Quantitative Finance

ISSN: 1469-7688 (Print) 1469-7696 (Online) Journal homepage: https://www.tandfonline.com/loi/rquf20

A PDE approach to jump-diffusions

Peter Carr & Laurent Cousot

To cite this article: Peter Carr & Laurent Cousot (2011) A PDE approach to jump-diffusions,
Quantitative Finance, 11:1, 33-52, DOI: 10.1080/14697688.2010.531042

To link to this article: https://doi.org/10.1080/14697688.2010.531042

Published online: 16 Dec 2010.

Submit your article to this journal

Article views: 260

Citing articles: 7 View citing articles

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

https://www.tandfonline.com/action/journalInformation?journalCode=rquf20
Quantitative Finance, Vol. 11, No. 1, January 2011, 33–52

A PDE approach to jump-diffusions

PETER CARRy and LAURENT COUSOT*z
yCourant Institute, NYU, 251 Mercer Street, New York, NY, 10012, USA
zBNP Paribas, 10 Harewood Avenue, London NW1 6AA, UK

(Received 21 December 2009; in final form 7 October 2010)

In this paper, we show that the calibration to an implied volatility surface and the pricing of
contingent claims can be as simple in a jump-diffusion framework as in a diffusion framework.
Indeed, after defining the jump densities as those of diffusions sampled at independent and
exponentially distributed random times, we show that the forward and backward Kolmogorov
equations can be transformed into partial differential equations. This enables us to (i) derive
Dupire-like equations [Risk Mag., 1994, 7(1), 18–20] for coefficients characterizing these
jump-diffusions; (ii) describe sufficient conditions for the processes they induce to be
calibrated martingales; and (iii) price path-independent claims using backward partial
differential equations. This paper also contains an example of calibration to the S&P 500
market.

Keywords: Martingales; Jump-diffusion processes; Partial differential equations; Calibration;

Options

1. Introduction strike, or, equivalently, marginal distributions at all times,

then solutions exist in the literature. First, Kellerer (1972)
When pricing derivative products, having a model showed that if the option prices satisfy simple necessary
calibrated to the most liquid products seems to be a conditions for no arbitrage, then there exists a calibrated
must for practitioners, since these very products will be martingale. However, the proof does not exhibit any
used for hedging. In many situations, one would actually explicit construction and is therefore not very useful for
be satisfied by a model calibrated to the available practitioners.
European option quotes. However, even this restricted The first explicit construction is due to Dupire (1994),
goal is difficult: very few models succeed in being who showed that, at most, one driftless diffusion can be
arbitrage-free and calibrated to a finite number of calibrated to a continuum of densities. The appeal of this
option quotes. so-called local volatility model is that the diffusion
To the best of our knowledge, only discrete time and coefficient is easy to recover from option prices, using
space Markov chain models have been proven to achieve the well-known Dupire’s formula. Nonetheless, this class
this goal so far. Indeed, Buehler (2006), Cousot (2007) of models has several flaws.{ In particular, the rendering
and Davis and Hobson (2007), who were inspired by the of the short-term smile is problematic. This is mainly due
earlier results of Carr and Madan (2005), independently to the fact that diffusions have continuous sample paths.
proved that, if the option quotes satisfy certain explicit To better fit the observed data, Andersen and
conditions, then there exists an arbitrage-free Markov Andreasen (2000) considered jump-diffusions instead of
chain model, which is calibrated exactly to the data. diffusions and explained how to recover the diffusion
However, these results are mostly theoretical since the so- coefficient from option prices, once the jump component
constructed models are particularly unrealistic.x is specified. An alternative approach, involving the
If one relaxes the assumptions by considering given a so-called local Le´vy process, is due to Carr et al. (2004),
continuum of European option prices in maturity and where the authors consider a stochastic differential

*Corresponding author. Email: [email protected]

xFor instance, the number of states can be decreasing with maturity. We refer to Cousot (2005) for an attempt to alleviate this
defect.
{We refer to Gatheral (2006) for an overview of the problems linked to local volatility models.
Quantitative Finance
ISSN 1469–7688 print/ISSN 1469–7696 online  2011 Taylor & Francis
http://www.informaworld.com
DOI: 10.1080/14697688.2010.531042
34 P. Carr and L. Cousot

equation, which is not driven by a Brownian motion but the value of European-style path-independent claims.
by a general Lévy process, and explain how to obtain the Section 5 gives an example of calibration to S&P 500
local speed function from option prices. option quotes. Section 6 concludes. Appendix G derives
Two different points of view, which are based respec- heuristically small time asymptotics for the implied
tively on the Azéma–Yor solution to the Skorokhod volatility when the underlying process is a generic jump-
embedding problem and on self-similar processes, can be diffusion with finite activity.
found in the work of Madan and Yor (2002), but will not
be pursued here.
We assume in this paper that the underlying process
2. Definition and properties of localizable
follows a jump-diffusion, as in Andersen and Andreasen
jump-diffusions
(2000) (or as in Carr et al. 2004 if the driving Lévy process
has finite activity). The originality of our approach comes
For the sake of simplicity, we assume throughout this
from the specification of the jump densities. Indeed, the
paper that the underlying asset does not pay dividends
above papers assume that the ratio of the asset values,
and that interest rates are zero. Moreover, we assume
after a jump and right before a jump, have a fixed density:
that, under a risk-neutral measure, this asset follows a
we instead assume that the jump densities are those of a
finite activity Markov process, with Lévy-type generator.
diffusion sampled at independent and exponentially
Consequently, its generator, Gt, has the following form:
distributed random times. As we shall show, this class
of models has numerous advantages for calibration and 1
Gt f ðxÞ  xðx, tÞ f 0 ðxÞ þ x2  2 ðx, tÞ f 00 ðxÞ
pricing. Z þ1 2 
As far as calibration is concerned, the forward þ ðx, tÞ ðx, t, yÞ f ð yÞdy  f ðxÞ , ð1Þ
Kolmogorov equation—which is a partial integro- 0
differential equation (PIDE)—can be transformed into a where  corresponds to the drift term,  to the diffusion
partial differential equation (PDE). As a consequence, a component,  to the intensity of the counting process and
Dupire-like formula, linking the different coefficients ((x, t,.)) to the densities of the jumps, or, more precisely,
characterizing the jump-diffusion and the partial deriva- of the values of the asset after a jump.
tives of the option prices, can be derived. Therefore, the Of course, the different coefficients of the generator Gt
calibration procedure resembles the one encountered need to satisfy some conditions for the above process to
when fitting local volatility models and is therefore exist. Fortunately, the well-posedness of the martingale
much less time-consuming than those described by problem associated with the generator Gt has been studied
Andersen and Andreasen (2000) or Carr et al. (2004). by Komatsu (1973) and Stroock (1975)y and we use the
Moreover, from a theoretical point of view, we state results of the latter to state, in the next proposition,
sufficient conditions for the underlying process to be a sufficient conditions on , ,  and  for the process (Ft)
calibrated martingale, which was not achieved in the to be well-defined.
above papers nor by Dupire (1994).
As far as pricing is concerned, the situation is similar. Proposition 2.1: There exists a positive strong Markov
First, the backward Kolmogorov PIDE, which governs process (Ft), which is right-continuous with left limits and
the price of path-independent claims, can also be trans- whose generator Gt is defined by equation (1) for

f 2 C1
0 ðRþ Þ—the space of infinitely differentiable functions
formed into a PDE and the associated numerical schemes
with compact support in Rþ  fx 2 R s:t: x 4 0g—if
are therefore much faster than those needed for a general
jump-diffusion. Second, pricing by Monte-Carlo meth- .  ,  and  are bounded and continuous functions
ods, which is more practical for the valuation of complex from Rþ  Rþ to, respectively, Rþ , R and Rþ;
claims or in a multinomial setting, is also possible since .  is a function from Rþ  Rþ  Rþ to Rþ such
the jump densities are easy to simulate. Consequently, for that ðx, t, Þ is a probability density for
pricing too, the situation is similar to that faced with a ðx, tÞ 2 Rþ  Rþ and such that
local volatility model, but this model does account for Z
jumps. This is why we call this kind of jump-diffusion lnðy=xÞ=½1 þ ðlnðy=xÞÞ2 ðx, t, yÞdy

localizable in this paper.
The rest of this paper is organized as follows. Section 2 is continuous in x and t for every Borel set  of
defines robustly localizable jump-diffusions and describes Rþ .
sufficient conditions for these processes to be martingales.
Proof: See appendix A. œ
Section 3 addresses the issue of calibration and derives
sufficient conditions for a localizable jump-diffusion to be Remark 1: Note that the second-order elliptic term is
calibrated. Section 4 derives a backward PDE governing dominating in the generator Gt since 40.

yNote that these papers generalize to processes with Lévy-type generators the results of the seminal papers of Stroock and Varadhan
(1969a, b) for diffusions.
 A PDE approach to jump-diffusions 35

Now that we have at our disposal sufficient conditions and apply this operator to equation (7), we obtain
for the process (Ft) to exist, we would like to describe
 
further constraints on Gt, ensuring that it is a martingale @2 a2 ð yÞ
ðb
pðx, y, Þ  ð y  xÞÞ ¼ 2 b
pðx, y, Þ : ð9Þ
and not just a strict local martingale.y This is the purpose @y 2
of the next proposition.
By taking  ¼ 1 in the above equation and defining
Proposition 2.2: Under the assumptions of proposition 2.1,
(Ft) is a local martingale if there exists A40 such that gðx, yÞ  b
pðx, y, 1Þ, ð10Þ
Z þ1
y
ðx, t, yÞdy4A, ð2Þ we see that this function solves
0 x

for ðx, tÞ 2 Rþ  Rþ and @2 a2 ð yÞ
Z  I 2 gðx, yÞ ¼ ð y  xÞ, ð11Þ
þ1
y @y 2
ðx, tÞ ¼ ðx, tÞ ðx, t, yÞdy  1 : ð3Þ
0 x and is therefore a Green’s function of the operator
If we further assume that there exists B40 such that [I  (@2/@y2)a2(y)/2], which will prove to be crucial later.
Z þ1    Moreover, g(x,.) is most likely a probability density.
y y 
ln ðx, t, yÞdy4B, ð4Þ Indeed, it can be interpreted as the density of Sx j S0 ¼ x,
0 x x where  x is an independent random variable with a mean
then (Ft) is a martingale. 1 exponential distribution. Furthermore, it would not be
Proof: The proof of this proposition is based on the senseless to hope that g(x,.) would have a mean equal to
extensive literature on conditions for a positive local x, since (St) is a local martingale.
martingale to be a martingale and in particular on the A final point that needs to be considered is of course
results of Lepingle and Mémin (1978). See appendix B for the boundary conditions associated with the operator
the actual proof. œ [I  (@2/@y2)a2(y)/2] or at least the domain of definition of
g(x,.). Since we are mainly interested in densities on Rþ to
So far, we have considered a general jump-diffusion. model the underlying asset jumps, we need to restrict the
But, as explained in the Introduction, the originality of function a2 for (St) to remain positive. This would
our approach comes from the specification of the jump undoubtedly force g(x,.) to be defined on a subset of
densities. Indeed, in the following, we consider jump Rþ and this is done by imposing the well-known Feller’s
densities ((x, t,.)), which are the densities of diffusions, conditions (Feller 1952).
sampled at independent and exponentially distributed The following lemma makes the above heuristic
random times. These distributions are called in this paper presentation precise and summarizes some of the prop-
generalized Laplace distributions because the Laplace erties of g that will be needed throughout the paper. The
distribution, whose density is defined, for 40, by proof is based on the results of Feller (1952) and McKean
 
1 jxj (1956).
p ðxÞ  exp  , ð5Þ
2  Lemma 2.3: If a2 is a continuous and positive function
pffiffiffi
can be interpreted as the distribution of ð 2W Þ, where on Rþ that satisfies
(Wt) is a Brownian motion and  is an independent Z 1 Z 1 
random time with an exponential distribution of mean 1. 2
J0  2
dy dx ¼ þ1, ð12Þ
To gain some intuition, let us first investigate infor- 0 x a ð yÞ
mally the properties of these densities. If a is a continuous
then, for every y 2 Rþ , there exists a unique function
and positive function on R, the driftless stochastic
differential equation x ! G(x, y) that satisfies

dSt ¼ aðSt ÞdWt ð6Þ a2 ðxÞ @2 a2 ð yÞ
I 2
Gðx, yÞ ¼ ðx  yÞ, ð13Þ
admits a solution that is unique in the sense of probability 2 @x 2
law, according to Engelbert and Schmidt (1984). If we
as well as the following boundary conditions:
denote by p(x, y, t) the associated transition probability
density, the forward Kolmogorov equation is @G
  ðx, yÞ ! 0: ð14Þ
@x x!0 or þ1
@p @2 a2 ð yÞ
ðx, y, tÞ ¼ 2 pðx, y, tÞ : ð7Þ
@t @y 2 Moreover, x ! G(x, y) is positive and continuous on Rþ as
well as convex, twice differentiable and increasing (respec-
If we define the Carson–Laplace transform of p by tively decreasing) on (0, y] (respectively [y, þ1)). It also
Z þ1 possesses the following symmetry:
b
pðx, y, Þ   eu pðx, y, uÞdu, ð8Þ
0 Gðx, yÞ ¼ Gð y, xÞ: ð15Þ

yWe refer to Cox and Hobson (2005) for the problems one might encounter when pricing with a strict local martingale such as
failure of put–call parity, etc.
36 P. Carr and L. Cousot
Table 1. Examples of functions g.
2
a (x) g(x, y)

a2x2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2ðminðx, yÞÞþ ðmaxðx, yÞÞ 1 1 þ ð8=a2 Þ
, with  ¼
ðayÞ2 ðþ   Þ 2
pffiffiffiffiffiffi pffiffiffi pffiffiffi
a2x2 (51) 2 xyI ð½ 2=ðað1  ÞÞðminðx, yÞÞ1 ÞK ð½ 2=ðað1  ÞÞðmaxðx, yÞÞ1 Þ
,
a2 y2 ð1  Þ
with ¼ 1/[2(1  )] and I (respectively K ) the modified Bessel function of the first (respectively second)
kind

Furthermore, if we define g(x, y)  2G(x, y)/a2(y), then with M 2 Rþ , then g is well-defined as a family of
y ! g(x, y) is a probability density and g satisfies generalized Laplace densities associated with a2.
 Moreover, g satisfies the assumptions of propositions 2.1
@2 a2 ð yÞ and 2.2, except for condition (4).
I 2 gðx, yÞ ¼ ð y  xÞ, ð16Þ
@y 2 If we further assume that there exists m40 s.t.
as well as m2 x2 4a2 ðx, tÞ, ð23Þ
 then condition (4) is also valid.
a2 ðxÞ @2
I gðx, yÞ ¼ ðx  yÞ: ð17Þ
2 @x2 Proof: See appendix D. œ
Finally, if we further assume that Remark 2: Depending on the nature of the underlying
Z asset, it could seem odd, for a practitioner, to use
1
xdx generalized Laplace distributions as jump densities.
K0  ¼ þ1, ð18Þ
0 a2 ðxÞ Indeed, these densities are centered, whereas jump den-
sities usually satisfy
Z þ1 Z þ1
xdx
Kþ1  ¼ þ1, ð19Þ ðx, t, yÞdy 5 x ð24Þ
1 a2 ðxÞ
0
then to account for credit risk, the fear of downward moves,
etc. However, note that we could have defined, along the
Gðx, yÞ ! 0, ð20Þ
x!0 or þ1 same lines, non-centered generalized Laplace distributions
by considering a stochastic differential equation with drift
@G instead of equation (6). Indeed, the main properties of
x ðx, yÞ ! 0, ð21Þ these densities, that is being the Green’s functions of a
@x x!0 or þ1
backward and a forward operator, would have been
and y ! g(x, y) is a probability density of mean x. preserved, and their means could have been computed
Proof: See appendix C. œ explicitly, had we taken this drift linear. Nonetheless, for
the sake of simplicity, we will not pursue this idea here.
Table 1 shows two examples of functions g for different
choices of function a2.
We now have all the elements to define localizable
jump-diffusions. Indeed, let us consider a function a2 on 3. Calibrating a localizable jump-diffusion to an implied
Rþ  Rþ , such that, for each t 2 Rþ, a2(, t) satisfies the volatility surface
assumptions of lemma 2.3. Then, for each t, it is possible
to associate with a2(, t) a (unique) family of generalized In this section, we determine, informally, a necessary
Laplace densities (g(x, t,.)). And, as expected, we assume condition on ,  and a2 for (Ft) to be calibrated to a
in the following that these densities are the jump densities surface of call prices (C(x, t)). Furthermore, we derive
of the process (Ft), that is, in symbols,  ¼ g. Our only sufficient conditions for this process to be a calibrated
concern is therefore to find sufficient conditions on this martingale.
function a2 for the densities g to satisfy the assumptions of
propositions 2.1 and 2.2. This is the aim of the following Remark 3: Note that we assume implicitly throughout
lemma. this section that the marginal distributions of (Ft) admit
smooth densities. Indeed, we were not able to find in the
Lemma 2.4: With the above notation, if a2 is a positive literature sufficient conditions on Gt for this to be true.
and continuous function that satisfies Actually, conditions are only available, in the discontin-
uous case, for strong solutions of stochastic differential
a2 ðx, tÞ4M2 x2 , ð22Þ equations with jumps (we refer to Bichteler and Jacod
 A PDE approach to jump-diffusions 37

1983, Komatsu and Takeuchi 2001, Takeuchi 2002 and Note that g(y, t, x) is the price of an out-of-the-money
Cass 2006). And, as remarked by Bichteler and Jacod option struck at x when the risk-neutral density is g(y, t, ).
(1983), to go from this framework to the framework of Consequently, by put–call parity,
the martingale problem for Lévy generators is not Z þ1
straightforward. gð y, t, zÞðz  xÞþ dz  ð y  xÞþ ¼ g ð y, t, xÞ: ð29Þ
However, this assumption seems intuitive since the 0

process has finite activity, corresponds to a diffusion Therefore, equation (27) becomes
between jumps (Stroock 1981a, b). and the jump distri- @C 1 @2 C
butions have densities. We are confident that sufficient ðx, tÞ  x2  2 ðx, tÞ 2 ðx, tÞ
@t 2 @x
conditions on  ,  and  will soon be available. Z þ1 2
@ C
Finally, notice that knowledge of the structure of the ¼ ð y, tÞð y, tÞg ð y, t, xÞdy: ð30Þ
0 @y2
marginal distributions of (Ft) is necessary to prove that
they are calibrated to a call price surface. Indeed, risk- Moreover, using the definition of g (equation (28)) and
neutral distributions can be derived from this surface and equation (16), it is fairly simple to show thaty
if one wants to reconcile them with the marginal 
a2 ðx, tÞ @2 a2 ð y, tÞ
distributions of (Ft), one needs to assume that I g ð y, t, xÞ ¼ ðx  yÞ: ð31Þ
2 @x 2 2
they have, at least, the same regularity. Accordingly, we
assume in the following that C(, t) corresponds to a Consequently, we apply the operator [I  (a2(x, t)/2) @2/
marginal density pMarket(, t). @x2] to get rid of the integral on the RHS of equation (30):
  
a2 ðx, tÞ @2 @C 1 2 2 @2 C
I ðx, tÞ  x  ðx, tÞ ðx, tÞ
3.1. A necessary condition for calibration 2 @x2 @t 2 @x2
The process (Ft) has the Markov property and we know @2 C a2 ðx, tÞ
¼ 2 ðx, tÞðx, tÞ : ð32Þ
its generator Gt. Consequently, we can write the forward @x 2
Kolmogorov equation:
This equation can be rewritten as
@
ðE½PðFt ÞÞ ¼ E½Gt PðFt Þ, ð25Þ
@t a2 ðx, tÞ ð@C=@tÞðx, tÞ  12 x2  2 ðx, tÞð@2 C=@x2 Þðx, tÞ
for P 2 C1  ¼ 8 2 9 , ð33Þ
0 ðRþ Þ. Even if a call payoff does not belong to 2 < ð@ =@x2 Þðð@C=@tÞðx, tÞ =
1 
C0 ðRþ Þ, let us write the above equation in the case  12 x2  2 ðx, tÞð@2 C=@x2 Þðx, tÞÞ
P(y) ¼ (y  x)þ to gain intuition: : ;
þðx, tÞð@2 C=@x2 Þðx, tÞ
@
ðE½ðFt  xÞþ Þ or
@t 
1

¼ E F2t  2 ðFt , tÞðFt  xÞ þ ðFt , tÞ ½I  ða2 ðx, tÞ=2Þ@2 =@x2 ðð@C=@tÞðx, tÞ
2
Z þ1   12 x2  2 ðx, tÞð@2 C=@x2 Þðx, tÞÞ
ðx, tÞ ¼ : ð34Þ
 gðFt , t, zÞðz  xÞþ dz  ðFt  xÞþ : ð26Þ ða2 ðx, tÞ=2Þð@2 C=@x2 Þðx, tÞ
0
The last two equations suggest different strategies for the
Therefore, if the process (Ft) is calibrated to the call price
calibration.
surface C, we have
. Strategy A: choose a2 according to equation
@C 1 @2 C
ðx, tÞ  x2  2 ðx, tÞ 2 ðx, tÞ (33) for fixed  and .
@t 2 @x . Strategy : choose  according to equation (34)
Z þ1 2  Z þ1
@ C for fixed  and a2.
¼ 2
ð y, tÞ ð y, tÞ gð y, t, zÞðz  xÞþ dz
0 @y 0
 Remark 4: The closed-form local Lévy model obtained
þ in section 4 of Carr et al. (2004) can be interpreted as a
 ð y  xÞ dy, ð27Þ
special case of strategy  for a given choice of a
since pMarket(, t) ¼ (@2C/@x2)(, t), in the spirit of Breeden non-centered generalized Laplace distribution
and Litzenberger (1978). Let us denote by g(y, t, x) the (see remark 2).
double tail associated with the density g(y, t,.) , that is, in Remark 5: Before describing sufficient conditions for
symbols: strategies A and  in the following sections, we would like
( R þ1 R þ1
ð gð y, t, vÞdvÞdu, if y4x, to emphasize the fact that the above equations are only
g ð y, t, xÞ  Rxx R u u ð28Þ necessary, just as Dupire’s formula is only necessary.
0 ð 0 gð y, t, vÞdvÞdu, if y 4 x: To illustrate what could go wrong in a diffusion setting,

yNote that a consequence of equation (31) is g(y, t, x) ¼ G(y, t, x)  [a2(x, t)/2] g(y, t, x) by uniqueness of the associated boundary
value problem proved in lemma 2.3. Therefore, the double tail associated with the density g(y, t,.) is simply obtained by a
multiplication of a2 and the density.
38 P. Carr and L. Cousot

let us assume that the underlying asset follows the 0 5 inf 4 sup  5 þ1, ð40Þ
Rþ Rþ Rþ Rþ
following risk-neutral dynamic:

dSt  
¼ rdt þ S1 dWt , ð35Þ  @ 
St t sup x  5 þ1, ð41Þ

Rþ Rþ @x

S0 ¼ x0 , ð36Þ and

with 51, until it eventually hits 0 and is then stopped. sup  5 þ1, ð42Þ
Rþ Rþ
This model is the celebrated CEV model, which was first
studied by Cox (1975) and subsequently by Schröder Z þ1
(1989) and Delbaen and Shirakawa (2002), among others. 2 ðx, tÞ
sup dx 5 þ1: ð43Þ
Note that the stopped process is a martingale and that call Rþ 0 x
prices are known in closed form. Consequently, if market
prices were coming from a CEV model, one could attempt If the equation
to calibrate these prices with a diffusion and one would a2 ðx, tÞ ð@C=@tÞðx, tÞ  12 x2  2 ðx, tÞð@2 C=@x2 Þðx, tÞ
obtain, using Dupire’s formula,  8 2 9 ð44Þ
2 < ð@ =@x2 Þðð@C=@tÞðx, tÞ =
2 ð@CCEV =@tÞðx, tÞ þ rxð@CCEV =@xÞðx, tÞ  12 x2  2 ðx, tÞð@2 C=@x2 Þðx, tÞÞ
loc ðx, tÞ  : ;
1 2 2 2 þðx, tÞð@2 C=@x2 Þðx, tÞ
2 x ð@ CCEV =@x Þðx, tÞ
¼  2 x2ð1Þ : ð37Þ defines a continuous function a2 on Rþ  Rþ , such that
a2(x, t)/x2 takes values in a compact interval of Rþ , then
However, this local volatility model is not arbitrage-free there exists a positive strong Markov martingale (Ft),
when  1/2 if nothing is stated about stopping the 
starting at x0, whose generator, for f 2 C1
0 ðRþ Þ, is
process at 0. Indeed, in this case, {0} is reached a.s. and is
instantaneously reflecting. Therefore, American calendar 1 @2 f
Gt f ðxÞ ¼ x2  2 ðx, tÞ 2 ðxÞ
spreads struck at 0 have a positive value, which is an 2 @x
Z þ1 
arbitrage.
þ ðx, tÞ gðx, t, yÞ f ð yÞdy  f ðxÞ , ð45Þ
0

3.2. Sufficient conditions for strategy A where g(x, t,.) is uniquely determined by

Once  and  are fixed and a2 is defined by equation (33), @2 a2 ð y, tÞ
I 2 gðx, t, yÞ ¼ ð y  xÞ, ð46Þ
we still need to show that the corresponding localizable @y 2
jump-diffusion defines a martingale and that the call
prices in this very model are equal to C(x, t). The first part  
@ a2 ð yÞ
is easy since we know sufficient conditions for a localiz- gðx, t, yÞ ! 0: ð47Þ
@y 2 y!0 or þ1
able jump-diffusion to be a well-defined martingale
(see propositions 2.1 and 2.2 as well as lemma 2.4). Moreover, if the marginal distributions of (Ft) admit
However, the second part seems a bit trickier (see differentiable densities (pModel(, t)) that satisfy
remark 3) and is related to the uniqueness of the solution Z þ1
of the forward equation in an ad hoc space. xp2Model ðx, tÞdx 5 þ1,
0
Proposition 3.1: Let us first consider a continuous func- Z þ1  2 ð48Þ
tion C(x, t) from Rþ  Rþ to Rþ, such that CjRþ Rþ is in 3 @pModel
x ðx, tÞ dx 5 þ1,
C4,1, ð@2 C=@x2 ÞjRþ Rþ is in C0,1 and, which satisfies, for a 0 @x
fixed x040, then, for ðx, tÞ 2 Rþ  Rþ , we have
Cðx,0Þ ¼ ðx0  xÞþ , Cðx, tÞ  ðx0  xÞþ !x!0 or þ1 0, E½ðFt  xÞþ  ¼ Cðx, tÞ: ð49Þ
@C @C
@x ðx, tÞ 5 0, @x ðx, tÞ þ 1fxx0 4 0g !x!0 or þ1 0,
@2 C @2 C
@x2
ðx,tÞ 40, @x2
ðx, tÞx2 !x!0 or þ1 0, Proof: See appendix E. œ
@C @C
@t ðx, tÞ 4 0 , @t ðx, tÞ !x!0 or þ1 0, A necessary condition on , which follows implicitly
R þ1 @2 C
0 xð @x2 ðx, tÞÞ2 dx 5 þ1, from the assumptions of proposition 3.1, is
R þ1 3 @3 C
x ð @x3 ðx, tÞÞ2 dx 5 þ1: ð@C=@tÞðx, tÞ
0
 2 ðx, tÞ 5 loc
2
ðx, tÞ  1 , ð50Þ
ð38Þ 2 x2 ð@2 C=@x2 Þðx, tÞ
Let us also introduce two continuous functions  and  from where  loc is the local volatility defined by Dupire (1994).
Rþ  Rþ to Rþ that satisfy, respectively, Actually, this condition is quite intuitive: the diffusion
component of a calibrated jump-diffusion must be less
 2 C2,0 ðRþ  Rþ Þ, ð39Þ volatile than that of a local volatility model because of the
presence of jumps.
 A PDE approach to jump-diffusions 39

Note that equation (50) suggests a possible choice for : then there exists a positive strong Markov martingale (Ft),

starting at x0, whose generator is, for f 2 C1
0 ðRþ Þ,
ðx, tÞ ¼ ðtÞloc ðx, tÞ,
1 @2 f
with (t) 2 (0, 1). To have a consistent at-the-money smile Gt f ðxÞ ¼ x2  2 ðx, tÞ 2 ðxÞ
2 @x
for short maturities, one should take (t) ! t!01 (see Z þ1 
appendix G, where short time expansions are derived þ ðx, tÞ gðx, t, yÞ f ð yÞdy  f ðxÞ , ð59Þ
0
heuristically for the smile of a generic jump-diffusion).
Moreover, in this case, equation (44) greatly simplifies to where g(x, t, ) is uniquely determined by

a2 ðx,tÞ ð@C=@tÞðx, tÞ @2 a2 ð y, tÞ
¼ 3 : I 2 gðx, t, yÞ ¼ ð y  xÞ, ð60Þ
2 ð@ C=@x2 @tÞðx, tÞ þ ½ðx,tÞ=ð1  2 ðtÞÞð@2 C=@x2 Þðx, tÞ @y 2
ð51Þ
 
@ a2 ð yÞ
gðx, t, yÞ ! 0. ð61Þ
@y 2 y!0 or þ1

3.3. Sufficient conditions for strategy  Moreover, if the marginal distributions of (Ft) admit
differentiable densities (pModel(, t)) that satisfy
In this section, we investigate the feasibility of strategy . Z þ1
Of course, the point of view is a bit different, but a xp2Model ðx, tÞdx 5 þ1,
proposition, similar to proposition 3.1, can be readily 0
stated. Z þ1  2
@pModel
x3 ðx, tÞ dx 5 þ1, ð62Þ
Proposition 3.2: Let us first consider a continuous func- 0 @x
tion C(x, t) from Rþ  Rþ to Rþ, such that CjRþ Rþ is in then, for ðx, tÞ 2 Rþ  Rþ , we have
C4,1, ð@2 C=@x2 ÞjRþ Rþ is in C0,1 and which satisfies, for a
E½ðFt  xÞþ  ¼ Cðx, tÞ: ð63Þ
fixed x040,
Cðx,0Þ ¼ ðx0  xÞþ , Cðx, tÞ  ðx0  xÞþ !x!0 or þ1 0, Proof: The proof of this proposition is in every way
@C @C similar to that of proposition 3.1 in appendix E and is left
@x ðx, tÞ 50, @x ðx, tÞ þ 1fxx0 4 0g !x!0 or þ1 0,
to the reader. œ
@2 C @2 C 2
@x2
ðx, tÞ4 0, @x 2 ðx, tÞx ! x!0 or þ1 0, 2
A specification of  and a , which greatly simplifies the
@C @C
@t ðx, tÞ 40, @t ðx, tÞ ! x!0 or þ1 0: definition of , is
R þ1 2 ð@C=@tÞðx, tÞ
0 xð@@xC2 ðx, tÞÞ2 dx 5þ1,  2 ðx, tÞ ¼ 2 ðtÞ 1 , ð64Þ
R þ1 3 2 x2 ð@2 C=@x2 Þðx, tÞ
0 x3 ð@@xC3 ðx, tÞÞ2 dx 5 þ1:
ð52Þ a2 ðx, tÞ ð@C=@tÞðx, tÞ
¼
2 ðtÞ 2 , ð65Þ
Let us also introduce two continuous functions  and a2 2 ð@ C=@x2 Þðx, tÞ
from Rþ  Rþ to Rþ that satisfy, respectively, with 052(t),
2(t)51 and 2(t) ! 01. In this case,
equation (57) becomes
 2 C2,0 ðRþ  Rþ Þ, ð53Þ
 
2 1 ð@3 C=@t @x2 Þðx, tÞ
0 5 inf 4 sup  5 þ1, ð54Þ ðx, tÞ ¼ ð1   ðtÞÞ 2  2 : ð66Þ
Rþ Rþ

ðtÞ ð@ C=@x2 Þðx, tÞ
Rþ Rþ

  Moreover, this specification could be appealing for

 @  practitioners since 2 allows one to control, in a simple
sup x  5 þ1,
 ð55Þ

Rþ Rþ @x way, how much of the smile should be explained by the
diffusion component and
2 allows one to control the size
and
of the jumps.
a2 ðx, tÞ
0 5 m2 4 4M2 5 þ1: ð56Þ
x2
If the equation

 4. Pricing path-independent claims
½I  ða2 ðx, tÞ=2Þ@2 =@x2 ðð@C=@tÞðx, tÞ
 12 x2  2 ðx, tÞð@2 C=@x2 Þðx, tÞÞ In this section, we investigate the valuation of path-
ðx, tÞ  ð57Þ
ða2 ðx, tÞ=2Þð@2 C=@x2 Þðx, tÞ independent claims and show how their valuation is
possible with PDEs. First, we consider the case of a
defines a non-negative, bounded and continuous function ,
European-style claim, whose payoff is denoted by f—
such that
which has sub-linear growth, say—and whose value is
Z þ1 2 denoted by
 ðx, tÞ
sup dx 5 þ1, ð58Þ
Rþ 0 x VFf ðx, tÞ  E½ f ðFT Þ j Ft ¼ x: ð67Þ
40 P. Carr and L. Cousot

09/16/05 10/21/05
0.6 0.24
0.5
0.4 0.18
0.3
0.2 0.12
0.1
0 0.06
1150 1200 1250 1100 1150 1200 1250 1300 1350

11/18/05 12/16/05
0.24 0.24
0.2 0.2
0.16 0.16
0.12 0.12
0.08 0.08
1100 1150 1200 1250 1300 1350 1100 1200 1300 1400
03/17/06 06/16/06
0.2 0.2
0.175 0.175
0.15 0.15
0.125 0.125
0.1 0.1
1100 1150 1200 1250 1300 1350 1100 1200 1300 1400

12/15/06 06/15/07
0.2 0.2
0.175 0.175
0.15 0.15
0.125 0.125
0.1 0.1
1100 1200 1300 1400 1500 1600 1100 1200 1300 1400 1500 1600

SVI Bi d Ask

Figure 1. SVI fits to the SPX quotes as of the close on September 15, 2005.

The terminal condition is clearly VFf ðx, TÞ ¼ f ðxÞ and the using equation (17). After some simplifications, we obtain
backward Kolmogorov equation
@VFf 1 @2 V f a2 ðx, tÞðx, tÞ @2
0¼ ðx, tÞ þ x2  2 ðx, tÞ 2F ðx, tÞ þ
@VFf 1 2 2 @ Vf
2
@t 2 @x 2 @x2 !!
0¼ ðx, tÞ þ x  ðx, tÞ 2F
ðx, tÞ f 2 f
@t 2 @x 1 @VF 1 @ V
Z þ1   VFf ðx, tÞ  ðx, tÞ þ x2  2 ðx,tÞ 2F ðx, tÞ :
þ ðx, tÞ gðx, t, yÞVFf ð y, tÞdy  VFf ðx, tÞ : ð68Þ ðx, tÞ @t 2 @x
0
ð71Þ
If we further assume that 40, we can rewrite this
equation as Therefore, the valuation of VFf has been localized since it
can be performed by a PDE scheme.
Z þ1
As far as American-style claims are concerned, the
gðx,t,yÞVFf ðy,tÞdy formalism is a bit more complicated, but similar manip-
0
! ulations are possible since the differential operator
1 @VFf 1 2 2 @2 VFf
f
¼ VF ðx,tÞ  ðx, tÞ þ x  ðx,tÞ 2 ðx,tÞ : [I  (a2(x, t)/2)@2/@x2] preserves inequalities. We refer to
ðx, tÞ @t 2 @x Bensoussan and Lions (1984) for more details concerning
ð69Þ optimal stopping for a jump-diffusion.
2 2 2
Applying the operator [I  (a x, t/2)@ /@x ] on both sides
of the above equation, we obtain
  5. A numerical example: SPX quotes
a2 ðx, tÞ @2 1
VFf ðx, tÞ ¼ I VFf ðx, tÞ 
2 @x2 ðx, tÞ In this section, we give an example of calibration to S&P
!! 500 option quotes using the simplified Strategy , for
@VFf 1 2 2 @ Vf 2
 ðx, tÞ þ x  ðx, tÞ 2F ðx, tÞ which  2, a2 and  are defined, respectively, by equations
@t 2 @x (64), (65) and (66). The input data are SPX quotes as of
ð70Þ the close on September 15, 2005. To construct an implied
volatility surface that is approximately consistent with
 A PDE approach to jump-diffusions 41

0.9
0.8 1.2
0.7 1
0.6
0.8 0.25
0.5
0.4 0.6
0.2
0.3 0.4
0.2 1.5 0.15
0.1
0.2
Time
0 0 0.1
1
−1.5 Time
−1.5 0.05
−1 −1
0.5 −0.5
−0.5 0
0
0.5
Log−strike 1
Log−strike 0.5 1.5
1 0

Figure 3. Local volatility surface from the first to the fourth

Figure 2. Implied volatility surface obtained from the SVI fits. maturity.

these quotes, we follow Gatheral (2006, p. 37) and

proceed in two steps.
First, at each of the eight maturities, we fit the SPX
implied volatilities with the Stochastic Volatility Inspired
(SVI) parametrization. The so-generated smiles are
displayed in figure 1 along with the SPX bid and ask
implied volatilities.y Note that these smiles only make
sense, a priori, in a risk-neutral measure, where interest
0.8
rates are not zero. Therefore, to apply the above
0.7
constructions, we need to work in the corresponding
0.6
forward measure. But this change of measure does not
0.5 1.8
require extra work. Indeed, due to the special form of the
0.4 1.6
SVI parametrization, the implied volatility smiles in the
forward measure are given by the parametrizations in the 0.3 1.4
risk-neutral measure. With this in mind, we then proceed 0.2 1.2
to the second step and interpolate the cumulated vari- 0.1 1 Time
ances along the time variable using the algorithm of 0.8
Stineman (1980), which has the advantage of producing a −1 0.6
−0.5
0 0.4
fit with the same monotonicity as the input data. The Log−strike
0.5
1
implied volatility surface is displayed in figure 2. 1.5
In the simplified strategy ,  is chosen proportional to
Figure 4. Local volatility surface from the fourth to the eighth
the local volatility function, defined by Dupire (1994) as
maturity.
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð@C=@tÞðK, tÞ
loc ðK, tÞ  1 2 2 : ð72Þ
2 K ð@ C=@K2 ÞðK, tÞ generalized Laplace distributions are defined in lemma
2.3 as the distribution of a driftless diffusion sampled at
This function is displayed as a function of ln(K/F0) in an independent and exponentially distributed random
figures 3 and 4. time of mean 1. But this mean is completely arbitrary and
The function a(K, t)/K is also chosen proportional to can be set to any positive value m. In this case, all the
 loc and a natural concern is the order of magnitude one properties of these distributions remain valid provided
should choose for the function
2. Remember that that the associated function a2 is replaced by a2/m.

yWe would like to thank Jim Gatheral for communicating the SPX quotes and the corresponding SVI fits used in his book
(Gatheral 2006).
42 P. Carr and L. Cousot

1 0.25
0.2
0 0.15
−1.5 0.1 Time
−1
−0.5 0 0.05
0.5
1 0
Logstrike 1.5

Figure 5. Intensity corresponding to strategy  from the first to the fourth maturity.

5.6

5.5

5.4

5.3

5.2

5.1 1.8
1.6
1.4
5 1.2
−1.5 1
−1 0.8 Time
−0.5
0 0.6
0.5 0.4
Log−strike 1
1.5

Figure 6. Intensity corresponding to strategy  from the fourth to the eighth maturity.

Consequently, if one considers it realistic that a jump with Ndays(t) 2 [0, 30], say.
corresponds to the activity of the diffusion component Figures 5 and 6 display the intensity corresponding to
during a few days, but compressed in one instant, then the case where 2(t) ¼ 0.99 and Ndays(t) ¼ 2/3. With these
one should set parameters, we obtain an intensity  that is close to 10
times the constant intensity G ¼ 0.54 obtained by
Ndays ðtÞ 2 Gatheral (2006, p. 69) when fitting a stochastic volatility

2 ðtÞ ¼  ðtÞ, ð73Þ
365 model with jumps to the same data. Note that the
 A PDE approach to jump-diffusions 43

0.12
gnum at T2
gapp at T2
glin at T2
0.1
gnum at T6
gapp at T6
glin at T6
0.08

0.06

0.04

0.02

0
1200 1205 1210 1215 1220 1225 1230 1235 1240 1245 1250
Spot

Figure 7. At-the-money forward jump distributions and the corresponding approximations.

µ10−3
−2

−3

−4

−5

−6

−7

−8

−9

−10
1.5
1
0.5
0
Log−strike −0.5 0
0.4 0.2
−1 0.8 0.6
1.2 1
−1.5 1.6 1.4 Time
1.8

Figure 8. Means of the jumps in the log-price.

function  was floored by f ¼ 0.1 in figure 5 to avoid densities using the definition of generalized Laplace
negative values. However, this should not have a signif- distributions as the Green’s functions of a differential
icant impact on the goodness of fit, since the S&P 500 operator and the algorithm described by Robertson
index needs to double or fall by a half in a few days to (1971) to numerically solve linear second-order boundary
reach the floored region. value problem on infinite intervals.
The jump densities corresponding to the intensity  do One can also imagine analytical approximations. For
not have analytical expressions because of our specifica- instance, if the function a(y, t)/y is approximated by the
tion of the function a2. However, it is always possible to constant a(x, t)/x, then the jump density g(x, t, ) has an
obtain a numerical approximation, gnum, of the jump analytical approximation, denoted by glin. Another
44 P. Carr and L. Cousot

analytical approximation can also be obtained if

2 is
0.15
specified via equation (73). Indeed, if a jump corresponds
roughly to the activity of the local volatility model in few
0.125
days, even weeks, then we can approximate its distribu-
tions by using the small time expansion of the heat kernel
0.1
introduced by Varadhan (1967) and later refined by
Molchanov (1975). Appendix G shows heuristically that,
0.075
under this hypothesis,
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
R
aðx, tÞ pffiffi2j y ½du=aðu, tÞj 0.05
1.5
gðx, t, yÞ ’ gapp ðx, t, yÞ  e x : ð74Þ 1
2a3 ð y, tÞ 0.5
0
Log−strike −0.5
To assess the quality of these analytical approximations, −1 0.2 0
−1.5 1.2 1 0.8 0.6 0.4
1.8 1.6 1.4
we display in figure 7 the functions gnum, glin and gapp Time

when x is at-the-money forward. Unfortunately, we note

Figure 9. Standard deviations of the jumps in the log-price.
that both gapp and glin do not approximate gnum very well
as they tend not to take into account the downward-
sloping smile. Our main idea was that the calibration procedure and
Jump distributions are often taken constant in the log- the pricing procedure are significantly simplified when the
price. For instance, taking the same data as input, jump-distributions are the Green’s functions of linear
Gatheral (2006, p. 69) uses, in the log-price, a normal differential operators. We could have obtained similar
distribution with mean mG ¼ 0.12 and standard devia- results had we decided to sample, at independent expo-
tion  G ¼ 0.1. To gain intuition, we display these quan- nential times, solutions of first-order ordinary differential
tities for the generalized Laplace distributions respectively equations or diffusions time changed by Brownian times.
in figures 8 and 9. In the first case, we would have obtained probability
The means of the jumps in the log-price are most of the distributions that would have been the Green’s functions
time in the interval [0.002, 0.004]. Thus, they are of first-order linear differential operators. An example of
significantly less than mG. This could be changed by the use such distributions are the generalized exponential distri-
of non-centered generalized Laplace distributions (see butions introduced by Cousot (2008). In the second case,
remark 2). As far as their standard deviations are we would have obtained Green’s functions of fourth-order
concerned, they tend to stay in the interval [0.05, 0.075]. linear differential operators (see Burdzy 1993, Allouba and
Loosely speaking, this jump-diffusion model jumps Zheng 2001 and Allouba 2002 and references therein).
10 times more than the stochastic volatility model Moreover, notice that this recipe is not only valid for
with the jumps of Gatheral (2006) and the standard one-dimensional Markov processes. Indeed using jump
deviations of the jumps are not one-tenth of  G, but a half. distributions of this type for two-dimensional Markov
One can imagine that the standard deviations of the jumps stochastic volatility models with jumps could also allow
are of this order to compensate for the fact that the model one to reduce the Kolmogorov equations to PDEs.
is Markov and does not have stochastic volatility. Therefore, the valuation of American-style path-indepen-
dent claims in this framework would be much more
efficient. Moreover, one can also expect that the calibra-
tion procedures available for stochastic local volatility
6. Conclusion models (see Ren et al. 2007 and references therein) could
also be adapted to incorporate jumps.
In this paper, we have shown how using generalized Finally, another application of generalized Laplace
Laplace distributions as the jump distributions of a jump- distributions can be found in Carr and Cousot (2007),
diffusion leads to efficient calibration procedures. For where the authors use these very distributions as transition
instance, strategy  allowed us to compute the intensity densities of a discrete time Markov martingale and explain
of a calibrated jump-diffusion using the partial derivatives how to calibrate them to given implied volatility smiles.
of the implied volatility surface. This strategy can be seen
as a generalization of the closed-form formula obtained
by Carr et al. (2004) (see remark 4) and has the advantage Acknowledgements
of giving rise to numerical schemes that are much faster
than those described by Andersen and Andreasen (2000) We are grateful to Robert Kohn and to Jim Gatheral for
or in section 6 of Carr et al. (2004), where the intensity is insightful discussions and comments.
expressed as an inverse Fourier transform. As far as
pricing is concerned, the backward Kolmogorov PIDE
References
can be transformed into a PDE. A consequence of this is
that the pricing of path-independent claims should be Allouba, H., Brownian-time processes: The PDE connection II
much more efficient when using generalized Laplace and the corresponding Feynman–Kac formula. Trans. Am.
distributions rather than generic ones. Math. Soc., 2002, 354(11), 4627–4637.
 A PDE approach to jump-diffusions 45

Allouba, H. and Zheng, W., Brownian-time processes: The PDE Hartman, P., On the derivatives of solutions of linear, second-
connection and the half-derivative generator. Ann. Probab., order, ordinary differential equations. Am. J. Math., 1953,
2001, 29(4), 1780–1795. 75(1), 173–177.
Andersen, L. and Andreasen, J., Jump-diffusion processes: Hartman, P., Ordinary Differential Equations, 2nd ed., 1973
Volatility smile fitting and numerical methods for option (Wiley: New York).
pricing. Rev. Deriv. Res., 2000, 4, 231–262. Henry-Labordere, P., A general asymptotic implied volatility for
Bensoussan, A. and Lions, J., Impulse Control and stochastic volatility models. Preprint, Barclays Capital, 2005.
Quasi-variational Inequalities, 1984 (Gauthier-Villars). Kallsen, J. and Shiryaev, A., The cumulant process and Escher’s
Berestycki, H., Busca, J. and Florent, I., Asymptotics and change of measure. Finance Stochast., 2002, 6, 397–428.
calibration of local volatility models. Quant. Finance, 2002, Kellerer, H., Markov-komposition und eine anwendung auf
2(1), 61–69. martingale. Math. Ann., 1972, 198, 99–122.
Bichteler, K. and Jacod, J., Calcul de Malliavin pour les Komatsu, T., Markov processes associated to certain integro-
diffusions avec sauts: Existence d’une densité dans le cas differential operators. Osaka J. Math., 1973, 10, 271–303.
unidimensionnel. Se´minaire de probabilite´s de Strasbourg, Komatsu, T. and Takeuchi, A., On the smoothness of pdf of
1983, 17, 132–157. solutions to SDE of jump type. J. Differ. Eqns Applic., 2001,
Black, F. and Scholes, M., The pricing of options and corporate 2, 141–197.
liabilities. J. Polit. Econ., 1973, 81, 637–659. Lepingle, D. and Mémin, J., Sur l’intégrabilité uniforme des
Breeden, D. and Litzenberger, R., Prices of state-contingent martingales exponentielles. Z. Wahrscheinlichkeitstheorie
claims implicit in options prices. J. Business, 1978, 51, Verwandte Gebiete, 1978, 42, 175–203.
621–651. Madan, D. and Yor, M., Making Markov martingales meet
Buehler, H., Expensive martingales. Quant. Finance, 2006, 6(3), marginals. Bernouilli, 2002, 8, 509–536.
207–218. McKean, H., Elementary solutions for certain parabolic partial
Burdzy, K., Some path properties of iterated Brownian motion. differential equations. Trans. Am. Math. Soc., 1956, 82(2),
Seminar on Stochastic Processes 1992, pp. 67–87, 1993 519–548.
(Birkhäuser). McKean, H., Stochastic Integrals, 1969 (Academic Press:
Carr, P. and Cousot, L., Explicit constructions of martingales New York).
calibrated to given implied volatility smiles. Preprint, Courant Medvedev, A. and Scaillet, O., Approximation and calibration
Institute of Mathematical Sciences, New York University, of short-term implied volatilities under jump-diffusion sto-
2007. chastic volatility. Rev. Financial Stud., 2007, 20(2), 427–459.
Carr, P., Geman, H., Madan, D. and Yor, M., From local Molchanov, S., Diffusion processes and Riemanian geometry.
volatility to local Lévy models. Quant. Finance, 2004, 4(5), Russ. Math. Surv., 1975, 30(1), 1–63.
581–588. Protter, P., Stochastic Integration and Differential Equations,
Carr, P. and Madan, D., A note on sufficient conditions for no 2nd ed., 2004 (Springer: Berlin).
arbitrage. Finance Lett., 2005, 2, 125–130. Ren, Y., Madan, D. and Qian-Qian, M., Calibrating and pricing
Cass, T., Smoothness of densities for solutions to stochastic with embedded local volatility models. Risk Mag., 2007,
differential equations with jumps. Preprint, University of 20(9), 138–143.
Cambridge, 2006. Revuz, D. and Yor, M., Continuous Martingales and Brownian
Cousot, L., When can given European call prices be met by a Motion. Grundlehren der Mathematischen Wissenschaften,
martingale? An answer based on the building of a Markov 3rd ed., Vol. 293, 1999 (Springer: Berlin).
chain model. Preprint, New York University, 2005. Robertson, T., The linear two-point boundary-value problem on
Cousot, L., Conditions on option prices for absence of arbitrage an infinite interval. Math. Comput., 1971, 25(115), 475–481.
and exact calibration. J. Bank. Finance, 2007, 31(11), Schröder, M., Computing the constant elasticity of variance
3377–3397. formula. J. Finance, 1989, 44, 211–219.
Cousot, L., Constructions of martingales and of increasing Stineman, R., A consistently well-behaved method of interpola-
processes with constrained marginal distributions. PhD thesis, tion. Creat. Comput., 1980, 6(7), 54–57.
Courant Institute of Mathematical Sciences, New York Stroock, D., Diffusion processes associated with Lévy gen-
University, 2008. erators. Z. Wahrscheinlichkeitstheorie Verwandte Gebiete,
Cox, A. and Hobson, D., Local martingales, bubbles and option 1975, 32, 209–244.
prices. Finance Stochast., 2005, 9(4), 477–492. Stroock, D., The Malliavin calculus and its application to
Cox, D., Notes on option pricing I: Constant elasticity of second-order parabolic differential equations: Part I. Theory
variance diffusions. Preprint, Stanford University, 1975.
Comput. Syst., 1981a, 14(1), 25–65.
Davis, M. and Hobson, D., The range of traded option prices.
Stroock, D., The Malliavin calculus and its application to
Math. Finance, 2007, 17(1), 1–14.
second-order parabolic differential equations: Part II. Theory
Delbaen, F. and Shirakawa, H., A note of option pricing for
Comput. Syst., 1981b, 14(1), 141–171.
constant elasticity of variance model. Asian-Pacif. Financial
Stroock, D. and Varadhan, S., Diffusion processes with
Mkts, 2002, 92, 85–99.
continuous coefficients I. Commun. Pure Appl. Math.,
Doléans-Dade, C., Quelques applications de la formule de
1969a, 22, 345–400.
changement de variables pour les semi-martingales. Z.
Stroock, D. and Varadhan, S., Diffusion processes with
Wahrscheinlichkeitstheorie Verwandte Gebiete, 1970, 16,
continuous coefficients II. Commun. Pure Appl. Math.,
181–194.
Dupire, B., Pricing with a smile. Risk Mag., 1994, 7(1), 18–20. 1969b, 22, 479–530.
Durrleman, V., Convergence of at-the-money implied volatilities Takeuchi, A., The Malliavin calculus for SDE with jumps and
to the spot volatility. Preprint, École Polytechnique, 2007. the partially hypoelliptic problem. Osaka J. Math., 2002,
Engelbert, H. and Schmidt, W., One-dimensional stochastic 39(3), 523–559.
differential equations with generalized drift. In Lecture Notes Varadhan, S., Diffusion processes in a small time interval.
in Control and Information Sciences, pp. 143–155, 1984 Commun. Pure Appl. Math., 1967, 10, 659–685.
(Springer: Berlin).
Feller, W., The parabolic differential equations and the Appendix A: Proof of proposition 2.1
associated semi-groups of transformations. Ann. Math.,
1952, 55(3), 468–519.
Gatheral, J., The Volatility Surface: A Practitioner’s Guide, 2006 The results of Stroock (1975) cannot be applied directly
(Wiley: New York). to prove the existence of (Ft), since this process needs
46 P. Carr and L. Cousot

to be positive. However, they imply that the martingale Moreover, using lemma 2.6 of Kallsen and Shiryaev
problem associated with the generator Ht, defined by e t with
(2002), we have Ft ¼ EðXÞ
X
1
Ht hðxÞ  mðx, tÞh0 ðxÞ þ s2 ðx, tÞh00 ðxÞ Xet  Xc þ 1 hXic þ ðexpðDXs Þ  1Þ, ðB1Þ
t t
2 2 s4t
Z  
u 0 since the number of jumps of (Xs) before t is almost surely
þ hðx þ uÞ  hðxÞ  h ðxÞ
R 1 þ u2 finite.
 l ðx, tÞ gðx, t, x þ uÞdu, Step 2: We prove that (Ft) is a local martingale by
for h 2 C1
0 ðRÞ, where showing that ðX~ t Þ is a martingale. Note that equation (B1)
sðx, tÞ  ðexpðxÞ, tÞ, can be transformed into
Z Z !
t X t
1 et 
X sðXs , sÞdWs þ ðexpðDXs Þ  1Þ þ mðXs ,sÞds ,
mðx, tÞ  ðexpðxÞ, tÞ   2 ðexpðxÞ, tÞ þ ðexpðxÞ, tÞ
Z 2 
0 s4t 0
u expðx þ uÞ where (Wt) is a Brownian motion. Since s is bounded, the
 ðexpðxÞ, t, expðx þ uÞÞdu ,
R 1 þ u2 stochastic integral is a martingale (we refer, for example,
l ðx, tÞ  ðexpðxÞ, tÞ, to corollary 3, p. 73, of Protter 2004). Furthermore, it is
fairly easy to show that the expression between paren-
gðx, t, yÞ  expð yÞðexpðxÞ, t, expð yÞÞ,
theses is also a martingale since the jumps are compen-
is well-posed and induces a càdlàg strong Markov process sated (by m), the expectation of the number of jumps in a
(Xt), whose generator is Ht. Finally, defining Ft  exp (Xt) finite interval is bounded (since  is bounded) and
gives the desired conclusion since simple computations the absolute value of the jumps is also bounded (see

show that, for f 2 C10 ðRþ Þ, equation (2)). Indeed, note that
Z
Gt f ðxÞ ¼ Ht hðlnðxÞÞ,
ðexpð y  xÞ  1Þ gðx, s, yÞdy
where h(x)  f (exp(x)). R
Z  0 
y gðlnðx0 Þ, s, lnð y0 ÞÞ 0
¼ 0
 1 dy
R x y0
Z  
Appendix B: Proof of proposition 2.2 y0
¼ 0
 1 ðx0 , s, y0 Þdy0 ,
R x
For this proof, we use the extensive literature concerning
with x ¼ ln(x0 ) and y ¼ ln(y0 ). The main consequence
positive local martingales and notably stochastic expo-
is that (Ft) is a local martingale—see, for example,
nentials, also known as Doléans-Dade exponentials
Protter (2004).
(Doléans-Dade 1970). For a local martingale (Mt), the
stochastic exponential is defined by Step 3: We show that (Ft) is a martingale if we further
 Y assume condition (4). Let us denote by (At) the process
1
EðMÞt  exp Mt  M0  hMict ð1 þ DMs ÞexpðDMs Þ,
2 1 ec X
s4t
At  hX it þ es Þ lnð1 þ DX
½ð1 þ DX es Þ  DX
es , ðB2Þ
where hMic denotes the continuous part of the quadratic 2 s4t
variation of (Mt) and DMs  Ms  Ms . This definition ec is the continuous part of X, e and by (Ct) its
where X
generalizes that of McKean (1969), which defined sto-
compensator. If E[Ct]5þ1 for t40, then (Ft) is a
chastic exponentials for Brownian stochastic integrals,
martingale using theorem 3.1 of Lepingle and Mémin
and that of Stroock and Varadhan (1969a,b), which es ¼ ðexpðDXs Þ  1Þ, we have
(1978). Since DX
introduced it for continuous martingales. Zt
1 X
Step 1: We express (Ft) as a stochastic exponential. At ¼ s2 ðXs , sÞds þ ½expðDXs ÞDXs  expðDXs Þ þ 1,
2 0 s4t
First, without loss of generality, we assume that F0 ¼ 1.
Using the proof of proposition 2.1, we know that Ft ¼ exp and its compensator can be written as
(Xt), where X0 ¼ 0 and the generator of (Xt) is, for Z
1 t 2
h 2 C10 ðRÞ,
Ct  s ðXs , sÞds
  2 0
1 2 1 Zt Z
Ht hðxÞ  mðx, tÞ  s ðx, tÞ h0 ðxÞ þ s2 ðx, tÞh00 ðxÞ þ l ðXs , sÞ ½expð y  Xs Þð y  Xs Þ
2 2
Z 0 R

þ ðhð yÞ  hðxÞÞl ðx, tÞ gðx, t, yÞdy,
R
 expð y  Xs Þ þ 1 gðXs , s, yÞdy ds:
  ðB3Þ
and
Since  and  are bounded,
mðx, tÞ  ðexpðxÞ, tÞ, Z t Z
sup  2
sðx, tÞ  ðexpðxÞ, tÞ, Ct 4 t þ ðsup Þ ½expð y  Xs Þ½ð y  Xs Þ  1 þ 1
2 0 R
l ðx, tÞ  ðexpðxÞ, tÞ, 
 gðXs , s, yÞdy ds: ðB4Þ
gðx, t, yÞ  expð yÞðexpðxÞ, t, expð yÞÞ:
 A PDE approach to jump-diffusions 47

Let us try to bound the following quantity uniformly in If we further assume conditions (18) and (19), then
s and x: 0 and þ1 are natural boundaries according to Feller’s
Z classification. Consequently, we have in this case
ðexpð y  xÞ½ð y  xÞ  1 þ 1Þ gðx, s, yÞdy u1 ðxÞ ! 0, ðC6Þ
R x!0
Z þ1  0  0 
y y
¼ ln ðx0 , s, y0 Þdy0 4B, u2 ðxÞ ! 0, ðC7Þ
0 x0 x0 x!þ1

with x ¼ ln(x0 ) and y ¼ ln(y0 ), using equation (4). and equation (20) follows.
Therefore, To prove equation (21), we obtain an estimate of u02 in
  terms of u2 using theorem 1 of Hartman (1953). Indeed, u2
sup  2 is positive, decreasing and satisfies equation (C2);
Ct 4 þ ðsup ÞB t, ðB5Þ
2 therefore,
x
and the conclusion follows. jxu02 ðxÞj48u2 , ðC8Þ
2
for x big enough and the conclusion follows, since the
equivalent result for u1 is a consequence of equation (C3).
Finally, to prove that the mean of y ! g(x, y) is x,
Appendix C: Proof of lemma 2.3 we multiply equation (16) by y, integrate over Rþ to
obtain
Z Z
The results of lemma 2.3 are mostly a consequence of the @2 G
well-known boundary classification due to Feller (1952) ygðx, yÞdy  y 2 ðx, yÞdy ¼ x,
Rþ Rþ @y
(see p. 487, in particular). Indeed, condition (12) ensures
that 0 is either a natural or an entrance boundary. and use the boundary conditions of equations (20) and
Moreover, since a2 is positive and continuous, we have (21) to show that the second integral is zero by using two
Z þ1 Z x  integrations by parts.
dy
Jþ1  2
dx ¼ þ1, ðC1Þ
1 1 a ð yÞ

which also ensures that þ1 is either a natural or an Appendix D: Proof of lemma 2.4
entrance boundary. Consequently, as proved by Feller
This proof has four main steps.
(1952, pp. 482–493) and summed up by McKean (1956,
p. 523), there exists two twice continuously differentiable Step 1: We prove that g is well-defined. This is actually a
solutions u1 and u2 of simple consequence of the fact that equation (22) implies
 equations (12), (18) and (19).
a2 ðxÞ @2 R
I f ðxÞ ¼ 0, ðC2Þ Step 2: We prove that 2
2 @x2 {ln(y/x)/[1 þ (ln(y/x)) ]}
g(x, t, y)dy is continuous in x and t for a given Borel set
that satisfy of Rþ , . With the proof of lemma 2.3 in mind (see
u01 ðxÞ ! 0, ðC3Þ appendix C), for each t 2 Rþ, there exists a positive
x!0 increasing (respectively decreasing) function u1,t (respec-
tively u2,t), the solution of
u02 ðxÞ ! 0: ðC4Þ 
x!þ1 a2 ðx, tÞ @2
I f ðxÞ ¼ 0, ðD1Þ
Moreover, u1 (respectively u2) is positive, convex and 2 @x2
increasing (respectively decreasing). Besides, u1 and u2 are such that
independent solutions and their Wronskian (denoted a2 ðy,tÞ u1,t ðminðx,yÞÞu2,t ðmaxðx,yÞÞ
W[ui]) is a positive constant. Consequently, it allows us Gðx, t,yÞ ¼ gðx,t,yÞ ¼ ,
2 W½ui,t 
to define
where W[ui,t] denotes the Wronskian of u1,t and u2,t.
u1 ðminðx, yÞÞu2 ðmaxðx, yÞÞ Moreover, it is possible to find versions of (u1,t(.)) and
Gðx, yÞ  , ðC5Þ
W½ui  (u2,t(.)), such that u1,.(.), u01,: ð:Þ, u2,.(.) and u02,: ð:Þ are
and all the properties of G follow easily from this continuous in x and t. Indeed, using the proof of
definition. corollary 6.4 of Hartman (1973, p. 357) for a given t,
Now let us focus on the properties of g. Equation (16) is u1,t and u2,t can be constructed, in a smooth way, from the
a direct consequence of the definition of g and of equation solution of the initial value problem

(13). Equation (17) is proved by using the definition of g a2 ðx, tÞ @2
(equation (13)) and the symmetry property of G (equation I f ðx, tÞ ¼ 0,
2 @x2
(15)). To prove that g integrates to 1, we just need to
f ð1, tÞ ¼ 1,
integrate equation (16) over Rþ and use the boundary
conditions of equation (14). f 0 ð1, tÞ ¼ 1,
48 P. Carr and L. Cousot

which is continuous in x and t, since a2 is positive and since u2,t (respectively u2,tn ) is decreasing in x. Likewise,
continuous (see theorem 2.1 of Hartman 1973, p. 94).  
To prove the continuity, let us now consider a max½mt , Mt  u1,t ðMx Þ 0
A
,n 4  ðu2,t ð
Þ þ u02,tn ð
ÞÞ:
sequence (xn, tn) that goes to a given (x, t) when n goes min½mt , Mt  W½ui,t 
to þ1, and prove that the following quantity goes to 0: Consequently, by taking  (respectively
) small (respec-
Z Z  tively large) enough, it is possible to make A,n (respec-
 lnðy=xn Þ lnðy=xÞ 
 gðx ,t , yÞdy  gðx,t, yÞdy
 1 þ ðlnðy=x ÞÞ2 n n 2  tively A
,n) as small as desired in the limit n ! þ1, since
 n  1 þ ðlnðy=xÞÞ
Z þ1   u01,t ðxÞ !x!0 0 (respectively u02,t ðxÞ !x!þ1 0) and
lnðy=xn Þ  
4 gðxn , tn , yÞ  gðx,t,yÞdy u01,tn ðÞ !n!þ1 u01,t ðÞ (respectively u02,tn ð
Þ !n!þ1
2
0 1 þ ðlnðy=xn ÞÞ 0
u2,t ð
Þ).
Z þ1  
 lnðy=xn Þ lnðy=xÞ  Finally, once  and
are fixed, A,
,n ! þ10 thanks

þ 1 þ ðlnðy=x ÞÞ2 1 þ ðlnðy=xÞÞ2  gðx,t,yÞdy

to the bounded convergence theorem, since g is contin-
0 n
Z uous on the compact set [mx, Mx]  [mt, Mt]  [,
].
1 þ1  
4 gðxn ,tn , yÞ  gðx, t,yÞdy
2 0 Step 3: We prove that condition (2) is satisfied. Using
Z þ1  
 lnðy=xn Þ lnðy=xÞ  equation (16),

þ 1 þ ðlnðy=x ÞÞ2 1 þ ðlnðy=xÞÞ2  gðx,t,yÞdy

Z
0 n
1 j y  xj gðx, t, yÞdy
 An þ Bn : Rþ
2 Z 2 
@ G
Using the dominated convergence theorem, Bn goes to 0. ¼ ðx, t, yÞ þ ðx  yÞ j y  xjdy
2
Indeed, Rþ @y
Z
  @2 G
 lnðy=xn Þ lnðy=xÞ  ¼ ðx, t, yÞj y  xjdy:
 1  2
1 þ ðlnðy=x ÞÞ2 1 þ ðlnðy=xÞÞ2  gðx, t, yÞ4gðx,t, yÞ 2 L ðRþ Þ:
 Rþ @y
n

As far as An is concerned, we have, with Moreover, using integration by parts, equations (20) and
0 4 5mx  min((xn), x) 4 Mx  max((xn), x)5
5þ1, (21), we have
Z þ1 2  þ1
Z þ1 @ G @G
ðx, t, yÞðy  xÞdy ¼ ðx, t, yÞðy  xÞ  Gðx,t,yÞ
j gðxn , tn , yÞ  gðx, t, yÞjdy x @y2 @y x
0
Z ¼ Gðx,t, xÞ:
4 ð gðxn , tn , yÞ þ gðx, t, yÞÞdy
0 Likewise,
Z

Z x
þ j gðxn , tn , yÞ  gðx, t, yÞjdy
 ð@2 G=@y2 Þðx, t, yÞð y  xÞdy ¼ Gðx, t, xÞ:
Z þ1 0
þ ð gðxn , tn , yÞ þ gðx, t, yÞÞdy Therefore, to sum up, we have

Z
 A,n þ A,
,n þ A
,n :
gðx, t, yÞj y  xjdy ¼ 2Gðx, t, xÞ: ðD2Þ
Rþ
Moreover, if u 5  and s 5 0, we have
Z Z Moreover, by definition,

u2,s ðuÞ 
2u1,s ð yÞ u2,s ðuÞu01,s ðÞ
gðu, s, yÞdy ¼ dy ¼ , u1,t ðxÞu2,t ðxÞ
0 W½ui,s  0 a2 ð y, sÞ W½ui,s  Gðx, t, xÞ 
u1,t ðxÞu2,t ðxÞ  u1,t ðxÞu02,t ðxÞ
0

and if u 4
and s 5 0, 1
¼ ,
Z Z ju01,t ðxÞ=u1,t ðxÞj þ ju02,t ðxÞ=u2,t ðxÞj
þ1
2u2,s ðyÞ
u1,s ðuÞ þ1 u1,s ðuÞu02,s ð
Þ
gðu,s,yÞdy ¼ dy ¼  ,


a2 ðy,sÞ
W½ui,s  W½ui,s  since u1,t (respectively u2,t) is positive and increasing
(respectively decreasing).
since u01,s ðxÞ !0 0, u02,s ðxÞ !þ1 0 and using equation The growth condition on a2 (equation (22)) allows us
(D1). Therefore, with mt  min((tn), t) and Mt  to control this quantity. Indeed, two independent solu-
max((tn), t), we obtain tions of the system

u2,t ðxÞu01,t ðÞ u2,tn ðxn Þu01,tn ðÞ M2 x2 @2
A,n 4 þ I f ðxÞ ¼ 0 ðD3Þ
W½ui,t  W½ui,tn  2 @x2
u2,t ðmx Þu01,t ðÞ u2,tn ðmx Þu01,tn ðÞ are
4 þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
W½ui,t  W½ui,tn 
   1 þ ð8=M2 Þ 1
max½mt , Mt  u2,t ðmx Þ v ðxÞ ¼ x , with  ¼ ,
4 ðu01,t ðÞ þ u01,tn ðÞÞ, 2
min½mt , Mt  W½ui,t  and, using equation (22) and corollary 6.5 of Hartman
 A PDE approach to jump-diffusions 49

(1973, p. 359) we have Consequently,

Z þ1   

u02,t ðxÞ  þ u01,t ðxÞ y y 

4 4 4 : ðD4Þ ln  gðx, s, yÞdy
u2,t ðxÞ x x u1,t ðxÞ 0 x x
Z þ1 Z1 
As a consequence, 2Gðx, t, xÞ ðlnðvÞÞ
ðlnð1=vÞÞ

4 dv þ dv
1 m2 x 1 v1 0 v1þ
Gðx, t, xÞ4Ax, with A  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi : ðD5Þ Z 1 Z 
1 þ ð8=M2 Þ 2A ðlnðvÞÞ
þ1
ðlnð1=vÞÞ

4 2 1þ
dv þ dv ,
m 0 v 1 v1
Therefore, using equation (D2),
ðD7Þ
Z
gðx, t, yÞj y  xjdy 4 2Ax, ðD6Þ using equation (D5). Since þ40 and 50, both
Rþ integrals on the RHS are finite.
and the conclusion readily follows since g(x, t, .) is a
Appendix E: Proof of proposition 3.1
probability density.
Step 4: We prove that condition (4) is satisfied. For The existence and the properties of (Ft) are mostly the

40, consequence of a2,  and  satisfying the assumptions of

Z þ1   
propositions 2.1 and 2.2 and lemma 2.4. The only
y y 
ln  gðx, t, yÞdy exception is the fact that (Ft) is calibrated to the call
0 x x price function C, i.e. equation (49). Since equation (44)
Z x   
has been obtained by applying the differential operator
y x
¼ ln gðx, t, yÞdy [I  (a2(x, t)/2)@2/@x2] to the forward Kolmogorov equa-
0 x y
tion, let us do some reverse engineering to show that the
Z þ1   

y y function pMarket(x, t) ¼ (@2C/@x2)(x, t) is a probability

þ ln gðx, t, yÞdy
x x x density for each t40 that satisfies the forward PIDE
Z x   
associated with the generator Gt with initial condition
y x 2 u1,t ð yÞu2,t ðxÞ pMarket(x, t) ¼ (x  x0).
¼ ln dy
0 x y a 2 ð y, tÞ W½ui,t  The definition of the function a2 (equation (44)) can be
Z þ1   
rewritten as follows:
y y 2 u1,t ðxÞu2,t ð yÞ  
þ ln 2 ð y, tÞ
dy ð@C=@tÞðx, tÞ  12 x2  2 ðx,tÞð@2 C=@x2 Þðx, tÞ
x x x a W½ui,t 
ða2 ðx, tÞ=2Þ
Z x   

y x 1 u1,t ð yÞ   
¼ 2Gðx, t, xÞ ln 2 ð y, tÞ u ðxÞ
dy @2 a2 ðx, tÞ ð@C=@tÞðx, tÞ  12 x2  2 ðx,tÞð@2 C=@x2 Þðx,tÞ
0 x y a 1,t  2
@x 2 ða2 x, t=2Þ
Z þ1   
!
y y 1 u2,s ð yÞ @2 C
þ ln dy : ¼ ðx,tÞ ðx, tÞ:
x x x a2 ð y, tÞ u2,s ðxÞ @x2
Moreover, the boundary value problem [I  (@2/
We already know how to control G(x, t, x) using equation @x2)a2(x, t)/2] f (x) ¼ 0 with the limit condition
(D5). As far as the term between brackets is concerned, f (x)(a2(x, t)/2) ! y!0 or þ1 0 has a unique solution
we have, for x 4 y, using lemma 2.3. Therefore,
u2,t ð yÞ y yþ u1,t ð yÞ @C 1 @2 C
4 4 4 ð y, tÞ  y2  2 ð y, tÞ 2 ð y, tÞ
u2,t ðxÞ x x u1,t ðxÞ @t 2 @y
Z 2  2 
@ C a ð y, tÞ
by integrating equation (D4). Therefore, we obtain, using ¼ ðx, tÞðx, tÞ gðx, t, yÞ dx,
equation (23), @x2 2
Z þ1   
since  is bounded and using the limit conditions
y y 1 u2,t ð yÞ
ln dy concerning @C/@t and @2C/@x2 of equation (38).
x x 2
a ð y, tÞ u2,t ðxÞ
x
Z þ1   
Differentiating twice in y, we obtain
y 1 y1þ  
4 ln dy @pMarket @2 1
x x a2 ð y, tÞ x ðy,tÞ  2 pMarket ðy,tÞ y2  2 ð y, tÞ
Z þ1   
@t @y 2
y 1 y1þ Z þ1  2 
4 ln dy 2
@ a ð y, tÞ
x x m 2 y2 x ¼ pMarket ðx,tÞðx,tÞ 2 gðx,t,yÞ dx:
Z þ1 @y 2
1 ðlnðvÞÞ
0
Z þ1
¼ 2 dv,
m x 1 v1 ¼ pMarket ðx,tÞðx,tÞðgðx,t, yÞ  ðx  yÞÞdx
0
and, likewise, Z þ1
Z x   
Z1 ¼ pMarket ðx,tÞðx,tÞgðx,t,yÞdx  pMarket ðy,tÞð y, tÞ,
y x 1 u1,t ð yÞ 1 ðlnð1=vÞÞ
0
ln 2
dy4 dv:
0 x y a ð y, tÞ u1,t ðxÞ m x 0 v1þ
2
using equation (16).
50 P. Carr and L. Cousot

On the other hand, if we denote by pModel(, t) the    

1 @d 
density of Ft, then it is possible to show, by integration by 4 sup s þ sup ssy kd ð, tÞkL2  ð, tÞ
2 

parts and interchange of integrals, that pModel also 2 @y L2

2
2
satisfies the above equation, since the forward inf s @d 
 ð, tÞ
 : ðE3Þ
Kolmogorov equation (equation (25)) is satisfied for all 2 @y L2

P 2 C1 0 ðRþ Þ. Moreover, both the functions pMarket and
pModel have the same initial condition. As a consequence, As far as the jump part is concerned, we have
if we succeed in showing that their difference is equal to Z þ1 
 
0 for all t 5 0, the conclusion will follow. Therefore,  ½d ð y, tÞ  d ðx, tÞðx, t, yÞdy

we consider their difference, but in the log-price 1
for convenience: Z þ1 Z y 
 @d 
4  ðu, tÞdu ðx, t, yÞdy
d ð y, tÞ  pModel ðey , tÞey  pMarket ðey , tÞey ,  @u 
1 x
which satisfies vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 
Z þ1 u uZ y @d2 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
   t   
@d @ 1 4 ðu, tÞdu  y  xðx, t, yÞdy
ð y, tÞ ¼  d ð y, tÞ  s2 ð y, tÞ 1  x @u 
@t @y 2
  Z
1 @2 @d  þ1 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
þ ðd ð y, tÞs2 ð y, tÞÞ 4 ð, tÞ
 j y  xjðx, t, yÞdy:
2 @y2 @x 
Z þ1 L2 1

þ d ðx, tÞl ðx, tÞðx, t, yÞdx  d ð y, tÞl ð y, tÞ, Moreover, with x0 ¼ ex and y0 ¼ ey,
1
Z þ1 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðE1Þ
j y  xjðx, t, yÞdy
d ð y, 0Þ ¼ 0, ðE2Þ 1
Z þ1 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 0 ffi
x y x y  
with l(x, t)  (e , t), (x, t, y)  e g(e , t, e ) and s(y, t) 
¼ ln y gðx0 , t, y0 Þdy0 4A 5 þ1,
(ey, t). Note that d(, t), (@d/@x)(, t) and l(, t) belong to  x0 
0
L2(R) thanks to the integral conditions of equations (38),
(43) and (48). using step 4 of the proof of lemma 2.4 in appendix D with
If we multiply equation (E1) by d(y, t) and integrate
¼ 1/2. Therefore,
over R, we obtain Z þ1 Z þ1 
Z þ1 
1@ 2
d ðy,tÞ dðx, tÞl ðx, tÞðx, t, yÞdx dy
d ð y, tÞdy 1 1
2 @t 1 Z þ1
2
Z þ1    d ðy, tÞl ð y, tÞdy
@ 1 2 1
¼ d ð y, tÞ s ð y, tÞd ð y, tÞ dy Z þ1  Z þ1  
@y 2  
4 ½dð y, tÞ  d ðx, tÞðx, t, yÞdy dx
1
  d ðx, tÞ l ðx, tÞ
Z þ1
@2 1 2 1
  1
þ d ð y, tÞ 2 s ð y, tÞd ð y, tÞ dy @d 
1 @y 2 4Akl ð, tÞkL2  
@x ð,tÞ kdð, tÞkL2 :
Z þ1 Z þ1  L2

þ d ð y, tÞ d ðx, tÞl ðx, tÞðx, t, yÞdx dy To sum up, we have

1 1
Z    2
þ1
1@   @d 
 d2 ð y, tÞl ð y, tÞdy: e@d ð, tÞ kd ð, tÞk  e
ðkd ð, tÞk2L2 Þ4A B  ð, tÞ  ,
2 @t @x  L2 @x 
1 L2 L2
Let us assume for now that d has compact support. Then with e ¼ A sup kl ð, tÞk þ ð1 sup s2 þ sup ssx Þ
A and
L2 2
the diffusion and drift parts can be bounded using
e 2 e e
B ¼ inf s =2. Note that 04A, B 5 þ1 and B 4 0 using e
integration by parts as follows:
equations (40), (41) and (43).
Z þ1  
1 2 @d Therefore, applying Young’s inequality, we obtain
d ð y, tÞ s ð y, tÞ ð y, tÞ dy for "40
1 2 @y
Z þ1    2 !
1 2 @2 d 1@ e 1
A @d 
þ d ð y, tÞ s ð y, tÞ 2 ð y, tÞ dy 2
ðkd ð, tÞkL2 Þ4 kd ð, tÞkL2 þ "
2
ð, tÞ
1 2 @y 2 @t 2 " @x 
L2
Z þ1    2
1 @d @d 
¼ d ð y, tÞ s2 ð y, tÞ ð y, tÞ dy eB 
1 2 @y @x ð, tÞ :
L2
Z þ1  2
1 2 @d
 s ð y, tÞ ð y, tÞ dy As a consequence, by taking " small enough, we obtain
2 1 @y
Z þ1  2
@d @ @d 
2 b 2 b
ðkd ð, tÞkL2 Þ4Akd ð, tÞkL2  B ð, tÞ

 ðsð y, tÞsy ð y, tÞÞd ð y, tÞ ð y, tÞdy @t @x  ,
1 @y L2
 A PDE approach to jump-diffusions 51

with b
0 5 A, b
B 5 þ1. This leads to these expansions are derived for t ! 0 at a fixed strike and
2 2
kd ð, tÞkL2 4kd ð, 0ÞkL2 ¼ 0 using Gronwall’s lemmay are therefore different from those described by Medvedev
and the conclusion follows since, by density, this inequal- and Scaillet (2007) for stochastic volatility models with
1
ity is not only true for d 2 C10 ðRÞ, but for any d 2 H (R). jumps. Indeed, in Medvedev and Scaillet (2007), the
expansions are derived for a strike that goes to the initial
value of the asset when t goes to 0.
Appendix F: Approximation of generalized Laplace The process of the underlying asset is assumed to be a
distributions for small diffusion coefficients generic positive jump-diffusion and a martingale. With
proposition 2.2 in mind, its generator can be written as
In this appendix, we derive an approximation of the
1
generalized Laplace densities when the diffusion coeffi- Gt fðxÞ ¼ x2  2 ðx,tÞf 00 ðxÞ þ ðx,tÞ
2 
cient is small. Remember that the generalized Laplace Z þ1
distribution g(x, .) associated with the function a2 is  ðx,t,yÞðf ðyÞ  fðxÞ  f 0 ðxÞðy  xÞÞdy ,
defined as the density of (S ), where (St) is a weak solution 0
of the following stochastic differential equation: ðG1Þ
dSt ¼ aðSt ÞdWt , ðF1Þ and its initial value is denoted by x0.
Our strategy to obtain a small time expansion is pretty
S0 ¼ x, ðF2Þ simple: since we know the infinitesimal generator of the
jump-diffusion, we can guess an expansion in small time
and  is an independent and exponentially distributed of its call price function that we will match with an
random time of mean 1. To obtain an approximation of expansion in small time of the call price function in the
this density when the function a2 is small, one can resort Black–Scholes model (Black and Scholes 1973). Let us
to the small time expansion of the heat kernel described denote the price of a call struck at x and expiring at t in
by Varadhan (1967) and perfected notably by Molchanov the jump-diffusion (respectively Black–Scholes) model by
(1975). Indeed, if we denote by p(x,, t) the marginal CJD(x, t) (respectively CBS(x, t)). Since the generator is
density of (St), we have defined as the ‘derivative’ of the transition function, we
sffiffiffiffiffiffiffiffiffi
obtain for the jump-diffusion model
1 aðxÞ d2 ðx, yÞ=2t
pðx, y, tÞ ’ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi e , ðF3Þ
2
2pa ð yÞt að yÞ CJD ðx, tÞ  ðx0  xÞþ
Ry t
where d ðx, yÞ  j x ðdu=aðuÞÞj is the geodesic distance of Z þ1
1 2 2
the metric associated with the above SDE. Therefore, we ¼ x0  ðx0 , 0Þðx0  xÞ þ ðx0 , 0Þ ðx0 , 0, yÞðð y  xÞþ
have, heuristically, 2 0
!
Z þ1
gðx, yÞ ¼ et pðx, y, tÞdt ðF4Þ  ðx0  xÞþ  Hðx0  xÞð y  x0 ÞÞdy þ OðtÞ, ðG2Þ
0

Z sffiffiffiffiffiffiffiffiffi where H is the Heaviside function. Denoting by  (x0, )

þ1
1 aðxÞ d2 ðx, yÞ=2t the double tail associated with (x0, 0, .), i.e.
’ et pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi e dt ðF5Þ
0 2
2pa ð yÞt að yÞ
 ðx0 , xÞ
sffiffiffiffiffiffiffiffiffiffiffiffiffiffi 8 R þ1 R þ1
> ð u ðx0 , 0, vÞdvÞdu, if x 4 x0 ,
aðxÞ pffiffi2d ðx, yÞ >
>
> x
¼ e : ðF6Þ >
> R R þ1
2a3 ð yÞ <1ð þ1
ðx0 , 0, vÞdvÞdu
2 x0 ð u
¼ R R
> þ 0 ð ðx , 0, vÞdvÞduÞ,
> x u
if x ¼ x0 ,
One can check that this approximation is consistent with >
> 0 0 0
>
>
the analytical formula for g when a(x) ¼ Ax and A is :Rx Ru
small. Finally, note that a more precise expansion could 0 ð 0 ðx0 , 0, vÞdvÞdu, if x 5 x0 ,
be derived easily since the second-order term in the we obtain by put–call parity
expansion of p is known. We refer to Henry-Labordere
(2005) and references therein. CJD ðx, tÞ ¼ ðx0  xÞþ
 
1 2 2
þ x0  ðx0 , 0Þðx0  xÞ þ ðx0 , 0Þ ðx0 , xÞ t
2
Appendix G: Small time expansion of the smile of þ Oðt2 Þ: ðG3Þ
a generic jump-diffusion
A rigorous reader would definitely object to the fact that
In this appendix, we derive heuristically short time the infinitesimal generator can be used safely only with
asymptotics of the implied volatility, when the underlying infinitely differentiable functions with compact
process is a generic positive jump-diffusion. Note that support and that a call payoff does not have any

ySee theorem 1.1 of Hartman (1973, p. 24).

52 P. Carr and L. Cousot

of these properties. But remember that, here, all argu- In the case  2(x0, 0) ¼ 0, we would like to obtain a more
ments are heuristic. precise expansion. This can be done by equating equa-
As far as the Black–Scholes model is concerned, we tions (G3) and (G10):
obtain similarly pffiffiffiffiffiffi !
pffiffi 2pðx0 , 0Þ ðx0 , x0 Þ
CBS ðx,tÞ  ðx0  xÞþ 1 2 2 I ðx0 , tÞ ¼ t þ OðtÞ : ðG14Þ
¼ x0 I ðx,tÞðx0  xÞ þ OðtÞ, ðG4Þ x0
t 2
but let us obtain another expansion in case this one is not
sufficient. To simplify computations, let us examine first
the Black–Scholes model with constant volatility  I: G.2. OTM options
dSt In the case of an out-of-the-money option, to reconcile
¼ I dWt , ðG5Þ the two models is a bit trickier. If we equate equations
St
(G3) and (G11), we obtain
S0 ¼ x0 : ðG6Þ    
pffiffiffiffiffiffiffiffi x lnðx=x0 Þ
By applying the Ito–Tanaka formula with the call payoff xx0 ln A pffiffi ¼ ððx0 , 0Þ ðx0 , xÞ þ OðtÞÞt:
x0 I ðx, tÞ t
function x ! (x  K)þ we obtain heuristicallyy
Zt Z ðG15Þ
1 t
ðSt  xÞþ ¼ ðx0  xÞþ þ 1fSs 4xg dSs þ ðSs  xÞdhSs i:
0 2 0 A quick study of the function A shows that this function
ðG7Þ is a bijection. Therefore, provided that (x0, 0)
 (x0, x) 6¼ 0, we have
Taking the risk-neutral expectation on both sides, we
obtain lnðx=x0 Þ
Z I ðx,tÞ ¼ pffiffi 1 pffiffiffiffiffiffiffiffi :
 2 x2 t tA ðtððx0 , 0Þ ðx0 , xÞ þ OðtÞÞ= lnðx=x0 Þ xx0 Þ
E½ðSt  xÞþ  ¼ ðx0  xÞþ þ I PðSs ¼ xÞds, ðG8Þ
2 0
Moreover, it is possible to obtain an asymptotic of A1
and it is well-known that around 0 using the asymptotics of the Lambert W
  function:
1 ðlnðx=x0 Þ þ ðI2 s=2ÞÞ2
PðSs ¼ xÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffi exp  : pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2pI2 sx 2I2 s
A1 ð yÞ
2 ln y: ðG16Þ
ðG9Þ y#0

Consequently, by expanding the square in the exponential Therefore,

in equation (G9) and by using an expansion of the
lnðx=x0 Þ
exponential, we transform equation (G8) into I ðx, tÞ
( pffiffipffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ) ,
pffiffi t#0 t 2 lnðtÞ  2 lnðððx0 , 0Þ ðx0 , xÞ
I tx0 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffiffiffiffiffiffi ffi
E½ðSt  x0 Þþ  ¼ pffiffiffiffiffiffi ð1 þ OðtÞÞ, ðG10Þ þOðtÞÞ= lnðx=x0 Þ xx0 Þ
2p
    and a final simplification gives
þ pffiffiffiffiffiffiffiffi x lnðx=x0 Þ
E½ðSt  xÞ  ¼ xx0 ln A pffiffi ð1 þ OðtÞÞ, if x0 5 x,
x0 I t lnðx=x0 Þ
I ðx, tÞ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi : ðG17Þ
ðG11Þ t#0 2t lnðtÞ

with
Remark G1: We did not tackle the case where
N0 ðxÞ (x0, 0) (x0, x) ¼ 0 for an OTM option. Indeed, in this
AðxÞ  þ NðxÞ  1, ðG12Þ
x case, the infinitesimal generator applied to the call payoff
and N the cumulative distribution function of a standard function does not allow us to guess easily an expansion
normal distribution. for the call price function for the jump-diffusion model.
However, in the case where (x0, 0) ¼ 0, or, more strongly,
where  ¼ 0 in a neighborhood of (x0, 0) containing (x, 0),
G.1. ATM options a natural guess would be
Using the definition of the implied volatility and by
equating equations (G2) and (G4), we obtain for at-the- lnðx=x0 Þ
I ðx, tÞ
R x 0 0
,
x0 ðdx =ðx , 0ÞÞ
money options thatz t#0

I ðx0 , tÞ ! ðx0 , 0Þ: ðG13Þ as proved by Berestycki et al. (2002) for diffusions.
t!0

ySee Revuz and Yor (1999, p. 223) for more details.

zNote that this result was proved rigorously for local volatility models by Berestycki et al. (2002) and for jump-diffusions by
Durrleman (2007).