HAL Id: hal-00979936
https://hal-upec-upem.archives-ouvertes.fr/hal-00979936v2
Submitted on 7 Jul 2016
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
The critical price of the American put near maturity in the jump diffusion model
Aych Bouselmi, Damien Lamberton
To cite this version:
Aych Bouselmi, Damien Lamberton. The critical price of the American put near maturity in the
jump diffusion model. SIAM Journal on Financial Mathematics, Society for Industrial and Applied
Mathematics 2016, 7 (1), pp.236-272. �10.1137/140965910�. �hal-00979936v2�
The Critical Price of the American Put Near Maturity in the Jump Diffusion Model
∗Aych Bouselmi
† andDamien Lamberton
†Abstract. We study the behavior of the critical price of an American put option near maturity in the jump diffusion model when the underlying stock pays dividends at a continuous rate and the limit of the critical price is smaller than the stock price. In particular, we prove that, unlike the case where the limit is equal to the strike price, jumps can influence the convergence rate.
Key words. American put, L´ evy processes, critical price, free boundary, jump diffusion, convergence rate AMS subject classifications. 60H30, 60J75, 91G80
DOI. 10.1137/140965910
1. Introduction. The behavior of the critical price of the American put near maturity has been deeply investigated. Its limit was characterized in the Black–Scholes model (see [5, 13]) by
b(T ) := lim
t→T
b(t) = min r
δ K, K
,
where r and δ denote the interest rate and the dividend rate and b(t) is the critical price at time t.
This result was generalized to more general exponential L´ evy models in [7]. In fact, denoting ¯ d = r − δ − R
(e
y− 1)
+ν (dy),
1with ν the L´ evy measure of the underlying L´ evy process, we have
b(T ) = K if d ¯ ≥ 0 and
b(T ) = ξ if d < ¯ 0, where ξ is the unique solution, in [0, K], of
(1) rK − δx −
Z
(xe
y− K)
+ν (dy) = 0.
In the Black–Scholes model, the quantity ¯ d reduces to ¯ d = r − δ and we distinguish, according as ¯ d > 0, ¯ d = 0, and ¯ d < 0, different behaviors of the critical price near maturity.
∗Received by the editors April 21, 2014; accepted for publication (in revised form) February 2, 2016; published electronically May 5, 2016.
http://www.siam.org/journals/sifin/7/96591.html
†Universit´e Paris-Est, Laboratoire d’Analyse et de Math´ematiques Appliqu´ees (UMR 8050), UPEM, UPEC, CNRS, Projet Mathrisk INRIA, F-77454, Marne-la-vall´ee, France (aych.bouselmi@gmail.fr,damien.lamberton@u-pem.fr).
1
The quantity ¯
dis denoted by
d+in [7].
236
In fact, Barles et al in [1] (see also Lamberton [6]) established, in the case where ¯ d > 0 (which implies b(T ) = K), that
(2) K − b(t)
σK ∼
t→Tp (T − t)| ln(T − t)|, where the expression f ∼
t→ag (or f ∼
ag) is equivalent to lim
t→af(t)g(t)
= 1. The cases ¯ d < 0 and ¯ d = 0 were investigated by Lamberton and by Villeneuve in [14] and they obtained as follows: If ¯ d = 0 (which also implies b(T ) = K)
K − b(t)
σK ∼
t→Tp
2(T − t)| ln(T − t)|.
If ¯ d < 0 (b(T ) < K), there exists y
0∈ (0, 1), which is characterized thanks to an auxiliary optimal stopping problem, such that
b(T) − b(t)
σb(T ) ∼
t→Ty
0p (T − t).
Note that y
0can also be characterized more explicitly as the solution of an equation (see [14]).
The critical price has also been studied in the jump diffusion model. In fact, Pham proved in [11] that the result (2), obtained in [1, 6], remains exactly the same in the jump diffusion model, in the case where ¯ d > 0 and δ = 0. This remains true if δ > 0 (see [10]).
The purpose of this paper is to study the convergence rate of the critical price of the American put, in the jump diffusion model, with ¯ d ≤ 0. Considering the results of Pham in [11], we expect to obtain the same results as the study performed by Lamberton and Villeneuve in the Black–Scholes model when ( ¯ d = r − δ ≤ 0 ), meaning that jumps do not have any influence on the convergence rate. Surprisingly, we obtain the expected result only for the case ¯ d = 0. Indeed, we obtain for ¯ d = 0 (see Theorem 5.1)
K − b(t)
σK ∼
t→Tp
2(T − t)| ln(T − t)|, and for ¯ d < 0 (see Theorem 4.7)
b(T ) − b(t)
σb(T ) ∼
t→Ty
λ,βp
(T − t),
where y
λ,βis a real umber satisfying y
λ,β≥ y
0, and depending on ν({ln(K/b(T ))}) we can have y
λ,β> y
0. This point will be discussed in more detail in section 4.3.
This study is composed of four sections. In section 1, we recall some useful results on the
American put which will be used throughout this study. In section 2, we give some results
on the regularity of the American put price and the early exercise premium. In section 3, we
investigate the case where the limit of the critical price is far from the singularity K (we refer
to this as the regular case). Indeed, we then have enough regularity to give an expansion of the
American put price near maturity from which the critical price behavior will be deduced. The
method is similar to the one used in [14] and is based on an expansion of the American put
price along parabolas. However, the possibility that the stock price jumps into a neighborhood of the exercise price produces a contribution of the local time in the expansion. Section 4 is devoted to the study of the case ¯ d = 0. In this case b(T) = K; hence we no longer have enough smoothness to obtain an expansion around the limit point (T, b(T )). Then we will study the behavior of the European critical price b
e(t) instead of b(t). Thereafter, we prove that b(t) and b
e(t) have the same behavior.
2. Preliminaries. In the jump diffusion model, under a risk-neutral probability which is used as a pricing measure, the risky asset price is modeled by (S
t)
t≥0, given by
S
t= S
0e
X˜twith ˜ X
t= (r − δ)t + σB
t− σ
22 t + Z
t− t Z
(e
y− 1)ν(dy),
where r > 0 is the interest rate, δ ≥ 0 the dividend rate, (B
t)
t≥0a standard Brownian motion, and (Z
t)
t≥0a compound Poisson process and ν its L´ evy measure. We then have
dS
t= S
t−γ
0dt + σdB
t+ d Z ¯
twith ¯ Z
t= Σ
0<s≤t(e
∆Zs− 1) and γ
0= r − δ − Z
(e
y− 1)ν(dy).
Denote by F the completed natural filtration of the process ˜ X
tand suppose throughout this paper that the following assumptions are satisfied:
σ > 0, ν( R ) < ∞, Z
e
yν(dy) < ∞, and d ¯ = r − δ − Z
y>0
(e
y− 1)ν(dy) ≤ 0.
The price of an American put with maturity T > 0 and strike price K > 0 is given, at t ∈ [0, T ], by P (t, S
t) with P defined for all (t, x) ∈ [0, T ] × R
+by
P (t, x) = sup
τ∈T0,T−t
E
e
−rτψ(xe
X˜τ)
with ψ(y) = (K − y)
+, where T
0,T−tis the set of all F -stopping times taking values in [0, T − t].
The value function P can also be characterized (see [7]) as the unique continuous and bounded solution of the variational inequality
max
ψ − P ; ∂P
∂t + AP − rP
= 0 (in the sense of distributions)
with the terminal condition P (T, .) = ψ. Here, A is the infinitesimal generator of the process S. The free boundary of this variational inequality is called the exercise boundary, and at each t ∈ [0, T ], the critical price is given by
b(t) = inf
x > 0 | P(t, x) > (K − x)
+. It was proved in [7] that if ¯ d ≤ 0, then
(3) lim
t→T
b(t) = ξ := b(T ),
where ξ is the unique solution, in [0, K], of rK = δx + R
(xe
y− K)
+ν(dy). Note that if ¯ d = 0, then b(T ) = ξ = K.
Finally, recall that the price of a European put with maturity T and strike price K is given, at time t, by
P
e(t, x) = E
e
−r(T−t)(K − S
T−t)
+| S
0= x
.
The quantity (P − P
e) is called the early exercise premium; we then have P (t, x) = P
e(t, x) + e(T − t, x). Setting θ = T − t, then the early exercise premium, e(θ, x), is characterized for the American put in the exponential Levy model as follows (see [10]):
e(θ, x) = E Z
θ0
e
−rs×
rK − δS
sx− Z
y>0
[P (t + s, S
sxe
y) − (K − S
sxe
y)] ν(dy)
1
{Sxs<b(t+s)}
ds
.
Here S
sx= xe
X˜s. We also define, for all t ∈ (0, T ), the European critical price, b
e(t), as the unique solution of
F(t, x) = P
e(t, x) − (K − x) = 0.
It easy to check that for all t ∈ (0, T ), b
e(t) is well defined, b
e(t) ∈ (0, K). It is also straight- forward that P
e≤ P ; therefore b(t) ≤ b
e(t) ≤ K.
3. Regularity estimate for the value function in the jump diffusion model. In this section, we study the spatial derivatives behavior of P, P
e, and e(θ, x) near (T, b(T)). We also give a lower bound for the second spatial derivative near (T, b(T )). These results will be proved in Appendix A.
Lemma 3.1. Under the model assumption, we have the following 1. For all x ∈ (0, b
e(t) ∧ b(T )], we have, as θ(= T − t) goes to 0,
∂e
∂x (θ, x)
= 1 x o( √
θ)
with o( √
θ) uniform with respect to x.
2. For all x ∈ (0, b(T ) ∧ b
e(t)], we have
∂P
∂x (t, x) + 1 =
1 + 1 x
o(
√ θ)
with o( √
θ) uniform with respect to x.
Lemma 3.2. According to the hypotheses of the model, we have, for all b(t) ≤ x < b(T) ∧ b
e(t) and for all θ = T − t small enough, the following inequality:
inf
b(t)<u<x
u
2σ
22
∂
2P
∂x
2(t, u) ≥
δ ˆ − (θ)
(b(T ) − x) − λK E
σB
θ− ln b(T )
x
++ o(
√ θ)
with lim
θ↓0(θ) = 0, ˆ δ = δ + R
{y>ln(K/b(T))}
e
yν(dy), λ = ν{ln(
b(TK))}.
4. Regular case. We begin this section by introducing an auxiliary optimal stopping problem which will be needed for deriving the expansion of the American put price near maturity along a parabolic branch. Once we have this expansion we will be able to derive the convergence rate of the critical price.
4.1. An auxiliary optimal stopping problem. Let β be a nonnegative number and (B
s)
s≥0be a standard Brownian motion with local time at x denoted by ˜ L
x. We denote by T
0,1the set of all σ(B
t; t ≥ 0)-stopping times with values in [0, 1]. Consider also a Poisson process (N
s)
s≥0, independent of B, with intensity λ; we denote by ˆ T
1its first jump time and by ˆ T
0,1the set of all σ ((N
t, B
t) ; t ≥ 0)-stopping times with values in [0, 1]. We define the functions υ
λ,βas follows:
v
λ,β(y) = sup
τ∈Tˆ0,1
E
e
λτ1
{Nτ=0}Z
τ 0f
λβ(y + B
s)ds + β
2 e
λτ1
{Nτ=1}L
−yτ(B) − L
−yˆT1
(B)
,
where f
a(x) = x + ax
+. Notice that v
λ,βis a nonnegative function. Moreover, we have the following.
Lemma 4.1. Define
y
λ,β= − inf{x ∈ R | v
λ,β(x) > 0}.
We have 0 < y
λ,β< 1 + λβ(2 + e
λ) and
∀y < −y
λ,β, v
λ,β(y) = 0.
We finish this subsection with an inequality, which will be used to derive a lower bound for the second derivative of P (see the proof of the upper bound in Theorem 4.7).
We define the function C on R by C(x) = x − λβ E(B
1− x)
+and we have the following lemma.
Lemma 4.2. For all x > y
λ,β, we have
C(x) > 0.
These results will be proved in Appendix B.
4.2. American put price expansion. Throughout this section, we assume ¯ d < 0, so that b(T ) < K . We then have enough regularity of the American put price to derive an expansion of P around b(T ) along a certain parabolic branch.
Theorem 4.3. Let a be a negative number (a < 0) and let b(T) denote the limit of b(t) when t goes to T , b(T ) = lim
t→Tb(t). If d < ¯ 0, we have
P (T − θ, b(T )e
a√
θ
) = (K − b(T )e
a√
θ
)
++ Cθ
32υ
λ,βa σ
+ o(θ
32), where C = σb(T )ˆ δ, with λ = ν{ln
b(TK)}, δ ˆ = δ + R
y>ln(K/b(T))
e
yν(dy) and υ
λ,β(y) as defined in the previous section with β =
Kb(T)ˆδ
.
Remark 1. Notice that if ν does not charge ln(K/b(T )), meaning that λ = 0 and ˆ T
1=
∞ a.s., then
υ
λ,β(a) = υ
0(a) = sup
τ∈T0,1
E Z
τ0
(a + B
s) ds
.
In this case, the American put price will have the same expansion as in the Black–Scholes model (see [9]).
The proof of Theorem 4.3 will require some estimates for the local time at K of the process (S
t)
t≥0, which we denote by (L
Kt)
t≥0. In fact, the main difference with the approach used in the Black–Scholes case lies in this analysis of the local time. The process L
Kcan be derived from the Ito–Tanaka formula (see [12]):
(K − S
t)
+= (K − S
0)
++ Z
t0
(−1
{Ss≤K})S
s(γ
0ds + σdB
s) (4)
+ X
0<s≤t
(K − S
s)
+− (K − S
s−)
++ 1
2 L
Kt.
The following proposition provides an expansion of E L
Kτfor stopping times τ with values in [0, θ] with θ close to 0.
Proposition 4.4. We assume b(T ) < K . Let a be a fixed negative number, a < 0, and S
0= b(T )e
a√θ
.
• We have lim sup
θ↓0
E(LKθ )
θ3/2
< ∞. Moreover, if ν{ln(
b(TK)} = 0, lim sup
θ↓0
E(LKθ) θ3/2
= 0.
• If ν{ln(K/b(T))} 6= 0, then we have, for all F -stopping time τ with values in [0, θ], E L
Kτ= 2K E h
(−a √
θ − σB
τ)
+− (−a √
θ − σB
Tˆ1
)
+1
{Tˆ1<τ}
i
+ o(θ
32), where T ˆ
1= inf{s ≥ 0 ; ∆X
s= ln(
b(TK))} and o(θ
3/2) is uniform with respect to τ . For the proof of Proposition 4.4, we will need an elementary estimate for the expectation of the local time of Brownian motion.
Lemma 4.5. For all real number a and for all t > 0, we have 0 ≤ E (a − B
t)
+− a
+≤ √
t e
−a2
√
2t2π .
Proof of Lemma 4.5. The first inequality follows from Jensen’s inequality. For the other inequality, we have
E (a − B
t)
+= Z
a/√t
−∞
(a − √
ty)e
−y2/2dy
√ 2π
= a Z
a/√t
−∞
e
−y2/2dy
√
2π + √ t e
−a2
√
2t2π . Then, if a ≤ 0,
E(a − B
t)
+≤ √ t e
−a2
√
2t2π . If a ≥ 0, we can write
E (a − B
t)
+− a = −a Z
+∞a/√ t
e
−y2/2dy
√
2π + √ t e
−a2
√
2t2π ≤ √ t e
−a2
√
2t2π .
Before proving Proposition 4.4, we state and prove an estimate for E (L
Kθ| S
0= x).
Lemma 4.6. There exists a positive constant C such that for all x > 0 and for all θ > 0, we have
E L
Kθ| S
0= x
≤ Cx √
θ exp
− (K − x)
22x
2σ
2θ
+ θ
.
Proof of Lemma 4.6. We will use the notation E
xfor E (. | S
0= x). Taking expectations in (4) and using the compensation formula (see, for instance, [3]), we have
1
2 E
xL
Kθ= E
x(K − S
θ)
+− (K − S
0)
++ E
xZ
θ 0γ
0S
s1
{Ss≤K}− Z
Ψ(S
s, y)ν(dy)
ds
,
where Ψ(x, y) = (K − xe
y)
+− (K − x)
+. We deduce easily from this equality that
1
2 E
xL
Kθ= E
x(K − S
θ)
+− (K − x)
++ xO(θ) with O(θ) independent of x. We have, with the notation ˜ Z
θ= Z
θ− θ R
(e
y− 1)ν(dy), E
x(K − S
θ)
+− (K − x)
+= E
x(K − xe
(r−δ−σ2
2 )θ+σBθ+ ˜Zθ
)
+− (K − x)
+. We also have
E
e
(r−δ−σ2
2 )θ+σBθ+ ˜Zθ
− e
σBθ= e
σ2θ/2E
e
(r−δ−σ2
2 )θ+ ˜Zθ
− 1
= O(θ).
Therefore
E
x(K − S
θ)
+− (K − x)
+= E (K − xe
σBθ)
+− (K − x)
++ xO(θ)
= E (K − x(1 + σB
θ))
+− (K − x)
++ xO(θ)
= xσ E
K − x xσ − B
θ +−
K − x xσ
+!
+ xO(θ).
Hence, using Lemma 4.5 above,
E
x(K − S
θ)
+− (K − x)
+≤ xσ p
θ/(2π) exp
− (K − x)
22x
2σ
2θ
+ xO(θ).
Proof of Proposition 4.4. Let T
1be the first jump time of the process Z . We will use the following decomposition:
L
Kθ= L
Kθ∧T1+ L
Kθ− L
Kθ∧T1= L
Kθ∧T1+ 1
{T1<θ}L
Kθ− L
KT1.
Estimating E L
Kθ∧T1
. In the stochastic interval [0, T
1[, the process (S
t) matches the process ( ˇ S
t) defined by
S ˇ
t= S
0e
(γ0−σ2
2 )t+σBt
.
We deduce (when observing that the process L
Kis continuous) that L
Kθ∧T1= ˇ L
Kθ∧T1≤ L ˇ
Kθ,
where ˇ L
Kis the local time at K of the process ˇ S. Note that 1
2
L ˇ
Kθ= (K − S ˇ
θ)
+− (K − S
0)
+− Z
θ0
(−1
{Sˇs≤K}
) ˇ S
s(γ
0ds + σdB
s).
As the process ( ˇ L
Kθ) increases only on { S ˇ
t= K}, we have L ˇ
Kθ= ˇ L
Kθ1
{ˇτK<θ},
where ˇ τ
K= inf {t ≥ 0; ˇ S
t> K}. Note that ˇ τ
K= inf {t ≥ 0; (γ
0−
σ22)t + σB
t> ln(K/S
0)}, so that, with our assumptions on S
0, we have P (ˇ τ
K≤ θ) = o(θ
n) for all n > 0. By H¨ older,
E L ˇ
Kθ≤ ( P (ˇ τ
K< θ))
1−1p|| L ˇ
Kθ||
p, p > 1.
We easily deduce that E L ˇ
Kθ= o(θ
n) for all n > 0, so that EL
Kθ∧T1= o(θ
n) for all n > 0.
Estimating E
1
{T1<θ}L
Kθ− L
KT1
. By the strong Markov property, we have E
1
{T1<θ}L
Kθ− L
KT1≤ E
1
{T1<θ}L
KT1+θ− L
KT1= E
1
{T1<θ}E
ST1(L
Kθ) , (5)
where E
xis the expectation associated to P
xand P
xdefines the law of S
twhen S
0= x.
Using Lemma 4.6, we deduce E
1
{T1<θ}L
KT1+θ− L
KT1≤ C E 1
{T1<θ}S
T1√
θ exp − (K − S
T1)
22S
T21
σ
2θ
! + θ
!!
. (6)
At this stage, we notice that P (T
1≤ θ) = O(θ) and that, conditionally on {T
1≤ θ}, T
1is uniformly distributed on [0, θ].
As Z
T1is independent of both T
1and B , we see that, conditionally to {T
1< θ}, S
T1has the same law as the random variable ζ
θdefined by
ζ
θ= K exp
(V − ln(K/b(T ))) +
√ θ
a +
γ
0− σ
22
√
θU + σg
√ U
,
where U , g, and V are independent random variables, U is uniform on [0, 1], g is standard Gaussian, and V has the same law as Z
T1. Therefore, the estimate (6) can be rewritten as follows:
E
1
{T1<θ}L
KT1+θ− L
KT1≤ C √
θ P (T
1< θ) E
ζ
θexp
− (K − ζ
θ)
22ζ
θ2σ
2θ
+ √
θ (7)
≤ C
√
θ P (T
1< θ) E
ζ
θ1 +
√ θ
.
(8)
Clearly, with probability one,
θ→0
lim ζ
θ= K exp
V − ln K
b(T )
,
and we easily deduce from (8) that E
1
{T1<θ}L
KT1+θ− L
KT1= O(θ
3/2). We can now conclude that E (L
Kθ) = O(θ
3/2).
Moreover, if we assume ν{ln(
b(TK))} = 0, we have V − ln(
b(TK)) 6= 0 a.s., so that lim
θ→0ζ
θ6=
K a.s. and, by dominated convergence, lim
θ↓0E
ζ
θexp
− (K − ζ
θ)
22ζ
θ2σ
2θ
= 0.
Therefore, using (7), E
1
{T1<θ}L
KT1+θ
− L
KT1
= o(θ
3/2), hence E
L
Kθ| S
0= b(T )e
a√ θ
= o(θ
3/2).
Expansion of E L
Kτ, in the case where ν{ln(
b(TK))} > 0. For the proof of the second part of the proposition, we assume ν{ln(
b(TK))} > 0, and we introduce the processes ˆ X and ˆ Z such that
Z ˆ
t= X
s<t
∆ ˜ X
s1
{∆ ˜Xs=lnb(T)K }
and X ˆ = ˜ X − Z, ˆ and ˆ T
1= inf {s ≥ 0, Z ˆ
t6= 0}.
For any stopping time with values in [0, θ], we have E L
Kτ= E L
Kτ∧T1+ E L
Kτ− L
Kτ∧T1= E L
Kτ− L
Kτ∧T1+ o(θ
3/2), where we have used the inequality E L
Kτ∧T1≤ E L
Kθ∧T1and the fact (observed in the first step of the proof) that E L
Kθ∧T1
= o(θ
3/2).
We now observe that since T
1≤ T ˆ
1and τ ≤ θ, 0 ≤ E
L
Kτ∧Tˆ1
− L
Kτ∧T1≤ E
L
Kθ∧Tˆ1
.
On the stochastic interval [0, T ˆ
1), the process ˜ X matches the process ˆ X whose L´ evy measure does not charge the point ln(K/b(T )). So, we have E (L
Kθ∧ˆT1
) = E ( ˆ L
Kθ∧ˆT1
) ≤ E ( ˆ L
Kθ), where L ˆ
Kis the local time at K of the process S obtained by replacing ˜ X with ˆ X in the definition of S. Since the L´ evy measure of ˆ X does not charge ln(K/b(T )), we deduce from the above discussion that E ( ˆ L
Kθ) = o(θ
3/2). Hence
E L
Kτ= E
L
Kτ− L
Kτ∧Tˆ1
+ o(θ
3/2).
Going back to (4) and using again the compensation formula, we have 1
2 E
L
Kτ− L
Kτ∧Tˆ1
= E
(K − S
τ)
+− (K − S
τ∧Tˆ1)
++E
Z
τ τ∧Tˆ1γ
0S
s1
{Ss≤K}− Z
Ψ(S
s, y)ν(dy)
ds
with Ψ(x, y) = (K − xe
y)
+− (K − x)
+. Note that
E
Z
τ τ∧Tˆ1γ
0S
s1
{Ss≤K}− Z
Ψ(S
s, y)ν(dy)
ds
≤ E
1
{Tˆ1<θ}Z
θ 0j(S
s)ds
with j(z) = |γ
0|z + R
|Ψ(z, y)|ν (dy). Since the function Ψ is bounded, we easily derive E (1
{Tˆ1<θ}
R
θ0
j(S
s)ds) = O(θ
2), so that 1
2 E
L
Kτ− L
Kτ∧ˆT1
= E
(K − S
τ)
+− (K − S
τ∧Tˆ1
)
++ O(θ
2)
= E h
1
{Tˆ1<τ}
(K − S
τ)
+− (K − S
Tˆ1
)
+i
+ O(θ
2).
We now argue that up to O(θ
2), we have at most one jump before θ. More precisely, let (N
t)
t≥0be the counting process of the jumps of Z, so that
N
θ= X
0<s≤θ
1
{∆ ˜Xs6=0}
= X
s≤θ
1
{∆Zs6=0}.
We have P (N
θ≥ 2) = O(θ
2), so that E
h 1
{Tˆ1<τ}
(K − S
τ)
+− (K − S
Tˆ1
)
+i
= E h
1
{Tˆ1<τ,Nθ≤1}(K − S
τ)
+− (K − S
Tˆ1)
+i
+ O(θ
2).
On { T ˆ
1< τ, N
θ≤ 1}, we have, for ˆ T
1≤ s ≤ θ,
S
s= S
0e
Xˆs+ ˆZTˆ1= S
0e
XˆsK/b(T ) = Ke
a√
θ+ ˆXs
= Ke
a√
θ+µs+σBs
,
where µ = γ
0−
σ22. Therefore 1
2 E
L
Kτ− L
Kτ∧Tˆ1
= K E h
1
{Tˆ1<τ}
(1 − e
a√
θ+µτ+σBτ
)
+− (1 − e
a√
θ+µTˆ1+σBTˆ1
)
+i
+ O(θ
2)
= K E h
(−a √
θ − µτ − σB
τ)
+−(−a √
θ − µ T ˆ
1− σB
Tˆ1
)
+1
{Tˆ1<τ}
i
+O(θ
2)
= K E
h (−a √
θ − σB
τ)
+− (−a √
θ − σB
Tˆ1
)
+1
{Tˆ1<τ}
i
+ O(θ
2).
The last two equalities follow from P ( ˆ T
1< τ ) = O(θ),
(1 − e
x+ x)1
{x≤0}≤
x22, and the fact
that E (B
τ)
2≤ θ.
Proof of Theorem 4.3. First of all, we recall our notation ˇ X
t= ˜ X
t− Z
t, ˇ S
t= ˜ S
t/e
Zt, and T
1the first jump time T
1= inf{t > 0|Z
t6= 0}, and from now on we consider S
0as a function of θ. More precisely, we denote by S
0θ= b(T )e
a√θ
= e
x0+a√θ
with a < 0 and x
0= ln(b(T )).
We deduce from (4) and the compensation formula that for all stopping times τ ∈ T
0,θ, E
e
−rτ(K − S
τ)
+− (K − S
0)
+= E Z
τ0
e
−rs1
{Ss≤K}−rK + δS
s+ S
sZ
(e
y− 1)ν(dy)
+ e
−rsZ
(K − S
se
y)
+− (K − S
s)
+ν(dy)
ds
+ 1
2 E Z
τ0
e
−rsdL
Ks= I
a(τ ) + J
a(τ ), (9)
where, with the notation Ψ(x, y) = (K − xe
y)
+− (K − x)
+, I
a(τ ) = E
Z
τ 0e
−rs1
{Ss≤K}−rK + δS
s+ S
sZ
(e
y− 1)ν(dy)
+ e
−rsZ
Ψ(S
s, y)ν(dy)
ds and
J
a(τ ) = 1 2 E
Z
τ 0e
−rsdL
Ks.
Note that since τ ≤ θ, E R
τ0
(1 − e
−rs)dL
Ks≤ rθE(L
Kθ), so that using Proposition 4.4, J
a(τ ) = 1
2 E L
Kτ+ O(θ
1+32)
= KE h
(−a √
θ − σB
τ)
+− (−a √
θ − σB
Tˆ1)
+1
{Tˆ1<τ}i
+ o(θ
32).
(10)
Estimating I
a. First of all, note that we have E
Z
τ 0e
−rs1
{Ss>K}Z
(K − S
se
y)
+− (K − S
s)
+ν(dy)
ds
≤ Kν(R) Z
θ0
P {S
s> K}ds
≤ Kν( R ) Z
θ0
P {S
s> K, T
1> θ} + P {S
s> K, T
1≤ θ}ds
≤ Kν( R ) Z
θ0
P { S ˇ
s> K}ds + θ P {T
1≤ θ}
= O(θ
2).
Here, we have used the fact that with the notation ˇ τ
K= inf{t ≥ 0; ˇ S
t> K }, P( ˇ S
s> K ) = P (ˇ τ
K≤ s) ≤ P (ˇ τ
K≤ θ) = o(θ
n) for all n > 0, as observed in the proof of Proposition 4.4.
We can now write I
a(τ ) = E
Z
τ 0e
−rs1
{Ss≤K}−rK + δS
s+ Z
(S
s(e
y− 1) + Ψ(S
s, y)) ν (dy)
ds
+ O(θ
2)
= E Z
τ0
e
−rs1
{Ss≤K}−rK + δS
s+ Z
(S
se
y− K)
+ν (dy)
ds
+ O(θ
2),
where the last equality follows from 1
{x≤K}x(e
y− 1) +
(K − xe
y)
+− (K − x)
+= (xe
y− K)
+1
{x≤K}.
We can also omit e
−rsin the expression as an error of the order of O(θ
2). Then we obtain, for all stopping times τ with values in [0, θ],
I
a(τ ) = E Z
τ0
1
{Ss≤K}−rK + δS
s+ Z
(S
se
y− K)
+ν (dy)
ds
+ O(θ
2).
We denote
h(x) = −rK + δe
x+ Z
(e
xe
y− K)
+ν(dy) and recall that S
t= S
0θe
X˜t= b(T )e
a√
θ+ ˜Xt
= b(T)e
X˜a√ θ
t
= e
x0+ ˜Xa√ θ
t
, where ˜ X
ty= y + ˜ X
t. We thus have
I
a(τ ) = E Z
τ0
1
{a√θ+ ˜Xs≤lnb(T)K }
h(x
0+ a
√
θ + ˜ X
s)ds
| {z }
(I)
+o(θ
32).
(11)
Now, we will try to express the quantity (I ) under a more appropriate form. The first step is to neglect the contribution of the finite variation part of the process ˜ X. Notice that
1
{x≤ln(K)}h(x)
≤ K(r ∨ | d|) ¯ and |h(x) − h(y)| ≤ |e
x− e
y|
δ + Z
e
yν(dy)
.
Moreover, for all (x, y) ∈ R
2, we have
1
{x≤ln(K)}h(x) − 1
{y≤ln(K)}h(y)
=
(h(x) − h(y)) 1
{x∨y≤ln(K)}+ h(x)1
{x≤ln(K)<y}− h(y)1
{y≤ln(K)<x}≤ A
0|e
x− e
y| 1
{x∨y≤ln(K)}+ A
11
{ln(K)<y}+ 1
{ln(K)<x}, where A
1= K(r ∨ | d|) and ¯ A
0= δ + R
e
yν(du). Let k
b= ln(
b(TK)) > 0 and recall that X ˜
t− σB
t= (γ
0−
σ22)t + Z
t; then
1
{x0+a√
θ+ ˜Xs≤lnK}
h(x
0+ a
√
θ + ˜ X
s) − 1
{x0+a√
θ+σBs≤lnK}
h(x
0+ a
√
θ + σB
s)
≤ A
0e
x0+a√
θ+ ˜Xs
− e
x0+a√ θ+σBs
1
{X˜s∨σBs≤kb−a√θ}
+ A
1(1
{kb−a√
θ<σBs}
+ 1
{kb−a√ θ<X˜s}
)
≤ A
0b(T )e
σBse
(γ0−σ2
2 )s+Zs
− 1
+ A
1(1
{kb<σBs}+ 1
{kb<X˜s}
), where the last inequality is due to a < 0 and e
x0= b(T ).
Note that for all s ∈ [0, θ],
P (k
b< σB
s) ≤ P k
bσ √ θ < B
1≤ C
√ θe
−k2 b 2σ2θ
.
Moreover, for θ small enough, we have
k2b< k
b− (γ
0−
σ22)s, so that
P (k
b< X ˜
s) ≤ P
k
b− (γ
0−
σ22)s σ √
θ < B
1!
+ P (T
1≤ θ) ≤ C
√ θe
−k2 b 8σ2θ
+ Aθ and
E
e
σBse
(γ0−σ2
2 )s+Zs
− 1
≤ e
σ2 2 s
e
(r−δ−σ2 2 )s
− 1
+ e
σ2 2 s
E e
Zs−sR(ey−1)ν(dy)
− 1 ≤ Dθ.
Hence,
Z
θ 0E
1
{x0+a√
θ+ ˜Xs≤lnK}
h(x
0+ a
√
θ + ˜ X
s)
− 1
{x0+a√
θ+σBs≤lnK}
h(x
0+ a
√
θ + σB
s)
ds = O(θ
2).
Thanks to this estimation, (11) becomes I
a(τ ) = E
Z
τ 01
{ξa,θs ≤ln K
b(T)}
h(x
0+ ξ
sa,θ)ds
+ o(θ
32), (12)
where
ξ
sa,θ= a √
θ + σB
s.
We will now use an expansion of h around x
0. When h is differentiable at x
0, we have h(x
0+ y) = h(x
0) + yh
0(x
0) + o(y). As we will see, when ν{ln(K/b(T ))} >, there is a jump in the derivative that modifies the expansion.
The function h is convex; therefore it has right-hand and left-hand derivatives everywhere.
Particularly, we have, for x < ln(K ), h
0g(x) = e
xδ +
Z
e
y1
{y>ln(K)−x}ν(dy)
and
h
0d(x) = e
xδ + Z
e
y1
{y≥ln(K)−x}ν (dy)
.
Hence, we can write
h
0d(x
0)(x − x
0)
+− h
0g(x
0)(x − x
0)
−≤ h(x) − h(x
0) ≤ h
0g(x)(x − x
0)
+− h
0d(x)(x − x
0)
−, and hence
0 ≤ h(x) − h(x
0) + h
0d(x
0)(x − x
0)
+− h
0g(x
0)(x − x
0)
−≤ h
0g(x) − h
0d(x
0)
(x − x
0)
++ h
0g(x
0) − h
0d(x)
(x − x
0)
−= h
0g(x ∨ x
0) − h
0d(x ∧ x
0)
|x − x
0|.
Thanks to the equation characterizing b(T ) when ¯ d < 0, we have h(x
0) = h(ln(b(T )) = 0. We thus obtain, by setting ∆h
0(x
0) = h
0d(x
0) − h
0g(x
0),
h(x
0+ x) = ∆h
0(x
0)x
++ h
0g(x
0)x + |x| R(x), ˜ where ˜ R(x) −→
x→00, and
0 ≤ R(x) ˜ ≤ h
0g(x
0+ x
+) − h
0d(x
0− x
−)
≤ L (1 + e
x) , with L a positive constant. We can then write
1
{ξa,θs ≤lnb(T)K }
h(x
0+ ξ
sa,θ)
=
∆h
0(x
0)(ξ
sa,θ)
++ h
0g(x
0)ξ
sa,θ1 − 1
{ξa,θs >lnb(T)K }
+
ξ
sa,θR(ξ ˜
a,θs)1
{ξa,θs ≤lnb(T)K }
. (13)
We claim that
E Z
θ0
ξ
sa,θR(ξ ˜
sa,θ)1
{ξa,θ s ≤ln Kb(T)}
ds = o(θ
32) (14)
and
E Z
θ0
∆h
0(x
0)(ξ
sa,θ)
++ h
0g(x
0)ξ
sa,θ1
{ξa,θs >lnb(T)K }
ds = o(θ
32).
(15)
Indeed, for (14), we have, by setting s = uθ and using the fact that ξ
θua,θhas the same distribution as √
θξ
a,1u, E
Z
θ 0ξ
sa,θR(ξ ˜
a,θs)1
{ξa,θ s ≤ln Kb(T)}
ds
= θ
32Z
10
E
ξ
ua,1R( ˜
√
θξ
ua,1)1
{√θξa,1u ≤ln K
b(T)}
du.
As | R(x)| ≤ ˜ L(e
x+ 1) and | R(x)| −→ ˜
x→00, we have by dominated convergence lim
θ↓0Z
1 0E h
ξ
a,1uR( ˜
√ θξ
a,1u)
i du
= 0.
This proves (14). For (15), we have, for some positive constant C, E
Z
θ0
∆h
0(x
0)(ξ
sa,θ)
++ h
0g(x
0)ξ
sa,θ1
{ξa,θs >lnb(T)K }
ds
≤ C
√ θ
Z
θ 0E
|a| + σ r s
θ |B
1|
1
{a+σ√
s/θB1>√1
θlnb(T)K }
ds
≤ Cθ
32q
E (|a| + |B
1|)
2s
P
a + σB
1> 1
√ θ ln K b(T )
= O(θ
n).
Going back to (13), we deduce from (14) and (15) I
a(τ ) = h
0g(x
0) E
Z
τ 0ξ
sa,θds + ∆h
0(x
0) E Z
τ0
ξ
sa,θ +ds + o(θ
32),
= b(T )ˆ δ E Z
τ0
a √
θ + σB
s+ λβ a √
θ + σB
s+ds
+ o(θ
32) (16)
with ˆ δ = δ + R
y>lnb(T)K
e
yν(dy), β =
Kb(T)ˆδ
, λ = ν{ln
b(TK)} and we recall that h
0g(x
0) = b(T)ˆ δ and ∆h
0(x
0) = Kν {ln
b(TK)}, so that λβ =
∆hh00(x0)g(x0)
.
Going back to (9) and using (10) and (16), we obtain E e
−rτ(K − S
τ)
+= (K − S
0)
++ E
b(T )ˆ δ Z
τ0
a √
θ + σB
s+ λβ(a √
θ + σB
s)
+ds
+K1
{Tˆ1<τ}(a
√
θ + σB
τ)
+− (a
√
θ + σB
Tˆ1)
++ o(θ
3/2) with o(θ
3/2) uniform with respect to τ . Hence
P (T − θ, b(T)e
a√
θ
) = (K − b(T )e
a√
θ
)
++ σb(T )ˆ δ v ¯
λ,β,θ(a/σ) + o(θ
3/2), where ¯ v
λ,β,θis defined by
¯
v
λ,β,θ(y) = sup
τ∈T0,θ
E Z
τ0
f
λβ(y
√
θ + B
s)ds + β1
{Tˆ1<τ}(y
√
θ + B
τ)
+− (y
√
θ + B
Tˆ1)
+with f
a(x) = x + ax
+. Recall that β = K/(b(T )ˆ δ). To simplify the expression of ¯ v
λ,β,θ, we notice first that if we set B
tθ= B
θt/ √
θ, we can write
¯ v
λ,β,θ=
√ θ sup
τ∈T0,θ
E Z
τ0
f
λβ(y + B
θs/θ)ds + β1
{Tˆ1<τ}
(y + B
τ /θθ)
+− (y + B
Tθˆ1/θ
)
+= √ θ sup
τ∈T0,θ
E θ Z
τ /θ0
f
λβ(y + B
sθ)ds + β1
{Tˆ1<τ}
(y + B
τ /θθ)
+− (y + B
Tθˆ1/θ
)
+! .
We also notice that τ ∈ T
0,θif and only if τ /θ ∈ T
0,1θ, where T
0,1θis the set of the stopping times of the filtration (F
θt)
t≥0, with values in [0, 1], so that
¯
v
λ,β,θ= √ θ sup
τ∈T0,1θ
E
θ Z
τ0
f
λβ(y + B
sθ)ds + β1
{Tˆ1<θτ}
(y + B
θτ)
+− (y + B
θˆT1/θ
)
+.
Note that ¯ v
λ,β,θ(y) does not change if we replace T
0,1θby ˆ T
0,1the set of the stopping times of the natural filtration of the couple (B
θt, N ˆ
θt), where ˆ N is defined by
N ˆ
t= X
0<s≤t
1
{∆Zs=ln(K/b(T))}.
The process ( ˆ N
θt)
t≥0is a Poisson process with intensity θλ, where λ = ν{ln(K/b(T ))}. Under the probability ˆ P , defined by
d P ˆ
dP = θ
Nˆ1e
−λ(θ−1),
the process (B
t, N ˆ
t)
0≤t≤1has the same law as (B
tθ, N ˆ
θt)
0≤t≤1. Observe that (θ
Nˆte
−λt(θ−1))
t≥0is a martingale. Hence,
¯
v
λ,β,θ(y) = √ θ sup
τ∈T0,1
E
θ
Nˆ1e
−λ(θ−1)θ Z
τ0
f
λβ(y + B
s)ds + β1
{Tˆ1<τ}
× ·
(y + B
τ)
+− (y + B
Tˆ1
)
+i
=
√ θ sup
τ∈T0,1
E
θ
Nˆτe
−λτ(θ−1)θ Z
τ0
f
λβ(y + B
s)ds + β
2 1
{Tˆ1<τ}L
−yτ(B) − L
−yˆT1
(B)
,
where L
−y(B) denotes the local time of B at −y. We have for τ ∈ T
0,1, E
θ
Nˆτe
−λτ(θ−1)θ Z
τ0
f
λβ(y + B
s)ds
= θ E
1
{Nˆτ=0}
e
−λτ(θ−1)Z
τ0
f
λβ(y + B
s)ds
+θR
τ, and if θ ≤ 1
|R
τ| ≤ θ E
1
{Nˆτ≥1}
e
−λτ(θ−1)Z
10
|f
λβ(y + B
s)|ds
= O(θ).
Hence, E
θ
Nˆτe
−λτ(θ−1)θ Z
τ0
f
λβ(y + B
s)ds
= θ E
1
{Nˆτ=0}
e
λτZ
τ0
f
λβ(y + B
s)ds
+ O(θ
2).
Besides, E
h
θ
Nˆ1e
−λ(θ−1)1
{Tˆ1<τ}
L
−yτ(B) − L
−yˆT1
(B) i
= E h
θ
Nˆτe
−λτ(θ−1)1
{Tˆ1<τ}
L
−yτ(B) − L
−yˆT1
(B ) i
= θ E h
e
λτ1
{Nˆτ=1}
L
−yτ(B ) − L
−yˆT1
(B) i
+ O(θ
2).
We then have
¯
v
λ,β,θ(y) = θ
3/2v
λ,β(y) + o(θ
3/2) with
v
λ,β(y) = sup
τ∈T0,1
E
e
λτ1
{Nˆτ=0}
Z
τ 0f
λβ(y + B
s)ds + β 2 e
λτ1
{Nˆτ=1}
L
−yτ(B) − L
−yˆT1
(B) .
Finally, we obtain P (T − θ, b(T )e
a√
θ
) − (K − b(T)e
a√
θ
) = θ
32(σb(T)ˆ δ)υ
λ,βa σ
+ o(θ
32).
4.3. Convergence rate of the critical price. Thanks to the expansion given in Theo- rem 4.3, we are now able to state the first main result of this paper.
Theorem 4.7. Under the hypothesis of the model and d < ¯ 0, we have as follows.
If ν{ln
b(TK)} = 0, then we have
t→T
lim
b(T ) − b(t) σb(T ) p
(T − t) = y
0with y
0= − sup{x ∈ R ; v
0(x) = sup
τ∈T0,1E ( R
τ0
(x + B
s)ds) = 0}.
If ν{ln
b(TK)} > 0, we then have
t→T
lim
b(T ) − b(t) σb(T) p
(T − t) = y
λ,βwith y
λ,βas defined in Lemma 4.1, with
λ = ν
ln K b(T )
, β = K
b(T )ˆ δ and δ ˆ = δ + Z
y>ln(K/b(T))
e
yν (dy).
Proof of Theorem 4.7.
According to Theorem 4.3, we have for all a < 0, P (T − θ, b(T )e
a√
θ
) = (K − b(T )e
a√
θ
)
++ Cθ
32υ
λ,βa σ
+ o(θ
32).
Lower bound for b(T ) − b(t). Specifically, we have for all a > −σy
λ,β, where y
λ,βis defined by Lemma 4.1,
υ
λ,βa σ
> 0;
we thus obtain for θ close to 0,
P(t, b(T )e
a√θ
) > (K − b(T )e
a√θ
), and then
ln(b(T)) + a
√
θ > ln(b(t)), hence
b(T ) − b(t) b(t) √
θ > −a.
Noting that since r > 0 we have b(T ) > 0, and by making t tend to T, then a to −σy
λ,β, we obtain
lim inf
t→T