Вы находитесь на странице: 1из 23

Chapter 1

Lvy Processes and Their Characteristics

Abstract We give an introductory review of Lvy processes and their properties with
emphasis on subordinators and spectrally positive Lvy processes. The -potentials
of these processes are given. Results on the times of first exit of such processes are
discussed. Several examples of such processes are given.
Keywords Lvy processes The Lvy-It decomposition formula Scale
functions Subordinators Spectrally positive Lvy processes Killed processes
First exit times Brownian motion Gamma processes Stable processes

1.1 Lvy Processes


Definition 1.1 A stochastic process X = {X t , t 0} is said to be a Lvy process if
the following hold:
(i) It has right continuous sample paths with left limits.
(ii) X has stationary increments, i.e., for every s, t 0, the distribution of X t+s
X t , is independent of t.
(iii) X has independent increments, i.e., for every t, s 0, X t+s X t is independent
of (X u , u t).
That is to say, a Lvy process is a process with stationary and independent increments
whose sample paths are right continuous with left-hand limits.
Any Lvy process X enjoys the following property: For all t 0
E[ei X t ] = et() .
The function  is known as the characteristic function of the process X , it has the
form

2 b
+ [exp(ix) 1 ix I{|x|<1} ](d x),
(1.1)
id
2
R
M. Abdel-Hameed, Lvy Processes and Their Applications in Reliability and Storage,
SpringerBriefs in Statistics, DOI: 10.1007/978-3-642-40075-9_1, The Author(s) 2014

1 Lvy Processes and Their Characteristics

where d R, b R+ and is a measure on R satisfying ({0}) = 0,


1)(d x) < .

R (x

The measure is called the Lvy measure, it characterizes the size and frequency
of the jumps. If this measure is infinite, then the process has an infinite number of
jumps of every small sizes in any small interval.
Definition
1.2 A Lvy process is said to be of bounded variation if b = 0 and

R (| x | 1)(d x) < .
For such processes, the characteristic function is of the form

() = ia +

[exp(ix) 1](d x),

(1.2)

where a = d


{|x|<1}

x(d x).

1.2 The Lvy-It Decomposition


The Lvy-It decomposition identify any Lvy process as the sum of four independent
processes, it is stated as follows:
Theorem
1.3 Given any d R, b R+ and measure on R satisfying ({0}) = 0,

2 )(d x) < , there exists a probability space (, , P) on which a Lvy
(1

x
R
process X is defined. Furthermore, for each t R+

X t = dt + Bt +


[0,t]{|x|>1}

x M(ds, d x) +

[0,t]{|x|1}

x(M m)(ds, d x),

where M is a Poisson
random measure on R+ R0 with mean measure m(ds, d x) =

ds(d x) and (x 2 1)(d x) < . Furthermore, B = {Bt : t 0} is a Brownian
motion with zero mean and variance coefficient b, and d is called the drift term.
The above theorem implies that the jump random measure defined, for any A
(R+ R0 ), by

M(A) =
I A (s, X s X s ),
s0

is a Poisson
random measure on R+ R0 with mean measure m(ds, d x) = ds(d x),

and (x 2 1)(d x) < . Furthermore, the process X is the sum of four independent
(1) (2) (3)

(4)

(1)

(2)

processes X, X, X, and X, where X is a constant drift, X is a Brownian motion with


(3)

zero mean and variance coefficient b, X is a compound Poisson process with arrival
I
(d x)
rate equal to (| x |> 1), jump magnitude distribution function F(d x) = {|x|1}
(|x|>1) ,

1.2 The Lvy-It Decomposition

(4)

and X is a pure jump martingale that has countably many jumps over every finite
interval; these jumps are of magnitude less than one almost surely. The characteristic
(1) (2) (3)

(4)

(1) (2) (3)

(4)

exponents of X, X, X, and X (denoted by , , and , respectively) are as


follows:
(1)

() = id
2 b
,
() =
 2

(2)

(3)

() =

(4)

(z) =

{|x|1}

(exp(ix) 1)(d x),

{|x|<1}

(exp(ix) 1 ix)(d x).

1.3 The Strong Markov Property for Lvy Processes


Definition 1.4 For any stopping time T with respect to  , the sigma algebra
generated by T is defined as follows:
T = {A  : A {T t} t , t R+ }.
The next theorem illustrates that the stationarity and the independence of the increments of Lvy processes hold even if the starting time in the increment is a stopping
time, instead of being fixed.
Theorem 1.5 Let L = {L t , t R+ } be a Lvy processes. For any stopping time T
with respect to  and for any t R+ , we define
Yt = L T +t L T .
Then, on the event {T < }, the process Y has the same distribution as the process
L and is independent of T .
Proof Since the indicator function of any event can be approximated by a sequence
of bounded continuous function, it suffices to show that, for m = 1, 2, ..., t1 , ..., tm
R+ , every bounded continuous function f : R m : R, and every A T
E[I A{T

<}

f (Yt1 , .., Ytm )] = P(A {T < })E[ f (L t1 , .., L tm )].

From Theorem 10 in the Appendix, it suffices to show that the above identity
holds for m = 2, and f = f 1 f 2 where f 1 , f 2 : R : R are bounded continuous
functions

1 Lvy Processes and Their Characteristics

For every n = 1, 2, ... , we define


Tn =

 k
I k1
k
2n { 2n <T 2n }.
k1

For every n = 1, 2, ..., t R+ , we define Ytn = L T n +t L T n . It is known


that T n T, as n almost surely. Since the process L is right continuous,
f 1 and f 2 are continuous functions, it follows that f i (Ytni ) f i (Yti ) almost surely,
as n , i = 1, 2. Using the bounded convergence theorem we have
lim E[I A{T n <} f 1 (Ytn1 ) f 2 (Ytn2 )] = E[I A{T

< } f 1 (Yt1 ) f 2 (Yt2 )]

For simplicity, for k = 1,2, ..., we will denote the set { k1


2n < T
A T , we write
E[I A{T n <} f 1 (Ytn1 ) f 2 (Ytn2 )]

I AAk f 1 (L
= E[
k1

= E[E[


k1

= E[

k
2n

+t1

I AAk f 1 (L

k
2n

+t1

k
2n

) f 2 (L

k
2n

k
2n

+t2

) f 2 (L

k
2n

+t2

k
2n

k
2n }

by Ak . For

)]

k
2n

) |  kn ]]
2

I AAk E[ f 1 (L t1 ) f 2 (L t2 )]]

k1

= P(A {T n < })E[ f 1 (L t1 ) f 2 (L t2 )],


where the third equation follows since the process L has stationary independent
increments and since, for k = 1, 2, ..., and A T the event A Ak  kn . . Our
2
assertion is proved by letting n , in both sides of the last equation above.

The following shows that every Lvy process is a strong Markov process in the
sense described in the Appendix.
Corollary 1.6 Let L be a Lvy process, then L is a strong Markov process.
Proof Let T be an arbitrary stopping time, we need to show that, for each t R+
and every bounded function f,
E[ f (L t+T ) | T ] =E[ f (L t+T ) | L T ].
Let the process Y = {Yt , t R+ } be as defined in the previous theorem. Observe
that L t+T = Yt + L T , hence E[ f (L t+T ) | T ] = E[ f (Yt + L T )) | T ]. From the
above theorem Yt has the same distribution as the L t and is independent of T .
Thus, E[ f (L t+T ) | T ] = E[ f (Yt + L T )) | L T ] = E[ f (L t + L T )) | L T ], L t
is independent fro L T . But, given L T , the random variable L t + L T has the same
distribution as L t+T and our assertion is proven. 

1.4 Subordinators

1.4 Subordinators
Definition 1.7 A Lvy process is called a subordinator if its sample paths are
increasing.
From Theorem 1.3 it follows that, for each t R+

X t = t +

x M(ds, d x),

[0,t)R+

where 0, and M is a Poisson random measure on R+ R+ with mean

measure m(ds, d x) = ds(d x), and 0 (x 1)(d x) < .


For such processes we have, for all 0,
E[e X t ] = e

t()

(1.3)

where


() = +

(1 ex )(d x),

and 0 is the drift term.


The function is called the Laplace exponent of the subordinator. It follows that
every subordinator is of bounded variation.
We now mention some examples of subordinators.
Example 1 Compound Poisson processes. A subordinator X with finite Lvy
measure is called a compound Poisson process with a positive drift. In this case,
(d x) = F(d x), where > 0, F is the distribution function with support R+ , and
the corresponding Poisson random measure M is finite. Let T = (Tn , n = 1, 2, ...)
be a sequence of independent identically distributed exponential random variables,
with mean 1 . For n N+ , let Sn = T1 + + Tn , S0 0, then, for every t R+
X t = t +

X n I(0,t] (Sn )

where {X n , n = 1, 2, ...} is a sequence of independent positive identically distributed


random variables with distribution function F, and independent of T .
In this case,

() = +

(1 ex )F(d x).

(1.4)

Assuming = 0, and that F has a density f, then the probability transition


function of this process is as follows:

1 Lvy Processes and Their Characteristics


 et (t)n f (n) (y x),
n!
p(t, x, y ) = n=0

0,

y>x
y x,

where f (n) is the nth convolution of f with itself.


Example 2 Inverse Brownian motion. A subordinator X with Lvy measure of the
form
1
x2
(d x) =
exp (
).
2 2
2 2 x 3
is called inverse Brownian process.
In this case,
() = +

1
( 2 2 + 2 ).
2

(1.5)

If = 0, then, for t,x,y R+ , the probability transition function of the process X


is as follows:

2
t
exp{ [(yx)t]
}, y > x
2(yx) 2
p(t, x, y ) = 2(yx)3
0,
y x.
In this case, E X 1 = 1 , and V ar (X 1 ) =

2
.
3

Example 3 Gamma processes. A subordinator X with Lvy measure


(d x) =

exp(x)d x,
x

x >0

where , > 0, is called a gamma process.


It follows that
() = + ln (1 + /).

(1.6)

Furthermore, if = 0, its probability transition function is of the form

p(t, x, y ) =

at (yx)
(at) e

0,

(y x)at1 ,

y>x
y x.

In this case, the mean term (E(X 1 )) and the variance term (V (X 1 )) are equal to
/ and / 2 , respectively.
Example 4 Stable processes with stability parameter , (0, 1). A subordinator
X with Lvy measure

1.4 Subordinators

(d x) =

(1 )x +1

x >0

is called a stable process with stability parameter .


In this case

() = + ,

(1.7)

and E[X t ] = , for all t 0.


We will discuss stable processes in general and the case when the index (1, 2)
in Sect. 1.6 of this chapter.
Example 5 Generalized stable processes with stability parameter , (0, 1). A
subordinator X with Lvy measure
(d x) =

ex
dx
(1 )x +1

x, > 0

is called a generalized stable subordinator.


It is easily seen that

() = + ( + ) ,

(1.8)

and E[X 1 ] = + 1 .

1.5 Spectrally Positive Processes


Definition 1.8 A non-subordinator is said to be spectrally positive (negative) if it
has no negative (positive) jumps.

For any spectrally positive process L, we let L = L, throughout. It is clear that

L is spectrally positive if and only if the process L is spectrally negative.


From (1.1), it follows that, for each R+ , the Laplace transform E[eL t )]
exists, furthermore,
E[eL t )] = et() ,
where

2 2

() = d +
2


0

(1 ex x1{x<1} )(d x).

(1.9)

The term d R is the drift term, 2 R+ is the variance of the Brownian motion

and is a positive measure on [0, ), ({0}) = 0, and (x 2 1)(d x) < .
0

1 Lvy Processes and Their Characteristics

The function is known as the Laplace exponent of the spectrally positive process.
The following gives some properties of the Laplace component above:
Lemma 1.9 Let be as defined in (1.9). Then
(0) = 0.
is a convex function in its argument.

If (0+) > 0, then is strictly increasing on R+ .

If (0+) 0, then there exists > 0 such that () < 0 if < , and
() 0 and increasing if .
(vi) lim () = .

(i)
(ii)
(iii)
(iv)

Proof (i) This follows immediately from the definition of .



(ii) From the definition of , it follows that () has the same sign as E[eL 1 ]
2
2
L
L
1 ] (E[L e
1 ]) . The fact that this term is positive, is easily seen
E[L 1 e
1
from Hlder inequality. This establishes the assertion.
(iii) This assertion also follows from (ii) above.
(iv) This assertion also follows from (i) and (ii) above.
(vi) Since the process L is spectrally positive, then there exists a t (0, ) such
that P{L t < 0} > 0. For such a t, et() = E[eL t ] E[eL t , L t < 0]. The
assertion follows by letting , in the last inequality.



It is clear that, (0+) = E[L 1 ]. For R+ , we define () = 1 (), i.e,


() = sup{ : () = },

(1.10)

It is seen that (0) = 0 if and only E[L 1 ] 0. Note that, E(L 1 ) =

x(d x) + d.

Furthermore, lim L t = if and only if E(L 1 ) > 0, and lim L t = if and


t

only if E[L 1 ] < 0. Also, if E(L 1 ) = 0, then L oscillates from to .


A version of the following theorem is given in Theorem 1 of [1], it is also included
in Theorem 8 p. 194 of [2].
Theorem 1.10 Let X be a spectrally positive process, with Laplace exponent ,
and is as defined in (1.10). Then, there exists an absolutely continuous increasing
function W such that,


ex W (x)d x =

1
, > (0).
()

(1.11)

Definition 1.11 For any spectrally positive process with Laplace component and
for 0, the -scale function W : R  R+ , W (x) = 0 for every x < 0, and
on [0, ) it is defined as the unique continuous increasing function such that

0

ex W () (x)d x =

1
, > ().
()

(1.12)

1.5 Spectrally Positive Processes

The existence of W () and its relation to W above is established as follows. Since


> () if and only if () > , then we have
1
1
1
=
[
]
()
() 1 /()

1 k+1
k [
]
=
()
k0


k
=
[
ex W (x)]k+1 d x
k0

ex W (k+1) (x)d x,

k0

where for k = 1, 2, ..., W (k) is the kth convolution of W with itself. Note that, since
W is increasing

(2)

(x) =

W (x y)W (y)dy

x
W (x)2 .
1!

By induction on k, it follows that for k 1,


W (k+1) (x)
Hence, for each x R+ , the series

xk
(W (x))k+1 .
k!
k W (k+1) (x) converges. Using Fubinis

k0

Theorem we have

 
k
x (k+1)

e W
(x)d x =
k0

ex

k W (k+1) (x)d x.

k0

From the uniqueness of the Laplace transform, we have, for > 0


W () (x) =

k W (k+1) (x).

(1.13)

k=0

If a spectrally positive Lvy process has bounded variation, then using (1.2) it
follows that

() =
(1 ex )(d x).
(1.14)
0

10

1 Lvy Processes and Their Characteristics

where
=

{|x|<1}

x(d x) d > 0.

(1.15)

In this case, we can write, for each t 0


X t = Yt t
where the process Y is a subordinator with drift term equal to zero,
Lemma 1.12 Let X be a spectrally positive process. Then, for each > 0
(a) W () (0) = 1 if and only if X is of bounded variation, where is given in (1.15).
(b) W () (0) = 0 if and only if X is of unbounded variation.
Proof (a) From the initial value theorem for the Laplace transform, and (1.12) we
have

W () (0) = lim
ex W () (x)d x
0

= lim

()

() 1
) .

= ( lim

Since, for x, R+ and large enough, (1 ex ) (x 1) < (x 1),

using the fact that 0 (x 1)(d x) < , (1.14) and the Lebesgue dominated
convergence theorem, we have
W () (0) =

if and only if X is of bounded variation.


(b) The assertion that W () (0) = 0 if the process L is of unbounded variation

= 0. 
follows, since in this case and from the definition of , lim ()
Furthermore, (see Lemma 8.2 of [3]), W () is right and left differentiable on
()
(0, ). By W+ (x), we will denote the right derivative of W () in x.
The adjoint -scale function associated with W () (denoted by Z () ) is defined
as follows:
Definition 1.13 For 0, the adjoint -scale function Z () : R+  [1, ) is
defined as

1.5 Spectrally Positive Processes

11

Z () (x) = 1 +

W () (y)dy.

(1.16)

Lemma 1.14 For > 0


(a)

e() x
, as x .

(() )

(1.17)

e() x
, as x .

() (() )

(1.18)

W () (x)
(b)
Z () (x)
()

Proof (a) Let W

(x) = e()x W () (x), then from (1.12) we have, for R+




()

ex W

1
.
( + ())

(x)d x =

From the final-value theorem of the Laplace transform we have


()

lim W


(x) = lim

0 0

()

ex W

(x)d x

0 ( + ())

= lim
= lim

0 (

Hence,
W () (x)

+ ()) (())

1
.

(())

e() x
, as x .

(() )

(b) From (1.16) and (1.17), it follows that as x , for > 0,


W () (x)

W () (x)

() ,

hence
Z () (x)

e() x
, as x .

() (() )

Z () (x)
W () (x)

12

1 Lvy Processes and Their Characteristics

1.6 Examples of Spectrally Positive Processes


Example 1 Brownian Motion. The Brownian motion with mean R, variance
term 2 , is an example of spectrally positive Lvy processes, where (R+ ) = 0.
2 2
From (1.1) we have, that for 0, () = + 2 . It follows that, for 0,

2 2 +2 +
() =
. For each t R+ , x, y R, the transition probability function
2
of this process is given as follows:
p(t, x, y) =
Let =

1
2b2 t

exp {

(y x t)2
}.
2 2

2 2 + 2 , then
W () (x) =

2 x/2
e
sinh (x/ 2 ),

Z () (x) = ex/ ( cosh(x/ 2 )


2

sinh(x/ 2 ))

(1.19)

Example 2 Stable processes with stability parameter (1, 2). A Lvy process
X is called stable process with stability parameter > 0, if its Lvy measure has
support [0, ) and for each t 0, X t has the same distribution as t (1/) X 1 . When
(0, 1), the process X is a subordinator with no drift, as discussed in Example 4
of Sect. 1.4. Here we will deal with the case where (1, 2), in this case the process
is spectrally positive. Let X be such a process, it follows that for t, 0,
E[e

Xt

] = E[e t

(1/)

X1

].

Since the left-hand side of the above equation is equal to et() , then we must have
E[e t

(1/)

X1

] = et() .

Clearly () = C , is the solution of the last equation. Since lim () = ,

(Lemma 1.9 (vi)), the constant C must be greater than zero. In summary
() = C ,

(1.20)

C > 0. In this case, the Lvy measure is of the form


(d x) =
where a is a positive real number.

a
x +1

(1.21)

1.6 Examples of Spectrally Positive Processes

13

It follows that, for all t 0, E[X t ] = 0 and the value of the term d in (1.9) is

equal x(d x). Furthermore,
1

() =

(ex 1 + x)(d x).

(1.22)

If C in (1.20) is taken to be 1, then the constant a in (1.21) is found to be


In this case,
Z () (x) = E (x )
W
where, for v > 0, E v (x) =

()

(x) = x

1
() .

(1.23)

E (x ),

(1.24)

x k / (1 + vk) is the MittagLeffler function. (see

k0

[3], p. 233)
The process X jumps upwards only and creeps downwards (in the sense that, for
every negative x, P{X Tx = x} = 1, where Tx is the first time the process X hits

x from above). Furthermore, = 0, and 0 (x 1)(d x) = , thus X is of
unbounded variation.
Example 3 Spectrally positive processes of bounded variation. Assume that X is
a spectrally positive
 process of bounded variation, with Laplace exponent given in
(1.14). Let = 0 x(d x) and assume that < . For every x R+ , we let

(x) = ((x, )), define the probability density function f (x) = (x))
, F(x) as
the distribution function corresponding to f , and = which is assumed to be
less than one. From (1.14), we have


() =
0

(1 ex )(d x)

ex (x)d x.

Thus,
1
1

=
()
[1 0 ex f (x)d x]


1 x  n (n)
=
e
f (x)d x
0
n=0


1 x  n (n)
=
e
F (x)d x.
0
n=0

14

1 Lvy Processes and Their Characteristics

Since, < 1, then (0) = 0, thus




ex W (x)d x =

Thus, we must have

1
, > 0.
()

W (x) =

1  n (n)
F (x).

(1.25)

n=0

The following are three examples of spectrally positive processes of bounded


variation.
Example 4 Spectrally positive processes of bounded variation with a gamma subordinator. If X is a spectrally positive process with gamma subordinator, then from
(1.14), for each t 0,
X t = Yt t,
> 0, and the process Y is a gamma process with drift term equal to zero, and
parameters , > 0, in the sense described in Example 3 of Sect. 1.4.
In this case the Laplace exponent of the process X is given as follows:
() = ln (1 + ).

(1.26)

Note that, = E[Y1 ] = < . Then, assuming that < , the scale
function W is computed using (1.25), where =



y ex p(y/)dy, F(x) = [0,x) (y)dy/.

and, for x > 0, (x) =

Example 5 Spectrally positive processes of bounded variation with a stable subordinator. From Example 4 of Sect. 1.4, for (0, 1), it follows that

() = ,

(1.27)

> 0.
(

In this case (0) = 1 , = . Thus, we cannot apply (1.25) to compute the


scale function. However, when = 1, then from (1.11) and (1.27), we have


ex W (x)d x =

, >

1
( 1
)

1.6 Examples of Spectrally Positive Processes

15

It can be shown that the solution of the last equation above is


W (x) =

1
x 1
E 1 (
),

(1.28)

where, for v > 0, E v (x) is the MittagLeffler function with parameter v, which is
defined in Example 2 of this section.
Example 6 Spectrally positive processes of bounded variation with a generalized
stable subordinator. Let X be a spectrally positive process, with generalized stable
subordinator. From Example 5 of Sect. 1.4 and (1.14), the Laplace exponent of X
is of the form

(1.29)
() = ( + ) + ,
where (0, 1), > 0, and > 0.
In this case, = 1 < . Assuming that 1 < , we can use (1.25)
1
to compute the scale function W , with the following ingredients: = and
 ey
1
F  (x) = (1)
x y +1 dy.

1.7 The Compensation Formula


For a proof of the following theorem the reader should consult [4], also see Chapter
II of [5].
Theorem 1.15 Let X be a Lvy process, defined on a probability space (, P). Let
M be a random measure on (R+ R0 ). Then, M is a Poisson random measure with
mean measure ds(d x) if and only if



E[
[0,t)R0

G s f (x)M(ds, d x)] = E[

G s f (x)ds(d x)],

[0,t)R0

for each t R+ , for every positive measurable function f on R0 , and every t


predictable process (G s ).
The following is an extension of the above theorem.
Theorem 1.16 Let g(t, x, ) be such that
(i) x g(t, x, ) is a positive bounded measurable function, and
(ii) t g(t, x, ) is predictable with respect to t .
Then, For each t R+ , we have

16

1 Lvy Processes and Their Characteristics

g(s, x)M(ds, d x) = E

[0,t)R0

g(s, x)ds(d x)

[0,t)R0

Proof We use the monotone class theorem. Take F to be the class of functions for
which the above equation holds. Let =
L = {g(s, x) : g(s, x) = G s f, where f : R
R+ is a measurable function, and (G s ) is t predictable}. From Theorem 1.15,
we have F L
=. It is clear that F is a vector space that contains the constant functions
and, by the monotone convergence theorem, is closed under taking monotone limits
of functions. From Theorem 10 of the appendix F contains every bounded (L
=)
measurable function. But (L
=) is nothing but the sigma algebra generated by functions satisfying conditions (i) and (ii) of this theorem. Thus the class of all functions
g satisfying the assumptions of this theorem are in F, this finishes the proof.


1.8 Non-homogeneous Lvy Processes


The classes of Lvy processes dealt with thus far are known as homogeneous Lvy
processes. Nonhomogeneous Lvy processes are encountered in practice. More than
one definition of such processes are found in the literature. The following definition
of such processes is suitable for our purposes.
A nonhomogeneous subordinator has the same properties as the homogeneous
subordinator with the exception that the increments are not stationary. In this case,
we have that, for each t, 0

X t
] = exp ( (t)
(1 ex )n(ds, d x)),
(1.30)
E[e
[0,t]R+


where n(ds, d x) = (ds)(d x), 0 (x 1)(d x) < ,  is an arbitrary positive
measure on R+ with 0 [0, t] < for every t 0, and (0) = 0. We assume
that the function t (t) [0, t] is continuous. It follows that a stochastic
process X is a nonhomogeneous subordinator, if and only if, for every t R+ , X t =
Y(t) , where the process Y is a homogeneous subordinator, and  is as defined
above.
In the same manner we define a nonhomogeneous Lvy process as a stochastic
process L, for every t R+ , L t = Y(t) , where Y is homogeneous Lvy process,
and  is as defined above. In this case we have


[0,t]R [exp(ix) 1 ix I{|x|<1} ](ds)(d x)
iL t
] = exp
E[e
, (1.31)
2
+ ia(t) 2b (t)
where

R0 (x

1)(d x) < .

1.9 Potentials

17

1.9 Potentials
We begin by defining the -potential measure R (x, .).
Definition 1.17 Let X be a stochastic process with state space S. For x S, any
Borel set A S, and 0


R (x, A) = E x


I{X t A} dt =

Pt (x, A)et dt.

(1.32)

Since every bounded measurable function can be approximated by a sequence of


simple functions, from the bounded convergence theorem, and (1.32) it follows that
for every bounded measurable function f on S
R f (x) = E x

et f (X t )dt =

f (y)R (x, dy).

(1.33)

We note that if X is a Levy process, then R (x, dy) = R (0, dy x). We will
denote R (0, dy) by R (dy) throughout.
Lemma 1.18 Let X be a subordinator, with Laplace exponent given in (1.3), then
1
with respect to .
R (dy) is obtained by inverting the function +()
Proof For > 0, let f (x) = ex , x > 0, in (1.33). Then,


R f (0) = E
et e X t dt
0

et E[e X t ]dt
=
 0
et et() dt
=
0

1
,
+ ()

where the second equation above followsfrom Fubinis theorem. The assertion fol
lows, since for f (x) = ex , R f (0) = 0 ey R (dy). 
Corollary 1.19 Let X be a compound Poisson process with no drift, rate , and

jump distribution function F whose support is R+ . For 0, let F = +


F, for
(n)

n =, 1, ..., F is the nth convolution of F , F (0) is the Dirac measure 0 (x), and
(n)
(n)
we write F (dy) instead of d F (y). Then, for each y 0,
R (dy) =


1
F(n) (dy).
( + )
n0

(1.34)

18

1 Lvy Processes and Their Characteristics

Proof Let the function


1.18. Note that
 f be as defined in the proof of Lemma

() = + 0 (1 ex )F(d x) = + 0 ex F(d x) = ( +

)(1 0 ex F (d x)). Thus,
1
+ ()
1
1

=
( + ) (1 0 ex F (d x))

R f (0) =

The result is immediate from Lemma 1.18 upon inverting the right-hand side of
the last equation with respect to .

Corollary 1.20 Assume that X is an inverse Gaussian process, as defined in
Example 2 of Sect. 1.4. Let be the density function of the standard normal random
variable, and erfc be the well-known complimentary error functions. Then R is
absolutely continuous with respect to the Lebesgue measure on R+ , for y R+
R (dy) = r (y)dy,
where

2 y( 2 )

2
2
)e
r (y) = ( y/) + (
erf c( y
).
y
2
2 2

(1.35)

Proof Let f be as defined in the proof of Lemma 1.18, then from (1.5) we have
R f (0) =

2

.
2 + { 2 2 + 2 }

(1.36)

Our assertion is proven using Lemma 1.18 and inverting the right-hand side of (1.36)
with respect to . 
We now introduce the so-called killed process.
Definition 1.21 Let L be Lvy process and be a stopping time. For t 0, let
X t = {L t , t < }.

(1.37)

The process X is obtained by killing the process L at time .


Let X be the process defined in (1.37) then, for every Borel set A contained in the
state space of X , t R+ , the probability transition function of this process is given
as follows:
Pt (x, A) = Px (L t A, t < }
and for each R+ its -potential is defined as follows:

1.9 Potentials

19

U (x, A) =

Pt (x, A)et dt = E x

For R+ , we define

et I{L t A} dt.

(1.38)

T+ = inf{t : L t }.

(1.39)

If the stopping time in (1.37) is taken to T+ , then the state space of the process X
is [0, ) if it is a subordinate and (, ) if it is spectrally positive.
Lemma 1.22 Assume that the process L is a subordinator, and the process X is
obtained by killing L at T+ . For any Borel set A [0, ), let R (x, A) be as
defined in (1.32), and U (x, A) be as defined in (1.38). Then, for x [0, )
U (x, A) = R (x, A).

(1.40)

Proof Write


U (x, A) = E x
0


= Ex


= Ex

et I{L t A,t<T + } dt

et I

{L t A,, L t <}

dt

et I{L t A} dt

= R (x, A),
where the second equation above follows from the definition of T+ , the third equa

tion follows since, for each t 0, L t = L t almost everywhere and A [0, ).


Furthermore, the last equation follows from (1.32). 
As an application of the above result we have the following.
Theorem 1.23 Let X be a positive compound Poisson process as defined in Example

1 of Sect. 1.4. For 0, let R be as given in (1.34). For x 0, let (x) =

([x, )) = F(x), where F 1 F. Then for any 0, v , u <


E{e

T+


, X T + > v, X T + u} =

(0,u]

(v y)R (dy).

(1.41)

Proof For n = 1, 2, .., let Yn = X 1 + + X n , Y0 = 0. Note that, for n = 0, 1, ..,


Yn is the value of the compound Poison process at time Sn = time of the nth jump
of the process X . Let N be the renewal process associated with {Yn , n = 0, 1, .},
i.e. N x = sup{n : Yn x}. Furthermore, for n = 1, 2, ..., Sn is a gamma random
variable with mean n/. We write

20

1 Lvy Processes and Their Characteristics


+

E{eT , X T + > v, X T + u}

E[eSk+1 , Yk+1 > v, Yk u, N = k}

k=0

E[eSk+1 , Yk+1 > v, Yk u}

k=0

E[eSk+1 , Yk + X k+1 > v, Yk u]

k=0

=
=


k=0

E[eSk+1 , X k+1 > v Yk , Yk u]


E[eSk+1 ]

n=0


[0,u]

P{X k+1 > v y}P{Yk dy}

k+1
)
=
(
F(v y)P{Yk dy}
+
[0,u]
k=0


=
F(v y)
F (k)
(dy)
+ [0,v]
nk=0


=
F(v y)R (dy)
[0,v]


(v y)R (dy).
=
[0,u]

where the second equation follows since for every k = 0, 1, ..., v > , u ,
{Yk+1 > v, Yk u} {N = k}, and the fifth equation follows since for k = 0, 1, ..,
the random variable Sk+1 is independent of X k+1 and Yk . 
We conclude this section by computing the potential for spectrally positive
processes. We start by computing the potential of a spectrally positive process
killed at time T +
. First, we let X be a spectrally positive Lvy process, and as

usual we define X = X . For any a R, we let Ta = inf{t 0 : X t a},

a+ = inf{t 0 : X t a}, and a = inf{t 0 : X t a}.


Lemma 1.24 Let X be a spectrally positive process, with -scale function W
(1)
-potential (U ) of the process

() .

X killed at time T = T+ Ta
For 0, a the
is absolutely continuous with respect to the Lebesgue measure on (a, ) and a version
of its density is given by

1.9 Potentials

21

(1)

u (x, y) = W () ( x)

W () (y a)
W () (y x), x, y (a, ).
W () ( a)

(1.42)

Proof For any Borel set A (a, )


(1)

U (x, A) = E x

T+ Ta

= E x

+

a

et I

{ I t A}


= E x

+

0 a

et I


=

et I{It A} dt

(A)

[W () ( x)

dt
dt

{ I t A}
W () ( a

y)

W () ( a)

W () (y x)]dy,

where the last equation follows from Theorem 8.7 of [3], this establishes our assertion.

Corollary 1.25 Let X be a spectrally positive Lvy process, with -scale function
W () . For 0 the -potential (U ) of the process killed at time T +
is absolutely
continuous with respect to the Lebesgue measure on (, ) and a version of its
density is given by
u (x, y) = W ( x)e(y)() W (y x), x, y (, ).

(1.43)

Proof The proof follows from (1.42) by letting a and since from (1.18),
() x
as x . 
for 0, W () (x) e
(() )

Corollary 1.26 Let X be a spectrally positive Lvy process, with -scale function
W () . For 0 the -potential (R ) is absolutely continuous with respect to the
Lebesgue measure on (, ) and a version of its density is given by
r (x, y) =

e(xy)()
W (y x), x, y (, ).

(() )

(1.44)

Proof The proof follows from (1.43) by letting and since, for 0,
()(x)
W () ( x) e 
as . 
(() )

The following is well known (see (8.8) of [3]), whose proof is outside the scope
of this book and is omitted.

22

1 Lvy Processes and Their Characteristics

Lemma 1.27 Let X be a spectrally negative Lvy process, a and a+ be the times
of first hitting level a from above and below, respectively. Then, for x a and
R+,
W () (x)
+
+
.
(1.45)
E x [ea , 
>

]
=
a
0
W () (a)

References
1. Takcs L (1968) On dams with finite capacity. J Aust Math Soc 8:161170
2. Bertoin J (1996) Lvy processes. Cambridge University Press, Cambridge
3. Kyprianou AE (2006) Introductory lectures on fluctuations of Lvy processes and their applications. Springer, Berlin
4. Jacod J (1975) Multivariate point processes: predictable projection, Randon-Nikodym derivatives, representation of martingales. Z Wahrscheinlichkeitstheorie Verw Gebiete 31:235253
5. Jacod J, Shiryaev A (2003) Limit theorems for stochastic processes. Springer, Berlin

Further Reading
6. Bernyk V, Danlang RC, Beskir G (2008) The law of the supremum of a stable Lvy process with
no negative jump. Ann Probab 36:17771789
7. Feller W (1971) An Introduction to Probability Theory and its Applications, vol 2. Wiley, New
York
8. Takcs L (1967) Combinatorial methods in the theory of stochastic processes. Wiley, New York

http://www.springer.com/978-3-642-40074-2

Вам также может понравиться