QUANTI TATI VE FI NANCE VOLUME 2 (2002) 6169 RESEARCH PAPER

I NSTI TUTE O F PHYSI CS PUBLI SHI NG quant.iop.org

Asymptotics and calibration of local
volatility models
H Berestycki
1
, J Busca
2
and I Florent
3
1
CAMS, Ecole des Hautes

Etudes en Sciences Sociales, 54 bd Raspail,
75270 Paris Cedex 06, France
2
Ceremade, Universit e Paris Dauphine, Pl. Mar echal de Lattre de Tassigny,
75775 Paris Cedex 16, France
3
HSBC-CCF, 103 av. des Champs-Elys ees, 75419 Paris Cedex 08, France
E-mail: hb@ehess.fr, busca@ceremade.dauphine.fr and
igor.orent@ccf.com
Received 20 September 2001
Published 4 February 2002
Online at stacks.iop.org/Quant/2/61
Abstract
We derive a direct link between local and implied volatilities in the form of a
quasilinear degenerate parabolic partial differential equation. Using this
equation we establish closed-form asymptotic formulae for the implied
volatility near expiry as well as for deep in- and out-of-the-money options.
This in turn leads us to propose a new formulation near expiry of the
calibration problem for the local volatility model, which we show to be well
posed.
In the BlackScholesMerton model [4, 24], it is assumed that
the price of a non-dividend paying stock S
t
follows the log-
normal stochastic differential equation
dS
t
= S
t
(dt + dW
t
), (1)
where t is time, and are constants and W
t
is a standard
Brownian motion. The parameter is called the volatility of
the stock S
t
. It is well-known that the price C(S
t
, t ; K, T ) of a
European call option written on S
t
with strike K and maturity
T satises the linear parabolic partial differential equation
C
t
+

2
2
S
2
C
SS
+ rSC
S
rC = 0 in (0, +) (0, T )
C(S, T ) = (S K)
+
,
(2)
where r is the risk-free short-term interest rate. Such options
are commonly traded on markets, however is not directly
observable. Therefore it is common practice to start from the
observed prices and invert the closed-form solution to (2) in
order to nd that constant called implied volatilityfor
which the solution to (2) agrees with the market price at todays
value of the stock. It is widely observed that calls having
different strikes and otherwise identical have different implied
volatilities. This phenomenon, usually referred to as the smile
effect, clearly violates the BlackScholesMerton model, since
in this framework a constant is supposed to determine the
dynamics of the underlying stock S
t
through (1) regardless of
options, strikes and maturities.
To overcome this difculty, the model has to be extended.
One widely used approach is to consider that the volatility
also follows a stochastic diffusion process. Another case is
when the volatility is not a constant any more but rather a
(deterministic) function of the underlying asset and the time.
Actually this can be seen as a particular case of the previous
approach. These types of models are called local volatility
models. For them, the dynamics of the underlying asset is
governed by the stochastic differential equation
dS
t
= S
t
dt + (S
t
, t )S
t
dW
t
. (3)
There are two problems which are relevant in practice. Firstly
one needs to compute accurately the implied volatilities of
option pricesthe pricing problem. If one follows a traditional
approach (e.g. solve the PDE corresponding to each model
and then invert BlackScholes formula), this is known to be
1469-7688/02/010061+09$30.00 2002 IOP Publishing Ltd PII: S1469-7688(02)32071-2 61
H Berestycki et al QUANTI TATI VE FI NANCE
computationally difcult, especially near expiry or far from
the money.
Secondly, one wants to recover the value of the parameter
of the model from market datathe calibration problem. One
remarkable result, due to Dupire [13], states the following:
should the call options corresponding to all possible strikes
and maturities be priced on the market in a consistent manner,
the local volatility (S, t ) would be uniquely determined by
the relation
(Dupires formula) (K, T ) =
_
2
_
C
T
+ rKC
K
K
2
C
KK
_
,
(4)
where C
K
, C
KK
are the rst- and second-order derivatives of
the call price with respect to its strike and C
T
the derivative
with respect to its maturity. For the readers convenience,
we present in the appendix a proof of this result using PDE
methods.
It turns out that in practice, this approach has two severe
shortcomings. Firstly, there is but a nite set of observations.
Hence some interpolation is needed in order to use (4). It is
nowadays generally acknowledged (see for instance [8, 26])
that it is by no means obvious to gure out how to interpolate
the data set in such a way that the radicand in (4) remains
positive and nite. Further, the result is overly sensitive to
the (arbitrary) choice of the interpolationespecially for short
maturitiesthis resulting in poor robustness of the method.
Secondly, a (related) difculty lies in the intrinsic
indeterminacy of formula (4) in the regions {T t 1},
{|ln(S
0
/K)| 1}, {T t 1}, where it assumes the form
0
0
.
This is the reason why several other approaches have
been proposed in the recent years. We shall not attempt
to give any comprehensive survey of this broad subject, but
rather focus on the results that are the closest to our point of
view. Let us mention in this respect the approach by Lagnado
and Osher [20], further extended by Jackson et al [18] and
Berestycki and Cr epey [3] (see also [9]). The basic idea is to
introduce a regularized cost functional in the form
J() =
_
||
2
+

i,j
_
(C C

)(S
0
, t
0
; K
i
, T
j
)
_
2
, (5)
where C

are observed prices, to be minimized over an

appropriate functional space. Here represents the trade-
off between accuracy and smoothness of the minimizer, this
follows the lines of Tychonovs method [28].
Avellaneda et al [1] and Bodurtha and Jermakyan [5]
also propose other interesting minimization-like methods that
we do not detail here. Let us mention nally the works by
Bouchouev and Isakov [6, 7] who propose a different approach
based on a representation formula that results in an integral
equation for the local volatility.
The purpose of this paper is to propose a completely new
point of view. We intend to show that there exists an explicit
link between the implied and the local volatilities, in the form
of a quasilinear degenerate parabolic PDE. We then examine
several consequences that we can derive from this equation.
Firstly, a particularly important one is an asymptotic formula
for the implied volatility near expiry. Furthermore, using
this limit, we then show that some new formulation of the
calibration problemis well posed. Another use of the equation
allows us to give asymptotics of the implied volatility for deep
out-of-the-money options. Most of the results stated here have
been announced in [2].
In a broader perspective we also hope that this nonlinear
PDEapproach will prove useful for the challenging problemof
calibrating the local volatility and understanding its qualitative
properties.
We adopt throughout the paper the reduced variables
x = ln(S/K)+r, = T t , so that from(2) the transformed
prices
v(x, ) = e
r
C(S, T ; K, T )/K
formally satisfy
v

=
1
2

2
(x, )(v
xx
v
x
) in
T
= R (0, T )
v(x, 0) = (e
x
1)
+
,
(6)
abusing somewhat the notation (see for instance [5]). We shall
assume that
BUC(
T
), 0 < (x, ) < , (7)
BUC being the space of (globally) bounded uniformly
continuous functions and , are constants. In order to make
precise statements let us now specify the technical conditions
that we impose, and state some denitions of functional spaces
that we shall need.
We require that v C(
T
) W
2,1,p
loc
(), =
R (0, +), satisfy the equation in (6) pointwise almost
everywhere in (strong solution). As is classical when
studying parabolic problems, we make use of the following
anisotropic Sobolev spaces:
W
2,1,p
() =
_
w

|w
xx
|
p
+ |w

|
p
+ |w|
p
<
_
W
2,0,
() = {w||w|

+ |w
xx
|

< }
(8)
endowed with their natural norms, together with their local
versions W
2,1,p
loc
and W
2,0,
loc
.
Let us denote by u the solution to (6) corresponding to
1, i.e. satisfying
u

=
1
2
(u
xx
u
x
) in
u(x, 0) = (e
x
1)
+
.
(9)
The explicit solution to (9) is readily seen to be
u(x, ) = e
x
N
_
x

+
1
2

_
N
_
x

1
2

_
, (10)
where
N(d) =
1

2
_
d

y
2
2
dy. (11)
Note that this is nothing but the celebrated BlackScholes
formula [4] written in our reduced variables.
The assumption (7) we made on turns out to ensure
that the pricing problem in (6), which determines the price
of all European options, is well posed. Namely, we have the
following proposition.
62
QUANTI TATI VE FI NANCE Asymptotics and calibration of local volatility models
Proposition 1 (existence). Under assumption (7), there
exists a unique solution v in the class
W
2,1,p
loc
() C() {w | C, > 0,
(x, ) |w(x, )| Ce
x
2
}
for all 1 < p < to
v

1
2

2
(x, )(v
xx
v
x
) = 0 a.e. in
= R (0, +)
v(x, 0) = (e
x
1)
+
.
(12)
Furthermore, v satises
0 < v(x, ) < e
x
v(x, ) (e
x
1)
+
v

(x, ) > 0 in
T
. (13)
A trivial yet crucial observation is that (x, ) u(x, )
( > 0) satises (12) with

. This implies that if one

denes 0 so that
v(x, ) = u(x,
2
(x, )) (14)
for all (x, ) , then is clearly the implied volatility of
the corresponding call option, in the sense described above.
That (14) uniquely determines (x, ) 0 follows easily
from the fact that (e
x
1)
+
v(x, ) < e
x
(see (13) above),
u(x, 0) = (e
x
1)
+
, lim

u(x, ) = e
x
(see (10)), together
with u

> 0 in .
Our main result gives the implied volatility as the unique
solution to a well-posed degenerate quasilinear parabolic
problem.
Theorem 1. Under assumption (7) suppose that the implied
volatility is dened by (14) where v and u are solutions
to (12) and (9) respectively. Then:
(i) The implied volatility lies in W
2,1,p
loc
() for all
1 < p < and satises
2

+
2

2
(x, )
_
1 x

_
2

2
(x, )
xx
+
1
4

2
(x, )
2

2
x
= 0 (15)
a.e. in .
(ii) In the limit 0, the implied volatility is the harmonic
mean of the local volatility, namely
lim
0
1
(x, )
=
_
1
0
ds
(sx, 0)
, (16)
uniformly in x R.
(iii) Conversely, if W
2,1,p
loc
() (for some p > 1), satises
(15) and (16) then .
Remark. The small time-to-maturity asymptotic in theorem 1
part (ii) has an interesting connection with a classical result
by Varadhan [29], although it does not readily follow from it.
In [29] the fundamental solution of the operator
t

1
2

2
(x)
2
x
,
is considered, namely
v

=
1
2

2
(x)v
xx
v(x, 0) = (x)
(17)
(here and throughout the paper (x) is the Dirac mass at the
origin), with satisfying (7). It is then proved that
2 ln v(x, )
__
x
0
ds
(s)
_
2
as 0. This corresponds to the Riemannian metric
associated with the inverse of the diffusion coefcient.
Denoting by U the fundamental solution of
t

1
2

2
x
, namely
U(x, ) = (2)
1/2
exp
_

x
2
2
_
, we can rephrase the result
as follows. If is such that v(x, ) = U(x, (x, )
2
), noting
that by (7) is bounded, we deduce that
1
(x, )

_
1
0
ds
(sx)
(18)
as 0. Hence this problemgives rise to the same change of
metric as in theorem 1 (ii). However, it is not straightforward
to derive from (18) the asymptotics in theorem 1 (ii), since we
are dealing with different initial conditions and composition
with a different function.
We also refer the reader to [14] for other large-deviation
results.
We next establish asymptotics for deep in- and out-of-the-
money options.
Theorem 2. Suppose satises (7) and
lim
x+
(x, ) =
+
() (respectively x ,

())
locally uniformly in , with

continuous. Then
lim
x
(x, ) =
_
1

_

0

(s) ds
_
1/2
. (19)
We point out that theorem 1 (ii) and theorem 2 clarify in
particular the indeterminacy in Dupires formula in the regions
{T t 1} and {|ln(S
0
/K)| 1} that we have mentioned
earlier.
It results from theorem 1 (ii) that can be extended up to
= 0 as a continuous function. That the limit of (x, )
at = 0 exists at all is by no means obvious from (14).
As a matter of fact, the value of the limit is quite specic to
the problem under consideration, as the following proposition
makes clear. Specically, if the payoff function is strictly
convex in the initial variable S, then the asymptotic behaviour
is dominated by the local volatility.
Proposition 2. If one replaces (e
x
1)
+
in (9) by any smooth
function (x) satisfying
xx

x
> 0, that is,
SS
> 0, then
lim
0
(x, ) = (x, 0).
This result is actually much simpler than theorem 1. The main
reason is that here u

> 0 throughout
T
, so that the implicit
function theorem applied to (14) gives bounds for together
with its derivatives near = 0. In other words, it is the very
degeneracy of the original problemnear = 0 that allows such
an instantaneous averaging as that in (16) to take place.
A important feature of (16) is the following. Suppose that
for x = 0 vanishes on some interval between 0 and x.
Then (x, 0) = 0. This is consistent with the probabilistic
63
H Berestycki et al QUANTI TATI VE FI NANCE
point of view. Indeed, in this limiting case the stock price
process starting at x will never cross and so never reaches the
convexity region (x = 0). This feature was absent from other
approximation formulae that were proposed, like weighted
linear means (see for instance [10] the poor mans model).
We refer to the book by Rebonato [26] for a very interesting
and detailed discussion on the qualitative properties of local
and implied volatilities.
The asymptotics in theorem1 (ii) exhibits a linear relation
between the inverse of the local and implied volatilities. This
leads us to propose a new regularization of the calibration
problem. Instead of (5) we introduce the following penalized
functional:
J

() =
_

_
1

2
+

i,j
__
1

_
(x
i
,
j
)
_
2
, (20)
where

are observed implied volatilities, to be minimized

over a suitable functional space. We suspect that this
minimization problem is well posed, at least for short time-
to-maturities
j
. Indeed in this case J

is close to a convex
functional. As a matter of fact we shall prove this property
in the limiting case, that is,
j
0. Specically, we
denote by (x) = (x, 0)
1
the inverse of the local volatility
and (x) =
_
1
0
(sx) ds the inverse of the implied volatility in
the limit 0. We can consider the functional in terms of ,
and write (abusing the notations)
J

() =
_

2
+

i
_
(x
i
)

(x
i
)
_
2
, (21)
where

=
1

.
We assume that these volatilities are consistent, i.e. that
there exists
0
(x, ) for which the solution to (12) with
(x)
0
(x, 0) asymptotically replicates market prices, i.e.
such that lim
0
(x
i
, ) =

(x
i
)

(x
i
). It follows
that J

(
0
)
|=0
= 0, with
0
=
1
0
. This means that,
by assumption, we have a solution to the exact asymptotic
calibration problem. As a consequence, there are in fact
innitely many of them, as can easily be seen from the
argument in the proof of theorem 3 below. The whole point
is to choose one of these solutions in a stable way. This is
question that the following result addresses.
Theorem 3.
(i) For all > 0 there exists a unique solution of the
minimization problem
inf
H
1
(R)
J

(), (22)
denoted by

.
(ii) When 0,

converges uniformly in R to a solution

of the exact asymptotic calibration problem, i.e. such that

(x
i
, 0)
_
1
0

(sx
i
) ds =

(x
i
).
1. Proofs
In the proof of theorem 1 we shall need the following series of
lemmas.
Lemma 4 (maximum principle). Suppose a, b, c are
globally bounded continuous coefcients dened in
T
, and
assume that a(x, ) a > 0. Suppose that
z W
2,1,p
loc
(
T
) C(
T
) satises
z
t
a(x, )z
xx
+ b(x, )z
x
+ c(x, )z 0 a.e. in
T
lim
0
z(x, ) 0 in D

(R)
z(x, ) Ce
|x|
2
for some , C > 0.
(23)
Then z 0 in
T
.
Proof. This is a straightforward adaptation of (a particular
case of) theorem 9, chapter 2, section 4 in [15] to the strong
solutions framework and to the distributional initial data.
Lemma 5. For any W
2,1,p
loc
(
T
) dene
w(x, ) = u(x,
2
(x, )) where u is the solution to (9)
given by (10). The following relation holds:
w

1
2

2
(x, )(w
xx
w
x
)=u

(x, t
2
)F(x, , , D, D
2
)
(24)
a.e. in
T
, where
F(x, , , D, D
2
) = 2

+
2

2
_
1 x

_
2

xx
+

2
4

2

2
x
. (25)
Proof. A straightforward computation gives
w

1
2

2
(w
xx
w
x
) = u

(x,
2
)
_
2

+
2

2
2
_
2 2
x
+ 2(
2
x
+
xx
)
+4
x
u
x
u

(x,
2
) + 4
2

2
x
u

(x,
2
)
__
. (26)
Besides, one easily derives from (10) the relations
u
x
u

(x, ) =
1
2

x

(x, ) =
1
2
+
x
2
2
2

1
8
.
Combining the above identities gives the result.
Lemma 5 applied to w v, hence to , yields
statement (i) in theorem 1.
An important tool in the subsequent argument is the fact
that this problem satises a certain comparison principle that
64
QUANTI TATI VE FI NANCE Asymptotics and calibration of local volatility models
we now state. For this purpose, let us denote by H the
quasilinear operator
H[] H(x, , , D, D
2
) =
_
1 x

_
2
+
xx

1
4

2
x
, (27)
and dene I(0, T ) to be the class of those functions
C
2,1
() for which the associated local volatility
[](x, ) =
_ _

2
_

H(x, , , D, D
2
)
_
1/2
(28)
is well dened, continuous in
T
, and satises there
[](x, ) ; (29)
furthermore, we require the growth condition at zero

2
(x, ) 0 as 0. (30)
We have the following useful result.
Lemma 6 (comparison principle). Given , I(0, T ),
suppose that
[](x, ) [](x, ) (31)
for all (x, )
T
= R (0, T ). Then there.
Note that, due to the degeneracy of (25), it is not necessary to
prescribe any initial ordering condition on , .
Proof. Let us dene u(x, ) = u(x,
2
(x, )) and u(x, ) =
u(x,
2
(x, )) in
T
. By (5), (25) and (27), (29), (30), the
functions u, u satisfy
u

1
2
[]
2
(x, )(u
xx
u
x
) = 0 in
T
u(x, 0) = (e
x
1)
+
(32)
and
u

1
2
[]
2
(x, )(u
xx
u
x
) = 0 in
T
u(x, 0) = (e
x
1)
+
.
(33)
The difference function w(x, ) = u(x, ) u(x, ) then
satises
w

[]
2
(x, )(w
xx
w
x
)
=
_
[]
2
[]
2
1
_
u

in
T
w(x, 0) = 0.
(34)
By (13) and (31), the right-hand side is non-negative. It then
follows from the maximum principle (lemma 4) that w 0 in

T
. Since u

> 0 this implies in

T
.
Proof of theorem 1 (ii). Note that the existence of the limit is
by no means obvious from(15) since this equation degenerates
near = 0. This implies loss of a priori estimates for in
this region. To circumvent this serious difculty, the essential
idea is to dene suitable sub and supersolutions of (12) from
the formal limiting solution of (15) and prove that actual
convergence takes place through the comparison principle.
For the sake of clarity we rst treat the case that satises
the additional regularity assumption
C
2,1
(
T
) and
xx
L

(
T
). (35)
Let us dene the formal limiting solution of (15)
0
as the
unique positive solution of F(x, 0,
0
, D
0
, D
2

0
) = 0, i.e.
(
0
)
2

2
(x, 0)
_
1 x

0
x

0
_
2
= 0. (36)
It is clearly given by

0
(x) =
__
1
0
ds
(sx, 0)
_
1
. (37)
We next dene
(x, ) =
0
(x)(1 + ), (38)
and, respectively, (x, ) =
0
(x)(1 ) in
T
, for > 0,
T < 1/. Let us compute the associated local volatilities [],
[] in the sense of lemma 6, see (28). A simple computation
yields
H[] =
_
1 x

0
x

0
_
2
+
0

0
xx
+ O(
2
). (39)
It follows from here and (28) that
[](x, ) = (x, )

_
1 +
_
4
2

0
xx

(x, 0)
__
+ O(
2
), (40)
where O = O(, ,
W
2,0,). Hence, taking rst large
enough and then small enough we have I(0, ) with
[] in

. Similarly, changing by , one sees that

I(0, ) and [] in

. Clearly I(0, ) as
well (note that (30) comes from the initial condition v(x, 0) =
(e
x
1)
+
), with [] , this resulting from uniqueness
(see lemma 4). An application of the comparison principle
(lemma 6) then yields
(x, ) (x, ) (x, ) (41)
for all (x, )
T
. In particular, the desired result (ii) in
theorem 1 holds, in the case (35) is met.
We nally remove assumption (35) by an approximation
procedure. Let us rst observe that assumption (7) gives the
existence of a sequence

satisfying

C
,/2
(
/2
) for
any (0, 1),

uniformly in
/2
. Indeed,
take (x, ) = (x, ) for all 0 as an extension of and
dene for any > 0

( to be chosen) a standard mollier

(see (73)). Dene then

by

( + ). (42)
65
H Berestycki et al QUANTI TATI VE FI NANCE
Now clearly
() +

() + (43)
in
/2
, if denotes a modulus of uniform continuity for
throughout

. We can then choose = () in such a way

that () < , so that

in
/2
and

uniformly
in
/2
.
Let us nowdene

0
(x) as the harmonic mean of

(x, 0)
i.e.

0
(x) =
__
1
0
ds

(sx, 0)
_
1
(44)
and, correspondingly,

(x, ) =

0
(x)(1 +) together with
v

(x, ) = u(x,

(x, )
2
).
Repeating the computations in the rst step, we get that

I(0, ) and
[

(45)
in

by xing = () large enough, and = () small

enough. Lemma 6 thus implies

in
T
. Similarly,
one constructs

,
0

= (1 ), and shows that

(x, ) (x, ) in

.
Summing up, we get that for all > 0 there exists
(, ) = (, )() such that
__
1
0
ds

(sx, 0)
_
1
(1 )
(x, )

__
1
0
ds

(sx, 0)
_
1
(1 + ) (46)
for all (x, ) R (0, ()). This yields
__
1
0
ds

(sx, 0)
_
1
liminf
0
(x, )
limsup
0
(x, )

__
1
0
ds

(sx, 0)
_
1
(47)
for all > 0, so that
lim
0
(x, ) =
__
1
0
ds
(sx, 0)
_
1
. (48)
That this limit is uniform in x R is easily seen through (46).
Part (iii) in theorem 1 (uniqueness) results from the
maximum principle (lemma 4) applied to the difference
function z(x, ) = u(x,
2
(x, ))u(x,
2
(x, )). Indeed,
by (16) z can be extended as a continuous function on
T
,
with z(x, 0) = 0. Furthermore, by lemma 5 this function
satises z

=
1
2

2
(x, )(z
xx
z
x
) a.e. in
T
. Since clearly
|z(x, )| 2e
x
, we can infer from the maximum principle
(lemma 4) that z 0, and thus since u

> 0.
Remark. Note that we retained from (16) only the fact that

2
(x, ) 0 as 0. In this sense equation (15) has a
built-in initial condition, due to its degeneracy at = 0.
Proof of proposition 2. We apply identity (26) to w v
solution of (6), hence to , the implied volatility. In this
case, u

(x, 0) =
1
2
(u
xx
(x, 0) u
x
(x, 0)) =
1
2
(
xx

x
) > 0,
so that the terms
u
x
u

and
u

in (26) remain bounded as 0.

Furthermore, applying the implicit function theorem to (14),
and using u

> 0, we get that ,

,
x
,
xx
remain
bounded as 0. Hence, sending to 0 in (26) clearly
yields (x, 0) = (x, 0).
Proof of theorem2. By an obvious symmetry in the argument,
we treat only the case x +. As in the proof above, we
shall construct for a given T > 0 a sub- and supersolution
, I(0, T ) that have the required behaviour at innity. To
this purpose, we shall need auxiliary functions whose relevant
properties are summarized in the following lemmas. We refer
the reader to the appendix for the proof.
Lemma 7. Given A > 0, (0, 0.1), > 1/(1 ), there
exists C
2
(R) satisfying the following properties:
(i) W
2,
(R) with
W
2, independent of A
(ii) (z)
1
1
= lim
+
() z R
(iii) (z) 2 z (, A)
(iv)

(z)
(z)

1
2
z R
(v)
z

(z)
(z)
0,

(z) 0,

(z) 0 as z +.
By assumption, (x, )
+
() as x +, uniformly
in (0, T ). Hence, given (0, 0.1) there exists

A such
that

+
()
(x, )

_
1 (x, ) (

A, +) (0, T ). (49)
Besides, using (7) we get > 1 for which
1

+
()
(x, )
in R (0, T ). (50)
Let us dene the quadratic mean limiting volatility

+
() =
_
1

_

0

2
+
(s) ds
_
1/2
(51)
and, for > 0
(x, ) =
+
()(x), (52)
where is given in lemma 7, the values of , > 0 being
dened as above, and A, to be xed. A simple computation
yields
H[](x, ) =
_
1 x

(x)
_
2
+
+

1
4

2
+

2
. (53)
Clearly by (iv) in lemma 7 we can choose =
(
W
2,, T,
+
) > 0 independent of A such that
1
4
H[] 2 (54)
66
QUANTI TATI VE FI NANCE Asymptotics and calibration of local volatility models
in
T
; now by (v) in lemma 7 there is B > 0 for which
x B, (0, T ) imply H[](x, ) <
1
1
. Setting
A = max(B,

A), we see that (54) holds for all x and
H[](x, ) <
1
1
(55)
for all z = x A, (0, T ).
We next compute the local volatility associated with in
the sense of proposition 6, that is

2
[](x, ) =
_

2
+
()(x)
2
_

H[](x, )
(56)
=
2
+
()
(x)
2
H[](x, )
. (57)
Since (x)
1
(1)
for z = x A and (x) 2 for
z (, A), it follows from (49), (50), (54), (55) that
(x, ) [] (58)
for all (x, ) R (0, T ). By applying our comparison
principle (lemma 6) we get that in
T
. By (ii) in
lemma 7 and (52) we get
limsup
x+
(x, )
1
1

+
(), (59)
for all (0, 0.1); this yields limsup
x+
(x, )
+
().
We nally claim that
liminf
x+
(x, )
+
(). (60)
To this aim, we resort to an appropriate subsolution of the
problem. Since its construction is quite similar to that of ,
the rest of the argument is only sketchy.
Given (0, 0.1), we have

A for which

+
()
(x, )

_
1 + (x, ) (

A, +) (0, T ). (61)
We shall make use of an auxiliary function dened in the
following lemma.
Lemma 8. Given A > 0, (0, 0.1), > 1/1 + , there
exists C
2
(R) satisfying the following properties:
(i) W
2,
(R) with
W
2, independent of A
(ii) (z)
1
1+
= lim
+
() z R
(iii) (z)
1
2
z (, A)
(iv)

(z)
(z)

1
2
z R
(v)
z

(z)
(z)
0,

(z) 0,

(z) 0 as z +.
We set
(x, ) =
+
()(x), (62)
with the values of , > 0 dened previously. A choice of
and A similar to the above argument ensures that
H[](x, )
1
1 +
(63)
for x A and
1
4
H 2 in
T
. Now, by (50), (61), (ii),
(iii) in lemma 8, it follows that the local volatility associated
with

2
[](x, ) =
_

2
+
()(x)
2
_

H[](x, )
(64)
=
2
+
()
(x)
2
H[](x, )
(65)
satises
[] (x, ) (66)
for all (x, )
T
. By applying our comparison principle
(lemma 6) we get that in
T
, and sending 0, that
liminf
x+
(x, )
+
(), hence the result.
Proof of theorem 3. That the inmum of J

is achieved over
H
1
follows fairly directly from convexity. The uniqueness of
the minimizer follows from the Euler equation:

i
_

(x
i
)

i
_
2
1
x
i
1
(0,x
i
)
= 0. (67)
Indeed, if and

are two solutions to (67) (with corresponding
quantities ,

respectively), multiplication by

results in

_
_

_
2
+

i
_

_
2
= 0.
Finally, it is not difcult to infer from (67) uniformH
1
bounds
and to pass to the limit 0.
Acknowledgments
This work is part of a research program at the Cr edit
Commercial de France (CCF, HSBC Group). We thank an
anonymous referee for his or her valuable comments and
suggestions.
Appendix
Sketch of the proof of (4). Let us sketch briey the proof
of this well known result [13]. Since

2
K
2
(S K)
+
=
(S K), being the Dirac mass at zero, G(S, t ; K, T ) =

2
K
2
C(S, t ; K, T ) is the Green function for the operator L =

t
+

2
2
(S, t )S
2

2
S
+ rS
S
rC, i.e. satises the backward
parabolic equation
G
t
+ LG = 0
G(S, T ; K, T ) = (S K).
(68)
It is then classical (see [15]) to infer from the Green identity
that G also satises the dual forward equation
G
T
= L

G
G(S, t ; K, t ) = (S K),
(69)
67
H Berestycki et al QUANTI TATI VE FI NANCE
where
L

=
2
K
_
1
2
K
2

2
(K, T )
_
r
K
(K) r. (70)
Integrating twice (69) with respect to K and integrating by
parts results in
C
T
=
1
2
K
2

2
(K, T )C
KK
rKC
K
C(S, T ; K, T ) = (S K)
+
,
(71)
so that
(Dupires formula) (K, T ) =
_
2
_
C
T
+ rKC
K
K
2
C
KK
_
.
(72)

Proof of proposition 1. In the case is H older continuous,

namely C
,/2
(
T
) (0 < < 1), it is theorem 12
chapter 1 section 7 in [15]. We shall construct the solution
in the general case by an approximation procedure. Take a
standard mollier

(x, ), that is, a function satisfying

0,

0
, supp

(0),
_

= 1,

(x, ).
(73)
Dene (x, ) = (x, ) for all 0 and set for all > 0

. By (7) clearly

C
,/2
(
T
) for all (0, 1).
For each > 0 by theorem 12 chapter 1 section 7 in [15]
there exists a unique solution to the approximate problem
v

1
2
(

)
2
(x, )(v

xx
v

x
) = 0 in
T
v

(x, 0) = (e
x
1)
+
(74)
in the class W
2,1,p
loc
(
T
) {w|C, > 0 (x, )
T
|w(x, )| Ce
|x|
2
}. (Note that, by classical Schauder
estimates [15], v

together with its rst time derivative and

second space derivative lie actually in a local H older space.)
Applying the maximum principle (lemma 4), we have
0 v

(x, ) e
x
(75)
in
T
. Since the

are uniformly continuous, uniformly

in , by [30] and (75) we have uniform interior estimates
in W
2,1,p
loc
(
T
). To see this, note that (75) together with
the initial condition control the boundary terms. Hence v

converges locally uniformly to v, viscosity solution to the

limiting equation. By [30] again v W
2,1,p
loc
(
T
) and satises
the limiting equation pointwise a.e. (strong solution).
Now from (75) clearly 0 v(x, ) e
x
, which implies
0 < v(x, ) < e
x
by the strong maximum principle (see [15]).
Let us next take the notation z

= v

and observe that it satises

z

=
1
2
(

)
2
(x, )(z

xx
z

) + 2

(x, )z

(x, 0) =
1
2
(

)
2
(0, 0)(x),
(76)
(x) being the Dirac mass at x = 0. It follows that z

0 in
by the maximum principle (lemma 4). This implies in the
limit v

0 in , hence v(x, ) v(x, 0) = (e

x
1)
+
.
Proof of lemma 7. Take
(z) =
_

_
1
1
(1 + 4) if z 0
1
1
_
1 + 4N
_

1
4
ln
_
z
A
_
__
if z > 0
(77)
where N is dened in (11). It is easily seen that C
2
(R)
and satises (i) and (ii). Then

(z)
(z)

=
1
(1 )

2
exp
_

1
4
ln
2
_
z
A
_
_

1
(1 )

1
2
.
Further, noting that N
_

1
4
ln
_
z
A
__
1/2 for z A
yields (iii).
Proof of lemma 8. One may follow the lines of the argument
above, with
(z) =
_

_
1
1 +
1
4
if z 0
1
1 +
__
1
1
4
_
N
__
1
1
4
_
1
ln
_
z
A
_
+ z
0
()
_
+
1
4
_
if z > 0,
(78)
where z
0
() solves
_
1
1
4
_
N(z
0
()) =
1
4
.
References
[1] Avellaneda M, Friedman C, Holmes R and Samperi D 1997
Calibrating volatility surfaces via relative entropy
minimization Appl. Math. Finance 4 3764
[2] Berestycki H, Busca J and Florent I 2000 An inverse parabolic
problem arising in nance C. R. Acad. Sci., Paris I 331 15
[3] Berestycki H and Cr epey S A regularized functional approach
to the calibration problem in nance In preparation
[4] Black F and Scholes M 1973 The pricing of corporate
liabilities J. Political Economy 81 63754
[5] Bodurtha J N and Jermakyan M 1999 Non-parametric
estimation of an implied volatility surface J. Comput.
Finance 2 2960
[6] Bouchouev I and Isakov V 1997 The inverse problem of option
pricing Inverse Problems 13 L117
[7] Bouchouev I and Isakov V 1999 Uniqueness, stability and
numerical methods for the inverse problem that arises in
nancial markets Inverse Problems 15 R95116
[8] Cont R and Le Duigou M 2000 CNRS Winter School on Model
Calibration (Aussois, 2000) Communication
[9] Cr epey S 2000 Contribution ` a des m ethodes num eriques
appliqu ees ` a la nance et aux jeux diff erentiels PhD
Dissertation
68
QUANTI TATI VE FI NANCE Asymptotics and calibration of local volatility models
[10] Derman E 1997 Forecasting implied volatility: rolling along
the local volatility surface Derivatives 97 (Brussels, 1997)
Communication
[11] Derman E and Kani I 1994 Riding on a smile Risk Mag. 7 329
[12] Derman E, Kani I and Chriss N 1996 Implied trinomial trees
of the volatility smile J. Derivatives 3
[13] Dupire B 1994 Pricing with a smile Risk Mag. 7 1820
[14] Fourni e E, Lasry J M and Lions P-L 1997 Some nonlinear
methods for studying far-from-the-money contingent claims
Publ. Newton Inst. (Cambridge: Cambridge University
Press) pp 11545
[15] Friedman A 1964 Partial Differential Equations of Parabolic
Type (Englewood Cliffs, NJ: Prentice-Hall)
[16] Hagan P and Woodward D E 1999 Equivalent black volatilities
Global Derivatives (Paris) Communication
[17] Isakov V 1991 Inverse parabolic problems with the nal
overdetermination Commun. Pure Appl. Math XLIV
185209
Isakov V 1998 Inverse problems for PDE Applied
Mathematical Sciences vol 127 (New York: Springer)
[18] Jackson N, S uli E and Howison S 1998 Computation of
deterministic volatility surfaces Report Oxford Univ. Comp.
Lab.
[19] Kirsch A 1996 An Introduction to the Mathematical Theory of
Inverse Problems (New York: Springer)
[20] Lagnado R and Osher S 1997 A technique for calibrating
derivative security pricing models: numerical solution of an
inverse problem J. Comput. Finance 1 1325
[21] Lagnado R and Osher S 1997 Reconciling differences Risk
Mag. 10
[22] Lasry J M and Lions P-L 1995 Grandes d eviations pour des
processus de diffusion coupl es par un processus de sauts
C. R. Acad. Sci., Paris I 321 84954
[23] Lee R 2001 Implied and local volatilities under stochastic
volatility Int. J. Theor. Appl. Finance 4 4589
[24] Merton R C 1973 Theory of rational pricing Bell. J. Econ.
Management Sci. 4
[25] Musiela M and Rutkowski M 1997 Martingale Methods in
Financial Modelling (Applications of Mathematics Series)
vol 36 (Berlin: Springer)
[26] Rebonato R 1999 Volatility and Correlation in the Pricing of
Equity, FX and Interest-Rate Options (Financial Eng.
Series) (Chichester: Wiley)
[27] Rubinstein M 1994 Implied binomial trees J. Finance LXIX
771818
[28] Tykhonov M 1963 Regularization of incorrectly posed
problems Sov. Math. 4 16247
[29] Varadhan S R S 1967 On the behaviour of the fundamental
solution of the heat equation with variable coefcients
Commun. Pure Appl. Math. 20 43155
[30] Wang L 1992 On the regularity theory of fully nonlinear
parabolic equations: I Commun. Pure Appl. Math. XLV
2776
[31] Wilmott P 1998 Derivatives (Chichester: Wiley)
69