1

Introduction to Levy Processes

Levy processes, i.e. processes in continuous time with stationary and inde-
pendent increments, are named after Paul Levy: he made the connection with
innitely divisible distributions (LevyKhintchine formula) and described
their structure (LevyIto decomposition).
I believe that their study is of particular interest today for the following
reasons
They form a subclass of general Markov processes which is large enough
to include many familiar processes such as Brownian motion, the Poisson
process, Stable processes, etc, but small enough that a particular member
can be specied by a few quantities (the characteristics of a Levy process).
In a sense, they stand in the same relation to Brownian motion as general
random walks do to the simple symmetric random walk, and their study
draws on techniques from both these areas.
Their sample path behaviour poses a variety of dicult and fascinating
questions, some of which are not relevant for Brownian motion.
They form a exible class of models, which have been applied to the study
of storage processes, insurance risk, queues, turbulence, laser cooling, . . .
and of course nance, where the feature that they include examples having
heavy tails is particularly important.
This course will cover only a part of the theory of Levy processes, and will
not discuss applications. Even within the area of uctuation theory, there are
many recent interesting developments that I wont have time to discuss.
Almost all the material in Chapters 14 can be found in Bertoin [12].
For related background material, see Bingham [19], Sato [90], and Sato [91].

1.1 Notation
We will use the canonical notation, and denote by X = (Xt , t 0) the
co-ordinate process, i.e. Xt = Xt () = (t), where , the space of real-
valued cadlag paths, augmented by a cemetery point , and endowed with
2 1 Introduction to Levy Processes

the Skorohod topology. The Borel -eld of will be denoted by F and the
lifetime by = () = inf{t 0 : (t) = }.
Denition 1. Let P be a probability measure on (, F) with P( = )
= 1. We say that X is a (real-valued) Levy process for (, F, P) if for every
t s 0, the increment Xt+s Xt is independent of (Xu , 0 u t) and
has the same distribution as Xs .
Note that this forces P(X0 = 0) = 1; we will later write Px for the measure
corresponding to (x + Xt , t 0) under P.
(Incidentally the name Levy process has only been the accepted termino-
logy for approximately 20 years; prior to that the name process with station-
ary and independent increments was generally used.)
From the decomposition



X1 = X n1 + X n2 X n1 + + X nn X n1
n

it is apparent that X1 has an innitely divisible distribution under P. The

form of a general innitely divisible distribution is given by the well-known
LevyKhintchine formula, and from it we deduce easily the following result.
Theorem 1. Let X be a Levy process on (, F, P); then
E(exp iXt ) = et () , t 0, R,
where, for some real , and measure on R {0} which satises

{x2 1}(dx) < , (1.1.1)


2 2  
() = i + + 1 eix + ix1(|x|<1) (dx). (1.1.2)
2
is called the Levy exponent of X, and we will call the quantities
the linear cecient, the Brownian coecient, and the Levy measure of
X : together they constitute the characteristics of X. There is an existence
theorem: given real , any 0 and measure satisfying (1.1.1) there is
a measure under which X is a Levy process with characteristics , and
. There is also a uniqueness result, as any alteration in one or more of the
characteristics results in a Levy process with a dierent distribution.
Examples
The characteristics of standard Brownian motion are = 0, = 1, 0,
2
and () = 2 .
The characteristics of a compound Poisson process with jump rate c and
step distribution F are

=c xF (dx), = 0, (dx) = cF (dx),
{|x|<1}

and () = c(1 ()), where () =
eix dF (x).
 1.2 Poisson Point Processes 3

The characteristics of a Gamma process are

= c(1 e1 ), = 0, (dx) = cx1 ex 1{x>0} dx,

and () = c log(1 i).

The characteristics of a strictly stable process of index (0, 1) (1, 2)
are
c x1 dx if x > 0,
arbitrary, = 0, (dx) = + 1
c |x| dx if x < 0.
If = 1, c+ 0 and c 0 are arbitrary, and

() = c|| {1 isgn() tan(/2)} i.

If = 1, c+ = c > 0, and () = c|| i; this is a Cauchy process

with drift.
Note that there is a fairly obvious generalisation of Theorem 1 to Rd , but
we will stick, almost exclusively, to the 1-dimensional case.
The rst step to getting a probabilistic interpretation of Theorem 1 is to
realise that the process of jumps,

= (t , t 0) where t = Xt Xt ,

is a Poisson point process, but rst we need some background material.

1.2 Poisson Point Processes

A random measure on a Polish space E (this means it is metric-complete
and separable) is called a Poisson measure with intensity if
1. is a -nite measure on E;
2. for every Borel subset B of E with 0 < (B) < , (B) has a Poisson
distribution with parameter (B); in particular (B) has mean (B);
3. for disjoint Borel subsets B1 , Bn of E, the random variables (B1 ), ,
(Bn ) are independent.
In the case that c := (E) < , it is clear that we can represent as a sum
of Dirac point masses as follows. Let y1 , y2 , be a sequence of independent
and identically distributed E-valued random variables with distribution c1 ,
and N an independent Poisson-distributed random variable with parameter
c; then we can represent as


N
= yj ,
1

where y denotes the Dirac point mass at y E. If (E) = , there is a

decomposition of E into disjoint Borel sets E1 , E2 , , each having (Ej )
4 1 Introduction to Levy Processes

nite, and we can represent as the sum of independent Poisson measures j

having intensities 1Ej , each having the above representation, so again can
be represented as the sum of Dirac point masses.
To set up a Poisson point process we consider the product space E [0, ),
the measure = dx, and a Poisson measure on E [0, ) with intensity
. It is easy to check that a.s. (E {t}) = 1 or 0 for all t 0, so we can
introduce a process (e(t), t 0) by letting (e(t), t) denote the position of the
point mass on E {t} in the rst case, and in the second case put e(t) = ,
where is an additional isolated point. Then we can write

= (e(t),t) .
t0

The process e = (e(t), t 0) is called a Poisson point process with character-

istic measure .
The basic properties of a Poisson point process are stated in the next
result.
Proposition 1. Let B be a Borel set with (B) < , and dene its counting
process by
NtB = #{s t : e(s) B} = (B [0, t]), t 0,
and its entrance time by
TB = inf{t 0 : e(t) B}.
Then
(i) N B is a Poisson process of parameter (B), which is adapted to the
ltration G of e.
(ii) TB is a (Gt )-stopping time which has an exponential distribution with
parameter (B).
(iii) e(TB ) and TB are independent, and for any Borel set A
(A B)
P(e(TB ) A) = .
(B)
(iv) The process e dened by e (t) = if e(t) B and e (t) = e(t) otherwise
is a Poisson point process with characteristic measure 1B c , and it is
independent of (TB , e(TB )).
The process (e(t), 0 t TB ) is called the process stopped at the rst
point in B; its law is characterized by Proposition 1.
If we dene a deterministic function on E [0, ) by Ht (y) =
1B(t1 ,t2 ] (y, t) it is clear that


E Ht (e(t)) = (t2 t1 )(B);
0t<

this is the building block on which the following important result is based.
 1.3 The LevyIto Decomposition 5

Proposition 2. (The compensation formula) Let H = (Ht , t 0) be

a predictable process taking values in the space of nonnegative measurable
functions on E {} and having Ht () 0. Then

  
E Ht (e(t)) = E dt Ht (y)(dy) .
0t< 0 E

A second important result is called the exponential formula;

Proposition 3. Let f be a complex-valued Borel function on E {} with
f () = 0 and 
|1 ef (y) |(dy) < .
E
Then for any t 0

   
E exp f (e(s)) = exp t (1 ef (y) )(dy) .
E
0st

1.3 The LevyIto Decomposition

It is important to get a probabilistic interpretation of the LevyKhintchine
formula, and this is what this decomposition does. Fundamentally, it describes
the way that the measure determines the structure of the jumps in the
process. Specically it states that X can be written in the form
Xt = t + Bt + Yt ,
where B is a standard Brownian motion, and Y is a Levy process which
is independent of B, and is determined by its jumps, in the following
sense. Let = {t , t 0} be a Poisson point process on R [0, ) with
characteristic
 measure , and note that since {x : |x| 1} < , then
st {|s |1} |s | < a.s. Moreover if we dene
1

(2)

Yt = 1{|s |1} s , t 0
st

(2)
then it is easy to see that, provided c = {x : |x| 1} > 0, (Yt , t 0)
is a compound Poisson process with jump rate c, step distribution F (dx) =
c1 (dx)1{|x|1} and, by the exponential formula, Levy exponent

(2)
() = {1 eix }(dx).
|x|1

If 
I= (1 |x|)(dx) < , (1.3.1)
6 1 Introduction to Levy Processes

then, by considering the limit of st 1{<|s |<1} |s | as 0, we see that

1{|s |<1} |s | < a.s. for each t < ,
st

(1) (2)
and in this case we set Yt = Yt + Yt , where
(1)

Yt = s 1{|s |<1}, t 0,
st

is independent of Y (2) . Clearly, in this case Y has bounded variation (on each
nite time interval), and its exponent is

(1)
() = {1 eix }(dx).
|x|<1

In this case we can rewrite the LevyKhintchine formula as

2 2
() = i + + (1) () + (2) (),
2

where = |x|<1
x(dx) is nite, and the LevyIto decomposition takes
the form
(1) (2)
Xt = t + Bt + Yt + Yt , t 0, (1.3.2)
where the processes B, Y (1) and Y (2) are independent. The constant is called
the drift coecient of X. 
However, if I = then a.s. st |s | = for each t > 0, and in this
case we need to dene Y (1) dierently: in fact as the a.s. limit as 0 of the
compensated partial sums,
 
(1)
Y,t = 1{<|s |1} s t x(dx).
st <|x|1

(1)
It is clear that {Y,t , t 0} is a Levy process, in fact a compensated compound
Poisson process with exponent

(1)
() = {1 eix + ix}1(<|x|<1) (dx),

and hence a martingale.

 The key point, (see e.g. [12] p14), is that the basic
assumption that (1 x2 )(dx) < allows us to use a version of Doobs
maximal inequality for martingales to show that the limit as 0 exists, has
stationary and independent increments, and is a Levy process with exponent

(1) () = {1 eix + ix}1(|x|<1) (dx).

 1.4 Levy Processes as Markov Processes 7

In this case the LevyIto decomposition takes the form

(1) (2)
Xt = t + Bt + Yt + Yt , t 0, (1.3.3)

where again the processes B, Y (1) and Y (2) are independent.

Since Y (2) has unbounded variation we see that X has bounded variation
= 0 and I < . All the examples we have discussed have bounded
variation, except for Brownian motion and stable processes with index (1, 2).
To conclude this section, we record some information about the asymptotic
behaviour of the Levy exponent.
Proposition 4. (i) In all cases we have

() 2
lim 2 = .
|| 2

(ii) If X has bounded variation and drift coecient ,

()
lim = i.
||

(iii) X is a compound Poisson process if and only if is bounded.

(Note that we reserve the name compound Poisson process for a Levy
process with a nite Levy measure, no Brownian component and drift coe-
cient zero.)

1.4 Levy Processes as Markov Processes

It is clear that any Levy process has the simple Markov property in the
stronger, spatially homogeneous form that, given Xt = x, the process
{Xt+s , s 0} is independent of {Xu , u < t} and has the law of {x+Xs , s 0}.
In fact
a similar form of the strong Markov property also holds. In particular this
means that the above is valid if the xed time t is replaced by a rst
passage time
TB = inf{t 0 : Xt B}
whenever B is either open or closed.
It is also the case that the semi-group of X has the Feller property and
it turns out that the strong Feller property holds in the important special
case that the law of Xt is absolutely continuous with respect to Lebesgue
measure.
In these, and some other circumstances, the resolvent kernel is absolutely
continuous, i.e. there exists a non-negative measurable function u(q) such
that
8 1 Introduction to Levy Processes
 
(q) qt
U f (x) := e Pt f (x)dt = f (x + y)u(q) (y)dy,
0

where
Pt f (x) = Ex (f (Xt )).
The associated potential theory requires no additional hypotheses; in par-
ticular if we write X = X for the dual of X we have the following
duality relations. Let f and g be non-negative; then
 
Pt f (x)g(x)dx = f (x)Pt g(x)dx, t > 0,
R R

and  
U (q) f (x)g(x)dx = f (x)U (q) g(x)dx, t > 0,
R R
The relation between X and X via time-reversal is also simple; for each
xed t > 0, the reversed process {X(ts) Xt , 0 s t} and the dual
process {Xs , 0 s t} have the same law under P.

In summary; X is a nice Markov process, and many of technical prob-

lems which appear in the general theory are simplied for Levy processes.