the
terms for n2 = 0 survive in the series representations of For- (ii) Log put:
mula 3.1, we obtain the value of the Digital and European call
in the FMLS(σ , α) model : Plog
 ∞

Formula 3.3 Digital and European calls The value at time t −rτ 1 1 1−n

of a call option in the FMLS(σ , α) model is: =e −k− +   k−

n
(−ω− τ ) α .
α n=0 n! 1 + 1−n
α
(i) Asset or nothing: (57)
∞
F 1 n1 −n3 (iii) Log contract:
Ca/n =  n1 −n3
 k−
n1
(−ω− τ )− α .
α n ,n =0 n1 ! 1 − α
Clog − Plog = e−rτ k− .
1 3
(52) (58)
 Closed-form option pricing for exponential Lévy models: a residue approach 261

Note that, letting α → 2 in (58), we get Figure 2 illustrates the application of Algorithm 1 for the
TS(−) model (left) and an analogous algorithm for the FMLS


−rτ St model (right), where we plot the user-supplied error tolerance
Clog − Plog = e log + (r − q − σ )τ ,
2
against the maximum error incurred over a range of strikes,
K
from 70% to 130% of ATM, where we truncate the partial sum
which is the formula originally derived by Neuberger in Neu- as soon as the supplied tolerance is achieved. The maximum
√ error (across strikes) for each tolerance level is computed rel-
berger (1994) for the price of a log contract in the BSM(σ 2)
model. ative to the prices found by truncating the partial sums with a
tolerance of e − 15. For both models, we can see clearly the
ability to control the error in a predictable manner using this
3.3. Numerical examples approach.

Here we present several numerical illustrations to demonstrate

the accuracy of our pricing formulas in the Tempered Stable
and Stable (FMLS) cases. Source code is made publicly and 4. Variance gamma process
freely available at: https://github.com/jkirkby3/PROJ_Option
_Pricing_Matlab. Recall that the Variance Gamma (VG) process of Madan
In Table 2, we provide comparisons between our closed et al. (1998) is a CGMY process (Carr et al. 2002) whose
pricing formulas truncated at n1 = n2 = n3 = N and a numer- Lévy measure satisfies
ical evaluation of the Lewis formula (Lewis 2001), for the θx

θ2 + 2
example of an asset or nothing call. We can observe that the e σ2 − σ2 ν
|x|
vg (dx) = e σ dx, (60)
convergence to the actual price is fast, in particular when ν|x|
options are close to the money.
where σ > 0 is the scale parameter, ν ∈ R the kurtosis param-
eter and θ ∈ R the asymmetry (or sknewness) parameter (see
3.4. Error control details in Madan et al. (1998); the symmetric case θ = 0 was
considered earlier in Madan and Seneta (1990)). When the
A key practical advantage of the closed-form series pricing process Xt in (2) is such a VG process, we will speak of the
formula is the direct control it provides over the accuracy VG(σ , ν, θ ) model. It is also possible to define the VG process
of price approximations. Recall the pricing formula for the as a time changed Brownian motion, more precisely as
European option under the TS(−) model in (45),
Xt := θ Gt + σ WGt , (61)
∞

Ceur = ζ · z(n1 , n2 , n3 ), (59) where Gt is a Gamma process with unit mean rate and vari-
n1 ,n2 =0
ance equal to ν. It follows from definition (61) that the VG
n3 =1 process is actually distributed according to a so-called Normal
variance-mean mixture (see Barndorff-Nielsen et al. (1982)),
αω τ
Feλ − λn2 (1−n1 +n3 )n2 where the mixing distribution is the Gamma distribution;
where ζ := α
, and z(n1 , n2 , n3 ) := n1 !n2 ! (1− n1 −n2 −n3 )
n1 −n2 −n3
α the subordination t → Gt allows for a nonlinear passage of
n1
kts− (−ω− τ )− α . Algorithm 1 summarizes the steps used time materializing the alternation of intense business periods
to price a European option to within a specified tolerance. (corresponding to jumps in the Gamma process) and calmer
The user inputs a maximum partial summation term N, and periods.
a desired tolerance, tol. The algorithm terminates after a full The VG(σ , ν, θ ) model is well known in financial model-
step for which the price changes by no more than tol. A similar ing and has been shown to perform better than Gaussian or
algorithm is implemented for the other models. jump-diffusion models many times, with successful calibra-
tions to equity index options (Lam et al. 2002) or currency
options (Madan and Daal 2005) for instance. Subordination
Algorithm 1 European Option Valuation For TS(−) to bivariate or multivariate Brownian motions (in order to
Input: N, tol price products written on more than one underlying assets)
Ceur ← 0, L ← 0 have also been considered (see Semeraro (2008) and refer-
for n1 = 0, . . . , N do ences therein), and calibrated to basket options (Linders and
for n2 = 0, . . . , N do Stassen 2015). Besides option pricing, the VG process has
for n3 = 1, . . . , N do also been employed to forecast default risk (see discussion
Ceur ← Ceur + z(n1 , n2 , n3 ) and references in Luciano (2009)), with calibrations on CDS
end for indices.
end for The probability density function in the VG(σ , ν, θ ) model
if n1 > 1 and |Ceur − L| < tol/ζ then break is obtained by integrating the normal density conditionally to
L ← Ceur the gamma time change, resulting in
end for   2νt − 14
Ceur ← ζ Ceur 2eθ σ 2
x
x2
if Put Option then Ceur ← Ceur − S0 e−qT + Ke−rT fvg (x, t) = t √
ν ν 2π σ ( νt ) 2σ 2
ν
+ θ2
262 J.-P. Aguilar and J. L. Kirkby

Table 2. Asset or nothing call options in the TS(−) (σ , α, λ) and FMLS(σ , α) models.

Tempered stable (σ = 20%, α = 1.7, λ = 0.1)

Formula 3.1
N =5 N = 10 N = 15 N = 20 Lewis (2001)

Deep OTM (St = 2500) 614.2682 743.9498 468.3819 468.2851 468.2842

OTM (St = 3500) 1794.2681 1792.7628 1792.7617 1792.7617 1792.7617
ATM (St = 4000) 2576.8876 2576.9478 2576.9478 2576.9478 2576.9478
ITM (St = 4500) 3347.1593 3347.3768 3347.3768 3347.3768 3347.3768
Deep ITM (St = 5500) 4763.7750 4764.5108 4764.5109 4764.5109 4764.5109

Stable (FMLS) (σ = 20%, α = 1.7)

Formula 3.3 Lewis (2001)

N =5 N = 10 N = 15 N = 20

Deep OTM (St = 2500) 596.1091 489.8641 486.4925 486.4458 486.4454

OTM (St = 3500) 1819.5538 1818.6977 1818.6973 1818.6973 1818.6973
ATM (St = 4000) 2598.7843 2598.8725 2598.8725 2598.8725 2598.8725
ITM (St = 4500) 3363.0481 3363.2797 3363.2797 3363.2797 3363.2797
Deep ITM (St = 5500) 4768.3709 4769.0757 4769.0758 4769.0758 4769.0758

Note: Parameters: K = 4000, r = 1%, q = 0%, τ = 2.

Figure 2. Error Control: user-supplied tolerance vs Maximum Error for strike range 70% to 130% of ATM; St = 4000, r = 2%, q = 1%.
Left: TS(−) example: σ = 20%, α = 1.7, λ = 0.1. Right: FMLS example: σ = 20%, α = 1.7.

  
1 2σ 2 Lemma 4.1. Let (c1 , c2 ) ∈ (0, ∞) × (|τν |, ∞); then the fol-
K νt − 12 + θ |x| ,
2 (62) lowing Mellin–Barnes representation holds true:
σ2 ν
 c1 +i∞  c2 +i∞
where Kμ (.) denotes the modified Bessel function of the sec- fvg (x, τ ) =
1
√
ond kind (see definition and properties in Abramowitz and 22+2τν π σν1+2τν ( τν )
c1 −i∞ c2 −i∞
Stegun (1972) and appendix A.4). The Lévy symbol is    
s1 − τν s1 + τν
  (s2 )
1 σ 2ν 2 2 2
ψvg (u) = − log 1 − iθ νu + u , (63) s1 +τν
ds1 ds2
ν 2 −
× (−θσ )−s2 qσ ,ν,θ2 |x|−s1 +τν x−s2
(2iπ )2
from which it follows that the martingale adjustment equals
  Proof With the above notations, we-write the density (62)
1 σ 2ν as:
ωvg := −ψvg (−i) = log 1 − θ ν − . (64)
ν 2
 τν
τ
ν θ
2eθσ x |x|
Notations. We let τν := − σν := σ
1
, , θσ := , fvg (x, τ ) = τ √ √
ν 2 2 σ2
ν ν 2π σ ( τν ) 2σ 2 qσ ,ν,θ
qσ ,ν,θ := 14 ( σ12 + θσ2 ). We assume τν ∈
/ Q.  √ 
ν
Kτν 2 qσ ,ν,θ |x| , (65)
 Closed-form option pricing for exponential Lévy models: a residue approach 263

and we introduce the two Mellin–Barnes representation (see √ ( −n1 +n2 +n2 3 +2τν +1 )(−n1 + n3 + 1)n2
Table A2): + π
2−n1 +n2 +n3 (1 + −n1 +n2 2 +n3 )
 √  −n1 +n2 +n3 +2τν +1

Kτν 2 qσ ,ν,θ |x| X n1 Y − 2 Z n2 , (67)
 c1 +i∞
s1 − τν s1 + τν
= 2s1 −2 ( ) ( ) where the Pochhammer symbol (rising factorial) is defined by
c1 −i∞ 2 2
(z)n := (z+n)
(z)
.
 √ −s ds1
2 qσ ,ν,θ |x| 1 , c1 ∈ (|τν |, ∞), Formula 4.1 Digital option: asset or nothing The value at
2iπ
time t of an asset or nothing call option in the asymmetric
and VG model VG(σ , ν, θ ) is:
 c2 +i∞ (i) (OTM price) If kvg < 0,
θσ x −s2 −s2 ds2
e = (−θσ ) (s2 )x , c2 ∈ (0, ∞).
c2 −i∞ 2iπ − F
Ca/n = √
22+2τν π σν1+2τν ( τν )
Inserting into (65) and simplifying completes the proof. 
∞
(−1)n1 +n2 (vg)
Given a contingent claim delivering a payoff P(ST , K1 , A (−kvg , qσ ,ν,θ , −θσ ).
,n =0
n1 !n2 ! n1 ,n2 ,n3
. . . , Kn ) at t = T, let us introduce the following function: n1 ,n2 3

(68)
∗
Pvg (s1 , s2 )
 (ii) (ITM price) If kvg > 0,
∞
(r−q+ωvg )τ +x −s1 −s2 +τν
:= P(St e , K1 , . . . , Kn )x dx, (66) F
+
0 Ca/n = St e−qτ − √ 1+2τν τ
22+2τν π σν ( ν )
and assume that it exists into some subset {(s1 , s2 ) ∈ ∞
(−1)n1 +n2 +n3 (vg)
C2 , Re(s1 , s2 ) ∈ A} where A ⊂ R2 . Then, as an immediate An1 ,n2 ,n3 (kvg , qσ ,ν,θ , θσ ).
consequence of Lemma 4.1 and of the pricing formula (10) n1 ,n ,n =0
n1 !n2 !
2 3
we have: (69)
Proposition 4.1. In the VG(σ , ν, θ ) model, the value at time t
of a contingent claim C delivering a payoff P(ST , K1 , . . . , Kn ) (iii) (ATM price) If kvg = 0,
at its maturity T and satisfying the support condition (12) is:
− + F
Ca/n = Ca/n =

c2 +i∞c1 +i∞  22+2τν σν1+2τν ( τν )
e−rτ
C= √ ∞
22+τν π σν1+2τν ( τν ) c1 −i∞ c2 −i∞ (n2 + n3 )! ( n2 +n3 +2τ ν +1
) n2 − n2 +n3 +2τ ν +1

   
2
θ q 2
.
s1 − τν s1 + τν ∗ n ,n =0
2n2 +n3 n2 !n3 ! (1 + n2 +n
2
3
) σ σ ,ν,θ
(s2 )Pvg (s1 , s2 ) 2 3

2 2 (70)
s +τ
− 12 ν ds1 ds2
× (−θσ )−s2 qσ ,ν,θ , Proof Following the methodology detailed in section 2.2.3,
(2iπ )2
we only prove (i); when kvg < 0 then −kvg > 0 and in that
for any (c1 , c2 ) ∈ ((0, ∞) × (|τν |, ∞)) ∩ A. case the integral (66) can be rewritten as
∗
Note that, like in the negative tempered stable case, Propo- Pvg (s1 , s2 ) = p∗a/n (−s1 − s2 + τν + 1),
sition 4.1 can be extended to payoffs supported by R− by
appropriately modifying the definition of the Mellin trans- and therefore, using Proposition 2.1 (representation (18)) and
form (66). Proposition 4.1, we can write the OTM option price as:

− F
Ca/n = √ 1+2τν τ
4.1. Asymmetric model 22+τν π σν ( ν )
 c1 +i∞  c2 +i∞  c3 +i∞
Let us assume that θ = 0, and let us now provide some appli-
cations of Proposition 4.1, by computing digital and European (−1)−s3 (−kvg )−s1 −s2 −s3 +τν +1
c1 −i∞ c2 −i∞ c3 −i∞
prices (and related options). For n1 , n2 , n3 ∈ N, X , Y ∈ R+ s1 +τν
−
and Z ∈ R, define the quantities (−θσ )−s2 qσ ,ν,θ2
( s1 −τ ν
) ( s1 +τ ν
)
A(vg)
n1 ,n2 ,n3 (X , Y , Z)
2 2

(s2 ) (s3 ) (1 − s3 ) (s1 + s2 + s3 − τν − 1)
(−n1 − τν ) (−2n1 − n2 − n3 − 2τν − 1) ×
:= 2 (s1 + s2 − τν )
(−2n1 − n2 − 2τν ) ds1 ds2 ds3
, (71)
X 2n1 +n2 +n3 +2τν +1 Y n1 (2iπ )3
264 J.-P. Aguilar and J. L. Kirkby

where the integrand exists and is analytic in the polyhedron It is easy to extend Formula 4.1 to the case of European
G = {s := (s1 , s2 , s3 ) ∈ C3 , Re(s1 ) > τν , Re(s2 ) > 0, 0 < options:
Re(s3 ) < 1, Re(s1 + s2 + s3 − τν − 1) > 0}. We denote c =
Formula 4.2 European option The value at time t of a Euro-
(c1 , c2 , c3 ) ∈ G. Like in the proof of Formula 3.1, let us
pean call option in the asymmetric VG model VG(σ , ν, θ )
now apply the 2-step procedure that has been detailed in
is:
section 2.2.3:
(i) (OTM price) If kvg < 0,
• Step 1: We remark that, outside G, the integrand
of (71) admits two series of poles: ∞
F (−1)n1 +n2
1 Poles in (s1 , s2 , s3 ) = (−2n1 − τν , −n2 , −n3 ), −
=
Ceur √
n1 , n2 , n3 ∈ N, corresponding to singularities of 22+2τν π σν1+2τν ( τν ) n1 ,n2 =0 n1 !n2 !
the functions ( s1 +τ 2
ν
), (s2 ) and (s3 ) and n3 =1
whose residues are: A(vg)
n1 ,n2 ,n3 (−kvg , qσ ,ν,θ , −θσ ). (74)
F
√ (ii) (ITM price) If kvg > 0,
22+τν πσν1+2τν ( τν )
(−1)n1 +n2 F
×2× +
Ceur = St e−qτ − F − √ 1+2τν τ
n1 !n2 !n3 ! 22+2τν π σν ( ν )
(−n1 − τν ) (1 + n3 ) ∞
(−1)n1 +n2 +n3 (vg)
(−2n1 − n2 − n3 − 2τν − 1) An1 ,n2 ,n3 (kvg , qσ ,ν,θ , θσ ).
× n1 !n2 !
(−2n1 − n2 − 2τν ) n ,n =0
1 2
n3 =1
2n1 +n2 +n3 +2τν +1
× (−kvg ) (−θσ )n2 qnσ1,ν,θ . (72) (75)

2 Poles in (s1 , s2 , s3 ) = (−n1 + n2 + n3 + τν + (iii) (ATM price) If kvg = 0,

1, −n2 , −n3 ), n1 , n2 , n3 ∈ N, corresponding to
singularities of the functions (s1 + s2 + s3 − F
− +
τν − 1), (s2 ) and (s3 ) and whose residues Ceur = Ceur =
22+2τν σν1+2τν ( τν )
are:
∞
F (n2 + n3 )! ( n2 +n3 +2τ ν +1
) n2 − n2 +n3 +2τ ν +1

n2 +n3 θσ qσ ,ν,θ
2 2
√ +n
.
22+τν π σν1+2τν ( τν ) n =0
2n2 3 n2 !n3 ! (1 + 2 )
2
n3 =1
n1 +n2
(−1) (76)
×
n1 !n2 !n3 !
( −n1 +n22 +n3 +1 ) Proof The proof is straightforward thanks to the contour
−n1 +n2 +n3 +2τν +1 deformation argument described in section 2.2.3. 
( ) (1 + n3 )
× 2
(−n1 + n3 + 1) Last, we easily deduce from 4.1 and 4.2 the value of the
−n +n2 +n3 +2τν +1
− 1 cash or nothing options:
× (−kvg ) n1
(−θσ )n2 qσ ,ν,θ 2
. (73)
Formula 4.3 Digital option: cash or nothing The value at
These poles define, by construction, a cone  in time t of a cash or nothing call option in the asymmetric VG
C3 whose edge can be arbitrarily chosen to be any model VG(σ , ν, θ ) is:
element in G - hence we can chose ∂0  = c.
• Step 2: The characteristic vector (see (A14) in (i) (OTM price) If kvg < 0,
appendix 2) associated with (71) is, by definition,
⎡ ⎤ ⎡ ⎤ − e−rτ
1+1+1−1 2 Cc/n = √
22+2τν π σν1+2τν ( τν )
= ⎣ 1 + 1 − 1 ⎦ = ⎣1⎦ ,
∞
1−1+1 1 (−1)n1 +n2 (vg)
A (−kvg , qσ ,ν,θ , −θσ ).
from which it follows that  ⊂  where  := n ,n =0
n1 !n2 ! n1 ,n2 ,0
1 2

{s ∈ C3 , Re( .s) < Re( .c)}. Simplifying (72) (77)

and (73), using the Legendre duplication formula
and the definition of the Pochhammer symbol (ii) (ITM price) If kvg > 0,
(see (A23) and (A21) in appendix 3) to write
√ + e−rτ
( −n1 +n22 +n3 +1 ) 2π (−n1 + n3 + 1)n2 Cc/n = e−rτ − √
= −n +n +n , 22+2τν πσν1+2τν ( τν )
(−n1 + n3 + 1) 2 1 2 3 (1 + −n1 +n2 +n3 )
2 ∞
(−1)n1 +n2 (vg)
and summing all residues in virtue of the residue A (kvg , qσ ,ν,θ , θσ ). (78)
n1 !n2 ! n1 ,n2 ,0
theorem (A16) yields (68).  n ,n =0
1 2
 Closed-form option pricing for exponential Lévy models: a residue approach 265

(iii) (ATM price) If kvg = 0, X 2n1 +n3 +2τν +1 Y n1


e−rτ √ ( −n1 +n32+2τν +1 ) −
−n1 +n3 +2τν +1
−
Cc/n = +
Cc/n = + π −n1 +n3 X
n1
Y 2 , (81)
22+2τν σν1+2τν ( τν ) 2−n1 +n3 (1 + 2
)

∞
( n2 +2τ2 ν +1 ) n2 − n2 +2τ2 ν +1 it follows from Formula 4.2 that the call price is given by the
θ q . (79)
n =0
2n2 (1 + n22 ) σ σ ,ν,θ following formula:
2

Formula 4.4 European call The value at time t of a Euro-

Remark 4.1 Power options Extending the European option pean call option in the symmetric VG model symVG(σ , ν)
price (Formula 4.2) to the case of a power call is straight- is:
(a)
forward. For instance, in the OTM situation (kvg < 0), we
write (i) (OTM price) If kvg < 0,
∗
Pvg (s1 , s2 ) = p∗pow (−s1 − s2 + τν + 1). − F
Ceur = √
and, using (20) and proceeding like in the proof of For- 22+2τν π σν1+2τν ( τν )
mula 4.2, we obtain ∞
(−1)n1 (svg)
A (−kvg , qσ ,ν,0 ). (82)
− F n1 ! n1 ,n3
Cpow = √ n =0
1
n3 =1
22+2τν π σν1+2τν ( τν )
∞
(−1)n1 +n2 an3 (vg) (ii) (ITM price) If kvg > 0,
(a)
An1 ,n2 ,n3 (−kvg , qσ ,ν,θ , −θσ ). (80)
n ,n =0
n1 !n 2 ! F
+
= St e−qτ − F −
1 2
n3 =1 Ceur √ 1+2τν τ
22+2τν π σν ( ν )
+ ∞
The ITM price Cpow can be obtained from the martingale (−1)n1 +n3 (svg)
property and the risk-neutral expectation of STa , and the ATM An1 ,n3 (kvg , qσ ,ν,0 ). (83)
(a) n1 !
price by letting kvg → 0 in (80). n =0
1
n3 =1

(iii) (ATM price) If kvg = 0,

4.2. Symmetric model
The symmetric Variance Gamma model corresponds to the − + F
Ceur = Ceur =
case θ = 0, that is 22+2τν σν1+2τν ( τν )
∞
symVG(σ , ν) := VG(σ , ν, 0). 1 ( n3 +2τ2 ν +1 ) − n3 +2τ2 ν +1
q . (84)
n =1
2n3 (1 + n23 ) σ ,ν,0
3
In this case, the Lévy measure (60) is symmetric around
the origin, and upward and downward jumps occur with Remark 4.2 ATMF approximations When St = F, the lead-
the same probability; it was first introduced in Madan ing term (n1 = 0, n3 = 1) in the series (82) is given by (recall
and Seneta (1990) and has been extensively studied in that qσ ,ν,0 = 1/(4σν2 )):
Aguilar (2020b). Recall, in particular, that the symmetric
Variance Gamma model degenerates into the Black–Scholes − St
Ceur = √ 1+2τν τ (1 + τν )(4σν ) ν .
2 1+τ
model in its low variance regime, that is symVG(σ , ν) −→ 22+2τν π σν ( ν )
BSM(σ ) when ν → 0. From the point of view of the martin-
gale adjustment (64), we have: Using notations (4.0.1.1) and simplifying, we get
 
1 σ 2ν σ2 − St ( 12 + τν ) √
ωvg −→ log 1 − −→ − , Ceur =√ τ σ ν.
θ→0 ν 2 ν→0 2 2π ( ν )

thus recovering the traditional Gaussian adjustment of the From the behavior of the gamma function at infinity (see e.g.
BSM(σ ) model. Abramowitz and Stegun 1972), we know that
Pricing formulas for the symVG(σ , ν) model are straight- 
forward to obtain by letting θ → 0 (and therefore θσ → 0) in ( 12 + τν ) τ
−→ ,
the formulas obtained above for the asymmetric model. For ( τν ) ν→0 ν
instance, in the European case, introducing the reduced series
terms and therefore
− St √
Ceur −→ √ σ τ ,
A(svg)
n1 ,n3 (X , Y ) ν→0 2π

(−n1 − τν ) (−2n1 − n3 − 2τν − 1) which is the Brenner–Subrahmanyam approximation for the

:= 2 ATMF call in the Black–Scholes–Merton model BSM(σ ).
(−2n1 − 2τν )
266 J.-P. Aguilar and J. L. Kirkby

4.3. Numerical examples sum necessary to achieve this tolerance (as described in
section 3.4). For example, for the asymmetric VG case, a user-
4.3.1. Comparison to Fourier methods. Given the rela-
supplied basis point tolerance (1e-04) is reached in 0.0056 s.
tive popularity of the VG model, we begin by comparing
Achieving this level of control over pricing accuracy is much
the analytical formulas for the symmetric case, Formula 4.4,
more difficult for competing methods, which is another reason
and the asymmetric case, Formula 4.2, with several well-
to favor the closed-form expressions in practical applications,
known Fourier pricing methods. Among the pricing algo-
such as model calibration. Figure 3 illustrates the direct error
rithms that exist for Lévy models, Fourier methods are con-
control provided by the analytical series formula for the VG
sidered among the most efficient and accurate available. The
model (and NIG, introduced later), where we can see that the
methods of Lewis (2001) and Carr and Madan (1999) are spe-
maximum error across the range of strikes is close to (and well
cific to European option pricing, while the CONV method
bounded by) the prescribed error tolerance.
Lord et al. (2008) and Hilbert Transform Feng and Linet-
sky. July (2008) are applicable to Bermudan and Barrier
options as well. The PROJ Kirkby (2015) and COS Fang 4.3.3. Time to maturity robustness. A key competitive
and Oosterlee (2008) methods are applicable in a wide vari- advantage of the Mellin transform is that it is robust for
ety of contexts, including exotic option pricing, risk measure all time to maturity ranges and even improves in accuracy
computation, and density estimation problems. for small τ , which is where many other methods struggle.
We note that the VG model is one of the more chal- Figure 4 demonstrates this property by comparing Mellin
lenging models for many Fourier methods, due to its slow against the widely used COS method for the asymmetric
(polynomial) characteristic function decay. Table 3 provides VG model, with maturity ranges between τ = 1/500 and
a comparison for each method, and we can see that the ana- τ = 4 (years). We also include the method of Lewis (2001)
lytical formula is highly accurate (using just N = 15 terms to provide an additional benchmark. All three methods are
for the symmetric case, and N = 24 in the asymmetric case). sufficiently accurate for large maturities (e-07 for Mellin vs
For illustration purposes, the number of terms has been set e-10 for COS and Lewis), but for small maturities the COS
quite large (set to achieve a user-supplied accuracy of at and Lewis methods rapidly deteriorate, while Mellin nears
least e − 10), and the resulting cpu time reflects the algorithm machine precision (smoothly). This maturity robustness is
complexity in the aysmmetric VG case, O(N 3 ). For practical very desirable for market surface calibration, for which a
accuracy levels, say e − 04, many fewer terms are required. model is calibrated to a wide range of maturities, including
The approach outlined in Algorithm 1 allows one to control tenors with τ = 1/252 as they near expiry.
the desired accuracy that is suitable for the application. In Table 4, we compare the European prices obtained via
We can also see that while the Mellin transform retains truncations of the pricing Formula 4.2 with a numerical eval-
high accuracy in the asymmetric case (albeit at a higher, uation of the Carr–Madan formula. We may observe, again,
yet still competitive, computational cost), the most widely the accelerated convergence of the series for short horizons;
used Fourier method, COS of Fang and Oosterlee (2008), this is because  
suffers a substantial loss in accuracy when the asymme-    
kvg  ∼ log St  ,
try is introduced, for the same numerical parameters (N = τ →0  K
29 Fourier basis elements). The inefficiencies of both the
which is smaller than 1 as soon as
Carr–Madan approach as well as the COS method, espe-
cially in risk management applications, are well documented. 1
Both methods are inefficient distortions of the more straight- K < St < eK, (85)
e
forward realization of the inverse Fourier transform, see
Boyarchenko and Levendorskii (2011, 2015), Boyarchenko and therefore, in that case, the presence of positive powers of
and Levendorskiǐ (2015, 2019), Levendorskii (2016), and kvg accelerate the overall convergence of the series. Note that
De Innocentis and Levendorskii (2014, 2016). In particu- the bounds in (85) allow for maintained accuracy in deeply
lar, the method of Boyarchenko and Levendorskiǐ (2019), OTM and ITM situations (and the series would converge any-
and Boyarchenko and Levendorskii (2020b) achieves accu- way, even outside of this interval): for instance for an option
racy of the order 10−15 in microseconds for European striked at K = 4000, this corresponds to 1471 < St < 10873.
options, and 10−5 in dozens of microsecond for barrier On the contrary, the Carr–Madan formula has more difficulties
options, respectively; the results in Boyarchenko and Lev- to converge for short-term options and deep OTM situations,
endorskii (2022a, 2022b) are better still. For a comprehen- in particular, due to severe oscillations of the integrand when
sive analysis of various realizations, including more efficient τ → 0.
ones, see Boyarchenko and Levendorskiǐ (2015, 2019) and
Levendorskii (2014).
5. Normal inverse Gaussian process

4.3.2. Error control. Due to the series nature of the closed- The Normal inverse Gaussian (NIG) process is a pure jump
form formulas, and the generally smooth convergence of Lévy process, whose Lévy measure is given by
prices (i.e. a smoothly decreasing error as a function of the
partial series truncation parameter), it is especially easy to αδ βx K1 (α|x|)
nig (dx) := e dx, (86)
specify a desired accuracy tolerance and compute the partial π |x|
 Closed-form option pricing for exponential Lévy models: a residue approach 267

Table 3. VG comparison to Fourier Methods for European Call Options.Parameters: K = St = 4000, r = 2%, q = 1%, τ = 1. Symmetric
VG Params: σ = 20%, ν = 0.85, θ = 0. Asymmetric VG Params: σ = 20%, ν = 0.85, θ = −0.1. Time is measured in seconds.

Symmetric VG Asymmetric VG
Method Price |Error| Time Price |Error| Time

Analytical (Mellin) 304.94314588 9.65e − 11 0.0005 322.84387077 7.60e − 10 0.0224

PROJ (2015), Kirkby (2015) 304.94314588 6.08e − 10 0.0014 322.84387077 1.33e − 09 0.0007
COS (2008), Fang and Oosterlee (2008) 304.94314588 1.68e − 09 0.0015 322.84389066 1.99e − 05 0.0008
Hilbert Transform (2008), Feng and Linetsky. July (2008) 304.94314805 2.17e − 06 0.0025 322.84387271 1.94e − 06 0.0028
Lewis (2001) 304.94316497 1.91e − 05 0.0013 322.84387199 1.22e − 06 0.0014
CONV (2008), Lord et al. (2008) 304.94312141 2.45e − 05 0.0095 322.84384741 2.34e − 05 0.0096
Carr and Madan (1999) 304.94318041 3.45e − 05 0.0119 322.84389278 2.20e − 05 0.0093

Notes: Parameters: K = St = 4000, r = 2%, q = 1%, τ = 1. Symmetric VG Params: σ = 20%, ν = 0.85, θ = 0. Asymmetric VG Params:
σ = 20%, ν = 0.85, θ = −0.1. Time is measured in seconds.

Figure 3. Error Control: user supplied tolerance vs Maximum Error for strike range 70% to 130% of St , where St = 4000, r = 2%, q = 1%.
Left: Asymmetric NIG example: α = 15, β = −1, δ = 0.5. Right: Asymmetric VG example: σ = 20%, ν = 0.85, θ = −0.1.

where K1 (.) is the modified Bessel function of index μ = 1 mean mixture, that is
(see appendix A.4). α > 0 is a tail or steepness parameter con-
trolling the kurtosis of the NIG distribution; the large α regime Xt := βδ 2 It + δWIt (87)
gives birth to light tails, while small α corresponds to heav-
ier tails. β ∈ (−α, α − 1) is the skewness parameter: β < 0 where the time subordination isachieved this time via a
(resp. β > 0) implies that the distribution is skewed to the left Gamma subordinator It of shape δ α 2 − β 2 and mean rate 1.
(resp. the right), and β = 0 that the distribution is symmetric. When the process Xt in (2) is a NIG process, we speak
δ > 0 is the scale parameter and plays an analogous role to the of the NIG(α, β, δ) model; note that a location parame-
σ parameter in the TS and the VG processes. It is interesting ter can also be included, but it has no impact on option
to note that, because of the large argument behavior (A29), prices, therefore, we can assume it to be equal to 0. The
3
the NIG measure (86) decays in 1/|x| 2 at infinity, making for NIG(α, β, δ) model has been introduced in financial model-
thinner tails than with the VG measure (60). Let us mention ing by Barndorff-Nielsen (1995, 1997), approximately two
that, like in the VG case, the NIG process is also a variance decades after its original consideration for physical purposes
(more precisely, the modeling of dunes and beach sands

Table 4. Short maturity behavior of European options in the VG model: Formula 4.2 vs Carr–Madan formula, where the integral is truncated
at u = umax and evaluated by a trapezoidal algorithm.

Formula 4.2 Carr and Madan (1999)

Time to maturity N =5 N = 10 N = 15 N = 20 umax = 102 umax = 103 umax = 104

1 year (τ = 1) − 1493.7221 48.5386 78.7757 78.8532 78.8531 78.8533 78.8533

6 months (τ = 0.5) − 705.7659 32.0777 36.8706 36.8752 36.8740 36.8752 36.8752
1 month (τ = 1/12) 13.8933 5.4528 5.4328 5.4328 5.5711 5.4320 5.4328
1 week (τ = 1/52) 2.6464 1.2203 1.2174 1.2174 1.3431 1.2159 1.2174
1 day (τ = 1/360) 0.3663 0.1748 0.1744 0.1744 0.2750 0.1765 0.1753

Note: Parameters: S = 3000, K = 4000, r = 1%, q = 0%, σ = 20%, ν = 0.85, θ = 0.1.

268 J.-P. Aguilar and J. L. Kirkby

Figure 4. Mellin vs COS and Lewis accuracy for varying time to maturity, τ ∈ {1/500, 1/252, . . . , 2, 3, 4}. Asymmetric VG example:
K = St = 4000, r = 2%, q = 1%, σ = 20%, ν = 0.85, θ = −0.1. We use N = 24 terms for Mellin, and N = 213 basis elements for COS.

(Barndorff-Nielsen 1977)). The model has proved to provide Lemma 5.1. Let (c1 , c2 ) ∈ R2+ ; then the following Mellin–
a very good fitting to financial data many times: let us men- Barnes representation holds true:
tion, among others, initial tests for daily returns on Danish
and German markets in Rydberg (1997) and subsequently on
α γ δt
the FTSE All-share index (also known as ‘Actuaries index’) fnig (x, t) = e
in Venter and de Jongh (2002). More recently, the impact of 2π
 c1 +i∞  c2 +i∞ s 
high-frequency trading has also been taken into account, and × (−1)−s2 β −s2
1
(s2 )
calibrations have been performed on intraday returns e.g. in c1 −i∞ c2 −i∞ 2
Figueroa-López et al. (2012) for different sampling frequen-   s21
2δt ds1 ds2
cies. Like in the VG case, multivariate extensions have also K s
1− 21 (αδt) |x|−s1 x−s2 . (90)
been considered (see Luciano and Semeraro (2010) and refer- α (2iπ )2
ences therein), and applications to credit risk have also been
provided (Luciano 2009). Proof Using Table A2 and the Mellin inversion for-
The probability density function in the NIG(α, β, δ) model mula (A4), we can write:
can be obtained, similarly to the VG case, by integrating
the normal density conditionally to the Inverse Gaussian   
subordinator and reads K1 α (δt)2 + x2
   
(δt)2 + x2
αδt δt√α2 −β 2 +βx K1 α (δt) + x
2 2
fnig (x, t) := e  . (88)  c1 +i∞     s1
π (δt)2 + x2 1 s1 2δt 2 −s1 ds1
= K1− s21 (αδτ ) |x| ,
2δτ c1 −i∞ 2 α 2iπ
The Lévy symbol is (91)
  
ψnig (u) = −δ α 2 − (β + iu)2 − α2 − β 2 . (89) for any c1 > 0, and

It follows that the martingale adjustment in the NIG(α, β, δ)  c2 +i∞

βx ds2
model is
  e = (−1)−s2 β s2 (s2 )x−s2 , (92)
c2 −i∞ 2iπ
ωnig = δ α 2 − (β + 1)2 − γ ,

where we have introduced γ := α2 − β 2. for any c2 > 0. Inserting these two representations into the
density (88) yields (90). 
 Closed-form option pricing for exponential Lévy models: a residue approach 269

Given a contingent claim delivering a payoff P(ST , K1 , (ii) Log put:

. . . , Kn ) at t = T, we introduce the function
 ∞  
 ∞ α (γ δ−r)τ (−1)1+n2 1 + n2
∗ Plog = e
Pnig (s1 , s2 ) := P(St e(r−q+ωnig )τ +x , K1 , . . . , Kn )x−s1 −s2 dx, 2π n2 ! 2
0 n2 =0
(93)  
(nig) 2δτ
and assume that it exists into some subset {(s1 , s2 ) ∈ A 1+n2 1−n2 knig , β, , αδτ
1,n2 , 2 , 2 α
C2 , Re(s1 , s2 ) ∈ A} where A ⊂ R2 . Then, as an immediate
consequence of Lemma 5.1 and of the pricing formula (10) ∞
1  
n2 (nig)
we have: + 1+ A0,n ,1+ n2 ,− n2
n!
n2 =0 2
2 2 2 2

Proposition 5.1. In the NIG(α, β, δ) model, the value at  

2δτ
time t of a contingent claim C delivering a payoff knig , β, , αδτ
P(ST , K1 , . . . , Kn ) at its maturity T and satisfying the support α
(nig) ⎤
condition (12) is: (−1)n1 +n2 A2n1 +n2 +2,n2 ,−n
 1 ,1+n1
∞
knig , β, 2δτ , αδτ ⎥
 c1 +i∞  c2 +i∞ α ⎥
α (γ δ−r)τ +2 ⎥.
C= e (−1)−s2 β −s2 n1 =n2 =0
n1 !n 2 !(2n 1 + n2 + 2) ⎦
2π c1 −i∞ c2 −i∞ (2n1 + n2 + 1)
s    s21
1 ∗ 2δτ ds1 ds2 (97)
(s2 )P (s1 , s2 )K1− s21 (αδτ ) ,
2 α (2iπ )2
(94) Proof To prove (i), we remark that the integral (93) can be
rewritten as
for any (c1 , c2 ) ∈ R2+ ∩ A.
∗
Again, this proposition can be easily extended to negatively Pnig (s1 , s2 ) = p∗log (−s1 − s2 + 1),
supported payoff functions, by modifying the definition of the
Mellin transform (93). and therefore, using Proposition 2.1 (representation (21)) and
Proposition 5.1, we can write the Log call price as:
 c1 +i∞  c2 +i∞
5.1. Asymmetric model α (γ δ−r)τ
Clog = e (−1)−s2 β −s2
Let us assume β = 0, and let us start our applications of 2π c1 −i∞ c2 −i∞
 
Proposition 5.1 with log options. To simplify the notations, s21 (s2 )
we introduce, for m1 , m2 , m3 , m4 ∈ N and X , Y , Z, W ∈ R, the × (−knig )−s1 −s2 +2
(s1 + s2 − 1)(s1 + s2 − 2)
quantities   s1
2δτ 2 ds1 ds2
A(nig) K1− s21 (αδτ ) ,
m1 ,m2 ,m3 ,m4 (X , Y , Z, W ) := X Y Z Km4 (W )
m1 m2 m3
(95) α (2iπ )2

Formula 5.1 Log options Assume that |knig /(δτ )| < 1. where inside the polyhedron {(s1 , s2 ) ∈ C2 , Re(s1 ) > 0, Re
Then the value at time t of a Log option in the asymmetric (s2 ) > 0, Re(s1 + s2 ) > 2} the integrand exists. Outside this
NIG(α, β, δ) model is given by: polyhedron, it admits the following series of poles:
(i) Log call:
• Poles in s1 + s2 = 1, s2 = −n2 , n2 ∈ N, induced
by the (s1 + s2 − 1) and (s2 ) terms, and whose
Clog
 ∞ residues are:
 
α (γ δ−r)τ 1 1 + n2
= e α (γ δ−r)τ β n2 ( 1+n 2
)
2π n!
n2 =0 2
2 e × 2
knig
  2π n2 !
(nig) 2δτ   1+n2 2
A 1+n2 1−n2 knig , β, , αδτ 2δτ
1,n2 , 2 , 2 α K 1−n2 (αδτ ) . (98)
α
1  n2  (nig)
2
∞
+ 1+ A0,n ,1+ n2 ,− n2
n!
n2 =0 2
2 2 2 2
• Poles in s1 + s2 = 2, s2 = −n2 , n2 ∈ N, induced
  by the (s1 + s2 − 2) and (s2 ) terms, and whose
2δτ
knig , β, , αδτ residues are:
α
(nig) ⎤ α (γ δ−r)τ β n2 (1 + n22 )
(−1)n1 +n2 A2n1 +n2 +2,n2 ,−n
 1 ,1+n1 e ×
∞
knig , β, 2δτ , αδτ ⎥ 2π n2 !
α ⎥
+2 ⎥ . (96)   n2
n1 =n2 =0
n1 !n 2 !(2n 1 + n2 + 2) ⎦ 2δτ 1+ 2
(2n1 + n2 + 1) K −n2 (αδτ ) . (99)
2 α
270 J.-P. Aguilar and J. L. Kirkby

• Poles in s1 = −2n1 , s2 = −n2 , n1 , n2 ∈ N, induced We can now use the multiplication theorem with
by the ( s21 ) and (s2 ) terms, and whose residues
⎧
are: ⎪ 1
⎪
⎪ μ=
α (γ δ−r)τ ⎪
⎪ 2
⎨
e z = αδτ
2π
⎪
⎪ 
(−1)n1 +n2 β n2 ⎪
⎪
⎩λ = 1 − β = γ
⎪ 2
× k 2n1 +n2 +2
n1 !n2 !(2n1 + n2 + 1)(2n1 + n2 + 2) nig α 2 α
 
2δτ −n1
K1+n1 (αδτ ) . (100) which results in
α
 
Summing (98), (99) and (100) and using notation (95) α π 2δτ γ
Clog − Plog = e(γ δ−r)τ knig K 1 (γ δτ ).
yields Equation (96). We finish by checking the convergence π α α 2
of the obtained series. Let us start by fixing n2 and considering
the case n1 → ∞; without loss of generality, we can choose Using the particular value of the Bessel function of index
n2 = 0, and therefore, we have to study the behavior of 1/2 (A30) and simplifying, we are left with

(−1)n1 Clog − Plog = e−rτ knig .

Rn1 := k 2(n1 +1) (102)
2n1 !(n1 + 1)(2n1 + 1) nig
 
2δτ −n1 Let us provide another application of Proposition 5.1, with
Kn1 +1 (αδτ ) . the determination of the price of the power call option.
α
(a)
Formula 5.2 Power option Assume that |knig /(δτ )| < 1.
Using the Stirling approximation (A20) and the large index Then the value at time t of a European power call in the
argument behavior of the Bessel function (A27), we have asymmetric NIG(α, β, δ) model is:
 n1 +1
2
δτ knig Kαe(γ δ−r)τ
|Rn1 | ∼ , Cpow = √
n1 →∞ 2α(2n1 + 1)(n1 + 1) (δτ )2 π
∞
an3 (−n1 + n3 + 1)n2
and therefore the series converges if and only if |knig /(δτ )| < ×
1. Using the same approximations, it is easy to show that the n1 2
n !n ! (1 + −n1 +n2 2 +n3 )
,n =0 1 2
series converges for all parameter values when n2 → ∞. n3 =1

The proof of (ii) is similar: by an appropriate modification (nig)

A −n +n +n +1 n −n −n +1
n1 ,n2 , 1 22 3 , 1 22 3
of the Mellin transform (93) and of Proposition 5.1, we can  
write the Log put price as: (a) δτ
knig , β, , αδτ . (103)
  2α
c1 +i∞ c2 +i∞
α (γ δ−r)τ
Plog = e β −s2
2π c1 −i∞ c2 −i∞ Proof In the case of a power payoff, the integral (93) can be
 s1  written as
2 (s2 )
× k −s1 −s2 +2
(s1 + s2 − 1)(s1 + s2 − 2) nig ∗
  s1 Pnig (s1 , s2 ) = p∗pow (−s1 − s2 + 1),
2δτ 2 ds1 ds2
K1− 21 (αδτ )
s ,
α (2iπ )2 and therefore, using Proposition 2.1 (representation (20)) and
Proposition 5.1, we can write
which can be evaluated with the same residue technique than
for (i).  Cpow
 c1 +i∞  c2 +i∞  c3 +i∞
Remark 5.1 The log contract The determination of the price Kα (γ δ−r)τ
= e (−1)−s2 −s3 β −s2 a−s3
of the log contract provides a nice application of the mul- 2π c1 −i∞ c2 −i∞ c3 −i∞
tiplication theorem for the Bessel function (see (A31) in ( s21 ) (s2 ) (s3 ) (1 − s3 ) (s1 + s2 + s3 − 1)
appendix 3). Indeed, the price of the log contract can be ×
(s1 + s2 )
obtained by subtracting (96) and (97); only even terms n2 =
  s1
2p of the first series survive and, using the particular values (a) −s1 −s2 −s3 +1 2δτ 2 ds1 ds2 ds3
(A19) for ( 12 + p), we obtain × (−knig ) K1− s21 (αδτ ) ,
α (2iπ )3
 ∞
π2δτ ( αδτ )p ( βα2 )p
2
α which converges in the subset {(s1 , s2 , s3 ) ∈ C3 , Re(s1 ) >
Clog − Plog = e(γ δ−r)τ knig 2
π α p=0 (2p)! 0, Re(s2 ) > 0, −1 < Res(s3 ) < 0, Re(s1 + s2 + s3 ) > 1}.
Outside this subset, if we consider the singularities induced
K 12 −p (αδτ ). (101) by (s2 ) at s2 = −n2 , n2 ∈ N, by (s3 ) at s3 = −n3 , n3 ∈ N
 Closed-form option pricing for exponential Lévy models: a residue approach 271

Figure 5. Mellin vs COS and Lewis accuracy for varying time to maturity, τ ∈ {1/500, 1/252, . . . , 2, 3, 4}. Asymmetric NIG example:
Parameters: K = St = 4000, r = 2%, q = 1%, α = 15, β = −1, δ = 0.5. We use N = 14 terms for Mellin, and N = 213 basis elements for
COS.

 
and by (s1 + s2 + s3 − 1) at s1 + s2 + s2 − 1 = −n1 , n1 ∈ (nig) δτ
N, then, the associated residues are: A −n +n +n +1 n −n −n +1 knig , β, , αδτ , (105)
n1 ,n2 , 1 22 3 , 1 22 3 2α
Kαe(γ δ−r)τ
(−1)n2 +n3 β n2 an3 that was derived in Aguilar (2021a).
2π
(−1)n1 (−1)n2 (−1)n3
n1 ! n2 ! n3 ! 5.2. Symmetric model
(1 + n3 ) ( −n1 +n−2+n3 +1
) (a) n1 The symmetric NIG model corresponds to the case β = 0, that
2
(−knig )
(−n1 + n3 + 1) is, symNIG(α, δ) :=NIG(α, 0, δ). In that case, the Lévy mea-
  −n1 +n22+n3 +1 sure (86) is symmetric around 0, and the model degenerates
δτ
× K1− −n1 +n2 +n3 +1 (αδτ ) (104) into the Black–Scholes model when α → ∞, more precisely
2 2α
for n1 , n2 ≥ 0 and n3 ≥ 1. Using the Legendre duplication symNIG(α, δ) −→ BS(σ ),
α→∞
formula (A23) and the definition of the Pochhammer sym-
bol (A21), we write:
for σ 2 := δ/α. From the point of view of martingale adjust-
√
( −n1 +n22 +n3 +1 ) π (−n1 + n3 + 1)n2 ments, we have
= −n +n +n .
(−n1 + n3 + 1) 2 1 2 3 (1 + −n1 +n2 2 +n3 )
 σ2 δ
Inserting into (104), simplifying and summing all residues for ωnig −→ δ( α 2 − 1 − α) −→ − , σ 2 := .
β→0 α→∞ 2 α
n1 , n2 , n3 yields the series (103). Last, convergence studies
goes exactly as in the proof of Formula 5.1.  Pricing formulas for the symNIG(α, δ) model can be eas-
Remark 5.2 European option Letting a → 1 in Formula 5.2, ily obtained by letting β → 0 in the formulas obtained for
we recover the formula for the European call price the asymmetric model; more precisely, introducing the series
terms
Kαe(γ δ−r)τ
Ceur = √ A(snig)
m1 ,m3 ,m4 (X , Z, W ) := X Z Km4 (W )
m1 m3
(106)
π
∞
then one immediately obtains:
(−n1 + n3 + 1)n2
× (a)
n !n ! (1 + −n1 +n2 2 +n3 ) Formula 5.3 Power option Assume that |knig /(δτ )| < 1.
,n =0 1 2
n1 2
n3 =1 Then the value at time t of a European power call in the
272 J.-P. Aguilar and J. L. Kirkby

Table 5. Comparison to Fourier Methods for European Call Options.

Symmetric NIG Asymmetric NIG

Method Price |Error| Time Price |Error| Time

Analytical (Mellin) 302.49271926 9.09e − 13 0.0007 302.67462883 1.71e − 13 0.0046

PROJ (2015), Kirkby (2015) 302.49271926 9.09e − 13 0.0007 302.67462883 2.73e − 12 0.0006
COS (2008), Fang and Oosterlee (2008) 302.49271926 9.09e − 13 0.0010 302.67462883 2.73e − 12 0.0007
Hilbert (2008), Feng and Linetsky. July (2008) 302.49271926 5.68e − 13 0.0022 302.67462883 5.68e − 13 0.0018
Lewis (2001) 302.49271926 4.09e − 12 0.0014 302.67462883 3.64e − 12 0.0011
CONV (2008), Lord et al. (2008) 302.49270956 9.70e − 06 0.0082 302.67461904 9.78e − 06 0.0072
Carr and Madan (1999) 302.49274365 2.44e − 05 0.0081 302.67465309 2.43e − 05 0.0066

Notes: Parameters: K = St = 4000, r = 2%, q = 1%, τ = 1. Symmetric NIG Params: α = 15, β = 0, δ = 0.5. Asymmetric NIG Params:
α = 15, β = −1, δ = 0.5. Time is measured in seconds.

symmetric NIG model symNIG(α, δ) is: Using the asymptotic behavior of the Bessel function for large
arguments (A29), we recover the fact that
∞
Kαe(αδ−r)τ an3
Cpow = √ St √
π n ! (1 + −n12+n3 )
n =0 1 Ceur −→ √ σ τ ,
1
n3 =1 α→∞ 2π
 
(snig) (a) δτ
A −n1 +n3 +1 n1 −n3 +1 knig , , αδτ . (107)
n1 , 2 , 2 2α where σ 2 := δ/α, which is the Brenner–Subrahmanyam
approximation for the ATMF call in the Black–Scholes–
Remark 5.3 European option and ATMF approximations Merton model BSM(σ ).
Taking a → 1 in (107) yields the European call price

∞ 5.3. Numerical examples

Kαe(αδ−r)τ 1
Ceur = √ −n1 +n3 This section provides numerical experiments for European
π n
n =0 1
! (1 + 2
)
1
n3 =1
options under the symmetric and asymmetric NIG model. We
  begin with a comparison of our analytical formulas (Equa-
(snig) δτ tions (105) and (108)) to existing Fourier methods in Table 5
A −n1 +n3 +1 n1 −n3 +1 knig , , αδτ . (108)
n1 , 2 , 2 2α for τ = 1, where we can see that the closed-form expres-
sion is both fast and accurate. To further demonstrate the
which is valid if knig /(δτ ) < 1. When St = F, then the leading robustness of our method as a function of time to maturity,
term (n1 = 0, n3 = 1) of (108) reads figure 5 compares the accuracy with COS and Lewis for
τ ∈ {1/500, 1/252, . . . , 2, 3, 4}. All methods are accurate for
Kαe(αδ−r)τ large τ , but as we saw for the VG model, the Fourier numer-
bCeur = √ ical methods quickly deteriorate for maturities of a month or
π ( 32 )
  less. This can be problematic when calibrating a full option
(snig) δτ St δτ eαδτ surface, as short maturities are already the hardest to fit in
A0,1,0 knig , , αδτ = K0 (αδτ ).
2α π practice, even without errors caused by the pricing function.

Figure 6. Exponential convergence of NIG for a call option: τ = 1, St = K = 4000, r = 0.02, q = 0.01. (Left) Symmetric NIG model with
β = 0, δ = 1.5. (Right) Asymmetric NIG with β = −1.2, δ = 1.5. Here N = n1 = n2 is the number of terms in the truncated series.
 Closed-form option pricing for exponential Lévy models: a residue approach 273

Note that the analytical formula converges within machine Aguilar, J.P., Some pricing tools for the variance gamma model. Int.
precision at slightly less than τ = 1 year to expiry. J. Theor. Appl. Finance, 2020b, 23(4), 2050025.
Figure 6 demonstrates the smooth convergence of the par- Aguilar, J.P., Explicit option valuation in the exponential NIG model.
Quant. Finance, 2021a, 21(8), 1281–1299.
tial sums as a function of the number of terms, N, for several Aguilar, J.P., The value of power-related options under spectrally
levels of α, which as discussed at the beginning of section 5 negative Lévy processes. Rev. Deriv. Res., 2021b, 24, 173–
controls tail heaviness (smaller α leads to heavier tails). Not 196.
surprisingly, we achieve higher accuracy with a lighter-tailed Aguilar, J.P. and Korbel, J., Simple formulas for pricing and hedging
model, for any fixed N. However, as shown in figure 3 (right European options in the finite moment log-stable model. Risks,
2019, 7(2), 36.
panel), it is easy to control the accuracy of the approximation Andersen, L. and Lipton, A., Asymptotics for exponential Lévy pro-
by terminating the partial sums as soon as they change by less cesses and their volatility smile: Survey and new results. Int. J.
than a prescribed tolerance, which is largely due to the smooth Theor. Appl. Finance, 2013, 16(1), 1350001.
convergence of the series. Asmussen, S. and Klüppelberg, C., Large deviations results for
subexponential tails, with applications to insurance risk. Stoch.
Process. Their Appl., 1996, 64(1), 103–125.
Avramidis, A.N. and L’Ecuyer, P., Efficient Monte Carlo and quasi–
Monte Carlo option pricing under the variance gamma model.
6. Concluding remarks Manag. Sci., 2006, 52(12), 1930–1944.
Barndorff-Nielsen, O., Exponentially decreasing distributions for
This work provides a unified pricing framework for pricing the logarithm of particle size. Proc. R Soc. Lond., 1977, 353,
401–419.
vanilla and exotic path-independent options under a large Barndorff-Nielsen, O., Normal inverse Gaussian distributions and
class of exponential Lévy models, based on the machin- the modeling of stock returns. Research report no 300, Department
ery of Mellin transforms. We derive closed-form pricing of Theoretical Statistics, Aarhus University, 1995.
formulas for the Variance Gamma (VG), Normal-Inverse Barndorff-Nielsen, O., Normal inverse Gaussian distributions
Gaussian (NIG), Finite Moment Log Stable (FMLS), and and stochastic volatility models. Scand. J. Stat., 1997, 24(1),
1–13.
negative Tempered Stable (TS) processes. Comparisons with Barndorff-Nielsen, O., Kent, J. and Sørensen, M., Normal variance-
six well-known Fourier transform techniques show that our mean mixtures and z-distributions. Int. Stat. Rev., 1982, 50(2),
closed-form series expressions are both accurate and efficient. 145–159.
Moreover, these expressions are robust at all time to maturity Barndorff-Nielsen, O.E. and Levendorskii, S.Z., Feller processes of
levels, in contrast to existing Fourier methods which suffer normal inverse Gaussian type. Quant. Finance, 2001, 1(3), 318–
331.
accuracy loss for short maturities. The method also facilitates Barndorff-Nielsen, O.E., Mikosch, T. and Resnick, S.I., Lévy Pro-
reliable error control at any prescribed tolerance level, given cesses: Theory and Applications, 2001 (Birkhäuser: Boston,
the series nature of the pricing formulas and their smooth con- MA).
vergence. The robustness at all maturity and strike levels and Bateman, H., Tables of Integral Transforms (vol. I and II). 1954
ease of error control make this method ideal for model cal- (McGraw-Hill Book Company: New York).
Bates, D.S., Jumps and stochastic volatility: Exchange rate processes
ibration in practical applications. Future extensions of this implicit in deutsche mark options. Rev. Financ. Stud., 1996, 9(1),
work to path-dependent option pricing are especially interest- 69–107.
ing, as well as possible applications to more general processes Benth, F.E., Groth, M. and Kettler, P.C., A quasi-Monte Carlo
such as the generalization of the NIG process constructed in algorithm for the normal inverse Gaussian distribution and val-
Barndorff-Nielsen and Levendorskii (2001). uation of financial derivatives. Int. J. Theor. Appl. Finance, 2006,
9(6), 843–867.
Bertoin, J., Lévy Processes, 1996 (Cambridge University Press:
Cambridge, New York, Melbourne).
Bertrand, J., Bertrand, P. and Ovarlez, J., The Mellin transform.
Disclosure statement In The Transforms and Applications Handbook: Second Edition,
edited by A.D. Poularikas, 2000 (CRC Press LLC: Boca Raton).
No potential conflict of interest was reported by the author(s). Boyarchenko, M., Fast simulation of Lévy processes. Available at
SSRN 2138661, 2012.
Boyarchenko, S. and Levendorskii, S., Generalizations of the Black–
Scholes equation for truncated Lévy processes. Working Paper,
University of Pennsylvania, 1999.
ORCID
Boyarchenko, S. and Levendorskii, S., Option pricing for truncated
Lévy processes. Int. J. Theor. Appl. Finance, 2000, 3(3), 549–552.
Jean-Philippe Aguilar http://orcid.org/0000-0002- Boyarchenko, S.I. and Levendorskiǐ, S.Z., Non-Gaussian Merton-
2646-2401 Black–Scholes Theory, volume 9 of Adv. Ser. Stat. Sci. Appl.
Probab. ed. 2002 (World Scientific Publishing Co: River Edge,
NJ).
Boyarchenko, S. and Levendorskii, S., Barrier options and touch-
References and-out options under regular Lévy processes of exponential type.
Ann. Appl. Probab., 2002a, 12(4), 1261–1298.
Abramowitz, M. and Stegun, I., Handbook of Mathematical Func- Boyarchenko, S.I. and Levendorskii, S.Z., Perpetual American
tions, 1972 (Dover Publications: Mineola, NY). options under Lévy processes. SIAM J. Control Optim., 2002b,
Aguilar, J.P., On expansions for the Black–Scholes prices and hedge 40(6), 1663–1696.
parameters. J. Math. Anal. Appl., 2019, 478(2), 973–989. Boyarchenko, S. and Levendorskii, S., New efficient versions of
Aguilar, J.P., Pricing path-independent payoffs with exotic features Fourier transform method in applications to option pricing. Avail-
in the fractional diffusion model. Fractal Fract., 2020a, 4(2), 16. able at SSRN 1846633, 2011.
274 J.-P. Aguilar and J. L. Kirkby

Boyarchenko, S. and Levendorskii, S., Efficient Laplace inversion, Fang, F. and Oosterlee, C.W., A novel pricing method for European
Wiener–Hopf factorization and pricing lookbacks. Int. J. Theor. options based on Fourier cosine series expansions. SIAM J. Sci.
Appl. Finance, 2013, 16(3), 1350011. Comput., 2008, 31, 826–848.
Boyarchenko, M. and Levendorskii, S., Pricing barrier options and Feng, L. and Linetsky, V., Pricing discretely monitored barrier
credit default swaps (CDS) in spectrally one-sided Lévy models: options and defaultable bonds in Levy process models: A fast
The parabolic Laplace inversion method. Quant. Finance, 2015, Hilbert transform approach. Math. Finance, 2008, 18(3), 337–384.
15(3), 421–441. Figueroa-López, J.E. and Forde, M., The small-maturity smile for
Boyarchenko, S.I. and Levendorskiǐ, S.Z., Efficient variations of exponential Lévy models. SIAM J. Financ. Math., 2012, 3(1), 33–
Fourier transform in applications to option pricing. J. Comput. 65.
Finance, 2015, 18(2), 57–90. Figueroa-López, J.E., Lancette, S.R, Lee, K. and Mi, Y., Estimation
Boyarchenko, S. and Levendorskiǐ, S., Sinh-acceleration: Efficient of NIG and VG models for high frequency financial data. In Hand-
evaluation of probability distributions, option pricing, and Monte book of Modeling High-Frequency Data in Finance, edited by F.
Carlo simulations. Int. J. Theor. Appl. Finance, 2019, 22(3), Viens, M.C. Mariani and I. Florescu, 2012 (John Wiley & Son:
1950011. Hoboken, NJ).
Boyarchenko, S. and Levendorskii, S., Conformal accelerations Flajolet, P., Gourdon, X. and Dumas, P., Mellin transforms and
method and efficient evaluation of stable distributions. Acta Appl. asymptotics: Harmonic sums. Theor. Comput. Sci., 1995, 144,
Math., 2020a, 169(1), 711–765. 3–58.
Boyarchenko, S. and Levendorskii, S., Static and semistatic hedg- Frontczak, R. and Schobel, R., On modified Mellin transforms,
ing as contrarian or conformist bets. Math. Finance, 2020b, 30(3), Gauss–Laguerre quadrature, and the valuation of American call
921–960. options. J. Comput. Appl. Math., 2010, 234(5), 1559–1571.
Boyarchenko, S. and Levendorskii, S., Efficient evaluation of expec- Fusai, G., Marazzina, D. and Marena, M., Pricing discretely moni-
tations of functions of a Lévy process and its extremum. arXiv tored Asian options by maturity randomization. SIAM J. Financ.
preprint arXiv:2207.02793, 2022a. Math., 2011, 2(1), 383–403.
Boyarchenko, S. and Levendorskii, S., Efficient evaluation of expec- Griffiths, P. and Harris, J., Principles of Algebraic Geometry, 1978
tations of functions of a stable Lévy process and its extremum. (Wiley & Sons: Hoboken, NJ).
arXiv preprint arXiv:2209.12349, 2022b. Guardasoni, C., Rodrigo, M.R. and Sanfelici, S., A Mellin transform
Boyarchenko, S. and Levendorskii, S., Lévy models amenable to approach to barrier option pricing. IMA J. Manag. Math., 2020,
efficient calculations. arXiv preprint arXiv:2207.02359, 2022c. 31(1), 49–67.
Boyarchenko, S., Levendorskii, S., Kirkby, J.L. and Cui, Z., SINH- Heston, S., A closed-form solution for options with stochastic
acceleration for B-spline projection with option pricing applica- volatility with applications to bond and currency options. Rev.
tions. Int. J. Theor. Appl. Finance, 2022, 24(8), 2150040. Financ. Stud., 1993, 6(2), 327–343.
Brenner, M. and Subrahmanyam, M.G., A simple approach to option Heston, S.L. and Rossi, A.G., A spanning series approach to options.
valuation and hedging in the Black–Scholes model. Financ. Anal. Rev. Asset Pricing Stud., 2016, 7(1), 2–42.
J., 1994, 50, 25–28. Jackson, K.R., Jaimungal, S. and Surkov, V., Fourier space time-
Brummelhuis, R. and Chan, R.T.L., A radial basis function scheme stepping for option pricing with Lévy models. J. Comput. Finance,
for option pricing in exponential Lévy models. Appl. Math. 2008, 12(2), 1–29.
Finance, 2014, 21(3), 238–269. Jarrow, R. and Rudd, A., Approximate option valuation for arbitrary
Carr, P., Geman, H., Madan, D. and Yor, M., The fine structure of stochastic processes. J. Financ. Econ., 1982, 10, 347–369.
asset returns: An empirical investigation. J. Bus., 2002, 75(2), Jondeau, E. and Rockinger, M., Gram–Charlier densities. J. Econ.
305–333. Dyn. Control, 2001, 25, 1457–1483.
Carr, P. and Madan, D., Option valuation using the fast Fourier Kalemanova, A., Schmid, B. and Werner, R., The normal inverse
transform. J. Comput. Finance, 1999, 2, 61–73. Gaussian distribution for synthetic CDO pricing. J. Deriv., 2007,
Carr, P. and Wu, L., The finite moment log stable process and option 14(3), 80–94.
pricing. J. Finance, 2003a, 58(2), 753–777. Kirkby, J.L., Efficient option pricing by frame duality with the fast
Carr, P. and Wu, L., What type of process underlies options? A Fourier transform. SIAM J. Financ. Math., 2015, 6(1), 713–747.
simple robust test. J. Finance, 2003b, 58(6), 2581–2610. Kirkby, J.L., Robust barrier option pricing by frame projection under
Chan, T., Some applications of Levy process in insurance and exponential Lévy dynamics. Appl. Math. Finance, 2017, 24(4),
finance. Finance, 2004, 25, 71–94. 337–386.
Chan, R.T.L., Adaptive radial basis function methods for pricing Kirkby, J.L., American and exotic option pricing with jump diffu-
options under jump-diffusion models. Comput. Econ., 2016, 47(4), sions and other Levy processes. J. Comput. Finance, 2018, 22(3),
623–643. 89–148.
Cont, R. and Tankov, P., Financial Modelling with Jump Processes, Kirkby, J.L. and Deng, S., Swing option pricing by dynamic pro-
2004 (Chapman & Hall: New York). gramming with b-spline density projection. Int. J. Theor. Appl.
Cont, R. and Voltchkova, E., A finite difference scheme for option Finance, 2019, 22(8), 1950038.
pricing in jump diffusion and exponential Lévy models. SIAM J. Kirkby, J.L., Mitra, S. and Nguyen, D., An analysis of dollar cost
Numer. Anal., 2005, 43(4), 1596–1626. averaging and market timing investment strategies. Eur. J. Oper.
Cui, Z., Kirkby, J.L. and Nguyen, D., A data-driven framework for Res., 2020, 286(3), 1168–1186.
consistent financial valuation and risk measurement. Eur. J. Oper. Kirkby, J.L. and Nguyen, D., Equity-linked guaranteed minimum
Res., 2021, 289(1), 381–398. death benefits with dollar cost averaging. Insur. Math. Econ.,
De Innocentis, M. and Levendorskii, S., Pricing discrete barrier 2021, forthcoming. https://ssrn.com/abstract=3833502
options and credit default swaps under Lévy processes. Quant. Koponen, I., Analytic approach to the problem of convergence of
Finance, 2014, 14(8), 1337–1365. truncated Lévy flights towards the Gaussian stochastic process.
De Innocentis, M. and Levendorskii, S., Calibration and backtest- Phys. Rev. E, 1995, 52, 1197–1199.
ing of the Heston model for counterparty credit risk. Available at Kou, S., A jump-diffusion model for option pricing. Manag. Sci.,
SSRN 2757008, 2016. 2002, 48, 1086–1101.
Embrechts, P., Actuarial versus financial pricing of insurance. J. Risk Kudryavtsev, O. and Levendorskiǐ, S., Fast and accurate pricing of
Finance, 2000, 1(4), 17–26. barrier options under Lévy processes. Finance Stoch., 2009, 13(4),
Embrechts, P. and Maejima, M., An introduction to the theory of 531–562.
selfsimilar stochastic processes. Int. J. Mod. Phys., 2000, B14, Kudryavtsev, O. and Zanette, A., Efficient pricing of swing options
1399–1420. in Lévy-driven models. Quant. Finance, 2013, 13(4), 627–635.
 Closed-form option pricing for exponential Lévy models: a residue approach 275

Kuznetsov, A., Wiener–Hopf factorization and distribution of Mittnik, S. and Rachev, S., Stable Paretian Models in Finance, 2000
extrema for a family of Lévy processes. Ann. Appl. Probab., 2010, (John Wiley & Sons: Hoboken, NJ).
20, 1801–1830. Muhle-Karbe, J. and Nutz, M., Small-time asymptotics of option
Kuznetsov, A., Kyprianou, A.E., Pardo, J.C. and van Schaik, K., prices and first absolute moments. J. Appl. Probab., 2011, 48(4),
A Wiener–Hopf Monte Carlo simulation technique for Lévy 1003–1020.
processes. Ann. Appl. Probab., 2011, 21(6), 2171–2190. Neuberger, A., The log contract. J. Portf. Manag., 1994, 20, 74–80.
Lam, K., Chang, E. and Lee, M.C., An empirical test of the variance Ortiz-Gracia, L. and Oosterlee, C.W., Robust pricing of European
gamma option pricing model. Pac. Basin Finance J., 2002, 10, options with wavelets and the characteristic function. SIAM J. Sci.
267–285. Comput., 2013, 35(5), B1055–B1084.
Lee, S.S. and Hannig, J., Detecting jumps from Lévy jump diffusion Panini, R. and Srivastav, R.P., Option pricing with Mellin transforms.
processes. J. Financ. Econ., 2010, 96(2), 271–290. Math. Comput. Model., 2004, 40(1–2), 43–56.
Leitao, Á., Ortiz-Gracia, L. and Wagner, E.I., SWIFT valuation of Passare, M., Tsikh, A., and Zhdanov, O., A multidimensional Jor-
discretely monitored arithmetic Asian options. J. Comput. Sci., dan residue lemma with an application to Mellin-Barnes integrals.
2018, 28, 120–139. In Contributions to Complex Analysis and Analytic Geometry.
Levendorskii, S., Method of paired contours and pricing barrier Aspects of Mathematics, edited by H. Skoda and J.M. Trépreau,
options and CDSs of long maturities. Int. J. Theor. Appl. Finance, vol E 26, 1994 (Vieweg + Teubner Verlag: Wiesbaden).
2014, 17(5), 1450033. Phelan, C.E., Marazzina, D., Fusai, G. and Germano, G., Fluctuation
Levendorskii, S., Pitfalls of the Fourier transform method in identities with continuous monitoring and their application to the
affine models, and remedies. Appl. Math. Finance, 2016, 23(2), pricing of barrier options. Eur. J. Oper. Res., 2018, 271(1), 210–
81–134. 223.
Levendorskii, S.Z. and Boyarchenko, S.I., On rational pricing of Robinson, G.K., Practical computing for finite moment log-stable
derivative securities for a family of non-Gaussian processes. distributions to model financial risk. Stat. Comput., 2015, 25,
Working Paper, 1998. 1233–1246.
Lewis, A.L., A simple option formula for general jump-diffusion Rodrigo, M.R., Pricing formulas for perpetual American options
and other exponential Lévy processes, 2001. Available online at: with general payoffs. IMA J. Manag. Math., 2021, 33(2), 201–228.
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=282110 Roper, M., Implied volatility: Small time to expiry asymptotics in
(accessed 08 June 2020). exponential Lévy models. PhD diss., Thesis, University of New
Li, L. and Zhang, G., Option pricing in some non-Lévy jump models. South Wales, 2009.
SIAM J. Sci. Comput., 2016, 38(4), B539–B569. Rosinski, J., Tempering stable processes. Stoch. Process. Their Appl.,
Li, T.R. and Rodrigo, M.R., Alternative results for option pric- 2007, 117(6), 677–707.
ing and implied volatility in jump-diffusion models using Mellin Rydberg, T., The normal inverse Gaussian Lévy process: Simulation
transforms. Eur. J. Appl. Math., 2017, 28(5), 789–826. and approximation. Commun. Stat. Stoch. Models, 1997, 13, 887–
Linders, D. and Stassen, B., The multivariate variance gamma 910.
model: Basket option pricing and calibration. Quant. Finance, Sato, K., Lévy Processes and Infinitely Divisible Distributions, 1999
2015, 16(4), 555–572. (Cambridge University Press: Cambridge).
Lipton, A., Assets with jumps. Risk Mag., 2002, 15(9), 143–153. Schoutens, W., Lévy Processes in Finance: Pricing Financial
Lord, R., Fang, F., Bervoets, F. and Oosterlee, C.W., A fast and accu- Derivatives, 2003 (John Wiley & Sons: Hoboken, NJ).
rate FFT-based method for pricing early-exercise options under Semeraro, P., A multivariate variance gamma model for financial
Levy processes. SIAM J. Sci. Comput., 2008, 30, 1678–1705. applications. Int. J. Theor. Appl. Finance, 2008, 11(1), 1–18.
Loregian, A., Mercuri, L. and Rroji, E., Approximation of the vari- Venter, J. and de Jongh, P., Risk estimation using the normal inverse
ance gamma model with a finite mixture of normals. Stat. Probab. Gaussian distribution. J. Risks, 2002, 2, 1–25.
Lett., 2012, 82(2), 217–224. Wang, W. and Zhang, Z., Computing the Gerber–Shiu function by
Luciano, E., Business time and new credit risk models. International frame duality projection. Scand. Actuar. J., 2019, 2019, 291–307.
Centre for Economic (Research Working Paper No. 16/2010). Wilmott, P., Paul Wilmott on Quantitative Finance, 2006 (Wiley &
2009. https://ssrn.com/abstract=1626726 Sons: Hoboken, NJ).
Luciano, E. and Semeraro, P., Multivariate time changes for Lévy Xie, J. and Zhang, Z., Recursive approximating to the finite-time
asset models: Characterization and calibration. J. Comput. Appl. Gerber–Shiu function in Lévy risk models under periodic obser-
Math., 2010, 223(8), 1937–1953. vation. Working Paper, 2020.
Madan, D., Carr, P. and Chang, E., The variance gamma process and Zhang, Z., Yong, Y. and Yu, W., Valuing equity-linked death bene-
option pricing. Eur. Finance Rev., 1998, 2, 79–105. fits in general exponential Lévy models. J. Comput. Appl. Math.,
Madan, D. and Daal, E.A., An empirical examination of the variance 2020, 365, 112377.
gamma model for foreign currency options. J. Bus., 2005, 78(6), Zhdanov, O.N. and Tsikh, A.K., Studying the multiple Mellin–
2121–2152. Barnes integrals by means of multidimensional residues. Siber.
Madan, D. and Schoutens, W., Break on through to the single side. Math. J., 1998, 39(2), 245–260.
Working Paper, Katholieke Universiteit Leuven, 2008. Zhou, C., A jump-diffusion approach to modeling credit risk and
Madan, D. and Seneta, E., The variance gamma (V.G.) model for valuing defaultable securities. Board of Governors of the Fed-
share market returns. J. Bus., 1990, 63(4), 511–524. eral Reserve System, Finance and Economics Discussion Series
Mainardi, F., Luchko, Y. and Pagnini, G., The fundamental solution 1997–15, 1997.
of the space-time fractional diffusion equation. Fract. Calc. Appl.
Anal., 2001, 4, 153–192.
Mandelbrot, B., The variation of certain speculative prices. J. Bus.,
1963, 36(4), 394–419. Appendices
Mantegna, R.N. and Eugene Stanley, H., Stochastic process with
ultraslow convergence to a Gaussian: The truncated Lévy flight. Appendix 1. Additional numerical experiments
Phys. Rev. Lett., 1994, 73, 2946–2949.
Matache, A.-M., Von Petersdorff, T. and Schwab, C., Fast deter-
ministic pricing of options on Lévy driven assets. ESAIM Math. In Table A1 we compare the digital prices obtained by various
Model. Numer. Anal., 2004, 38(1), 37–71. truncations of the pricing Formulas 4.1 and 4.3 with a numerical
Merton, R., Option pricing when underlying stock returns are dis- evaluation of the Lewis method. Here we consider the asymmetric
continuous. J. Financ. Econ., 1976, 3, 125–144. VG model.
276 J.-P. Aguilar and J. L. Kirkby

Table A1. Digital options in the Variance Gamma model.

Formula 4.1
Asset or nothing N =5 N = 10 N = 15 N = 20 Lewis (2001)

Deep OTM (St = 3500) − 16218.3559 743.3252 1300.2662 1301.7552 1301.7559

OTM (St = 4500) 3029.2700 3071.5888 3071.9434 3071.9464 3071.9464
ATM (kvg = 0, St = 5050.24) 4174.4195 4174.4195 4174.4195 4174.4195 4174.4195
ITM (St = 5500) 4975.2793 4973.4963 4973.5001 4973.5001 4973.5001
Deep ITM (St = 6500) 6054.7046 6348.2679 6343.6084 6343.6108 6343.6108

Formula 4.3
Cash or nothing N =5 N = 10 N = 15 N = 20 Lewis (2001)

Deep OTM (St = 3500) − 4.4493 0.0960 0.2463 0.2467 0.2467

OTM (St = 4500) 0.5524 0.5575 0.5575 0.5575 0.5575
ATM (kvg = 0, St = 5050.24) 0.7257 0.7288 0.7288 0.7288 0.7288
ITM (St = 5500) 0.8321 0.8309 0.8309 0.8309 0.8309
Deep ITM (St = 6500) 0.8546 0.9376 0.9364 0.9364 0.9364

Note: Parameters: K = 5000, r = 1%, q = 0%, τ = 2, σ = 20%, ν = 0.85, θ = 0.1.

Table A2. Mellin pairs used throughout the paper. 3. The inversion of the Mellin transform is performed via an inte-
gral along a vertical line in the strip of convergence: if c ∈ (c− , c+ )
f (x) f ∗ (s) Convergence strip and f ∗ (c + it) is integrable, then :
 c+i∞
e−ax a−s (s) 0, ∞ ds
e−ax − 1 a−s (s) −1, 0 f (x) = f ∗ (s)x−s . (A4)
c−i∞ 2iπ
Kμ (ax) −s
a 2 ( s−μ s+μ
2 ) ( 2 ) |Re(μ)|, ∞
s−2
√ In particular, for the exponential function, one gets the so-called
Kμ (a x2 +b2 )
a 2 2 2 −1 b 2 −μ ( 2s )Kμ− 2s (ab)
s s s
μ 0, ∞ Cahen–Mellin integral:
(x2 +b2 ) 2
 c+i∞
ds
e−x = (s)x−s , c > 0. (A5)
c−i∞ 2iπ
Appendix 2. Review of the Mellin transform 4. When f ∗ (s) is a ratio of products of gamma functions of linear
arguments:
The theory of the one-dimensional Mellin transform is explained in (a1 s + b1 ) . . . (am s + bm )
full detail in Flajolet et al. (1995), and table of Mellin transforms can f ∗ (s) = (A6)
(c1 s + d1 ) . . . (cl s + dl )
be found in any monograph on integral transforms (see e.g. Bate-
man (1954)). An introduction to multidimensional complex analysis then one speaks of a Mellin–Barnes integral, whose characteristic
can be found in the classic textbook Griffiths and Harris (1978), quantity is defined to be
and applications to the specific case of Mellin–Barnes integrals are
m
l
developed in Passare et al. (1994) and Zhdanov and Tsikh (1998). = ak − cj . (A7)
k=1 j=1

governs the behavior of f ∗ (s) when |s| → ∞ and thus the pos-
A.1. One-dimensional Mellin transform
sibility of computing (A4) by summing the residues of the analytic
1. The Mellin transform of a locally continuous function f defined continuation of f ∗ (s) right or left to the convergence strip:
on R+ is the function f ∗ defined by   ∗ 
−s ,
< 0 f (x) = − Re(s)>β Res  ∗ f (s)x
 ∞ (A8)
> 0 f (x) = Re(s)<α Res f (s)x−s .
f ∗ (s) := f (x)xs−1 dx. (A1)
0 For instance, in the case of the Cahen–Mellin integral, one has = 1
The region of convergence {c− < Re(s) < c+ } into which the inte- and therefore:
∞
gral (A1) converges is often called the convergence (or sometimes   (−1)n n
fundamental) strip of the transform, and denoted c− , c+ . e−x = Res (s)x−s = x (A9)
2. The Mellin transform of the exponential function is, by n!
Re(s)<0 n=0
definition, the Euler gamma function:
 ∞ as expected from the usual Taylor series of the exponential function.
(s) = e−x xs−1 dx (A2)
0
with strip of convergence 0, ∞. Outside of this strip, it can be ana-
A.2. Elements of the multidimensional Mellin transform
lytically continued, except at every negative s = − n integer where Let n ∈ N. Let us denote the vectors in Cn by s = (s1 , . . . , sn ), and
it admits the singular behavior let ‘.’ represent the Euclidean scalar product.
(−1)n 1 •Polyhedron: a polyhedron in Cn is a subset  = 1 × · · · × N
(s) ∼ , n ∈ N. (A3) where
s→−n n! s + n ! "
k = s ∈ Cn , Re(hk (sk )) < ρk , k = 1, . . . , N, (A10)
In Table A2, we summarize the main Mellin transforms used in this
paper, as well as their convergence strips (see further details, e.g. in for N ∈ N, ρk ∈ R and hk : C → C a linear application. When
Bateman (1954)). N = n then one speaks of a polyhedral cone or simply of a cone.
 Closed-form option pricing for exponential Lévy models: a residue approach 277

Note that the notion of polyhedron (resp. cone) is simply the multidi- Appendix 3. Special functions and remarkable identities
mensional generalization of the notion of strip (resp. half plane) that
has been introduced in the former section for the one-dimensional A.3. Gamma function
Mellin transform.
•Distinguished boundary: If  = 1 × · · · × n is a cone in Cn Particular values of the gamma function. The gamma function (s)
then its distinguished boundary (or edge, or vertex) is has been defined in (A2) for Re(s) > 0; integrating by parts shows
that it satisfies the functional relation (s + 1) = s (s); as (1) = 1,
∂0  = ∂1 × · · · × ∂n . (A11) it follows that

•Mellin–Barnes integral: A Mellin–Barnes integral in Cn , which (n + 1) = n!, n∈N (A18)

generalizes the definition (A6), is an integral of the type and that the analytic continuation of (s) to the negative half-plane

is singular at every negative integer −n with residue (−1)
n

√ n! . Other
ω (A12)
c+iRn useful identities include ( 21 ) = π and, more generally,
⎧  
where ω is a complex differential n-form that reads ⎪
⎪ 1 (−1)n 4n n! √
⎪
⎨ − n = π
2 (2n)!
(a1 .s + b1 ) . . . (am .s + bm ) ds1 . . . dsn   (A19)
ω= x−s −sn
1 . . . xn
1
, ⎪
⎪
⎪ 1 (2n)! √
(ã1 .s + b̃1 ) . . . (ãl .s + b̃l ) (2iπ )n ⎩ +n = n π.
2 4 n!
x1 , . . . xn ∈ R, (A13) for n ∈ N.
Stirling approximation. We recall the well-known Stirling approx-
where ak , ãj ∈ Rn , bk , b̃j ∈ R, and where c belongs to the polyhe- imation for the factorial:
dron of convergence of the integral, which generalizes the notion of √
convergence strip in C. Note that we can, without loss of generality, n! ∼ 2πnnn e−n . (A20)
n→∞
assume that c ∈ Rn .
•Characteristic vector: The generalization of the characteristic Pochhammer symbol. The Pochhamer symbol (a)n , sometimes
quantity (A7) is the characteristic vector associated to the Mellin– denoted by the Appell symbol (a, n), and also called rising factorial,
Barnes integral (A13). It is defined by: is defined by
(a + n)

m
l (a)n := , a∈
/ Z− . (A21)
= ak − ãj (A14) (a)
k=1 j=1 The definition (A21) extends continuously to negative integers
thanks to the functional relation (s + 1) = s (s):
•Admissible subspace: The admissible subspace associated to the ⎧
Mellin–Barnes integral (A13) is defined to be ⎨ (−1)n k!
0≤n<k
(−k)n =
⎩ 0(k − n)! n > k
(A22)
 := {s ∈ Cn , Re( .s) < Re( .c)}. (A15)

Note that, being a real vector,  can also be written as the direct where k ∈ N.
sum  = π ⊕ iCn where π = {x ∈ Rn , .x < .c}. In dimen- Legendre duplication formula. For any s ∈ C, we have:
sion 1 and if = 0, the only possibilities for the admissible subspace   √
2s π 1
are the left and right half planes {s ∈ C, Re(s) < c} if > 0, and = s−1  . (A23)
{s ∈ C, Re(s) > c} if < 0. The admissible subspace is a key con- (s) 2 s+12
cept for contour closing as it acts as the multidimensional analog
of the well-known Jordan lemma (the proof involves the Stirling
approximation for the gamma functions and can be found in full
details in Passare et al. (1994) and Zhdanov and Tsikh (1998)).
• Residue theorem: If = 0 and if  ⊂  is a cone whose
A.4. Bessel functions
distinguished boundary is c + iRn , then† The modified Bessel function of the second kind, also called the
 MacDonald function, can be defined by the Mellin integral

ω= Resω (A16) 1  z μ ∞ −t− z2 −μ−1
c+iR2  Kμ (z) := e 4t t dt (A24)
2 2 0
π
and the series is absolutely convergent. The residues in (A16) are to for |argz| < 4. It follows that Kμ (z) has the symmetry property:
be understood as the ‘natural’ generalization of the Cauchy residue,
that is: Kμ (z) = K−μ (z) (A25)

 and has monotonous absolute values:
1 ds1 . . . dsn
Res f (s 1 , . . . , sn ) n n
(2iπ )n s11 . . . snn 0 ≤ μ1 < μ2 =⇒ |Kμ1 (z)| < |Kμ2 (z)|. (A26)

1 ∂ n1 +···+nn −n f  Large index. When μ → ∞, one has the following behavior:
=  . (A17)   −μ
(n1 − 1)! . . . (nn − 1)! ∂sn1 −1 . . . ∂snnn −1  π ez
1 s1 =...sn =0 Kμ (z) ∼ . (A27)
μ→∞ 2μ 2μ

† To be fully rigorous, one should group the gamma functions in the Large argument (Hankel’s expansion). Define the following
numerator of (A13) into n subfamilies called divisors and ensure that sequence:
⎧
each face of  is intersected by at most one of each divisors. We
deliberately omit these technical details which are automatically sat- ⎨ a0 (μ) = 1
⎪
isfied by construction in all our proofs. Interested readers are invited ⎪ (4μ2 − 12 )(4μ2 − 32 ) · · · (4μ2 − (2k − 1)2 )
⎩ ak (μ) = , k ≥ 1.
to refer to earlier papers such as Aguilar (2020b) or to the theoretical k!8k
papers Passare et al. (1994) and Zhdanov and Tsikh (1998). (A28)
278 J.-P. Aguilar and J. L. Kirkby

Then, for large z and fixed µ, we have: where cz > 0. The Mellin transform (13) can therefore be written as
 ∞  cz +i∞  ∞
π −z ak (μ) dz
Kμ (z) = e . (A29) p∗a/n (s) = K (−1)−z (z) (kX + x)−z xs−1 dx .
z→∞ 2z zk cz −i∞ −kX 2iπ
k=0
(A35)
The x integral is a particular case of Beta function; it exists on the
In particular, when 4μ2 − 1 = 0, i.e. when μ = 12 , all the ak (μ) are
condition that −kX > 0 is satisfied, and, in this case, we get
null in definition (A28) when k ≥ 1, and we are left with:
  cz +i∞
(z) (1 − z) (−s + z) dz
π −z p∗a/n (s) = K (−1)−z (−kX )s−z
K 1 (z) = e (A30) cz −i∞ (1 − s) 2iπ
2 2z
(A36)
for all z. with overall convergence condition 0 < Re(z) < 1, Re(−s + z) > 0.
Multiplication theorem. If |λ2 − 1| < 1 then the modified The proof of (iii) is similar: we write peur (St er−q+ωX )τ +x , K) =
Bessel function satisfies the multiplication theorem (see details in K(ekX +x − 1)1{x>−kX } , and we introduce the representation (see
Abramowitz and Stegun (1972)) Table A2 and the inversion formula (A4))
∞ z k
 cz +i∞

2
( 2 ) (1 − λ2 )k ekX +x − 1 = (−1)−z (z)(kX + x)−z
dz
Kμ−k (z) = λμ Kμ (λz). (A31) 2iπ
, (A37)
k! cz −i∞
k=0
for cz ∈ (−1, 0). We conclude by performing the Mellin trans-
form (13) like in (ii).
Last, to prove (iv), we write ppow (St er−q+ωX )τ +x , K) =
(a)
Appendix 4. Proofs K(ea(kX +x) − 1)1{x>−k (a) } , and, similarly to (iii), use the represen-
X
tation
A.5. Proof of Proposition 2.1 
(a) cz +i∞ dz
+x) (a)
We start by proving (i); we note first that pc/n (St er−q+ωX )τ +x , K) = ea(kX −1= (−1)−z a−z (z)(kX + x)−z , (A38)
cz −i∞ 2iπ
1{x>−kX } , and therefore the Mellin transform (13) reduces to
 ∞ for cz ∈ (−1, 0). We conclude by performing the Mellin trans-
(−kX )s
p∗c/n (s) = xs−1 dx = − (A32) form (13) like in (ii), and the proof is complete.
−kX s

where the integral converges for −kX > 0 and Re(s) < 0. The proof
of (v) is similar, by remarking that plog (St er−q+ωX )τ +x , K) = (kX +
x)1{x>−kX } and performing the integral Appendix 5. Table of characteristic exponents
 ∞ (−kX )s+1 In Table A3 we summarize the characteristic exponents (Fourier
p∗log (s) = (kX + x)xs−1 dx = − (A33) space) and their cumulant generating functions counterpart (Laplace
−kX s(s + 1)
space) for the various models under consideration in the paper.
that converges for −kX > 0 and Re(s) < −1.
To prove (ii), we note that pa/n (St er−q+ωX )τ +x , K) = KekX +x
1{x>−kX } , and we introduce the representation (see Table A2 and the
inversion formula (A4))
 cz +i∞
dz
ekX +x = (−1)−z (z)(kX + x)−z , (A34)
cz −i∞ 2iπ

Table A3. Characteristic exponents and cumulant generating functions.

Model Characteristic exponent ψX (u) Cumulant generating function ψX (−ip)

α α
TS(−) (σ , α, λ) − cosσ πα ((λ + iu)α − λα ) − cosσ πα ((λ + p)α − λα )
2 2
σα α
FMLS(σ , α) − cos πα (iu)α − cosσ πα pα
2 2
σ ν 2 2
σ ν 2 2
VG(σ , ν, θ ) − ν1 log(1 − iθ νu + 
 2 u ) − ν1 log(1 − θνp − 
 2 p )
NIG(α, β, δ) −δ( α 2 − (β + iu)2 − α 2 − β 2 ) −δ( α 2 − (β + p)2 − α 2 − β 2 )