Está en la página 1de 35

CARF Working Paper

CARF-F-273

A Remark on Approximation of the Solutions to Partial Differential Equations in Finance

Akihiko Takahashi The University of Tokyo Toshihiro Yamada Mitsubishi UFJ Trust Investment Technology Institute Co.,Ltd. (MTEC)
First version: February 2012 Current version: March 2012

CARF is presently supported by Bank of Tokyo-Mitsubishi UFJ, Ltd., Citigroup, Dai-ichi Mutual Life Insurance Company, Meiji Yasuda Life Insurance Company, Nippon Life Insurance Company, Nomura Holdings, Inc. and Sumitomo Mitsui Banking Corporation (in alphabetical order). This financial support enables us to issue CARF Working Papers.

CARF Working Papers can be downloaded without charge from: http://www.carf.e.u-tokyo.ac.jp/workingpaper/index.cgi

Working Papers are a series of manuscripts in their draft form. They are not intended for circulation or distribution except as indicated by the author. For that reason Working Papers may not be reproduced or distributed without the written consent of the author.

A Remark on Approximation of the Solutions to Partial Dierential Equations in Finance


Akihiko Takahashi and Toshihiro Yamada March 13, 2012

Abstract This paper proposes a general approximation method for the solution to a second-order parabolic partial dierential equation(PDE) widely used in nance through an extension of L eandres approach(L eandre (2006,2008)) and the Bismut identiy(e.g. chapter IX-7 of Malliavin (1997)) in Malliavin calculus. We present two types of its applications, approximations of derivatives prices and short-time asymptotic expansions of the heat kernel. In particular, we provide approximate formulas for option prices under local and stochastic volatility models. We also derive short-time asymptotic expansions of the heat kernel under general timehomogenous local volatility and local-stochastic volatility models in nance, which include Heston (Heston (1993)) and (-)SABR models (Hagan et.al. (2002), Labordere (2008)) as special cases. Some numerical examples are shown. Keywords: Malliavin calculus, Bismut indentity, Integration-by-parts, Semigroup, Asymptotic expansion, Short time asymptotics, Heat kernel expansions, Derivatives pricing, Stochastic volatility, Local volatility, SABR model, -SABR models, Heston model

Introduction

This paper proposes a new method for the approximation to the solutions of second-order parabolic partial dierential equations (PDEs), which has been widely used for pricing and hedging derivatives in nance since Black and Scholes (1973) and Merton (1973). In particular, we derive an approximation formula as Theorem 2.1 based on an asymptotic expansion of the solutions to the second-order parabolic PDEs by L eandres Approach (L eandre (2006, 2008)) and an application of Malliavin calculus eectively: the approximation formula is derived through an extension of L eandres elementary integration by parts formula (Theorem 2.2 in L eandre (2006)) presented in Proposition 2.1, and an application of the Bismut identity (e.g. chapter IX-7 of Malliavin (1997)). Also, this derivation can be regarded as an extension of the PDE weight method in Malliaivn-Thalmaier (2006) to an asymptotic expansion of the solutions of the PDEs. As for explanation of L eandres approach and its connection with our method, please see Takahashi and Yamada (2010). Moreover, our method has an advantage in a sense that our computational scheme can be applied in a unied way to obtaining derivatives prices and Greeks under various (multi-dimensional) diusion models. In addition, we apply this method to deriving a short-time asymptotic expansion of the heat kernel under the general diusion setting which includes general time-homogenous local volatility, Heston and (-)SABR models as special cases; for the local volatility model, we also show how to compute the coecients in the expansion by using the Lie bracket. Furthermore, we note that the similar method can be applied to a certain class of non-linear parabolic partial dierential equations though this paper explicitly deals with the linear PDEs. (Please see Remark 2.1.) There are many approaches for approximations of heat kernels through certain asymptotic expansions: for instance, there are recent works such as Baudoin (2009), Gatheral, Hsu, Laurence, Ouyang and Wang
Forthcoming in Recent Advances in Financial Engineering 2011.This research is supported by CARF (Center for Advanced Research in Finance) and the global COE program The research and training center for new development in mathematics. All the contents expressed in this research are solely those of the authors and do not represent the view of any institutions. Graduate Mitsubishi

School of Economics, the University of Tokyo UFJ Trust Investment Technology Institute Co.,Ltd. (MTEC).

(2009), Ben Arous and Laurence (2009), Ilhan, Jonsson and Sircar (2004), Takahashi, Takehara and Toda (2012) and Takahashi and Yamada (2012). Related to our work on approximation of the solutions to second-order parabolic equations and its applications to option pricing, Cheng et al. (2010, 2011) have been developing a new method called DysonTaylor Commutator method. Moreover, Kato, Takahashi and Yamada (2012) has developed an asymptotic expansion for solutions of Cauchy-Dirichlet problem for second-order parabolic PDEs; as an application, they have derived a new approximation formula for pricing barrier options under stochastic volatility setting. (Please see Remark 2.2 and Remark 3.1 below.) Furthermore, Fujii and Takahashi (2011) has developed a new approximation method for the solutions to the nonlinear PDEs associated with the four step scheme for solving forward backward stochastic dierential equations (FBSDEs). The organization of the paper is as follows: the next section derives an integration by parts formula as an extension of a L eandres theorem and then provides an approximation to the solution of second-order linear parabolic PDEs. Section 3 applies the method developed in the previous section to nance; we derive approximate formulas for the option price and vega under local/stochastic volatility environment. Section 4 derives a short-time asymptotic expansion using integration by parts formula. Section 5 shows examples of the short-time asymptotic expansion under general time-homogeneous local volatility, stochastic volatility model with log-normal local volatility and general local-stochastic volatility models. We also provide numerical examples of the short-time asymptotic expansion under Heston model. Finally, Appendix summarizes the calculation of the second order approximation in Section 5.1.

2 Integration by Parts Formula and Asymptotic Expansion of the Solution to Parabolic PDEs
L eandre (2006, 2008) reveals the connections between the semigroup theory and Malliavin calculus. In particular, his elementary integration by parts formula (Theorem 2.2 in L eandre (2006)) provides a nice idea for an approximation of the solutions to second-order parabolic PDEs. In this section, we will extend L eandres elementary integration by parts formula to Proposition 2.1 below, and present an approximation formula ((2.20) in Theorem 2.1) of the solution to a second-order linear parabolic partial dierential equation: Let X () be the unique solution to the following n-dimensional perturbed SDE: for [0, 1], dXt
( )

= =

d k=1

Vk (, Xt ) dWtk + V0 (, Xt )dt,

()

( )

(2.1)

X0 or
( ) dXt

x Rn ,
d k=1

= =

() ( ) 0 (, Xt Vk (, Xt )dWtk + V )dt,

(2.2)

X0

x Rn ,

1 n where Vk = (Vk , , Vk ) (k = 0, 1, , d) have bounded derivatives of any orders in the variables (, x) and n d i l 0i (, x) = V0i (, x) + 1 l Vk (, x)Vk (, x). V 2 l=1 k=1

Here, indicates the stochastic dierential in the Stratonovich sense. ( ) (),i Also, consider the following n n matrix-valued process, {Ut : (Uj (t)), 1 i, j n, 0 t}, dUj Uj where Ak,l (s) Bk
(),i (),i (),i

(t)

= =

d n l=1 k=1

Ak,l (s)Uj

(),i

(),k

(s) dWtl +

n k=1

Bk

(),i

(s)Uj

(),k

(s)ds,

(2.3)

(),i

(0)

i j ,

= =

() k Vli (, Xs ), () k V0i (, Xs )

(s)

i and j is the Kroneckers delta, that is U0 (0),i

( )

= I (the identity matrix). Specically, for Ut , = =

(0)

Ak,l (s)
(0),i Bk (s)

[ [

( ) k Vli (, Xs )

]
=0 =0

( ) k V0i (, Xs )

, .

Let Xt Then, we have


(1,)

:=

() X . t

(1,) Xt

( ) Ut 0

() 1 Us

( d
k=1

)
( ) ) Vk (, Xs

k dWs

() )ds V0 (, Xs

(2.4)

where means Xt
(1)

In particular, Xt
(1,0)

:= =

:=
t 0

(0) Ut

( ) X t Us

=0 ( d ) [ 1 (0) k=1

( ) Vk (, Xs )

]
=0

k dWs

() V0 (, Xs )

]
=0

)
ds .

Next, let a (s)i k , 1 i n, 1 k d, be the process;


1 ( ) i a (s)i Vk (, Xs )) . k = (U (s)

Then, the reduced Malliavin covariance V (t) = {(V (t))ij }i,j is expressed as (V (t))ij =
d k=1 0 t j a (s)i k a (s)k ds.

( )

(2.5)

Throughout this section, we assume the following non-degeneracy of the reduced Malliavin covariance: [A1] sup E[(det(V() (t)))p ] < for 1 < p < .
[0,1]

(2.6)

Then, by Theorem 9.2 in Ikeda and Watanabe (1989), we obtain a smooth density, y p (t, x, y ) associated with (2.1)((2.2)). Moreover, according to Remark 2.2 and Remark 2.3 in Watanabe (1987) as well as Proposition 2.2 in Ikeda and Watanabe (1989), we can see p (t, x, y ) is smooth in x and as well.
k We next dene V as n i=1

k V =

i Vk (, x)

, xi

k = 0, 1, , d.

and L = Next, for f Cb (R ), let


n d 1 2 (Vk ) + V0 . 2 k=1

u (t, x) :=

P t f (x)

[
:= E

( ) f (Xt )

]
=

f (y )p (t, x, y )dy.
Rn

(2.7)

Then, u (t, x) is the solution to the following PDE:

L u (t, x) t u (0, x)

= =

0, f (x).

(2.8)

Also, let u0 (t, x) := P0 t f (x) := E f (Xt ) =


Rn

(0)

f (y )p0 (t, x, y )dy, (2.9)

where p0 (t, x, y ) is the smooth density for (2.1) with = 0. Then, u0 (t, x) is the solution to the following PDE:

L0 u0 (t, x) t u0 (0, x)

= =

0, f (x).

(2.10)

2.1

Integration by Parts Formula


( ) u (t, x) =0 ,

In this subsection, we will give the formula for u1 (t, x) = the following PDE:

and show that u1 (t, x) satises

L0 u1 (t, x) t u1 (0, x)

= =

L1 u0 (t, x), 0,

(2.11)

where L1 := L =
=0 n d ] 1 [ i j (, x) Vk (, x)Vk 2 i,j =1 k=1 n i=1

=0

2 xi xj

(2.12)

V0i (, x)
=0

. xi

k For Xt D2 1 , we denote Ds,k Xt as the Malliavin derivative acting on the Brownian motion W , k = 1, , d. (Please see p.97 in Takahashi and Yamada (2012) for the details.) Then, we obtain the following proposition.

Proposition 2.1 Let 0,(1) (t) be the process given by 0,(1) (t)l = V0 (t)1 U 0 (t)1 Xt Then, the following formula holds:

(1)

)l

, 1 l n.

u (t, x) =
1

t 1 0 P0 ts L [Ps f ](x)ds =

(2.13)

[
E f (Xt )
(0)

n d { l=1 k=1

0,(1) (t)l
0

t k a0 (s)l k dWs

}]
Ds,k 0,(1) (t)l a0 (s)l k ds = (2.14) (2.15)

Rn

f (y )w(y )p0 (t, x, y )dy, where y w(y ) is a smooth function given by

w(y ) = E

n d { l=1 k=1

0,(1)

a
0 k (s)l k dWs

}
Ds,k
0,(1)

]
(0) |X t

(t)

l 0

(t) a

l 0

(s)l k ds

=y .

(Proof) () Let {fn }n Cb (Rn ) be a sequence such that fn f as n . For E [fn (Xt )], we can dierentiate with respect to (and set = 0) as follows; =
n i=1

=0

E [fn (Xt )] () (0) fn (Xt ) X xi it

( )

]
=0

= = =

E fn (Xt ) Xt

[ [

(0)

(1)

]
(1)

1 E fn (Xt ) Ut V0 (t)V0 (t)1 Ut Xt (0) n n n i=1 m=1 l=1 n n n i=1 m=1 l=1

] ( )l ) ( )l ] ]

[
E

(0) 0 m 1 (1) fn (Xt )U (t)i (V0 (t))1 Ut Xt m (V (t))l xi (0) fn (Xt )U (t)i m xi

( d
k=1

1 (0) m 1 (0) l (Us Vk (Xs )) (Us Vk (Xs )) ds

1 (1) (V0 (t))1 Ut Xt

n n d i=1 l=1 k=1 n n d i=1 l=1 k=1

[
E

(0) fn (Xt ) xi

t ( n
0 t 0

)
(0) l 1 )) ds Vk (Xs (Us

(0) m 1 U (t)i m (Us Vk (Xs ))

1 (1) Xt (V0 (t))1 Ut

)l

n d l=1 k=1 n d l=1 k=1

(0) E fn (Xt ) xi
t 0

m=1 (0) l 1 (0) i 1 )) ds Vk (Xs )) (Us Vk (Xs (U (t)Us

1 (1) Xt (V0 (t))1 Ut

)l ] ]


n i=1

xi

(0) 1 (0) i )) Vk (Xs fn (Xt )(U (t)Us

1 (1) (V0 (t))1 Ut Xt

)l

1 (0) l )) ds Vk (Xs (Us

[
E
0

]
.

(0) [Ds,k fn (Xt )] 0,(1) (t)l a0 (s)l k ds

In the above equality, U U 0 , and we used the following relation. Ds,k fn (Xt ) = For g = (g , , g ), g =
1 n l (0) n i=1 (0) fn (Xt ) 0,(1) (t)l ,

xi

1 (0) i fn (Xt )(U (t)Us Vk (Xs )) . (0)

we have

n d l=1 k=1

[
E g
l 0

]
a
0 k (s)l k dWs

n d l=1 k=1

[
E
0

]
Ds,k g a
l 0

(s)l k ds

and Ds,k g l = Ds,k [fn (Xt ) 0,(1) (t)l ] = [Ds,k fn (Xt )] 0,(1) (t)l + fn (Xt )[Ds,k 0,(1) (t)l ]. Then,
n d l=1 k=1 (0) (0) (0)

[
E
0

t (0)

]
[Ds,k fn (Xt )] 0,(1) (t)l a0 (s)l k ds

n d l=1 k=1

[
E fn (Xt )
(0)

{
0,(1) (t)l

t k a0 (s)l k dWs

}]
Ds,k 0,(1) (t)l a0 (s)l k ds .

Therefore, we obtain the following formula. E [fn (Xt )]


( )

= =

n d l=1 Rn k=1

[
E fn (Xt )
(0)

{
0,(1) (t)l

t k a0 (s)l k dWs

}]
Ds,k 0,(1) (t)l a0 (s)l k ds

=0

fn (y )w(y )p0 (t, x, y )dy,

where w(y ) = E

n d { l=1 k=1

0,(1)

a
0 k (s)l k dWs

}
Ds,k
0,(1)

]
(0) |X t

(t)

l 0

(t) a

l 0

(s)l k ds

=y .

The following estimates hold: |E [f (Xt )] E [fn (Xt )]| f fn , ( ) (0) E [fn (Xt )] E [f (Xt ) ] f fn L1 , =0
(0) (0)

where =
n d { l=1 k=1

0,(1)

a
0 k (s)l k dWs

}
Ds,k
0,(1)

(t)

l 0

(t) a

l 0

(s)l k ds

Therefore, we obtain as n , u1 (t, x) = = =


Rn Alternatively, let s = P0 ts Ps f (x). Then, we have

=0

E [f (Xt )]

()

n d

[
E
(0) f (Xt )

0,(1)

(t)
l 0

a
0 k (s)l k dWs

}]
Ds,k
0,(1)

(t) a

l 0

(s)l k ds

l=1 k=1

f (y )w(y )p0 (t, x, y )dy.

0 P t f (x) Pt f (x) = t 0 = 0

d (s )ds = ds

t 0 P0 ts [L L ]Ps f (x)ds.

Hence, using (2.12), we obain u1 (t, x) = u (t, x) =


=0

lim
0

] 1[ Pt f (x) P0 t f (x)
1

(2.16)

t 0

= =

lim

0 P0 P ts L L s f (x)ds

t 1 0 P0 ts L Ps f (x)ds. 0

Also, we easily see that t and hence, F (t, x) :=

(
0

)
1 0 P0 ts L Ps f (x)ds

= L0
0

t 1 0 1 0 P0 ts L Ps f (x)ds + L Pt f (x),

t
0

1 0 P0 ts L Ps f (x)ds satises (2.11) with starting condition 0.

2.2

Asymptotic Expansion
(),x

Let H(1) (Xt , ) : D D be the divergence operator (Malliavin weight) dened by the Bismut identity (pp.247-248 in Malliavin (1997)):
(),x H(1) (Xt , t )

n d [ i=1 k=1

i 0

a
k (s)i k dWs

]
Ds,k (t) a
i

(t)

(s)i k ds

(2.17)

where t is a smooth functional in the Malliavin sense, t D , and (t)i = V (t)1 U (t)1 t

)i

The iterated Malliavin weight Hk is recursively dened as follows: Hk (Xt with H0 (Xt
(),x (),x

, t ) = H(1) (Xt

(),x

, Hk1 (Xt

(),x

, t )),

, t ) t .

The next theorem is our main result in this section.

Theorem 2.1 Consider the following PDE with its initial condition f Cb (Rn ):

L u (t, x) t u (0, x)

= =

0, f (x).

(2.18)

Then, its solution u (t, x) = P t f (x) = E f (Xt ) =


Rn

( )

f (y )p (t, x, y )dy, (2.19)

has an asymptotic expansion in R:

{
P t f (x) where aj (x) =
j

P0 t f (x)

N j =1

}
aj (x)
j

+ O(N +1 ),

(2.20)

t1

(2.21)
tk1 k 0 2 1 0 Ptk f (x)dtk dt2 dt1 P0 P0 tk1 tk L tt1 L Pt1 t2 L k l=1

t
0

k=1 1 ++k =j,i 1

0 (j ) k (0),x Hk (Xt ,

]
0, Xl tl )

= =

(0),x f (Xt )

(2.22)

Rn

f (y )wj (y )p0 (t, x, y )dy, k N, and


(j ) k

with Lk :=

1 dk L |=0 , k! dk

k
l=1

k=1 1 ++k =j,i 1 (0),x

1 . k!

Here, the so called Malliavin weight Hk (Xt Malliavin weight wj S is given by wj (y ) = E


(0),k () 1 dk Xit |=0 , k! dk

Xl tl ) is dened by (2.17) and the push-down of the

0,

[ (j )
k

(0),x Hk (Xt ,

k l=1

]
(0),x 0,l X )|Xt lt

=y ,

(2.23)

where Xit

:=

k N, i = 1, , n. Moreover, we obtain a heat kernel expansion in R: = p0 (t, x, y ) +


N j =1

p (t, x, y )

j wj (y )p0 (t, x, y ) + O(N +1 ).

(2.24)

(Proof) We can recursively apply the integration by parts in Proposition 2.1 u (t, x) where wj (y ) = E
j

:=

1 j Pt f (x)|=0 = j ! j

f (y )wj (y )p0 (t, x, y )dy,


Rn

[ (j )
k

(0),x Hk (Xt ,

k l=1

]
(0),x 0, Xl tl )|Xt

=y .

Then, we have
0 P t f (x) = Pt f (x) + N j =1

j uj (t, x) + N +1 RN ().

where the remainder terms RN (),

RN () which satses =
0

(N +1) k (1 v )N (v ) (v ) (v ), E f (Xt ) Hk (Xt , Xl t l )) dv, N! k l=1

E [|RN ()|] C (T )f E [(det(V() (t))) ] < ,

for some C (T ), , . (See P.102 in Nualart (2006) for instance.) Alternatively, we can recursively obtain the following expression of uj (t, x) in the similar way for obtaining (2.16) in the proof of Proposition 2.1: uj (t, x) =
j

t
0 0

t1

tk1

1 0 2 k 0 P0 P0 Ptk f (x)dtk dt2 dt1 . tt1 L Pt1 t2 L tk1 tk L

k=1 1 ++k =j,i 1

Also, it is easily seen that uj (t, x) satises the following equation:

L0 uj (t, x) = L1 uj 1 (t, x) + + Lj u0 (t, x). t

Moreover, if we take a sequence {fn }nN such that fn S , fn y as n , we have


P t fn (x) = S fn , p (t, x, )S S y , p (t, x, )S = p (t, x, y ), n .

Then, the following heat kernel expansion holds : p (t, x, y ) = p0 (t, x, y ) + Therefore, we obtain the results. Remark 2.1 Let us consider the solution of the PDE:
N j =1

j wj (y )p0 (t, x, y ) + O(N +1 ),

{ (

+ L u (t, x) = 0, u (T, x) = f (x).


t

(2.25)

Suppose u (t, x) is expanded by a perturbation method as u (t, x) = u0 (t, x) + u1 (t, x) + 2 u2 (t, x) + . In order to obtain ui (t, x), i = 0, 1, 2 for instance, we formally expand the PDE:

(
i

)( ) + L0 + L1 + 2 L2 + u0 (t, x) + u1 (t, x) + 2 u2 (t, x) + = 0, t

1 L where Li = i | . ! i =0 i Then, u (t, x), i = 0, 1, 2 satisfy the following PDEs:

{ { {

( t + L0 )u0 (t, x) = 0, 0 u (T, x) = f (x), ( t + L0 )u1 (t, x) = L1 u0 (t, x), 1 u (T, x) = 0, + L0 )u2 (t, x) = (L1 u1 (t, x) + L2 u0 (t, x)), ( t 2 u (T, x) = 0.

Theorem 2.1 provides a solution to this problem. We note that the same method can be applied, at least formally to a certain class of non-linear parabolic partial dierential equations although Theorem 2.1 explicitly deals with the linear ones. A simple example is as follows: (t + L )u (t, x) = 0, (t < T ); u (T, x) = f (x) 1 L = (u , x u )2 xx , 2 (u , x u ) = 1 + (u + x u ), In this case, we have 1 xx , 2 1 L = (u0 (t, x) + x u0 (t, x))xx , } 1{ 0 L2 = (u (t, x) + x u0 (t, x))2 + 2(u1 + x u1 ) xx . 2 L0 = Hence, 1 xx )u0 (t, x) = 0; u0 (T, x) = f (x), 2 1 (t + xx )u1 (t, x) = (u0 (t, x) + x u0 (t, x))xx u0 (t, x); u1 (T, x) = 0, 2 1 (t + xx )u2 (t, x) = (L1 u1 (t, x) + L2 u0 (t, x)); u2 (T, x) = 0. 2 (t + u0 (t, x) is easily solved by (2.29): (2.29) (2.30) (2.31) (2.26) (2.27) (2.28)

u0 (t, x) =

2 (T t)

(z x)2 2(T t)

f (z )dz.

Then, given u0 (t, x), the right hand side of (2.30) is easily computed and so u1 (t, x) is solved, too:

[
u (t, x)
1

]
g (s, Ws )ds (2.32)
(z x)2 2(st)

(t,x)

=
t

)
g (s, z )dz ds,

2 ( s t )

where g (s, z ) = (u0 (s, z ) + x u0 (s, z ))xx u0 (s, z ).

Recursively, given u0 (t, x) and u1 (t, x), u2 (t, x) is obtained by (2.31). Moreover, please see Fujii and Takahashi (2011) for the details, which has developed a new general approximation method for the solutions to the nonlinear PDEs associated with the four step scheme for solving forward backward stochastic dierential equations (FBSDEs). Remark 2.2 If we try to derive the closed form approximation of the solution to Cauchy-Dirichlet problem for second order parabolic PDEs, which is an expectation including the exit time of a domain D such as E [f (Xt )1{ >t} ], the Malliavin calculus approach fails because Malliavin derivative Dt does not exist (see Fourni e et al. (2001)). Therefore, we cannot approximate analytically the solution to Cauchy-Dirichlet problem by applying the Malliavins integration by parts (2.22). However, Kato, Takahashi and Yamada (2012) has developed an asymptotic expansion for solutions of Cauchy-Dirichlet problem for second order parabolic PDEs and showed a similar formula as (2.20) with (2.21) still holds.

3 Perturbations around Closed Form Solutions : Application to Options


In this section, we derive approximation formulas for an options vega and price in local/stochastic volatility models using the expansion methods of semi-group developed in Section 2. Hereafter, we use the notation T (x)p(x)dx for T S (Rn ) and p S (Rn ) meaning that S T, pS .

3.1

Vega Weight

Fourni e et al. (1999) derive the greeks weights using Malliavin calculus. In this subsection, we obtain the Malliavin weight for the plain-vanilla options Vega(Vega weight) by the Bismut identity and show how to derive the analytic approximation of option price using the Vega weight. Let us consider the following asset price dynamics: dSt = (St )dWt , (3.1)

where S0 is a constant and (x) > 0. We also consider the perturbed diusion with () (x) = (x) + (x), where (x) = c (x) for some positive constant c : dSt
( )

= =

() (St )dWt , S0 .

( )

(3.2)

( ) S0

Then, the vega of the plain-vanilla (call) option is dened as vegaLV := () E [(ST K )+ ]|=0 . (3.3)

Under appropriate conditions, vegaLV is given by vegaLV = = = E (ST K )+

(0)

E (ST K )+ H(1) ST , (z K )+ (z )pS


R
(0)

() S |=0 T

]
(3.4)

(0)

(0)

() S |=0 T

)]

(T t, s, z )dz,
(0)

where H(1) ST ,

(0)

() S | T =0

is the Malliavin weight for vegaLV , (z ) is its push-down, and pS


(0) St

(T t, s, z )

is the density function of given = s. Hence, a European call option price for its underlying asset price S () with maturity T and strike K is approximated as follows:

(0) ST

C (T t, s, K ) = E(t,s) [(ST K )+ ] (z K )+ pS
R
(0)

()

(T t, s, z )dz + vegaLV ,

(3.5)

where we assume zero interest and dividend rates. We illustrate this by using a simple case, () (x) = ( + )x: dSt
( )

= =

( + )St dWt , S0 .

( )

(3.6)

( ) S0

The logarithmic process of St

()

is given by, dXt


( )

= =

( + )dWt log S0 .

1 ( + )2 dt. 2

X0

( )

The associated partial dierential equation is given by (t + L )u (t, x) u where L is the generator of Xt , i.e. L =
and f Cb . () ()

= =

0, f (ex ),

(T, x)

1 ( + )2 2

2 x2 x

)
,

10

The Vega is calculated in the following way. Let us consider the process, dUu Ut and introduce the process a(u),
1 (0) a ( u) = U u Su ,

= =

Uu dWu , 1.

with
(0) Su = se(Wu Wt ) 2
1 2

(ut)

Let C (T ) be the reduced Malliavin covariance,

C (T ) :=
t

a(u)2 ds =
t

s Su

)2
du = (s )2 (T t).

(0) Su (0)

Next, we dierentiate the underlying asset price at time T with respect to at = 0: ( ) S |=0 T = ST (WT Wt (T t)).
(0)

We dene the process 0,(1) (t) and 0,(1) (t) as 0,(1) (T ) 0,(1) (T ) := := s (0) () S |=0 = (0) ST (WT Wt (T t)) = s(WT Wt T ), T ST 1 C (T )1 0,(1) (T ) = (WT Wt (T t)). s 2 (T t)
1 UT

Then, the Malliavin derivative of 0,(1) (t) is given by Du,1 0,(1) (T ) = 1 1t<uT . s 2 (T t)

By the integration by parts derived in section 3.2, Vega is calculated as follows. () E(t,s) [(ST K )+ ]|=0

[ [

{ {

(T )
t

a(u)dWu
t

}]
Du,1
0,(1)

= = = =

E(t,s)

(0) (ST

K)

0,(1)

(T )a(u)du

E(t,s) (ST K )+

(0)

[
E(t,s) (ST K )+
(0)

{ [

1 (WT Wt (T t)) s 2 (T t)

sdWu

T t

(ez K )+ E

1 1 (WT Wt )2 (WT Wt ) (T t)

}]

1 sdu s 2 (T t)

}]

1 1 t,x 1 2 WT |X = z e t WT t 2 (T t) T 2 (T t)

(z x 1 2 (T t))2 2 2 2 (T t)

dz

(ez K )+ (z )
R

1 2 2 (T t)

(z x 1 2 (T t))2 2 2 2 (T t)

dz,

where (z ) := = E

1 t,x 1 2 WT |X = z t WT t (T t) T

]
(3.7) 1 2 1 (T t) . 2

1 1 z x + 2 (T t) 3 (T t) 2

)2

zx+

Equivalently, we can calculate Vega by dierentiating the semi-group. Recall that L() is the generator ( ) of Xt , L = 1 ( + )2 2

2 x2 x

11

We dene the dierential operators L0 , L1 as follows; L


0

= =

1 2 2

2 x2 x

)
,

L1

L |=0 =

2 x2 x

)
.

Using semi-group theory, the Vega is given by

( ) u (t, x)|=0
T 1 0 x P0 ut L PT u f (e )du t T

(
T R R

(
p
X (0)

(u t, x, y )

t T

(
=
t

2 x2 x

)
pX
R R

2 y 2 y

)
pX
(0)

(T u, y, z )dyf (ez )dzdu

(0)

(u t, x, y )pX
(0)

(0)

(T u, y, z )dyf (ez )dzdu

) (
R

du (T t)L1

2 x2 x
(0)

)
f (ez )pX
R

(T t, x, z )dz

= Note that

f (ez )pX

(T t, x, z )dz.

= = and

X (0) p (T t, x, z ) x { ( )2 } 1 1 1 exp 2 z x + 2 (T t) x 2 (T t) 2 t 2

2 (T t)

exp

1 1 (z x + 2 (T t))2 2 2 (T t) 2

1 1 (z x + 2 (T t)), 2 (T t) 2

= =

2 X (0) p (T t, x, z ) x2 { ( )2 } 2 1 1 1 2 exp 2 z x + (T t) x2 2 (T t) 2 (T t) 2

{(

2 (T t)

exp

1 1 z x + 2 (T t) 2 2 (T t) 2

)2 }

1 1 (z x + 2 (T t)) 2 (T t) 2

)2

}
.

1 2 (T t)

Then, we obtain ( ) u (t, x)|=0

{(
(0)

(T t)

f (ez )pX
R
(0)

(T t, x, z )

=
R

{
(T t, x, z )

1 2 zx+ 2 (T t)) 2 (T t)

)2 )2

zx+ 1 2 (T t) 1 2 2 (T t) 2 (T t) 1 1

)}
dz

f (ez )pX

1 1 z x + 2 (T t) 3 (T t) 2

zx+

1 2 (T t) 2

)}

dz.

We obtain the Malliavin weight for the Vega, (z ) = 1 3 (T t)

zx+

1 2 (T t) 2

)2

1 1

zx+

1 2 (T t) . 2 (3.8)

Finally, we remark that the Vega we have just evaluated is equivalent to the well-known Black-Scholes Vega.

12

3.2

Pricing Options under Stochastic Volatility Model

This subsection derives an approximate solution of the partial dierential equation (PDE) in stochastic volatility model by a perturbation method. We consider the following stochastic volatility model (St , t ): dSt
( )

= t St dW1,t , =
() t (dW1,t

( )

( )

( ) dt

(3.9) 2 dW2,t ),

S0 = 0 =

(0) S0 (0) 0

> 0, > 0,

where [0, 1]. The purpose of this subsection is to evaluate a European option price: C SV (T t, s, K ) = E(t,s) [(ST K )+ ], given St
( ) () ()

= s.

Let (Xt ) denotes the logarithmic process of the underlying asset (St ). We also dene P t f (x) = E [f (Xt )], and a generator L() =
2 2 1 2 2 1 2 2 21 2 + + . 2 x2 2 x x 2 2 ( ) , f Cb

( )

We decompose the generator in three parts, i.e. L = L0 + L1 + 2 L2 , where L0 L1 L2 = = = 1 2 2 1 2 , 2 x2 2 x 2 2 , x 1 2 2 . 2 2

x Note that L0 is the (logarithmic) Black-Scholes operator. For f Cb , u (t, x) = P T t f (e ) satises the following PDE:

{ (

+ L u (t, x) = 0, u (T, x) = f (ex ).


t (0) S , s u

(3.10)

Let (Uu ) be the rst variation process dened by Uu := dUu Ut = =

i.e.

(0) Uu u dW1,u ,

1,
( )

and C (T ) be the reduced Malliavin (co)variance of St

at = 0, i.e.,

C (T ) =
t

a(u)2 du,

where a(u) We introduce the following expressions:


0 SkT

(0) (0) (Uu )1 u Su .

= = =

k ( ) S |=0 , k T (UT )1
k i=1 0 S , iT

0,(1 ,,k ) (T ) 0,(1 ,,k ) (T )

C (T )1 0,(1 ,,k ) (T ),

13

where l 1 satisfy
k l=1 Theorem 3.1 For f Cb , we have an asymptotic expansion of the solution to the PDE (3.10) : x P T t f (e ) = x P0 T t f (e ) + N j =1 x P0 T t f (e ) +

l = j,

j N,

1 k j.

j t

P0 tk
tk

t1

t L

1 0 x N +1 P0 )= t1 t2 L PT t1 f (e )dtkj dt2 dt1 + O (

N j =1

j
R

f (ey )wj (t, T, x, y )pX

(0)

(T t, x, y )dy + O(N +1 ),

where
j

1 ++k =j,i =1or 2,kj 1

wj (t, T, x, y ) (j )
0,(1 ,,k ) (0)

= =

1 (0) (0),t,x = y ], E [(j ) |XT j!


j

k ( 0,(1 ,,k ) (T )),

k=1 1 ++k =j,i 1

1 (

(T ))

= =

0,(1 ,,k )

(T )

a(u)dW1,u (T )),

[Du,1 0,(1 ,,k ) (T )]a(u)du,

k ( 0,(1 ,,k ) (T )) and pX L0 .


(0)

1 k1 (

t 0,(1 ,,k )

(t, x, y ) is the transition density of X (0) and P0 is the Black-Scholes semigroup with the generator

(Proof) (0) x Under the condition of 0 = 0 > 0, P T t f (e ) has an asymptotic expansion around = 0. The result follows from Takahashi and Yamada (2012). The expansion coecients are obtained by the following way. 0 The limiting (0 -order) term, P0 T t is the (logarithmic) Black-Scholes semi-group with the generator L . The coecients of the asymptotic expansion of the solution to PDE are calculated as following way. First, PT t f (ex )|=0 = = E(t,s) [f (ST )S1T ]
(0) (0)

E(t,s) [f (ST )UT C (T ) 0,(1) (T )].


1 (0) (0) f (ST )UT Uu u Su , (0)

By the chain rule of Malliavin calculus, for u [t, T ], we have Du,1 f (ST ) Du,2 f (ST )
(0) (0)

= =

0.

Then, following the same way in the proof of Proposition (2.1), we obtain PT t f (ex )|=0 Note that, for u T ,

T (0) t

E(t,s)

{[Du,1 f (ST )] 0,(1) (T )}a1 (u)du .

Du,1 f (ST ) 0,(1) (T ) By the integration by parts formula,

(0)

= [Du,1 f (ST )] 0,(1) (T ) + f (ST )[Du,1 0,(1) (T )].

(0)

(0)

[ [

]
(0)

E
t

{[Du,1 f (ST )] 0,(1) (T )}a(u)du


(0)

{
0,(1) (T )

a(u)dW1,u
t

}]
[Du,1 0,(1) (T )]a(u)du ,

E f (eXT )

14

and we obtain PT t f (ex )|=0

a(u)dW1,u
t

]
[Du,1
0,(1) (0),t,x (T )]a(u)du|XT

=
R

f (e )E

0,(1)

(T )

= y pX

(0)

(t, x, y )dy.

The higher order approximation terms of the expansion is given as follows; uj (t, x) :=
(0)

1 i PT t f (ex )|=0 i! i

1 i!

f (ey )E [(i) |XT


R

(0)

(0),t,x

= y ] pX

(0)

(T t, x, y )dy,

where (i) D . Then, we obtain an asymptotic expansion formula of the solution to PDE of the stochastic volatility model around the Black-Scholes solution,
x x 0 P T t f (e ) = PT t f (e ) + N j =1

j uj (t, x) + O(N +1 ).

uj (t, x) satises

+ L0 uj (t, x) t uj (T, x)

= =

L1 uj 1 (t, x) L2 uj 2 (t, x), 0.

Therefore, we have uj (t, x) =


j t T

P0 tk
t k 1

t1

k j t L

x 1 0 L 2 P0 t1 t2 L PT t1 f (e )dtkj dt2 dt1 .

Specically, Corollary 3.1 below derives the rst order approximation formula of European option under the stochastic volatility model. Corollary 3.1 The following approximation formula holds. C SV (T t, ex , K ) = C BS (T t, ex , K ) + C1 (T t, ex , K ) + O(2 ),
BS

(3.11)

where C (T t, z, K ) denotes the Black-Scholes European option price (with time-to-maturity T t, spot price z and strike price K ) and

C1 (T t, e )
x

=
R

(ez K )+ w1 (t, T, x, z )pX (T t) 0 ex n(d1 )(d2 ), 2

(0)

(T t, x, z )dz

(3.12)

= with w1 (t, T, x, z ) = 3 (T t)2 2

2 (T t))3 2 (T t)) 2 (T t))2 (z x + 1 3(z x + 1 (z x + 1 1 2 2 2 2 2 3 2 2 2 ( (T t)) ( (T t)) ( (T t))2 (T t)

)
,

d1 =

2 log (ex /K ) + 0 (T t)/2 , 0 T t d2 = d1 T t,

1 n(d1 ) = exp 2 (Proof) By Theorem 3.1 ,

d2 1 2

PT t f (ex )|=0
T 1 0 x P0 ut L PT u f (e )du t

[
f (ez )E 0,(1) (T )

a(u)dW1,u
t

]
[Du,1 0,(1) (T )]a(u)du|XT
(0),t,x

= z pX

(0)

(T t, x, z )dz.

15

The conditional expectation above is evaluated as follows;

[
E 0,(1) (T ) =
3 (T

t)2 2
t (

a(u)dW1,u
t

]
[Du,1 0,(1) (T )]a(u)du|XT
(0),t,x

=z

2 (T t))3 3(z x + 1 2 (T t)) (z x + 1 2 (T t))2 (z x + 1 1 2 2 2 2 2 3 2 2 2 ( (T t)) ( (T t)) ( (T t))2 (T t)

)
.

Equivalently, we can proceed as follows: Note rst that

(
x 3 L1 P 0 T t f (e ) = 0 (T t)

3 2 3 x x2

)
x P0 T t f (e ).

x We also remark that L1 P0 T t f (e ) is closely related to one of the Greeks in Black-Scholes model, Vanna which is a second order derivative of the option value, once to the underlying spot price and once to volatility. Therefore,

T 1 0 x P0 ut L PT u f (e )du

(
R

= = = = = =

2 0 T

2 y X (0) P0 (s t, x, y )dydu T u f (e ) p y0

((

3 0

(T u)

3 0

t T

(
R R

3 2 3 y y 2
(0)

y X P0 T u f (e ) p

(0)

(u t, x, y )dydu,
(0)

(T u)

pX

(u t, x, y )

3 0

t T

(T u)

(
3 0 t 3 (T 0

2 3 3 x x2

)
pX
R

3 2 3 y y 2

)
pX (T u, y, z )dyf (ez )dzdu (T u, y, z )dyf (ez )dzdu

(0)

(u t, x, y )pX
(0)

(0)

)(

(T u)du t)2 2

2 3 3 x x2

)
pX
R

(T t, x, z )f (ez )dz

3 2 3 x x2

x P0 T t f (e ).

Take a sequence {fn }n such that fn S , fn ( K )+ in S (n ) , we have

fn (ez )pX

( )

(T t, x, z )dz (T t, x, z )dz (T t, x, z )dz

C SV (T t, x, K ), C BS (T t, ex , K ),
S (e

R
fn (ez )pX

(0)

z R

fn (e )w1 (t, T, x, z )pX as n . Then, in sum, we obtain

(0)

K )+ , w1 (t, T, x, )pX

(0)

(T t, x, )S ,

C SV (T t, ex , K ) = = =
3 C BS (T t, ex , K ) + 0

(T t)2 2

3 2 3 x x2

)
C BS (T t, x, K ) + O(2 )

C BS (T t, ex , K ) + C BS (T t, ex , K ) +

(T t) 0 ex n(d1t )(d2t ) + O(2 ) 2 (ez K )+ w1 (t, T, x, z )pX


R
(0)

(T t, x, z )dz + O(2 ).

Remark 3.1 Kato, Takahashi and Yamada (2012) has derived a new approximation formula for pricing barrier options under stochastic volatility setting as an application of an asymptotic expansion for solutions of Cauchy-Dirichlet problem for second order parabolic PDEs. To summarize Kato, Takahashi and Yamada ( ) (2012), consider the above stochastic volatility model and dene := inf {t; Xt = log B }, where B is a

16

constant barrier, and P f (ex ) := E [f (Xt )1{ >T } ] = C SV.Barrier (t, x, B ). Kato, Takahashi and Yamada (2012) has derived the following formula:

( )

C = C
SV.Barrier

T 1 0 x P0 T s L Ps f (e )dx t

(t, x) P f (e ) +
0 x

{
BS.Barrier

(t, x) +

(T t) 1 BS.Barrier L C (t, x) + 2
(0)

(T s) 1 BS.Barrier L C (s, x)h(s)ds 2

}
,

where h(s) is the density of the rst hitting-time of Xt

to log B .

Short-Time Heat Kernel Asymptotic Expansion

This section derives a short-time asymptotic expansion under multi-dimensional diusion setting: in particular, the asymptotic expansion formula developed in Theorem 2.1 is eectively applied. Consider the following SDE on Rn over the d-dimensional Wiener space (W , P ):
i dXt i X0

= =

i (Xt ) dWtk + V0i (Xt )dt, Vk

(4.1)

k=1 xi 0

R, i = 1, , n,

or
i dXt i X0

= = and

i 0i (Xt )dt, Vk (Xt )dWtk + V

(4.2)

k=1 xi 0

R, i = 1, , n.

where Vk =

1 n (V k , , Vk )

with

i l 0i (x) = V0i (x) + 1 V l Vk (x)Vk (x). 2 d


l k k=1

i Vk

Cb

(4.3)

We assume that (x) = [ ij (x)] where ij (x) = k as dene V k = V and L=


m n i=1 i Vk (x)

j i Vk (x)Vk (x) is positive denite at x = x0 . We also

, xi

k = 0, 1, , d.

(4.4)

d 1 2 Vk + V0 . 2 k=1

Let i = (i1 , , im ) {0, 1, , d} , we set (i) = #{i : il = 0} and i = (i) + m. The following stochastic Taylor expansion holds (e.g. p.4 in Baudoin (2009)): Xt = x0 +
N ( k=1 i,i=k

i V i2 (Vi1 )(x0 ) V k
0

t
1 dWti1

t1
2 dWti2

tm1

m dWtim

+RN (t, x), for some remainder term RN (t, x) which satises
xRn

sup E [RN (t, x)2 ]1/2 CN t(N +1)/2

sup
i,k+(i)=N +1orN +2

i V i2 (Vi1 ) . V k

We rst consider the scaling SDE in order to obtain a short-time heat kernel expansion:
dXt

= =

d l=1

Vk (Xt ) dWtk + 2 V0 (Xt )dt,

(4.5)

X0

x0 Rn ,

17

1 where (0, 1]. Note that Xt is equivalent in law to X 2 t , i.e. 1 Xt L X 2t, has an asymptotic expansion: and that X1 X1 x0 + k=1

k X1

(k )

in

D (Rn ),

where t = 1, i.e.

(k ) X1

(k ) (k ) (X11 , , Xn1 ),

k N is expressed as the coecient in the stochastic Taylor expansion at

Xi1 = Next, set

(k )

(
i,i=k

im V i2 (Vii )(x0 ) V 1
0

1
1 dWti1

t1
2 dWti2

tm1

m dWtim .

Y1 Then, we have

1 = ft (X1 t ) := (X1 t x0 ). t

(4.6)

pX (t, x0 , x) = pX

(1, x0 , x) = pY

(
1, 0,

x x0 t1/2

tn/2 .

(4.7)

Note also that the (i, j )-element of the Malliavin covariance matrix of Y10 = as:
ij Y 0
1

d
k=1

1
0

Vk (x0 ) dWtk is given

d k=1 d k=1 0

1 0 Dt,k Yi0 1 Dt,k Yj 1 dt

(4.8)

i j Vk (x0 )Vk (x0 ) = ij (x0 ).

Since Y1 p
Y
t

is uniformly non-degenerate by the assumption that (x0 ) is positive denite, the smooth density,
t

(1, y0 , y ) for the law of Y1


Y
t

exists.

Thus, p

(1, y0 , y ) has an asymptotic expansion by setting = Y1 :=

t for Y , where (4.9)

i1 (i) (X1 x0 ) X1 in D (Rn ). i=1

In particular, Y10 =
(1) X1

d k=1 0

Vk (x0 ) dWtk .

(4.10)

0,k Let Yi 1 denotes the i-th element of Y1 , that is Y1 = (Y11 , Y21 , , Yn1 ), and dene Yi1 , k N, i = 1, , n as ,k Yi0 1

1 dk (k+1) . Yi1 |=0 = Xi1 k! dk

(4.11)

Then, applying Theorem 2.1 especially, (2.24), we obtain an asymptotic expansion of pY (1, 0, y ): p
Y

(1, 0, y )

Y0

(1, 0, y )

(N
j =0

(j ) k

E [Hk (Y10 ,

k l=1

)
0, Yl 1 l )|Y10

= y]

(4.12)

+O(N +1 ), where
(j ) k

k=1 1 ++k =j,i 1

1 . k!

18

Here, it is easily seen that the density of Y10 is given by pY (1, 0, y ) = (2 )N/2 det( (x0 ))1/2 e where (x0 ) = (
0 y T (x0 )1 y 2

(4.13)

d
k=1

j i Vk (x0 )Vk (x0 ))(1i,j n) .

Consequently, by (4.7), we obtain the following theorem that presents a short-time o-diagonal heat kernel expansion. Theorem 4.1 As t 0, we have a short-time asymptotic expansion of the density pX (t, x0 , x):
(xx0 )T (x0 )1 (xx0 )) 1 1/2 2t p (t, x0 , x) det ( x ) e 0 (2t)n/2

(N
j =0

j/2

j t

1/2

(x x0 )

)
, (4.14)

j i where (x0 ) = ( k=1 Vk ( 1/2 ) (x0 )Vk (x0 ))(1i,j n) , and j t (x x0 ) is the j -th push-down of the Malliavin weights dened by

j t

1/2

(x x0 )

[j ] k (j ) k

[
E
(1) Hk (X1 ,

k l=1

]
(l ) (1) Xl 1 )|X1

=t

1/2

(x x0 )

(4.15)

[
E Hk (Y10 ,

k l=1

]
=t
(1) 1/2

0, Yl 1 l )|Y10

(x x0 ) .
(k)

,k Here, Y10 and Yi0 are given by (4.10) and (4.11), respectively, and X1 1 d k=0 (k ) Xi1 0 1

and Xi1 are given by

X1

(1)

Vk (x0 ) dWtk , im V i2 (Vii )(x0 ) V 1


0

1
1 dWti1

t1
2 dWti2

tm1

m dWtim .

i,i=k

Also,
[j ] k

k=1 1 ++k =j,i 2

1 , k!

and
(j ) k

k=1 1 ++k =j,i 1

1 . k!

Remark 4.1 In the diagonal case, the diagonal heat kernel pX (t, x0 , x0 ) is approximated by 1 det (x0 )1/2 (2t)n/2 where j (0) =
[j ] k (j ) k

(N
j =0

)
t j (0)
j

[
E
(1) Hk (X1 ,

k l=1

]
(l ) (1) Xl 1 )|X1

=0

[
E Hk (Y10 ,

k l=1

]
=0 .

0, Yl 1 l )|Y10

19

Next, we provide alternative methods to obtain the coecients of the expansion. Let A be the perturbed generator associated with (4.5): A=
d 1 2 0 . (Vk ) + 2 V 2 k=1
(Xt x0 )

Then, the generator L associated with the process after the transformation, Yt = L = where L k =
n i=1 i Vk (x0 + y ) d 1 2 (Lk ) + L 0. 2 k=1

is given by (4.16)

, yi

k = 0, 1, , d.

(4.17)

Hence, by applying (2.20) in Theorem 2.1, we have for f Cb (Rn ),


0 P 1 f (0) = P1 f (0) + N j =1

j j (y ) + N +1 RN (y ),

(4.18)

where j (y ) =
j

t1

tk

(4.19)

k=1 1 ++k =j,i 1

1 0 2 k 0 P0 P0 Ptk f (y0 )dtk dt2 dt1 |y0 =0 , (1t1 ) L P(t1 t2 ) L (tk tk1 ) L

with Lk :=

1 dk L |=0 , k! dk

k N, i = 1, , n.

Applications of Short-Time Asymptotic Expansion

This section shows three examples of Theorem 4.1 in the previous section. In particular, we explicitly derive short-time asymptotic expansions under stochastic volatility model with log-normal local volatility and general local-stochastic volatility models. Moreover, we applies (4.15) and (4.19) in Section 4 to computing the coecients in the expansions. In addition, for local volatility model in Section 5.1 and Appendix, we compute the expansion coecients j (y ) (j N), j = 1, 2 in (4.19) by using Lie brackets.(Lie bracket [A, Z ] stands for [A, Z ] = AZ ZA where A and Z are vector elds.)

5.1

Short-Time Asymptotic Expansion for Local Volatility Model


dXt X0 = = (Xt )dt + (Xt )dWt , x0 . (5.1)

Consider the following time-homogenous local volatility model.

Proposition 5.1 We t 0, we have 1 p(t, x0 , x) exp 2 (x0 )2 t where 1 (t, x0 , x) =

(x x0 )2 (x0 )2 t

1+

t1 (t, x0 , x) + t2 (t, x0 , x)

h1 ((x x0 )/ t, (x0 ) h3 ((x x0 )/ t, 2 (x0 )) 1 3 (x0 ) + ( x ) ( x ) , 0 0 2 (x0 ) 2 ( 2 (x0 ))3


2

(5.2)

20

and 2 (t, x0 , x) = h6 ((x x0 )/ t, 2 (x0 )) 1 (x0 )2 6 (x0 ) 8 ( 2 (x0 ))6

) h4 ((x x0 )/ t, ( 2 (x0 ))) 1( 2 (x0 ) (x0 )5 + 4 (x0 )2 (x0 ) + 3(x0 ) (x0 ) 6 ( 2 (x0 ))4 ) 1( 2 + (x0 ) (x0 )3 + 2 (x0 ) (x0 )2 + 2(x0 ) (x0 ) (x0 ) + (x0 )2 (x0 )2 + 2(x0 )2 4 h2 ((x x0 )/ t, 2 (x0 )) . ( 2 (x0 ))2 + (5.3) Here, hn (x, ) stands for the Hermite polynomial of degree n with , that is hn (x; ) = ()n ex
2

/(2)

dn x2 /(2) e . dxn

(Proof) We apply (4.15) and (4.19) in computation of the coecients of the expansion. First, we have the following stochastic Taylor expansion Xt where X1t X2t X3t = = x0 + X1t + X2t + X3t + R3 (t), (5.4)


=
0

(x0 )dWs ,
t

(x0 )ds +

(x0 )
0

(x0 )dWu dWs .

0 t

(x0 )

0 s

(x0 )dWu ds
0

+ +

1 2

0 t

0 s

2 (x0 )(

(x0 )dWu )2 dWs

(x0 )
0

(x0 )
0

(x0 )

(x0 )dudWs +

(x0 )dWv dWu dWs ,

and R3 (t) is a remainder term. Let Xt be the solution of the following scaling SDE.
dXt X0

= =

2 (Xt )dt + (Xt )dWt ,

x0 .

Let A be the generator of

Xt

dened by 1 2 A = 2 2 (x) 2 + 2 (x) . 2 x x

Consider a transform
Yt = f (Xt )=

1 (Xt x0 ),

(5.5)

then the generator L of Yt has the following form L = 2 1 2 (x0 + y ) 2 + (x0 + y ) . 2 y y

First, we apply the push-down of the Malliavin weights to computing the coecients of the expansion. Note that Xt and Yt are expanded in D as follows.
Xt

= =

x0 + X1t + 2 X2t + 3 X3t + O(4 ), Y0t + Y1t + 2 Y2t + O(3 ),

Yt

21

where Y0t = X1t = Xt |=0 =

(x0 )dWs ,

Y1t

X2t =

1 2 Xt |=0 = 2! 2

(x0 )ds +
0

(x0 )
0

(x0 )dWu dWs .

Y2t

= =

X3t =
t

1 3 Xt |=0 3! 3

(x0 )
0

0 t

1 (x0 )dWu ds + 2
s

t 2

(x0 )(

(x0 )dWu )2 dWs


0 s

(x0 )
0

(x0 )
0

(x0 )
0

(x0 )dudWs +
0

(x0 )dWv dWu dWs .

Note that Yt is uniformly non-degenrate. The following relation holds, 1 pX (t, x0 , x) = pY (1, ft (x0 ), ft (x)) . t 1. Using Bismut identity, Y p (1, y0 , y ) [ ( )] () E y Y1 |=0 (y () is a delta function at y .) [ ( ) ( )] (0) (0) () E y Y1 H1 Y1 , Y1 |=0 (5.6)

= = = = =

E y

(0) Y1

[
E y

(0) Y1

1 2 (x0 ) 1 2 (x0 )

{(

( ) Y |=0 1

(x0 )dWu
0

0 1

( ) Du,1 Y1 |=0 (x0 )du

}] }]

(x0 ) + (x0 ) (x0 )

Ws dWs

(x0 )W1 (x0 ) (x0 )W1


2

y ( v )
R

1 E (x0 ) + (x0 ) (x0 ) 2 (x0 ) 1 2 1 e 2 v dv 2 = 1 E 2 (x0 )

{(

)
Ws dWs (x0 )W1 (x0 ) (x0 )W1
2

}
| (x0 )W1 = v

{(

)
Ws dWs (x0 )W1 (x0 ) (x0 )W1
2

}
| (x0 )W1 = y

(x0 ) + (x0 ) (x0 )

1 e 2 2 2 (x0 ) Note that Du,1


0

( 1

y (x0 )

)2

(x0 )x0 (x0 )Ws dWs = (x0 )x0 (z ){Wu +


u

dWs } = (x0 ) (x0 )W1

Note that

]
Ws dWs | (z )W1 = y =

(
0

)(
sds

E
0

y2 1 (x0 )4 (x0 )2

1 = 2

y2 1 (x0 )4 (x0 )2

)
.

22

Then, we obtain ( ) E y Y1 |=0 ( 3 )} { 1 y 3y y 3 (x0 ) + (x0 ) (x0 ) (x0 )2 2 (x0 )6 (x0 )4 1 1 e 2 2 2 (x0 )

)]

(5.7)

y (x 0 )

)2

2. Alternatively, we can evaluate the coecients of the expansion in the following way. Let L0 L1 L2 then |=0 P 1 f (y0 )|y0 =0 Let h be a map y h(y ) such that h( y ) = = =
1 0 L1 P 0 s f (y ) = L P1(1s) f (y )

= = =

2 1 2 (x0 ) 2 , 2 y (x0 ) (x0 )y 2 + (x0 ) , y 2 y

1 2 (( (x0 ))2 + (x0 ) 2 (x0 ))y 2 2 + (x0 )y , 2 y y

=
0

1 1 0 P0 (1s) L Ps f (y0 )ds|y0 =0 .

E [f (Y10 )|Y10 s

= y] = L

1 R

pY (s, y, z )f (z )dz

2 (x0 ) (x0 )y 2 + (x0 ) y y

)
R

pY (s, y, z )f (z )dz.

Then, we explicitly evaluate (4.19) for j = 1.


1 0 P0 (1s) L Ps f (y0 )|y0 =0

pY (1 s, y0 , y )h(y )dy |y0 =0

R
=

(
pY (1 s, y0 , y ) L1
0

)
pY (s, y, z )f (z )dz
0

dy |y0 =0

((
pY (1 s, y0 , y )
0

(x0 ) (x0 )(y y0 )

2 + (x0 ) y 2 y

)
R

)
pY (s, y, z )f (z )dz
0

dy |y0 =0 .

Note that pY (1 s, y0 , y )(y y0 ) = (1 s) (x0 )2 pY (1 s, y0 , y ) Therefore, we have


1 0 P0 (1s) L Ps f (y0 )|y0 =0
0 0

. y0

(5.8)

Y0 p (1 s, y0 , y ) y0

((
(1 s) (x0 )3 (x0 )

( ) 0 3 pY (1 s, y0 , y ) (1 s) (x0 )3 (x0 ) 2 y y0
Y0 p (1 s, y0 , y ) (x0 ) y

2 + (x0 ) y 2 y

)
R

)
pY (s, y, z )f (z )dz
0

dy |y0 =0

)
dy |y0 =0

pY (s, y, z )f (z )dz
R

)
dy |y0 =0

pY (s, y, z )f (z )dz
R

(1 s) (x0 )3 (x0 )
R

3 Y 0 p (1 s, y0 , y ) 3 y0

(
pY (s, y, z )f (z )dz
R
0

)
dy |y0 =0

23


+(x0 )
R

Y0 p (1 s, y0 , y ) y0

(
p
R Y0

)
(s, y, z )f (z )dz
Y0

dy |y0 =0

(1 s) (x0 ) (x0 )
3

+(x0 )
R

y0
3

(
R

R
0

3 3 y0

(
R

)
(1 s, y0 , y )p
0

Y0

(s, y, z )dy

f (z )dz |y0 =0

)
f (z )dz |y0 =0

pY (1 s, y0 , y )pY (s, y, z )dy

= =

3 (1 s) (x0 ) (x0 ) 3 y0
3

p
R

Y0

(1, y0 , z )f (z )dz |y0 =0 + (x0 ) y0

pY (1, y0 , z )f (z )dz |y0 =0

3 0 (1 s) (x0 ) (x0 ) 3 P0 P1 f (y0 )|y0 =0 . 1 f (y0 )|y0 =0 + (x0 ) y0 y0

Then, the rst order approximation term is given by

1 1 0 P0 (1s) L Ps f (y0 )ds|y0 =0

(
0

)
(1 s)ds (x0 )3 (x0 )

=
0 {

=
R

1 (x0 )3 (x0 ) 2

0 3 0 P f (y0 )|y0 =0 + (x0 ) P1 f (y0 )|y0 =0 3 1 y0 y0

z3 3z (x0 )6 (x0 )4

+ (x0 )

z (x0 )2

pY (1, 0, z )f (z )dz.

3. Moreover, the coecient is computed by using the Lie bracket.

1 1 0 P0 (1t1 ) L Pt1 f (y0 )dt1 |y0 =0 0 1 0 1

(1 t1 )i i=0

i!

[L0 , [L0 , , [L0 , L1 ]]])P0 1 f (y0 )dt1 |y0 =0

(
=

L +

1 0 1 [L , L ] P0 1 f (y0 )|y0 =0 , 2

because [L0 , [L0 , L1 ]] = 0 and hence all the terms in (5.9) for i 2 are equal to 0. The Lie bracket [L0 , L1 ] is explicitly computed as follows. L0 L1 = = = L1 L0 = = Then [ L0 , L1 ] Then we have = = L0 L1 L 1 L0 (x0 )3 (x0 ) 3 ,

( ) 1 (x0 )2 2 (x0 ) (x0 )y 2 2 ( ) 1 3 (x0 ) (x0 ) ( 2 + y 3 ) 2 ( ) 1 3 4 (x0 ) (x0 ) 2 3 + yx . 2 ( ) 1 (x0 ) (x0 )y 2 (x0 )2 2 2 1 3 4 (x0 ) (x0 )y . 2

L1 +

{
=

1 0 1 1 1 e 2 [L , L ] 2 2 (x0 )2

( yy0 )2
(x0 )

|y0 =0

(5.9)

1 y + (x0 )3 x0 (x0 ) (x0 ) (x0 )2 2 1 2 (x0 )2 e


1 2

y (x0 )

)2

y3 3y (x0 )6 (x0 )4

)}

The calculation of the second term approximation is given in Appendix.

24

5.2 Short Time Asymptotics for Stochastic Volatility Model with Lognormal Local Volatility
Consider the following stochastic volatility model with log-normal local volatility which includes the Heston type model: dSt = rSt dt + vt St dWt , S0 dvt v0 = = = s0 > 0, a(vt )dt + b(vt )dZt , v > 0,

where Wt and Zt are two standard Brownian motions with correlation . We have a short-time expansion of density for the logarithmic process. Proposition 5.2 When t 0, we have
X

(x x0 )2 1 p (t, x0 , x) exp 2v0 t 2v0 t where x = log s , x0 = log s0 , w1 (t, x0 , x) =

1+

tw1 (t, x0 , x) ,

(5.10)

( ) h ((x x )/t, v ) h3 ((x x0 )/ t, v0 ) 1 1 1 0 0 v0 b(v0 ) + r v . 0 3 4 2 v0 v0

Also, the following approximation formula of the option price holds:

C (t, K )

(x x0 )2 1 (e K ) exp 2v0 t 2v0 t log(K )


x

)
dx

(x x0 )2 1 + t (e K )w1 (t, x0 , x) exp 2v0 t 2v0 t log(K )


x

)
dx.

(Proof) We will apply (4.15) and (4.19) in computation. First, we have the following stochastic Taylor expansion Xt where X1t X2t = = x0 + X1t + X2t + R2 (t), (5.11)

v0 dWs ,

(5.12)

1 1 v0 )ds + 2 2 0 Next, we introduce a time scaling parameter = t, = (r


dXt dvt

1 v0

b(v0 )dZu dWs .

= =

2 ( r

1 vt )dt + vt dWt , 2 2 a(vt )dt + b(vt )dZt .

The generator of the above diusion is 1 2 1 2 1 2 A = 2 v 2 + 2 ( r v ) + 2 vb(v ) + 2 b(v ) 2 + 2 a(v ) . 2 x 2 x xv 2 v v Consider a transform Y = f (X ) = 1 (X x0 ), then the generator of (Y, v ) is given by, L = 1 2 1 2 1 2 v 2 + ( r v ) + vb(v ) + 2 b(v ) 2 + 2 a(v ) . 2 y 2 y yv 2 v v

X and Y are expanded in D ,


Xt

= =

x0 + X1t + 2 X2t + O(3 ), Y0t + Y1t + O(3 ),

Yt

25

where Y0t Y1t Note that = = X1t =


X2t =
0 0

v0 dWs , 1 1 v0 )ds + 2 2

(r

1 v0

b(v0 )dZu dWs .

pX (t, x0 , x) = pY

(1, f

(x0 ), f

1 (x)) . t

(5.13)

1. Using the Bismut identity, the rst order approximation term is given as Y p (1, 0, y )|=0 = = = E [H1 (Y01 , Y11 )|Y01 = y ]pY (1, 0, y )
0

1 E [Y11 v0

1 0

1 0

v0 dW1,t y3 y 3 3 v0 v0

0 Dt,1 Y11 v0 dt|Y01 = y ]pY (1, 0, y )

1 v0 b(v0 ) 4

))

+ r

1 v0 2

)(

y v0

)}

pY (1, 0, y ).

Then we have a approximation formula of the density

pY where

(1, 0, y )

1 exp 2v0

y2 2v 0

1+

t1 (1, 0, y ) ,

(5.14)

1 (1, 0, y )

h3 (y, v0 ) 1 1 v0 b(v0 ) + r v0 3 4 2 v0

) h (y, v ) 1 0
v0

By (5.13) and (5.14), we have the formula (5.10). 2. Alternatively, we have |=0 P0 1 f (y0 )|y0 =0 = with L1 Note that Y0 p (s, y, z ) v Let g be a map y g (y ) such that g (y ) = =
1 0 L1 P0 s f (y ) = L P1(1s) f (y )

1 1 0 P0 (1s) L Ps f (y0 )ds|y0 =0

1 2 L |=0 = (r v ) + vb(v ) . 2 y yv 1 2 Y 0 s p (s, y, z ). 2 y 2

(5.15)

E [f (Y10 )|Y10 s

= y] = L

1 R

pY (s, y, z )f (z )dz.

Then, we explicitly evaluate (4.19) for j = 1.

1 0 P0 (1s) L Ps f (y0 )|y0 =0

pY (1 s, y0 , y )g (y )dy |y0 =0

R
= p

(
Y0

1 R

)
p
Y0

(1 s, y0 , y ) L (1 s, y0 , y )

(s, y, z )f (z )dz

dy |y0 =0

((
p
Y0

1 3 1 s vb(v ) 3 + (r v ) 2 y 2 y

)
R

)
p
Y0

(s, y, z )f (z )dz

dy |y0 =0

1 0 3 1 s vb(v ) 3 P0 v) P1 f (y0 )|y0 =0 . 1 f (y0 )|y0 =0 + (r 2 2 y0 y0

26

Therefore, we have

1 1 0 P0 (1s) L Ps f (y0 )ds|y0 =0

{
=
R

1 v0 b(v0 ) 4

y3 y 3 3 v0 v0

))

1 + r v0 2

)(

y v0

)}

pY (1, 0, y )dy.

5.3

Short Time Asymptotics for Local-Stochastic Volatility Model


dXt X0 dt 0 = = = = t c(Xt )dWt , x0 > 0, a(t )dt + b(t )dZt , > 0, (5.17) (5.16)

Consider the following diusion:

where Wt and Zt are two standard Brownian motions with correlation . Proposition 5.3 When t 0, we have (x0 x)2 1 p(t, x0 , x) exp 2 2 20 c(x0 )2 t 20 c(x0 )2 t where

1+

t (t, x0 , x) ,

) 1( (t, x0 , x) = b( ) + 2 c(x0 ) 2 c(x0 )3 2

} h3 ((x x0 )/ t, 2 c(x0 )2 ) . ( 2 c(x0 )2 )3

(Proof) We compute the coecient of the rst order in the expansion by applying (4.15) and (4.19). First, we introduce the time scaling parameter = t, dXt dt = = t c(Xt )dWt , 2 a(t )dt + b(t )dZt . (5.18) (5.19)

The generator A associated with X is given by 1 2 2 1 2 A = 2 2 c(x)2 2 + 2 b( )c(x) + 2 a( ) + 2 b( )2 2 . 2 x x 2 When 0, A is degenerate. We consider the following transform, Y = X x0 . (5.20)

Then the generator L associated with Y is elliptic under 0 and is given by L = and u (t, y0 ) = P t f (y0 ) =
R

1 2 2 2 1 2 c(x0 + y )2 2 + b( )c(x0 + y ) + 2 a( ) + 2 b( )2 2 , 2 y y 2

f (y )pY (t, y0 , dy ),

(5.21)

is the fundamental solution to the following equation,

L u (t, y0 ) t u (t0 , y0 )

= =

0, f (y0 ).

(5.22)

By dierentiating (5.21) at t = 1, we have |=0 P 1 f (y0 ) =

f (y )w1 (y )pY (1, y0 , y )dy,

(5.23)

where the map y w1 (y ) is the rst order PDE weight.

27

1. Using the integration by pats formula, we derive the rst order PDE weight w1 (y ), |=0 P 1 f (y0 ) =

1 1 0 P0 (1s) L Ps f (y0 )ds

=
0 1

1 f (y ) 2 E Y11 c ( x0 )2 R

0 [

c(x0 )dW1,t
0

]
Dt,1 Y11 c(x0 )dt|Y01 = y p(1, y0 , y )dy.

L and L are given as follows; L0 L1 Note that Y0 p (s, y, z ) pY (1 s, y0 , y )y Let g be a map y g (y ) such that g (y ) = L
1
0

= =

1 2 2 c(x0 )2 2 , 2 y 2 c(x0 )c(x0 )y 2 2 + b( )c(x0 ) . y 2 y

= =

2 Y 0 p (s, y, z ), y 2 Y0 (1 s) 2 c(x0 )2 p (1 s, y0 , y )|y0 =0 . y0 sc(x0 )2

=L
1

P0 s f (y )

=L

P0 1(1s) f (y )

E [f (Y10 )|Y10 s

= y] = L

1 R

pY (s, y, z )f (z )dz.

Then, we explicitly evaluate (4.19) for j = 1.


1 0 P0 (1s) L Ps f (y0 )|y0 =0

pY (1 s, y0 , y )g (y )dy |y0 =0

R
=
R

(
pY (1 s, y0 , y ) L1
0

)
pY (s, y, z )f (z )dz
0

dy |y0 =0

(
=
R R

(1 s) 4 c(x0 )3 c(x0 )

(
R

Y0 p (1 s, y0 , y ) y0

2 Y 0 p (s, y, z )dy y 2

)
f (z )dz |y0 =0

+
R (

pY (1 s, y0 , y )s 2 b( )c(x0 )3

(
4 3

3 Y 0 p (s, y, z )dy y 3

f (z )dz |y0 =0

=
R

(1 s) c(x0 ) c(x0 )
R ( (

0 0 3 pY (1 s, y0 , y ) pY (s, y, z )dy y 2 y0

)
f (z )dz |y0 =0

+
R R

0 3 Y 0 p (1 s, y0 , y ) s 2 b( )c(x0 )3 pY (s, y, z )dy y 3

f (z )dz |y0 =0 f (z )dz |y0 =0

(1 s) 4 c(x0 )3 c(x0 ) +s 2 b( )c(x0 )3 3 3 y0

R (

3 3 y0

pY (1 s, y0 , y )pY (s, y, z )dy


R

)
0

pY (1 s, y0 , y )pY (s, y, z )dy 3 y0 p


R 3 R 3

f (z )dz |y0 =0

R
R Y0

(1 s) 4 c(x0 )3 c(x0 ) 3 +s b( )c(x0 ) 3 y0


2 3

pY (1, y0 , z )f (z )dz (1, y0 , z )f (z )dz |y0 =0

(1 s) 4 c(x0 )3 c(x0 )

3 P f (y0 )|y0 =0 + s 2 b( )c(x0 )3 3 P1 f (y0 )|y0 =0 . 3 1 y0 y0

28

Therefore, we have

1 1 0 P0 (1s) L Ps f (y0 )ds|y0 =0

= and

1 2

(
4 c(x0 )3 c(x0 )

3 3 + 2 b( )c(x0 )3 3 3 y0 y0

)
P0 1 f (y0 )|y0 =0 .

|=0 pY (1, y0 , y )

= =

1 2

(
4 c(x0 )3 c(x0 )

3 3 + 2 b( )c(x0 )3 3 3 y0 y0

)
pY (1, y0 , y )
0

) h3 (y, 2 c(x0 )2 ) Y 0 1( p (1, y0 , y ). b( ) + 2 c(x0 ) 2 c(x0 )3 2 ( 2 c(x0 )2 )3

Setting y0 = 0, we obtain the result. 2. Next, we compute the rst order PDE weight by applying (4.15) for j = 1 in the following way. First, X is approximated by stochastic Taylor expansion, Xt where X1t X2t = = x0 + X1t + X2t + R3 (t),


=
0 0

c(x0 )dWs ,
t

c(x0 )
0

b( )dZu dWs +
0

c(x0 )
0

c(x0 )dWu dWs .

X () and Y () are expanded in D , Xt Yt where Y0t = X1t = Xt |=0 =


( )

= =

x0 + X1t + 2 X2t + O(3 ), Y0t + Y1t + O(2 ),

( )

c(x0 )dWs ,

Y1t

X2t =

1 2 Xt |=0 = 2 2

c(x0 )
0

b( )dZu dWs +
0

c(x0 )
0

c(x0 )dWu dWs .

Then, the PDE weight is caluculated as follows, w1 (y ) = = 1 E Y11 2 c(x0 )2

)
0

c(x0 )dW1,t
0

]
Dt,1 Y11 c(x0 )dt|Y01 = y
s

y 2 c(x0 )2 y
4 3

t
( b( )c(x0 )
2 3 0 s 0 0

duds

t
0

h2 (y, 2 c(x0 )2 ) ( 2 c(x0 )2 )2

+ c(x0 ) c(x0 ) = =

h2 (y, 2 c(x0 )2 ) duds ) ( 2 c(x0 )2 )2

h3 (y, 2 c(x0 )2 ) h3 (y, 2 c(x0 )2 ) 1 1 2 b( )c(x0 )3 + 4 c(x0 )3 c(x0 ) . 2 2 3 2 ( c(x0 ) ) 2 ( 2 c(x0 )2 )3

) h3 (y, 2 c(x0 )2 ) 1( . b( ) + 2 c(x0 ) 2 c(x0 )3 2 ( 2 c(x0 )2 )3

The following formula holds, 1 p (1, x0 , x) = p(1, 0, f (x)) ,

29

and we have p(t, x0 , x) = p


t

(1, x0 , x).

Then, we obtain a short time o-diagonal asymptotic expansion of heat kernel, p(t, x0 , x) where (t, x0 , x) =

2 20 c(x0 )2 t

exp

(x0 x)2 2 20 c(x0 )2 t

1+

t (t, x0 , x) ,

) h3 ((x x0 )/ t, 2 c(x0 )2 ) 1( b( ) + 2 c(x0 ) 2 c(x0 )3 . 2 ( 2 c(x0 )2 )3

5.4

Numerical Example

This subsection provides an numerical example for option pricing under the short-time asymptotic expansion. In particular, we use the following Heston model: dSt = vt St dW1,t , (5.24) dvt = ( vt )dt + vt (dW1,t + 1 2 dW2,t ), with parameters S0 = 100, v0 = 0.16, = 1.0, = 0.16, = 0.1, = 0.5. A call option price with strike K and maturity t is approximated as follows; C (t, K ) = E [(St K )+ ] C0 (t, K ) + tC1 (t, K ) + tC2 (t, K ),

(5.25) (5.26)

where C0 (t, K ) C1 (t, K ) C2 (t, K ) =

(ex K )+ p(t, x0 , x)dx,

(ex K )+ w1 (t, x0 , x)p(t, x0 , x)dx,

R
=
R

(ex K )+ w2 (t, x0 , x)p(t, x0 , x)dx, (5.27)

and p(t, x0 , x), w1 (t, x0 , x), w2 (t, x0 , x) are obtained in the similar manner as in Subsection 5.2. Also, put option prices are computed by the Put-Call parity.

Strike 70 Put 80 Put 90 Put 100 Call 110 Call 120 Call 130 Call

Table 1: Short time asymptotics T = 0.1 Benchmark HKE order 2 HKE order 1 0.01 0.01(-17.17%) 0.00 (-55.56%) 0.19 0.19 (-1.35%) 0.18 (-4.47%) 1.38 1.38 (-0.35%) 1.37 (-0.91%) 5.04 5.03 (-0.14%) 5.02 (-0.30%) 1.70 1.69 (-0.38%) 1.68 (-0.84%) 0.44 0.43 (-0.83%) 0.43 (-2.33%) 0.09 0.09 (-1.48%) 0.08 (-6.04%)

HKE order 0 0.81 (8079.80%) 0.95 (396.41%) 2.02 (46.06%) 5.47 (8.66%) 1.93 (13.95%) 0.53 (22.82%) 0.12 (36.91%)

Second Order Approximation in Section 5.1


1 2 Y p (1, 0, y )|=0 = 2 2

1. Applying Bismut identity, the weights of second order approximations are calculated as follows.

0 1 2 E [H2 (Y01 , Y11 )] + E [H1 (Y01 , Y21 )] pY (1, 0, y ). 2

30

Strike 70 Put 80 Put 90 Put 100 Call 110 Call 120 Call 130 Call

Table 2: Short time asymptotics T = 0.2 Benchmark HKE order 2 HKE order 1 0.15 0.14 (-5.86%) 0.12 (-18.18%) 0.86 0.84 (-1.39%) 0.82 (-4.09%) 2.94 2.93 (-0.59%) 2.90 (-1.43%) 7.12 7.09 (-0.31%) 7.07 (-0.66%) 3.46 3.44 (-0.64%) 3.42 (-1.35%) 1.50 1.48 (-1.17%) 1.46 (-2.75%) 0.59 0.58 (-1.91%) 0.56 (-5.41%)

HKE order 0 1.72 (1041.94%) 2.29 (167.57%) 4.14 (40.75%) 8.02 (12.71%) 4.07 (17.67%) 1.87 (24.73%) 0.79 (34.48%)

Strike 70 Put 80 Put 90 Put 100 Call 110 Call 120 Call 130 Call

Table 3: Benchmark 0.44 1.63 4.26 8.70 4.93 2.61 1.31

Short time asymptotics T = 0.3 HKE order 2 HKE order 1 0.42 (-4.64%) 0.38 (-14.33%) 1.61 (-1.63%) 1.56 (-4.60%) 4.22 (-0.84%) 4.18 (-2.00%) 8.66 (-0.49%) 8.61 (-1.06%) 4.89 (-0.90%) 4.84 (-1.91%) 2.58 (-1.50%) 2.53 (-3.38%) 1.27 (-2.31%) 1.23 (-5.83%)

HKE order 0 2.74 (516.85%) 3.71 (127.14%) 6.02 (41.33%) 10.10 (16.01%) 5.96 (20.89%) 3.33 (27.30%) 1.77 (35.59%)

Iterating the Bismut identity, the terms of

E H2 Yt =

1 , ((x0 )t)2 ) | (x0 )Wt = y 2 2 h ( 1 2 y, (x0 )) ((x0 )t)2 , 2 ( 2 (x0 ))2


(0)

2 1 H (Y01 , Y11 ) 2 2

are calculated as follows:

[
E H2 =

(
(0) Yt , (x0 )t (x0 ) (x0 )

)
Ws dWs | (x0 )Wt = y

h4 (y, 2 (x0 )) 1 2 t (x0 )t 3 (x0 ) (x0 ) , 2 ( 2 (x0 ))4

[
E H2 =

(
(0) Yt ,

1 ( (x0 ) (x0 ))2 2

(
0

)2 )
Ws dWs | (x0 )Wt = y

h6 (y, 2 (x0 )) 1 2 t ( (x0 ))2 6 (x0 ) 8 ( 2 (x0 ))6 +

} h4 (y, 2 (x0 )) 1{3 t ( (x0 ))2 4 (x0 ) 2 ( 2 (x0 ))4

h2 (y, 2 (x0 )) 1 . + t2 ( (x0 ) (x0 ))2 4 ( 2 (x0 ))2 The terms of H1 (Y01 , Y21 ) are calculated as follows,

[
E H1 =

(
Yt
(0)

, (x0 ) (x0 )
0

)
Ws ds | (x0 )Wt = y

h2 (y, 2 (x0 )) 1 (x0 ) 2 (x0 )t2 , 2 ( 2 (x0 ))2

[
E H1

(
Yt
(0)

1 , 2 (x0 ) 2 (x0 ) 2

t 2 Ws dWs

)
| (x0 )Wt = y

31

Strike 70 Put 80 Put 90 Put 100 Call 110 Call 120 Call 130 Call

Table 4: Benchmark 0.82 2.41 5.41 10.04 6.21 3.68 2.09

Short time asymptotics T = 0.4 HKE order 2 HKE order 1 0.79 (-4.46%) 0.71 (-13.68%) 2.37 (-1.92%) 2.29 (-5.29%) 5.35 (-1.09%) 5.27 (-2.61%) 9.97 (-0.69%) 9.88 (-1.51%) 6.14 (-1.17%) 6.06 (-2.49%) 3.61 (-1.84%) 3.53 (-4.07%) 2.03 (-2.71%) 1.95 (-6.49%)

HKE order 0 3.83 (364.08%) 5.13 (112.42%) 7.74 (43.03%) 11.94 (18.94%) 7.69 (23.79%) 4.77 (29.87%) 2.87 (37.41%)

h4 (y, 2 (x0 )) 1 2 (x0 ) 2 (x0 ) 3 (x0 )t3 6 ( 2 (x0 ))4 h2 (y, 2 (x0 )) 1 , + 2 (x0 ) 2 (x0 ) (x0 )t2 4 ( 2 (x0 ))2

[
E H1 =

(
(0) Yt , (x0 ) (x0 )

)
sdWs | (x0 )Wt = y

h1 (y, (x0 )) 1 (x0 ) (x0 ) (x0 )t2 , 2 ( 2 (x0 ))2

0 2

[
E H1 =

(
(0) Yt , ( (x0 ))2 (x0 )

t
0 0

)
Wu dWu dWs | (x0 )Wt = y

h4 (y, 2 (x0 )) 1 (x0 )2 (x0 )4 t3 6 ( 2 (x0 ))4

Therefore, we obtain the second approximation term 2 () E [y (Y1 )]|=0 2 h6 (y, 2 (x0 ))) 1 { (x0 )2 6 (x0 ) 8 ( 2 (x0 ))6

) h4 (y, 2 (x0 )) 1( 2 (x0 ) (x0 )5 + 4 (x0 )2 (x0 ) + 3(x0 ) (x0 ) 6 ( 2 (x0 )))4 1 2 + ( (x0 ) (x0 )3 4 h2 (x, 2 (x0 )) +2 (x0 ) (x0 )2 + 2(x0 ) (x0 ) (x0 ) + (x0 )2 (x0 )2 + 2(x0 )2 ) } ( 2 (x0 ))2 p(1, x0 , y ).
+ 2. Using the Lie bracket, the second order term is calculated as follows. 1 2 |=0 P t f (y0 )|y0 =0 2 2

t
=
0

t1 1 0 1 0 P0 (tt1 ) L P(t1 t2 ) L Pt2 f (y0 )dt2 dt1 |y0 =0

0 t 2 0 P0 (tt1 ) L Pt1 f (y0 )dt1 |y0 =0 .

+ The rst term is given by

t
0 0

t1

(L1 + (t t1 )[L0 , L1 ])(L1 + (t t2 )[L0 , L1 ])P0 t f (y0 )dt2 dt1 |y0 =0 , since [L0 , [L0 , L1 ]] = 0.

32

The second term is given by

(L2 + (t t1 )[L0 , L2 ] + because [L , [L , [L0 , L2 ]]] = 0. Then we have


0 0

1 (t t1 )2 [L0 , [L0 , L2 ]])P0 t f (y0 )dt1 |y0 =0 , 2

1 2 |=0 pY (1, 0, y ) = 2 2 1 1 (L2 + [L0 , L2 ] + [L0 , [L0 , L2 ]] + (L1 )2 2 6 0 1 1 1 + L1 [L0 , L1 ] + [L0 , L1 ]L1 + [L0 , L1 ]2 )pY (1, y0 , y )|y0 =0 , 2 6 8

(A.1)

where
0 1 1 pY (1, y0 , y ) = e 2 2 (x0 )2

( y y )2
0 (x0 )

Each terms are calculated as follows. L2 pY (1, y0 , y )|y0 =0 = 0,


0 1 0 2 Y0 1 1 1 [L , L ]p (1, y0 , y )|y0 =0 = (x0 )2 (x0 ) + (x0 )2 + 2 (x0 ) (x0 ) 2 pY (1, y0 , y )|y0 =0 , 2 2 2 2 0

( ) 0 0 1 0 1 [L , [L0 , L2 ]]pY (1, y0 , y )|y0 =0 = (x0 )4 (x0 )2 + 2 (x0 ) (x0 ) 4 pY (1, y0 , y )|y0 =0 , 6 6
0 1 ( (x0 ) (x0 )(x0 ) + (x0 )2 ) 2 pY (1, y0 , y )|y0 =0 2

(L1 )2 p(1, y0 , y )|y0 =0 =

0 1 1 1 0 1 Y0 L [L , L ]p (1, y0 , y )|y0 =0 = (x0 ) (x0 )3 (x0 ) 4 pY (1, y0 , y )|y0 =0 , 2 3

0 1 0 1 1 Y0 1 [L , L ]L p (1, y0 , y )|y0 =0 = ((x0 ) + 3 (x0 ) (x0 )) (x0 )3 (x0 ) 4 pY (1, y0 , y )|y0 =0 , 6 6

0 1 0 1 2 Y0 1 [L , L ] p (1, y0 , y )|y0 =0 = (x0 )6 (x0 )2 6 pY (1, y0 , y )|y0 =0 . 8 8

Hence, we have
1 2 |=0 pY (1, 0, y ) 2 2 h6 (y, 2 (x0 )) 1 { (x0 )2 6 (x0 ) 8 ( 2 (x0 ))6

+ +

) h4 (y, 2 (x0 )) 1( 2 (x0 ) (x0 )5 + 4 (x0 )2 (x0 ) + 3(x0 ) (x0 ) 6 ( 2 (x0 ))4 ) h2 (y, 2 (x0 )) 1( 2 } (x0 ) (x0 )3 + 2 (x0 ) (x0 )2 + 2(x0 ) (x0 ) (x0 ) + (x0 )2 (x0 )2 + 2(x0 )2 4 ( 2 (x0 ))2
0

pY (1, 0, y ). Therefore, we obtain the result.

(A.2)

33

References
[1] Baudoin, F.: Stochastic Taylor Expansions and Heat Kernel Asymptotics. (2009) [2] Ben Arous, G. and Laurence, P.: Second order expansion for implied volatility in two factor local stochastic volatility models and applications to the dynamic -SABR model.(2009) [3] Black, F. and Scholes, M. (1973). : The Pricing of Options and Corporate Liabilities, Journal of Political Economy 81 (3): 637-654. (1973) [4] Cheng, W., Costanzino,N., Liechty,J., Mazzucato,A. and Nistor,V.: Closed-form Asymptotics and Numerical Approximations of 1D Paraolic Equations with Applications to Option Pricing, Preprint, (2010). [5] Cheng, W., Mazzucato,A. and Nistor,V.: Approximate Solutions to Second Order Parabolic Equations II: Time-dependent Coecients, Preprint, (2011). [6] Ilhan A., Jonsson, M. and Sircar, K.R.: Singular Perturbations for Boundary Value Problems arising from Exotic Options, SIAM Journal on Applied Mathematics, 64 (2004) 1268-1293. [7] Fourni e, E., Lasry, J.-M., Lebuchoux, J., Lions, P.-L. and Touzi, N. : Applications of Malliavin calculus to Monte-Carlo methods in Finance, Finance Stoch. 3 (1999), no. 4, 391-412. [8] Fourni e, E., Lasry, J.-M., Lebuchoux, J. and Lions, P.-L. : Applications of Malliavin calculus to MonteCarlo methods in Finance II, Finance Stoch. 5 (2001), no. 2, 201-236. [9] Fujii,M. and Takahashi,A.: Analytical Approximation for Non-linear FBSDEs with Perturbation Scheme, Working paper, CARF-F-248, University of Tokyo (2011), forthcoming in International Journal of Theoretical and Applied Finance. [10] Gatheral, J., Hsu, E.P. and Laurence, P., Ouyang, C. and Wang, T-H.: Asymptotics of implied volatility in local volatility models, forthcoming in Mathematical Finance (2009) [11] Hagan, P.S., Kumar, D., Lesniewskie, A.S., and Woodward, D.E.: Managing Smile Risk, Wilmott magazine, (2002). [12] Ikeda, N., and Watanabe, S.: Stochastic Dierential Equations and Diusion Processes, Second Edition, North-Holland/Kodansha. (1989) [13] Kato, T., Takahashi, A., and Yamada, T.: An Asymptotic Expansion for Solutions of Cauchy-Dirichlet Problem for Second Order Parabolic PDEs and Its Applications to Pricing Barrier Options, Working paper, CARF-F-272, University of Tokyo. (2012) [14] Labordere, P.H.: Analysis, Geometry and Modeling in Finance : Advanced Methods in Options Pricing, Chapman and Hall, (2008). [15] L eandre, R.: Malliavin Calculus of Bismut type without probability (2006) [16] L eandre, R.: Malliavin Calculus of Bismut type in semi-group theory (2008) [17] Malliavin, P. : Stochastic Analysis, Springer. (1997) [18] Malliavin, P. and Thalmaier, A.: Stochastic Calculus of Variations in Mathematical Finance, Springer. (2006) [19] Merton, R. C. :Theory of Rational Option Pricing, Bell Journal of Economics and Management Science (The RAND Corporation) 4 (1): 141-183. (1973) [20] Takahashi, A., Takehara, K. and Toda, M.: A General Computation Scheme for a High-Order Asymptotic Expansion Method, Working paper, CARF-F-272, University of Tokyo. (2012), forthcoming in International Journal of Theoretical and Applied Finance. [21] Takahashi, A. and Yamada,T.: An Asymptotic Expansion with Push-Down of Malliavin Weights, SIAM Journal on Financial Mathematics Volume 3, pp 95-136, (2012) [22] Takahashi, A. and Yamada,T.: On Approximation of the Solutions to Partial Dierential Equations in Finance, CARF Working Paper Series CARF-F-249, University of Tokyo. (2010) [23] Watanabe, S.: Analysis of Wiener Functionals (Malliavin Calculus) and its Applications to Heat Kernels, The Annals pf Probability, Vol.15-1, 1-39. (1987)

34