You are on page 1of 44

D

R
A
F
T
bk13abstract
2006/5/24
page 1
i
i
i
i
i
i
i
i
Applied Stochastic Processes and Control for
Jump-Diusions: Modeling, Analysis and
Computation
Floyd B. Hanson
University of Illinois
Chicago, Illinois, USA
Chapter 13 Applied Guide to Abstract Stochastic Processes
Copyright c 2006 by the Society for Industrial and Applied Mathematics.
May 24, 2006
D
R
A
F
T
bk13abstract
2006/5/24
page 482
i
i
i
i
i
i
i
i
482
D
R
A
F
T
bk13abstract
2006/5/24
page 483
i
i
i
i
i
i
i
i
Chapter 13
Applied Guide to Abstract
Theory of Stochastic
Processes:
Mathematicians are like Frenchmen: Whatever you say
to them they translate into their own language
and forthwith it is something entirely dierent.
Johann Wolfgang von Goethe.
Since the mathematicians have invaded the theory of
relativity, I do not understand it myself anymore.
Albert Einstein (1879-1955),
http://en.wikiquote.org/wiki/Mathematics.
Martingale (1589): Any of several systems of betting
in which a player increases his stake,
usually by doubling each time he loses a bet.
Digital Webster, denition 3, 1992.
Martingales are treated because of their great importance,
but they are not used as a tool in this book.
William (Willy) Feller (1906-1970), p. 209 in [84].
The concept of martingales is due to P. Levy,
but it was J. L. Doob who realized its unexpected potential
and developed the theory.
William (Willy) Feller (1906-1970), p. 210 in [84].
Our view of Brownian motion never focused too closely
on the underlying measure space, and, by and large,
we have proted from keeping a respectful distance.
J. Michael Steele, p. 218 in [251].
483
D
R
A
F
T
bk13abstract
2006/5/24
page 484
i
i
i
i
i
i
i
i
484 Chapter 13. Applied Guide to Abstract Stochastic Processes
This chapter briey introduces more of the abstract analytical methods, such
as measure theoretic methods, Martingale methods, Radon-Nikod ym derivatives,
Girsanovs theorem, It o processes, Levy processes, characteristic functions and ex-
ponents, Levy-Klintchine formula, jump-diusion process comparisons and other
topics from the applied point of view as a bridge to more abstract methods.
The purpose of this chapter is to supply some insightful and useful background
to make the more abstract literature on stochastic processes and control more ac-
cessible to general students in applied mathematics, statistics, computer science,
applied science and engineering. The overall approach in this book is designed to
start from the common calculus and analysis constructs of entry level graduate
students in these applied areas by evolving these constructs to those of applied
stochastic processes and control, much as genes have evolved by small but power-
ful changes. However, students still need to understand the important results that
come from using more abstract methods.
The applied motivation is to solve problems with a combination of analytical
and computational methods. These problems may have great complexity in terms
on nonlinearities in the state and other dependencies. It is necessary to train both
students and researchers from a broad range of areas in science and engineering to
solve large scale problems. In the abstract approach the emphasis is not necessarily
to solve applied problems, but to prove existence, uniqueness and convergence, often
in very abstract language. However, sometimes the conditions of the proofs are too
restrictive, so as to exclude many complex and large scale applications. Proofs as
such are not given in this chapter, but some formal applied derivations are given
and readers can refer to the list of references at the end of the chapter for more
rigorous treatments.
13.1 Very Basic Probability Measure Background
In order to keep things simple and concise, it will be necessary to compromise on
completeness, but keep sucient detail for a coherent story. The notation will be
somewhat dierent from the usual, if there is such a thing as usual notation, so
that we can avoid conict with the stochastic process notation where possible. The
symbols are also selected so that they are related to what the quantity signies,
where possible.
13.1.1 Mathematical Measure Theory Basics
The starting point will be some notions of measure theory and its algebra, called
-algebra. Measure theory provides an abstract generalization of integration the-
ory including expectations, and distributions that are based on counts, intervals,
areas, volumes and mass to that of general sets. The ultimate goal is probability
measure, but the presentation begins with the foundations in the more general
mathematical measure theory.
D
R
A
F
T
bk13abstract
2006/5/24
page 485
i
i
i
i
i
i
i
i
13.1. Very Basic Probability Measure Background 485
Measure -Algebra Denition:
Let U be a nonempty set called the universe, but really is only the principal set of
interest. Let be a collection of subsets on U.
Denition 13.1. is a -algebra if
, i.e., the empty set is included.
U , i.e., the universe U is included.
The set S = S
c
, i.e., its complement S
c
with respect to U is
included too, i.e., verifying that S S
c
= U.
If {S
i
: i = 1 : n} is a sequence of subsets, then the union

n
i=1
S
i
,
i.e., additive closure under unions.
If so, then {U, } is called a measurable space.
Often the symbol is used for the general universe U and the symbol F is
used for the -algebra . Recall that the union of two sets
S
1
S
2
= {points X : X S
1
OR X S
2
},
the logical OR playing an important role when translated to probability measures.
A Borel set = B = B(R
nx
) is the -algebra of open sets on U = R
nx
, so
B(R
nx
) automatically contains all closed sets of R
nx
by complementarity.
Measure Denition:
Denition 13.2. The measure M is a function on the measurable space
{U, } that maps [0, ), such that
M() = 0, i.e., the empty set has measure zero.
If for any subset S , then M(S) 0, i.e., nonnegativity, as in mass.
If {S
i
: i = 1, 2, . . . ,} is any countable sequence of disjoint subsets (i.e.,
S
i

S
j
= , i = j, the intersection is empty), then the measure of the union
is the sum of the measures,
M
_

_
i=1
S
i
_
=

i=1
M(S
i
), (13.1)
i.e., countable additivity, as in preserving mass under partitioning.
D
R
A
F
T
bk13abstract
2006/5/24
page 486
i
i
i
i
i
i
i
i
486 Chapter 13. Applied Guide to Abstract Stochastic Processes
The triplet {U, , M} is called a measure space. Often the symbol is used
for the general measure symbol M used here, but the former conicts with the use
of as the mean or drift in this book. Recall that the intersection of two sets
S
1
S
2
= {points X : X S
1
AND X S
2
},
the logical AND playing an important role when translated to probability mea-
sures.
The nonnegativity measure property M(S) 0 means that positive mea-
sure has been dened. Positive meaures, among other things, facilitate convergence
proofs, i.e., monotone convergence. However, if for any subset S and M(S) 0,
then M(S) would be a negative measure and negative measure may be needed
for some applications in spite of the added awkwardness of the proofs.
Lebesgue Measure Introduction:
If the set S is measurable, the M(S) is called the total mass of the set, e.g., if
S is an interval [a, b] then it is the length (b a), if a rectangle [a, b] [c, d] then
it is its area (b a) (d c), or if a cube [a, b] [a, b] [a, b] then it is its volume
(b a)
3
. The closed intervals [a, b], open intervals (a, b) and semi-open intervals
[a, b) or (a, b], have the same measure or mass or length of (b a), since they dier
only by points of zero measure.
In general, a Lebesgue measure is a measure on an n
x
dimensional space of
real vectors, so the universe is U = R
nx
, a representative set is a hypercube
S = (a, b) (a
1
, b
1
) (a
2
, b
2
) (a
nx
, b
nx
),
such that < a
i
< b
i
< + and the measure has the form
M(S) =
nx

i=1
(b
i
a
i
).
Alternatively,
M(S) =
_
S
dx.
Lebesgue measure is a special case of Borel measure specialized to R
nx
.
Often, the innitesimal hypercube measure between vector positions from x
to x + dx is abbreviated as
M(dx) = M((x, x + dx)),
for compact notation, letting dx represent the vector-interval set (x, x + dx). This
also recognizes the translation invariance of the measure of a generalized interval
(x, x+dx), since its generalized length

nx
i=1
dx
i
is independent of the interval start
at x.
D
R
A
F
T
bk13abstract
2006/5/24
page 487
i
i
i
i
i
i
i
i
13.1. Very Basic Probability Measure Background 487
Dirac Measures:
Another measure that complements the Lebesgue measure is the Dirac measure

x
, for some point in U, having the properties that for some set S U,

x
(S) =
_
1, x S
0, x / S
_
. (13.2)
This is the set version of the Dirac delta function and apparently the same basic
denition as the indicator function 1
xS
, except without the measure infrastructure.
Counting Measures:
For Poisson processes and other discrete applications, there are also counting mea-
sures, i.e., when
M(S) = N(S) number of elements in set S. (13.3)
This includes the points of zero measure that do count.
Some Properties of Measures:
The measure space {U, , M} is nite if M(U) < and real.
The measure space {U, , M} is -nite if there exists a countable sequence
of measurable sets {S
i
: i = 1, 2, . . . ,} such that M(S
i
) < and real for
all i, i.e., sets of nite measure, and
U =

_
i=1
S
i
,
the union of a countable number of sets of nite measure. Note that -nite is
not necessarily nite, since the set of real intervals [i, i +1], have unit measure
which is nite (a Lebesgue measure), but their union is the real line, U = R
1
,
which is innite, so U is -nite while not nite.
The measure M is a montonic function since if measurable sets S
1
and S
2
ordered S
1
S
2
then M(S
1
) M(S
2
).
If {S
i
: i = 1, 2, . . . ,} is any countable sequence of subsets that are not
necessarily disjoint, then the measure of the union is only bounded by the
sum of the measures,
M
_

_
i=1
S
i
_

i=1
M(S
i
),
unlike the lack of redundancies of disjoint sets given in (13.1).
D
R
A
F
T
bk13abstract
2006/5/24
page 488
i
i
i
i
i
i
i
i
488 Chapter 13. Applied Guide to Abstract Stochastic Processes
If {S
i
: i = 1, 2, . . . ,} is any countable sequence of subsets that are
forward nested so that S
i
S
i+1
, then the limit of the measure of the
union has the limiting measure,
M
_

_
i=1
S
i
_
= lim
n
M(S
n
),
noting that M(S
i
S
i+1
) = M(S
i+1
).
If {S
i
: i = 1, 2, . . . ,} is any countable sequence of subsets that are
backward nested so that S
i+1
S
i
, then the limit of the measure of the
intersection has the limiting measure,
M
_

i=1
S
i
_
= lim
n
M(S
n
),
noting that M(S
i
S
i+1
) = M(S
i+1
).
A null set N is a measurable set such that M(N) = 0, a negligible set
is a subset of a null set and a measure M is complete if every negligible set
is measurable. A -algebra can always be completed by adding any missing
null sets to it.
A property P holds almost everywhere (a. e.) if the set of elements S in
for which the property does not hold is a null set, i.e., S = N is a set with
measure zero such that M(N) = 0.
Given the measure space {U, , M
1
}, another measure M
2
on the measur-
able space {U, } is absolutely continuous with respect to M
1
if for any
measurable set S
M
1
(S) = 0 =M
2
(S) = 0,
Absolute continuity is written symbolically as M
2
(S) M
1
(S) (or as M
2
(S)
M
1
(S), but this conicts with asymptotic notation). This property permits
dening the ratio M
2
(S)/M
1
(S) for comparison between two measures of a
set.
If M
2
(S) M
1
(S) and M
1
(S) M
2
(S), i.e., both are mutually absolutely
continuous with respect to the other, then the measures M
1
and M
2
are said
to be equivalent (M
1
(S)
a.c.
M
1
(S)). As Cont and Tankov [59] suggest, the
term equivalence is perhaps misleading and should be called something like
comparable.
Many of these properties are needed for proofs of existence and convergence, as well
as for constructing stochastic processes.
D
R
A
F
T
bk13abstr
2006/5/24
page 489
i
i
i
i
i
i
i
i
13.1. Very Basic Probability Measure Background 489
Measurable Functions:
A prerequisite that a function f is integrable is that f is a measurable function.
Denition 13.3. Given two measurable spaces, (U
1
,
1
) and (U
2
,
2
), a mapping
of the function f from U
1
to U
2
is measurable with respect to (
1
,
2
) if the
inverse (preimage) f
1
(S
2
)
1
for all S
2

2
, i.e., there is a S
1

1
such
that f(S
1
) = S
2
.
Just as in Riemann integration for general Riemann integrable functions, the
integral is built up from the limit of nite Riemann sums, the integral with respect
to a measurable function is built-up from sums of step functions called a simple
function.
Denition 13.4.
A simple function is a nite linear combination of set indicator functions
{1
xSi
} of measurable sets S
i
for i = 1: n on a measurable space (U, ), with
real coecients (could also be complex) c
i
, having the form
f(x) =
n

i=1
c
i
1
xSi
,
where x U.
The integral with respect to the measure Mfor such a simple function
is
I
M
[f] =
n

i=1
c
i
M(S
i
),
provided all the measures M(S
i
) are nite, i.e., providing the analogy to the
Riemann sums.
For a general, positive measurable function f, integrability can be extended
to f by comparison to simple measurable functions on U, such as
I
M
[f] = sup
g
_
I
M
[g] : g(x) =
n

i=1
c
i
1
xSi
, g(x) f(x), x U
_
,
provided I
M
[f] is nite. For functions that change sign, i.e., signed func-
tions, the positive-negative decomposition f(x) = f
+
(x) f

(x) with the


f

(x) (|f|(x) f(x))/2 for x U, such that


I
M
[f] = I
M
[f
+
] I
M
[f

],
provided the I
M
[f

] are nite. (The positive-negative decomposition is used


in Chapt. 9 for numerical up-winding to ensure stability.)
D
R
A
F
T
bk13abstract
2006/5/24
page 490
i
i
i
i
i
i
i
i
490 Chapter 13. Applied Guide to Abstract Stochastic Processes
If M is a Lebesgue measure, then the Lebesgue of the measure function f on
S U can be written,
I
M
[f] =
_
S
f(x)M(dx) =
_
U
1
xS
f(x)M(dx),
where recall dx symbolizes the set (x, x + dx).
Monotone Convergence Theorem:
Given the measure space (U, , M), if {f
n
(x), f
n
(x) 0 for n = 1, 2, . . .} is
a countable sequence of 1-dimensional (non-negative) measurable functions on
U that is a. e. monotone increasing and converging pointwise to f(x) a. e.,
then
lim
n
_
U
f
n
(x)M(dx) =
_
U
f
n
(x)M(dx).
This basic convergence theorem leads to several others.
13.1.2 Change of Measure: Radon-Nikodym Theorem and
Derivative:
The abstract analog of the change of variables, chain rule and Jacobian techniques
for Riemann or Riemann-Stieltjes integral is the change of measures and the Radon-
Nikod ym derivative.
Theorem 13.5. Radon-Nikodym Change of Measures:
Given the measure space {U, , M
1
} with -nite measure M
1
, if M
2
is a nite
measure that is absolutely continuous with respect to M
1
(M
2
M
1
) then there
exists a measurable real function D(x) > 0 for x U such that for each measurable
set S
M
2
(S) = I
M1
[D1
S
] =
_
U
D(x)1
xS
dM
1
(x) =
_
S
D(x)dM
1
(x), (13.4)
where dM
i
(x) = M
i
(dx) is equivalent notation for i = 1: 2. The function D is the
Radon-Nikodym derivative of M
2
with respect to M
1
, i.e.,
D(x) =
dM
2
dM
1
(x) or D(S) =
dM
2
dM
1
(S). (13.5)
Further, if is integrable with respect to the measure M
2
, then
I
M2
[] =
_
U
(x)dM
2
(x) =
_
U
(x)
dM
2
(x)
dM
1
(x)
dM
1
(x)
= I
M2
[ D] =
_
U
(x)D(x)M
1
(x),
i.e., using the Radon-Nikodym derivative in a measure-theoretic chain rule.
D
R
A
F
T
bk13abstr
2006/5/24
page 491
i
i
i
i
i
i
i
i
13.1. Very Basic Probability Measure Background 491
Thus, the Radon-Nikod ym derivative is the analog of the Jacobian of the
transformation (10.56) in an integral change of variables and leads to the absolutely
continuous measure chain rule, symbolically substituting for D,
dM
2
=
dM
2
dM
1
dM
1
.
If dM
2
and dM
1
are mutually absolutely continuous, i.e., equivalent (M
1
(S)
a.c.

M
1
(S)), the Radon-Nikod ym derivatives are mutual reciprocals,
dM
1
dM
2
= 1
_
dM
2
dM
1
,
formally justied by common null sets.
See the probability measure Examples 13.13 illustrations of applied-oriented
calculations for Radon-Nikod ym derivatives in Subsect. 13.2.1.
13.1.3 Probability Measure Basics
Since the probability distribution function for the real random variable X on the
real set S R
nx
has the property that

X
(S) = Prob[X S] [0, 1],
it is a natural candidate for a measure and the density
X
(x) could play the role of
the Radon-Nikod ym derivative. According to convention, we reset the universe as
U = , the -algebra as = F and the measure as M = P. For the jump part of
jump-diusions, counting or jump measures will also be needed.
Denition 13.6. Probability Measure:
A probability space (, F, P) is a measure space with elements called
sample points of random events in the sample space , elements F
i
F called
random events and any probability measure P on the measurable space (, F)
has total mass of one, i.e., P() = 1.
Summarizing the Kolmogorov axioms [33] of a probability space (, F, P):
P() = 0 and P() = 1.
P(S) 0 for all S .
P(

i=1
S
i
) =

i=1
P(S
i
), assuming the {S
i
} are disjoint and countable, i.e.,
there is countable additivity, so that if SS
c
= , then the complementarity
property also holds, P(S
c
) = P() +P(S).
If S
2
S
2
and P(S
1
) = 0, then P(S
2
) = 0, i.e., the probability space is
complete.
Some additional properties and nomenclature:
D
R
A
F
T
bk13abstr
2006/5/24
page 492
i
i
i
i
i
i
i
i
492 Chapter 13. Applied Guide to Abstract Stochastic Processes
The are also called scenarios as well as outcomes, the underlying or
background random variables, e.g., like the mark variable of the compound
Poisson process or Poisson random measure.
An event set S with probability P(S) = 1 is said to happen almost surely
(a.s.) or with probability one (w.p.o.), equivalent to almost everywhere
(a.e.) for mathematical measures. If an event S has probability P(S) = 0, the
event is said to be impossible.
Given a probability space (, F, P), then a (real) random variable X() is a
measurable mapping from to R
nx
such that the inverse (preimage) X
1
(S) =
{ : X() S} is F-measurable for Borel (open) sets S B(R
nx
), i.e.,
X() is the realization X upon event . If f is a (real) measurable function,
then f(X()) will also be a random variable.
If the problem involves only a single probability measure P for the single ran-
dom variable , then we can write in more usual notation,
X , Prob[X S] = Pr[X S] P(S),
i.e., the probability measure is the distribution

(S) = P(S) for S .


In general, if X = X() R
nx
for , then let S

, X()
S
X
= X(S

) and assuming the preimage S

= X
1
(S
X
) exists, then the
distribution of X is the probability measure

X
(S
X
) = P(X
1
(S
X
)),
so
X
(x) = P({ X x}), the inequality (X x) meant element-wise.
The expectation for a measurable real function f of X R
nx
with is
then
E[f(X)] =
_

f(X())P(d) =
_

f(X())dP() =
_
R
nx
f(x)
X
(dx),
provided f is absolutely integrable,
_

|f(X())|P(d) < ,
noting that the d argument of P is an abbreviation for the interval (, +d)
and that dP() and P(d) will be assumed to be equivalent notation.
Almost Sure Equivalence: Let X
1
() and X
2
() be two random variables
for , then X
1
a.s.
= X
2
if
P({ , X
1
() = X
2
()}) = 1.
D
R
A
F
T
bk13abstract
2006/5/24
page 493
i
i
i
i
i
i
i
i
13.1. Very Basic Probability Measure Background 493
Equivalence in Distribution: Let X
1
() and X
2
() be two random vari-
ables for . If the distribution satisfy

X1
=
X2
,
then X
1
() and X
2
() are called equal in distribution and we write
X
1
dist
= X
2
.
(Also called equal in law or identically distributed; the notation X
1
d
= X
2
is also used.)
The set of n random variables {X
i
} are independent with respect to the
measurable sets S
i
for i = 1: n if the probability of the union is the product of
the probabilities,
P
_
n
_
i=1
{X
i
S
i
}
_
=
n

i=1
P({X
i
S
i
}),
where the underlying random variable has been suppressed. A more concrete
and useful form as distribution in the vector X = [X
i
]
n1
is

X
(x) = P
_
n
_
i=1
{X
i
x
i
}
_
=
n

i=1
P({X
i
x
i
}) =
n

i=1

Xi
(x
i
).
An immediate corollary is the multiplication rule for the expectation of a set
of independent random variables,
E
_
n

i=1
X
i
_
=
n

i=1
E[X
i
],
assuming nite expectations, E[|X
i
|] < for i = 1: n.
For more background information, see Applebaum [12], Billingsley [32], Bing-
ham and Kiesel [33], Cont and Tankov [59]. Cyganowski, Kloeden and Ombach [66],
ksendal [218] and ksendal and Sulem [219].
Much of the further results, such as conditional expectations, follow the ap-
plied path in this book, except that matters like that of positivity and changes
in sign have to be treated with care to account for particular abstract constructs
and conditions that are designed to facilitate proofs rather than the wide variety of
problem applications.
13.1.4 Stochastic Processes in Continuous Time on Filtered
Probability Spaces
Since the emphasis of this book is on jump-diusions, stochastic processes in con-
tinuous time are treated and the relatively simpler, but not simple, discrete time
D
R
A
F
T
bk13abstr
2006/5/24
page 494
i
i
i
i
i
i
i
i
494 Chapter 13. Applied Guide to Abstract Stochastic Processes
stochastic processes are omitted (see Pliskas [221] book or Bingham and Kiesels [33,
Chapt. 3] chapter devoted to discrete time processes). The main additional di-
culty treating stochastic processes in continuous time is extending the notion of a
single probability space to a family of probability spaces over the continuous time
variable t which often has innite range.
Denition 13.7. Filtered Probability Space:
Based upon a probability space (, F, P), a ltration is a family of increasing
-algebras
F = {F
t
: t 0; F
s
F
t
, 0 s t < }
and the extended space (, F, P, F) is called a ltered probability space.
The sub--algebra F
t
represents the known information of the system on (0, t]
at time t.
The usual ltration conditions (with jump-diusions in mind) are
The initial sub--algebra F
0
the P-null-sets of F.
The ltration F is right-continuous with left limits (RCLL or
c adl ag in French), i.e., F
t
= F
t
+ = lim
0
+ F
t+
for the RC part
and F
t
= lim
0
+ F
t
for the LL part exists. The jump in the sub--
algebra at time t is [F]
t
= F
t
+ F
t
. If only continuous processes such
as diusions are under consideration, then right continuity (RC) is
sucient.
Denition 13.8. Stochastic Process:
Given the ltered probability space (, F, P, F), a stochastic process in con-
tinuous time X = {X(t) : t 0} and X is F
t
-adapted to the ltration F
if X(t) is F
t
-measurable (X(t) F
t
) for each t.
The natural ltration for the stochastic process X(t) can be written as
F
t,X
= (X(s), 0 s t),
with signifying the -eld of X(t), or more loosely the information or history
of the process X(t) up until time t.
Including the dependence on the underlying random variable, , the
X(t; ) denes a random function of time, called the sample path and is
a mapping from [0, t] to R
nx
. Usually, X(t; ) is denoted by X
t
() or just
X
t
, however in this book real subscripts are reserved to denote partial deriva-
tives, except for algebraic quantities like F
t
that are not genuine functions.
If X is adapted, i.e., F
t
-adapted to F, for t 0, then the conditional expec-
tation satises
E[X(t) | F
t
]
a.s.
= X(t),
D
R
A
F
T
bk13abstr
2006/5/24
page 495
i
i
i
i
i
i
i
i
13.1. Very Basic Probability Measure Background 495
since X(t) is known from F
t
(recall the symbol
a.s.
= denotes equals almost
surely). Saying that X or X(t) is F
t
-adapted to F means the same as
saying that X(t) is nonanticipating.
Two stochastic processes X
1
and X
2
are the same with respect to a set of
nite-dimensional distributions if for some positive integer n and discrete
time points {t
i
: i = 1 : n}, the random vectors X
j
= [X
i,j
]
n1
for j = 1 :
2 have the same n-dimensional distribution, corresponding to the stochastic
processes X
j
for j = 1: 2, respectively.
13.1.5 Martingales in Continuous Time
Martingales are processes with the property that the best predictor of the process
future value is the present value given present knowledge, i.e., it represents a fair
game of gambling, rather than a favorable or unfavorable one.
Denition 13.9. MartingaleProperties in Continuous Time:
Given a ltered probability space (, F, P, F) and F
t
-adapted process X(t) on
[0, T], T < , then X(t) is a martingale if
E[X(t) | F
s
]
a.s.
= X(s), t > s 0, (13.6)
provided X(t) is absolutely integrable, E[|X(t)|] < on [0, T], i.e., the best
predictor of X(t) with respect to the lter F
s
is X(s).
If instead of (13.6),
E[X(t) | F
s
]
a.s.
X(s), t > s 0,
then X(t) is a supermartingale,
but if
E[X(t) | F
s
]
a.s.
X(s), t > s 0,
then X(t) is a submartingale. (The submartingale corresponds to the fa-
vorable game and the supermartingale corresponds to the unfavorable game,
provided X(t) X(s) represents the gain.)
Two martingales M
1
(t) and M
2
(t) which are also equivalent or mutually ab-
solutely continuous measures, i.e., M
1
(t)
a.c.
M
2
(t), are called equivalent
martingale measures (EMM) and they play an important role in mathe-
matical nance.
Examples 13.10. Diusion, Jump and other Martingales:
For this set of examples, the time interval [0, T] as well as the coecients will
be nite, so there is no question that the stochastic processes will be absolutely
integrable.
D
R
A
F
T
bk13abstr
2006/5/24
page 496
i
i
i
i
i
i
i
i
496 Chapter 13. Applied Guide to Abstract Stochastic Processes
1. Let X(t) be a {
0
,
0
}-constant coecient, diusion process with SDE,
dX(t) =
0
dt +
0
dW(t),
and it is of interest to know for what values of
0
is X(t) a martingale, a
supermartingale or a submartingale.
The solution by integrating over [s, t] is
X(t) = X(s) +
0
(t s) +
0
(W(t) W(s)),
noting that (W(t) W(s))
dist
= W(t s) by stationary property and is indepen-
dent of W(s) and E[W(t s)|W(s)] = 0, so with F
t
= (X(r), 0 r t),
the natural ltration for X(t),
E[X(t) | F
s
] = X(s) +
0
(t s),
0 s < t. Hence, X(t) is a martingale if
0
= 0 (the case of the zero
innitesimal mean diusion, d

X(t) =
0
dW(t)), a supermartingale if
0
< 0
or a submartingale if
0
> 0. Alternatively, the translated process

X(t) = X(t)
0
t
dist
=

X(s) +
0
W(t s)
is a martingale.
2. Let X(t) be a {
0
,
0
}-constant coecient, geometric diusion process,
dX(t) = X(t)(
0
dt +
0
dW(t)),
which has the It o calculus solution,
X(t)
dist
= X(s) exp((
0

2
0
/2)(t s) +
0
W(t s)),
so
E[X(t) | F
s
] = X(s) exp(
0
(t s)),
0 s < t. Again, X(t) is a martingale if
0
= 0, a supermartingale if
0
< 0
or a submartingale if
0
> 0. Alternatively, the scaled process

X(t) = exp(
0
t)X(t)
dist
=

X(s) exp(
0
(W(t s)
0
(t s)/2)
is martingale, or more specically an exponential martingale [22] and the
scaling corresponds to the Girsanov transformation of W(t) that will be dis-
cussed in Subsect. 13.2.2.
3. Let X(t) be a simple Poisson process P(t) with additional drift and con-
stant coecients, {
0
,
0
,
0
},
dX(t) =
0
dt +
0
dP(t),
D
R
A
F
T
bk13abstr
2006/5/24
page 497
i
i
i
i
i
i
i
i
13.1. Very Basic Probability Measure Background 497
where E[dP(t)] =
0
dt = Var[dP(t)]. The solution is
X(t) = X(s) +
0
(t s) +
0
(P(t) P(s)),
where (P(t) P(s))
dist
= P(t s) and the conditional expectation is
E[X(t) | F
s
] = X(s) + (
0
+
0

0
)(t s),
so X(t) is a martingale if
0
=
0

0
(the zero innitesimal mean Poisson,
d

X(t) =
0
d

P(t), where d

P(t) = dP(t)
0
dt), a supermartingale if
0
<

0
or a submartingale if
0
>
0

0
. Alternatively, the translated process

X(t) = X(t) (
0
+
0

0
)t
dist
=

X(s) +
0
(P(t s)
0
(t s))
is a martingale.
4. Let X(t) be a compound Poisson process with additional drift and constant
coecients, {
0
,
0
,
0
,
Q
},
dX(t) =
0
dt +
0
dP(t)

i=1
Q
i
,
where E[dP(t)] =
0
dt = Var[dP(t)] and the Q
i
are IID random marks with
mean
Q
and variance
2
Q
which will not be needed. The solution is
X(t)
dist
= X(s) +
0
(t s) +
0
P(ts)

i=1
Q
i
and the conditional expectation, by iterated conditional expectations between
the Poisson counting process and the marks, is
E[X(t) | F
s
] = X(s) + (
0
+
0

Q
)(t s),
so X(t) is a martingale if
0
=
0

Q
(the zero innitesimal mean com-
pound Poisson, d

X(t) =
0

Q
d

P(t) +
0

dP(t)
i=1

Q
i
where

P(t) (P(t)
0
t)
and

Q
i
(Q
i

Q
)), a supermartingale if
0
<
0

Q
or a submartingale
if
0
>
0

Q
. The alternative process

X(t) = X(t) (
0
+
0

Q
)t =

X(s) +
0

P(t s) +
0
P(t)

i=1

Q
i
is a martingale, such that the dierence

X(t)

X(s) is a linear combination
of zero-mean random processes or variables.
D
R
A
F
T
bk13abstract
2006/5/24
page 498
i
i
i
i
i
i
i
i
498 Chapter 13. Applied Guide to Abstract Stochastic Processes
5. As an exercise the reader can nd the similar martingale properties as a func-
tion of the additional drift for the geometric jump diusion problem with
constant coecients,
dX(t) = X(t)
_
_

0
dt +
0
dW(t) +
0
dP(t)

i=1
(exp(Q
i
) 1)
_
_
,
where again the marks are IID with mean
Q
and variance
2
Q
, with the am-
plitude in the log-ready exponential form.
6. The simplest, but trivial, example is the constant process X(t) = c
0
for
t 0, i.e., dX(t) = 0, so X(t) is a martingale since E[X(t)|F
s
] = c
0
= X(s)
for s < t.
7. Another example is the closed martingale that is constructed from an abso-
lutely integrable random variable Y , independent of t, on the ltered probability
space, such that a stochastic process is dened as
X(t) E[Y | F
t
], t 0.
Thus, by the tower law ([22, p. 34], [205, Rule 6, p. 72]),
E[X(t) | F
s
] = E[E[Y | F
t
] | F
s
] = E[Y | F
s
] X(s),
for s < t, since the conditioning on F
t
followed by the conditioning on F
s
is
the same as the original conditioning on F
s
, i.e., dependence is on the smaller
of the conditioning lters.
13.1.6 Jump-Diusion Martingale Representation:
For hedging in nancial applications, martingale representations are heavily relied
upon. There are many versions of martingale representation in the literature. Some
have useful and elementary presentations. Many are restricted to diusions except
for a mention of jump processes. A selected sample is given by the references:
Baxter and Rennie [22], Due [74], Glasserman [96], ksendal [218] and Steele [251].
Here, a form of the martingale repesentation theorem is given for marked-jump-
diusion processes following Applebaum [12] and, particularly, Runggaldier [235].
Their formulation, after Jacod and Shiryaev [149], and Kunita and Watanabe [168],
respectively, uses Poisson random measure P(dt, dq) dened beginning in (5.1) on
page 206 and whose integrals are related to compound Poisson processes (5.3) on
mark-sample-space Q by
_
Q
h(t, q)P(dt, dq) =
dP(t)

i=1
h(T

i
, Q
i
), (13.7)
D
R
A
F
T
bk13abstract
2006/5/24
page 499
i
i
i
i
i
i
i
i
13.1. Very Basic Probability Measure Background 499
where the T

i
are the pre-jump-times and the Q
i
are the IID sampled marks, but
often found in martingale form by using the centered or mean-zero Poisson random
measure,

P(dt, dq) P(dt, dq) E[P(dt, dq)] = P(dt, dq)


Q
(q; t)dq(t)dt,
where
Q
(dq; t) =
Q
(q; t)dq is jump-amplitude probability measure and (t) is
the Poisson jump-rate. The mean-zero relationship corresponding to the original
relationship (13.7) is then
_
Q
h(t, q)

P(dt, dq) =
dP(t)

i=1
h(T

i
, Q
i
) E
Q
[h(t, Q)](t)dt, (13.8)
where E
Q
[h(t, Q)] =
_
Q
h(t, q)
Q
(q; t)dq.
Theorem 13.11. Marked-Jump-Diusion Martingale Representation
Theorem:
Given the Wiener process W(t) and compound Poisson triplet
{dP(t), (t),
Q
(q; t)}
or else a Poisson random measure P(dt, dq) on the sigma-eld
F = F
(W,P,Q)
t
= {W(s), P(t), S
Q
, S
N
: 0 s t, S
Q
Q, S
N
N
1
},
N
1
is the collection of null-sets of P. Then, any (P, F)-martingale M(t) has the
representions
M(t) = M(0) +
_
t
0

(D)
(s)dW(s) +
_
t
0
_
Q

(MJ)
(s, q)

P(ds, dq)
= M(0) +
_
t
0

(D)
(t)dW(s) +

dP(t)
i=1

(MJ)
_
T

i
, Q
i
_
E
Q
_

(MJ)
(t, Q)

(t)dt,
(13.9)
where
(D)
(t) is a predictable (measurable with respect to P), square-integrable pro-
cess, while
(MJ)
(t, q) is a F
(W,P,Q)
t
-predictable, Q-marked process, such that
E
Q
_

(MJ)
(t, Q)
_
=
_
Q

(MJ)
(t, q)
Q
(q; t)dq < .
The martingale representation theorem is used in the following Subsect. 13.2.2
for two versions of Girsanovs stochastic process transformation theorem, one for
the diusion process alone, i.e., without the Poisson terms in (13.9), and another
for marked-jump-diusion processes using the full form in (13.9).
The martingale approach may be a favored approach to solving SDE problems,
but Heath and Schweizer [132] show the equivalence of the martingale and PDE
approaches for a number of nancial applications. The Feynmann-Kac formula (see
(8.71) in the item on p. 329 here or the appendix of Due [74, Appendix E.] for
more background) is used to solve the corrresponding PDE problem that is derived
from the SDE.
D
R
A
F
T
bk13abstr
2006/5/24
page 500
i
i
i
i
i
i
i
i
500 Chapter 13. Applied Guide to Abstract Stochastic Processes
13.2 Change in Probability Measure:
Radon-Nikodym Derivatives and Girsanovs
Theorem
13.2.1 Radon-Nikodym Theorem and Derivative for Change of
Probability Measure:
Here, a version of the Radon-Nikod ym Theorem 13.5 and derivative is formulated
especially for probability measures and expectations. The abstract analog of the
change of distribution corresponding to a change in random variables presented
in Eq. (0.5) for the distribution and (0.6) for the density on p. 4 in preliminaries
Chapt. 0.
Theorem 13.12. Radon-Nikodym Change of Probability Measures:
Given a ltered probability space (, F, P, F) with -nite measure P
1
, if P
2
is a
nite measure that is mutually absolutely continuous with P
1
(equivalent, P
2
a.c.
P
1
)
then there exists a positive measurable real function
D(x) =
dP
2
dP
1
(x) or D(S) =
dP
2
dP
1
(S). (13.10)
called the Radon-Nikodym derivative of P
2
with respect to P
1
, for x such
that for each measurable set S F
P
2
(S) = E
P1
[D(X)1
XS
] =
_

D(x)1
xS
dP
1
(x) =
_
S
D(x)dP
1
(x), (13.11)
where dP
i
(x) = P
i
(dx) is equivalent notation for i = 1: 2.
Further, if is absolutely integrable with respect to the measure P
2
, then
E
P2
[(X)] =
_

(x)dP
2
(x) =
_

(x)
dP
2
(x)
dP
1
(x)
dP
1
(x)
= E
P2
[(X)D(X)] =
_

(x)D(x)P
1
(x),
i.e., using the Radon-Nikodym derivative in a measure-theoretic chain rule.
Thus, the Radon-Nikod ym derivative is the analog of the Jacobian of the
transformation (10.56) in an integral change of variables and leads to the absolutely
continuous measure chain rule, symbolically substituting for g,
dP
2
=
dP
2
dP
1
dP
1
.
If dP
2
and dP
1
are mutually absolutely continuous, i.e., equivalent (P
1
(S)
a.c.

P
1
(S)), the Radon-Nikod ym derivatives are mutual reciprocals,
dP
1
dP
2
= 1
_
dP
2
dP
1
,
D
R
A
F
T
bk13abstr
2006/5/24
page 501
i
i
i
i
i
i
i
i
13.2. Change in Probability Measure: Radon-Nikodym and Girsanovs 501
formally justied by common null sets.
Examples 13.13. Radon-Nikodym Derivative Calculations:
Normal distributions:
Suppose a transformation from a standard normal distribution with density

1
(x) = exp
_
x
2
/2
_
_

2
to a mean-, variance-
2
normal distribution with density,

2
(x) = exp
_
(x )
2
/
_
2
2
__
_

2
2
.
The change in measure coincides with a change of drift and a change of scale.
Thus, P
1
(x) =
1
(x) is the rst probability measure and the second is
P
2
(x) =
2
(x) =
_

D(y)
1
(y)dy =
_

D(y)
1
(y)dy,
or
2
(x) = D(x)
1
(x) upon dierentiating according to the fundamental the-
orem of integral calculus and the Radon-Nikodym derivative is
D(x) =
dP
2
(x)
dP
1
(x)
=
d
2
(x)
d
1
(x)
=

2
(x)

1
(x)
=
exp
_
(x)
2
/
_
2
2
__
_

2
2
exp (x
2
/2)
_
2
=
1

exp
_

_
1
2
_
x
2
2x +
2
2
2
_
. (13.12)
Hence, under measure P
1
the random variable X has mean 0 and variance 1,
but under measure P
2
the random variable X has mean and variance
2
. If
= 1, then there is only a change of drift and the Radon-Nikodym derivative
is simpler:
D(x) =
dP
2
(x)
dP
1
(x)
= exp
_
2x
2
2
_
.
The more general form (13.12), formally justied here, can be transformed
to the form in a proposition of Cont and Tankov [59, p. 306, Prop. 9.7] for
two diusion or Brownian motion processes, both denoted by X = X(T),
with parameters
j
T for the drifts and
2

2
1
T =
2
T =
2
2
T for a
common variance on (, F
T
, P
j
, F) for j = 1 : 2. Hence, using the fact the
D
R
A
F
T
bk13abstr
2006/5/24
page 502
i
i
i
i
i
i
i
i
502 Chapter 13. Applied Guide to Abstract Stochastic Processes
Radon-Nikodym derivative is the ratio of the two densities,
D(X(T), T) =
dP
2
(X(T), T)
dP
1
(X(T), T)
=
exp
_
(X(T)
2
T)
2
/
_
2
2
2
T
__
_
_
2
2
2
T
exp((X(T)
1
T)
2
/ (2
2
1
T))
_
_
2
2
1
T
= exp
_
2(
2

1
)X(T) (
2
2

2
1
)T
2
2
_
. (13.13)
This corrects an error in [59, p. 306, Prop. 9.7]. They also convert this to
the Cameron-Martin theorem form, by letting X(T) =
1
T + W
1
(T), in the
notation here, so
D(T) =
dP
2
(T)
dP
1
(T)
= exp
_
2(
2

1
)W
1
(T) (
2

1
)
2
T
2
2
_
, (13.14)
which is correct in [59, p. 306, following Prop. 9.7].
Sets of Independent Random Variables:
Let X = [X
i
]
n1
be a set on n independent random variables with vector mean

(1)
=
_

(1)
i
_
n1
and variance vector V
(1)
=
_

(1)
i
_
n1
, with product density

(1)
(x) =
n

i=1

(1)
i
(x
i
),
due to the independence property. The relationship between the measure, the
distribution

(1)
(x) = Prob
P1
[X x]
and the density can be written formally as
dP
1
(x)
dx
=
_
n

i=1

x
i
_

(1)
(x) =
(1)
(x),
where X x means X
i
x
i
for i = 1: n and dx =

n
i=1
dx
i
is the innitesi-
mal n-dimensional Euclidean measure, not a vector dierential.
Let there be a function D(x) that generates a second distribution or measure,

(2)
(x) = Prob
P2
[X x] =
_
n

i=1
_
xi

dy
i

(2)
i
(y
i
)
_
=
_
n

i=1
_
xi

dy
i

(1)
i
(y
i
)
_
D(y),
D
R
A
F
T
bk13abstr
2006/5/24
page 503
i
i
i
i
i
i
i
i
13.2. Change in Probability Measure: Radon-Nikodym and Girsanovs 503
so
dP
2
(x)
dx
=
_
n

i=1

x
i
_

(2)
(x) =
n

i=1
_
xi

dy
i

(2)
i
(x
i
) =
(2)
i
(x)
= D(x)
n

i=1
_
xi

dy
i

(1)
i
(x
i
) = D(x)
(1)
(x).
Solving produces
D(x) =
dP
2
(x)
dP
1
(x)
=

(2)
i
(x)

(1)
(x)
=
n

i=1

(2)
i
(x
i
)

(1)
i
(x
i
)
. (13.15)
This result is important for stochastic processes X(t) for t [0, T], since a
Radon-Nikodym derivative cannot be computed for a random variable over an
innite-dimensional interval, but it is possible to sample X(t) at sample times
t
i
= (i 1)T/n using X
i
= X(t
i
) for i = 1: n, assuming the process of interest
has independent increments.
As a more concrete example, suppose that the X
i
have a standard normal
distribution, i.e., IID with
(1)
i
= 0 and
_

(1)
i
_
2
= 1, and a nonstandard dis-
tribution is sought with mean
(2)
i
=
i
and
_

(2)
i
_
2
=
2
i
, then using (13.12),
D(x) =
dP
2
(x)
dP
1
(x)
=
1

n
j=1

j
exp
_

i=1
_
_
1
2
i
_
x
2
i
2
i
x
i
+
2
i
2
2
i
__
. (13.16)
This example is similar to one in Glasserman [96], except there the
i
1.
Poisson Distribution, a Discrete Analogy:
Next consider a Poisson cumulative distribution with parameter
1
for the
discrete variable N
1
,

(1)
n
= Prob[N
1
< n] = e
1
n

k=0

n
1
n!
which has increment (discrete derivative analog)

(1)
n1

(1)
n

(1)
n1
= e
1

n
1
n!
,
the numerical forward dierence notation, corresponding to a discrete density
and consistent with It o rules. The change of measure from variable N
1
with
parameter
1
to variable N
2
with parameter
2
is given by

(2)
n
= Prob[N
2
< n] = e
2
n

k=0

n
2
n!
= e
1
n

k=0
D
n

n
1
n!
,
D
R
A
F
T
bk13abstr
2006/5/24
page 504
i
i
i
i
i
i
i
i
504 Chapter 13. Applied Guide to Abstract Stochastic Processes
with the Radon-Nikodym discrete derivative satisfying

(2)
n1
= e
2

n
2
n!
= D
n
e
1

n
1
n!
,
and solving yields
D
n
=
P
2
(n 1)
P
1
(n 1)
=

(2)
n1

(2)
n1
= e

2
_
2
1
_
n
= e

2
+ nln(
2
/
1
)
. (13.17)
Thus, with the change in measure from P
1
to P
2
, the mean or average jump
count changes from
1
to
2
.
Poisson Distribution with Fixed Size Jumps:
Now, consider a Poisson distribution for discrete variable N
1
with parameter

1
and constant jump size
1
= 0, so
X =
1
N
1
.
Given the primary measure
P
1
(x) = Prob[X x] = Prob[N
1
x/
1
] = e
1

k=0

k
1
k!
1
{kx/1}
,
a change in measure with parameters {
2
,
2
} is sought such that
P
2
(x) = Prob[X x] = Prob[N
2
x/
2
] = e
2

k=0

k2
2
k
2
!
1
{k2x/2}
= e
1

k=0

k1
1
k
1
!
1
{k1x/1}
D
k1
.
In lieu of a proper derivative for the indicator functions 1
{kjx/j}
for j = 1: 2,
consider the increment at x = (n 1)
2
,
P
2
((n 1)
2
) = P
2
((n 1)
2
+ x) P
2
((n 1)
2
)
= e
2

k=0

k2
2
k
2
!
1
{n1<k2n1+x/2}
= e
1

k=0

k1
1
k
1
!
1
{(n1)2/1<k1(n1)2/1+x/1}
D
k1
.
Aside from the coupling of the potential Radon-Nikodym discrete derivatives
D
k1
, as Cont and Tankov [59, Prop. 9.5] state that the two measures will not
D
R
A
F
T
bk13abstract
2006/5/24
page 505
i
i
i
i
i
i
i
i
13.2. Change in Probability Measure: Radon-Nikodym and Girsanovs 505
be equivalent since their null sets will in general not coincide unless the jump
sizes are the same,
2
=
1
.
Thus, with
2
=
1
and x =
1
for a semi-open unit step (n 1, n], the new
measure increment becomes
P
2
((n 1)
1
) = P
2
(n
1
) P
2
((n 1)
1
)
= e
2

k=0

k2
2
k
2
!
1
{n1<k2n}
= e
2

n
2
n!
= e
1

k=0

k1
1
k
1
!
1
{n1<k1n}
D
k1
= e
1

n
1
n!
D
n
,
so obtaining the same Radon-Nikodym discrete derivative as in the previous
unit step example (13.17)
D
n
=
P
2
(n 1)
P
1
(n 1)
= e

2
+ nln(
2
/
1
)
. (13.18)
Note that although the original measures P
j
(n
j
) are RCLL as they should
be, inherited from the indicators 1
{kn}
, the increment P
1
((n 1)
1
) is
LCRL (left continuous, right limits) due to the indicator increments
1
{n1<k2n}
, but they precisely allow the selection of just the nth jump term
in the Poisson distribution sum since the indicator increments are closed at n
and open at n 1.
This Poisson distribution example is an applied justication of the proposition
in Cont and Tankov [59, Prop. 9.5] for two Poisson processes n = N
j
=
P(T) = N(T) with parameters
j
=
j
T on (, F
T
, P
j
, F) for j = 1: 2, i.e.,
P
2
(N(T) 1)
P
1
(N(T) 1)
= e
(
1

2
)T + N(T) ln(
2
/
1
)
, (13.19)
but only for the same size,
2
=
1
, which has an explicit form as given here.
13.2.2 Change in Measure for Stochastic Processes: Girsanovs
Theorem
There are many versions of Girsanovs theorem for changing a probability measure
to change the drift of a stochastic diusion process and some of these variants are
not very distinguishable from the Radon-Nikod ym theorem. Here, a modication
of Runggaldiers [235] (see also Bremaud [43] for even more details) version will
be followed since it has been found to be the most useful, the Radon-Nikod ym
derivative being relatively easy to calculate and comes with an extension to jump-
diusions. The application of this theorem is determining the measure change for
a relative change (t) of the drift from
1
(t) to a drift
2
(t) appropriate for the
problem of interest, e.g., the change of the drift coecient
1
(t) = in the Black-
Scholes [34] method to the current market rate
2
(t) = r.
D
R
A
F
T
bk13abstract
2006/5/24
page 506
i
i
i
i
i
i
i
i
506 Chapter 13. Applied Guide to Abstract Stochastic Processes
Diusion Girsanov Transformations
Let the reference P
1
-SDE for a state diusion process X(t) be
dX(t) =
1
(t)dt + (t)dW
1
(t) (13.20)
with time-dependent coecients {
1
(t), (t)}, whose integrabilities are implied by
the following Girsanov diusion theorem, on a nite time interval [0, T] on the
ltered probability space (, F
t
, P
1
, F) with W
1
(t) being a P
1
-Wiener process. In
addition, let the target P
2
-SDE objective for this state diusion process X(t) be
dX(t) =
2
(t)dt + (t)dW
2
(t) (13.21)
with the same volatility (t) but changed to drift
2
(t), integrability also implied,
on the nite time interval [0, T] on the ltered probability space (, F
t
, P
2
, F) with
W
1
(t) being a corresponding P
2
-Wiener process.
Theorem 13.14. Girsanovs Theorem for Changing the Probability Mea-
sure of a Diusion Process to Change the Drift:
Let (, F
t
, P
1
, F) be a ltered probability space with F =
t
F
t
, symbolically over
t. Let (t) be a square integrable predictable (measurable with respect to P
1
, i.e.,
knowable given F
t
) drift process
_
t
0

2
(s)ds <
for all t [0, T]. Then, the Radon-Nikodym derivative D(t) at time t for the process
X(t) is given by the martingale representatiion (13.9),
dD(t) = D(t)(t)dW
1
(t), D(0)
w.p.o.
= 1, (13.22)
supposing that E
P1
[D(t)] = 1 and there exists a second probability measure P
2
on F
that is equivalent to P
1
(mutually absolutely continuous, P
2
a.c.
P
1
), such that
dP
2
= D(t)dP
1
and
dW
2
(t) = dW
1
(t) (t)dt, (13.23)
where W
1
(t) is a P
1
-Wiener process as in (13.20) while W
2
(t) is a P
2
-Wiener process
as in (13.21).
The Radon-Nikodym derivative is explicitly given by
D(t) =
dP
2
(t)
dP
1
(t)
= exp
__
t
0
(s)
_

1
2
(s)ds + dW
1
(s)
__
(13.24)
and the relative drift change is
(t) =

2
(t)
1
(t)
(t)
. (13.25)
D
R
A
F
T
bk13abstract
2006/5/24
page 507
i
i
i
i
i
i
i
i
13.2. Change in Probability Measure: Radon-Nikodym and Girsanovs 507
If the ltration
F = F
(W1)
t
= {W
1
(s), S
N
: 0 s t, S
N
N
1
},
N
1
is the collection of null-sets of P
1
, then conversely every probability measure
P
2
a.c.
P
1
has the same Radon-Nikodym derivative structure.
Substituting (13.23) the Wiener process shift into the original SDE,
dX(t) =
1
(t)dt + (t)dW
1
(t) = (
1
(t) + (t)(t))dt + (t)dW
2
(t),
so comparing to the second SDE
2
(t) =
1
(t) + (t)(t) and (13.25) for (t) is
immediate, given common volatilities
1
(t) = (t) =
2
(t).
Upon applying the It o stochastic chain rule to solve the D-SDE (13.22),
d ln(D(t))
dt
=
DdW
1
D

(D)
2
dt
2D
2
=
_

1
2
dt + dW
1
_
,
integrating with D(0) = 1 and inverting the logarithm, the answer for D(t) in (13.24)
follows. Note that the assumption of common volatility is essential for obtaining the
simple linear SDE in D(t) given in (13.22), since from just one of the independent
example terms i in (13.16) it is seen that there is a quadratic term in x of the
ith exponent unless
(2)
i
=
i
=
(1)
i
= 1, the common in this example. Hence,
this Girsanov theorem is quite simple and special. The crudely derived constant
coecient case in (13.14), as an example for Radon-Nikod ym derivatives, can be
properly recovered from the Girsanov form (13.24) by setting t = T and replacing
the time-dependent coecients by constants, i.e.,
j
(s)
j
for j = 1 : 2 and
(s) .
Note that the relative drift shift (13.25), being state independent, is also the
same for the linear diusion case,
dX(t) = X(t) (
1
(t)dt + (t)dW
1
(t)) , (13.26)
which is important for applications in nance. This is a linear SDE for geometric
Brownian motion (GBM) or multiplicative diusion noise of the Black-
Scholes-Merton[34, 197] option pricing model, while the reference SDE 13.26 for
the Theorem 13.14 is for arithmetic Brownian motion (ABM) or additive
diusion noise of the historic 1900 Bachelier [16] model. The multiplicative model
is better for compounded eects, while the additive model is better for strictly cu-
mulative eects. It is well-known that the multiplicative model can be transformed
into an additive one by the logarithmic transformation using It o rules,
d ln(X(t)) =
_

2
(t)/2
_
dt + (t)dW
1
(t). (13.27)
Since the diusion coecient shift,
2
(t)/2, of the drift would be the same for the
GBM target model (2) as for the GBM reference model (1), it is clear that the
diusion Girsanov transformation of the drift will be the same as for the ABM
model, i.e.,

(GBM)
(t) =

2
(t)
1
(t)
(t)
. (13.28)
D
R
A
F
T
bk13abstr
2006/5/24
page 508
i
i
i
i
i
i
i
i
508 Chapter 13. Applied Guide to Abstract Stochastic Processes
Marked-Jump-Diusion Girsanov Transformations
Now consider the case of marked-jump-diusions or compound-jump-diusion. Let
the reference P
1
-SDE for a state marked-jump-diusion process X(t) be
dX(t) =
1
(t)dt + (t)dW
1
(t) +
dP1(t)

i=1
h
1
(T
i
, Q
i
) (13.29)
with P
1
-Wiener process W
1
(t), P
1
-Poisson process P
1
(t), E
P1
[dP
1
(t)] =
1
(t)dt
denes the jump-rate, integrable time-dependent coecients {
1
(t), (t),
1
(t)},
(time, mark)-dependent jump-amplitude h
1
(t, q), whose integrability is implied by
the following theorem, P
1
-Poisson jump-times T
i
and IID sample marks Q
i
dis-
tributed with density
(1)
Q
(q; t) on the ltered probability space (, F
t
, P
1
, F) with
W
1
(t) over on a nite time-interval [0, T]. In addition, let the target P
2
-SDE objec-
tive for this state marked-jump-diusion process X(t) be
dX(t) =
2
(t)dt + (t)dW
2
(t) +
dP2(t)

i=1
h
2
(T
i
, Q
i
) (13.30)
with P
2
-Wiener process W
2
(t), P
2
-Poisson process P
2
(t), E
P2
[dP
2
(t)] =
2
(t)dt de-
nes the jump-rate, the same volatility (t) but changed to drift
2
(t) and changed
jump-rate
2
(t), (time, mark)-dependent jump-amplitude h
2
(t, q), integrability also
implied, P
2
-Poisson jump-times T
i
and IID sample marks Q
i
distributed with den-
sity
(2)
Q
(q; t) on the nite time interval [0, T] on the ltered probability space
(, F
t
, P
2
, F).
The following theorem follows Runggaldier [235], but is also presented more
in the notation of this book.
Theorem 13.15. Girsanovs Theorem for Changing the Probability Mea-
sure of a Jump-Diusion Process to Change the Drift and the Jump-
Rate:
Let (, F
t
, P
1
, F) a ltered probability space on the nite time-interval [0, T] with
the mark space Q = R. and the (jump-rate, mark-density)-characteristics
(
1
(t),
Q
(dq; t) =
Q
(q; t)dq).
Let
(D)
(t) be the square integrable diusion drift change given in (13.25)

(D)
(t) =

2
(t)
1
(t)
(t)
(13.31)
of Theorem 13.14. Let
(J))
(t) be a nonnegative, F
t
-predictable jump-rate scaling
process such that
_
t
0

1
(s)
(J))
(s)ds <
D
R
A
F
T
bk13abstract
2006/5/24
page 509
i
i
i
i
i
i
i
i
13.2. Change in Probability Measure: Radon-Nikodym and Girsanovs 509
for all t [0, T] and let
(M)
(q; t) be a nonnegative, F
t
-predictable, Q-space depen-
dent mark-distribution scaling process such that
_
Q

(M)
(q; s)
(1)
Q
(q; t)ds = 1,
i.e., mark-space probability is conserved.
Let
D(t) = D
(D)
(t)D
(MJ)
(t),
where the diusion martingale representation factor D
(D)
(t) is given in (13.24) with
stochastic dierential in (13.22) and the marked-jump factor D
(MJ)
(t). Letting
K
(MJ)
(q; t)
(J))
(t)
(M)
(q; t) 1 (13.32)
K
(MJ)
(t) =
_
Q
K
(MJ)
(q; t)
(1)
Q
(q; t)ds, (13.33)
then D
(MJ)
(t) factor is given by the marked-jump martingale representation (13.9),
dD
(MJ)
(t) =
dP1(t)

i=1
K
(MJ)
(Q
i
; T
i
)D
(MJ)
(T

i
) K
(MJ)
(t)D
(MJ)
(t)
1
(t)dt, (13.34)
where E
P1
[D
(MJ)
] = 1 is assumed and the changed quantities are
dW
2
(t) = dW
1
(t)
(D)
(t)dt, (13.35)

2
(t) =
(J))
(t)
1
(t) (13.36)

(2)
Q
(q; t) =
(M)
(t)
(1)
Q
(q; t). (13.37)
Thus, the explicit form of the marked-jump-diusion Radon-Nikodym deriva-
tive is
D(t) =
dP
2
(t)
dP
1
(t)
= exp
_
_
t
0

(D)
(s)
_
dW
1
(s)
(D)
(s)ds/2
_
K
(MJ)
(s)
1
(s)ds
_

P1(t)

i=1
_
K
(MJ)
(Q
i
; T
i
) + 1
_
.
(13.38)
If the ltration
F = F
(W1,P1,Q)
t
= {W
1
(s), P
1
(t), S
Q
, S
N
: 0 s t, S
Q
Q, S
N
N
1
},
N
1
is the collection of null-sets of P
1
, then conversely every probability measure
P
2
a.c.
P
1
has the same Radon-Nikodym derivative structure.
D
R
A
F
T
bk13abstract
2006/5/24
page 510
i
i
i
i
i
i
i
i
510 Chapter 13. Applied Guide to Abstract Stochastic Processes
Note that the Wiener process W
1
is independent of the marked Poisson process
double (P
1
, Q), but the mark random variables Q are only conditionally independent
of P
1
and that condition is that there exists a jump of the state X in time, so the
factoring D(t) = D
(D)
(t)D
(MJ)
(t) into only two parts makes sense. Also, using It os
stochastic chain rule,
dD(t) = dD
(D)
(t)D
(MJ)
(t) +D
(D)
(t)dD
(MJ)
(t) + dD
(D)
(t)dD
(MJ)
(t)
dt
= dD
(D)
(t)D
(MJ)
(t) +D
(D)
(t)dD
(MJ)
(t)
Since (13.34) for D
(MJ)
is linear,
d ln(D
(MJ)
(t) = K
(MJ)
(t)
1
(t)dt +
dP1(t)

i=1
ln
_
K
(MJ)
(Q
i
; T
i
) + 1
_
,
since if dP
1
jumps the jump [D
(MJ)
] = K
(MJ)
D
(MJ)
, then the jump
[ln(D
(MJ)
)] = ln(D
(MJ)
+ K
(MJ)
D
(MJ)
) ln(D
(MJ)
) = ln(K
(MJ)
+ 1),
so
D
(MJ)
(t) = exp
_
_

_
t
0
K
(MJ)
(s)
1
(s)ds +
P1(t)

i=1
ln
_
K
(MJ)
(Q
i
; T
i
) + 1
_
_
_
. (13.39)
Finally, combining equations (13.22), (13.24), (13.34) and (13.39), along with con-
verting the exponential of a sum to a product yields the result (13.38) for D(t) for
the marked-jump-diusion change from measure P
1
to P
2
according to the recipe
(13.35) to (13.37).
For the geometric or linear marked-jump-diusion,
dX(t) = X(t)
_
_

1
(t)dt + (t)dW
1
(t) +
dP1(t)

i=1
h
1
(T
i
, Q
i
)
_
_
, (13.40)
the logarithmic change of variable can transform the geometric model to an arith-
metic one like (13.29),
d ln(X(t)) =
_

1
(t)
2
(t)/2
_
dt+(t)dW
1
(t)+
dP1(t)

i=1
ln (h
1
(T
i
, Q
i
)+1) . (13.41)
Again assuming a common volatility (t), the It o rule diusion coecient shift of
the drift coecient will be common in both target (2) and reference (1) models,
while the jump-rate
1
(t) and jump-amplitude distribution is unchanged, then the
Girsanov tranformation triplet,
dW
2
(t) = dW
1
(t)
(D)
(t)dt,

2
(t) =
(J))
(t)
1
(t)

(2)
Q
(q; t) =
(M)
(t)
(1)
Q
(q; t),
(13.42)
D
R
A
F
T
bk13abstract
2006/5/24
page 511
i
i
i
i
i
i
i
i
13.3. It o, Levy and Jump-Diusion Comparisons 511
will be preserved for the geometric case.
Refer to Runggaldiers jump-diusion handbook article [235] for more infor-
mation on the multidimensional case, Poisson random measure formulation and
nancial applications.
13.3 It o, Levy and Jump-Diusion Comparisons
13.3.1 Ito Processes and Jump-Diusion Processes
Many authors, Bingham and Kiesel [33] Due [74], Glasserman [96], Hull [144],
Merton [199], Mikosch [205], ksendal [218] and others, mostly refer to Brown-
ian motion or Wiener-driven processes with Wiener scaling by a factor (t) and
translated by drift (t),
dX(t) = (t)dt + (t)dW(t), (13.43)
at least as basic denition of an Ito process. Some such as Glasserman [96],
Hull [144], Merton [199] and Mikosch [205] would explicitly allow the composite
interpretation of the coecient functions in the basic denition (13.43) to include
dependence on the state X(t), such that (t) = f(X(t), t), (t) = g(X(t), t) and
dX(t) = f(X(t), t)dt + g(X(t), t)dW(t). (13.44)
Others extend the basic It o process (13.43) to include (13.44) by application of the
It o chain rule using a transformation like

X(t) = F(X(t), t) to obtain
d

X(t) = f(X(t), t)dt + g(X(t), t)dW(t),


where
f(X(t), t) =
_
F
t
+ (t)F
x
+
1
2

2
(t)F
2
xx
)
_
(X(t), t)
and
g(X(t), t) = (t)F
x
(X(t), t).
Thus, state dependent formula (13.44) will be taken as an acceptable denition of
the Ito process.
However, in his stochastic dierential equation classic 1951 memoir [146], It o
also correctly includes jumps in his discussion of simple Markov processes. It o
referred to simple Markov processes, specied by a stochastic dierential equation,
which for general Poisson noise with distributed jump-amplitudes might be called
stochastic integral dierential equations,
dX(t) = f(X(t), t)dt + g(X(t), t)dW(t) +
_
Q
h(X(t), t, q)P(dt, dq) , (13.45)
in our notation, or preferably by a stochastic integral equation,
X(t) = X(t
0
) +
_
t
t0
_
f(X(s), s)ds + g(X(s), s)dW(s)
+
_
Q
h(X(s), s, q)P(ds, dq)
_
, (13.46)
D
R
A
F
T
bk13abstr
2006/5/24
page 512
i
i
i
i
i
i
i
i
512 Chapter 13. Applied Guide to Abstract Stochastic Processes
again in our notation. Hence, there is a historical basis for calling the jump-diusion
processes that are the focus of this book as Ito processes.
Still others, for instance, Tavella and Randall [259] refer to a jump-diusion
processes as a superposition of an It o process and a Poisson jump process, while
ksendal and Sulem [219] refer to a similar combination as an It o-Levy process, but
see the next subsection on Levy processes for the dierences between jump-diusion
and Levy processes. Applebaum [12] and others more precisely call diusion pro-
cesses like (13.44) Ito diusion processes.
Although diusion processes are easier to treat since they have continuous
sample paths, jump processes and jump-diusion processes have discontinuous sam-
ple paths so are relatively more dicult to prove theorems for. Some of the most
signicant changes occur with jumps, such as extreme nancial crashes and natural
disasters.
Hence, according to the more of less standard It o process usage (13.44),
It o processes Jump-diusion processes. (13.47)
13.3.2 Levy Processes and Jump-Diusion Processes
Levy processes are essentially jump-diusion processes, but extended to processes
with innite jump rates. There have been much recent eorts in the literature study-
ing and applying Levy processes, such as Carr, Geman, Madan and Yor (CGMY
model) [46], Carr and Madan (VG model) [47] and Rydberg (NIG model) [239].
Sometimes the term non-Gaussian processes is used as in Barndor-Nielsen and
Shepherd (GIG model) [20], but may not necessarily mean strict Levy processes.
There also several recent books on Levy processes such as that of Applebaum [12],
as well as others on Levy processes but with jump processes or jump-diusions in
the titles such as those of Cont and Tankov [59] and ksendal and Sulem [219]. As
with other abstract concepts, there are many dierent denitions of a Levy process,
and some attempt has been made to merge them within the spirit of this book.
Denition 13.16. Basic Levy Process Conditions:
A Levy process satises the following conditions:
RCLL stochastic process: {X(t), t 0} on the probability space (, F, P)
with values in R
nx
(the term c adl ag means RCLL in French but is used in
English probability texts too).
Initial condition: X(0)
a.s.
= 0.
Independent increments: for every partition 0 = t
0
< t
1
< t
2
< <
t
nt
< , the increments
X(t
j
) X(t
j+1
) X(t
j
), for j = 0: n 1 (13.48)
are independent.
D
R
A
F
T
bk13abstract
2006/5/24
page 513
i
i
i
i
i
i
i
i
13.3. It o, Levy and Jump-Diusion Comparisons 513
Stationary increments: Together with independence,
X(t
j
)
dist
= X(t
j
), (13.49)
where t
j
t
j+1
t
j
.
Stochastic Continuity: The increments of X(t) satisfy,
lim
t0
Prob[X(t + t) X(t)| ] = 0, > 0 and t 0. (13.50)
All but the last condition (13.50) are standard for the processes dealt with
here when the coecients are constant, so it is usually sucient to show stochas-
tic continuity (note that continuous in probability is not the same as continuous).
However, when the process coecients are not constant, then the process will in
general not be stationary as Levy condition (13.49) requires. For many real prob-
lems the process coecients, as in nancial markets, time-dependence is important
(for instance, see Hanson and Westman [126]), so (13.49) will not be valid in these
problems. It is clear that the IID Wiener vector process W(t) or the Wiener driven
vector Gaussian process with constant coecients
G(t) =
0
t +
0
W(t)
and the Poisson vector process P(t) with constant jump rates (t) =
0
will all
be Levy processes, as well as any linear combination that is the simple constant
coecient jump-diusion n
x
-vector process,
X(t) =
0
t +
0
W(t) +
0
P(t),
where
0
R
nxnw
and
0
R
nxnp
consistent with IID W(t) R
nw
and IID
P(t) R
np
. Adding the compound Poisson process to the combination will be
discussed in the sequel.
There are some preliminary denitions that are important for further proper-
ties of Levy processes.
Denition 13.17. Innitely Divisible Distribution:
A probability distribution
X
on R
nx
is innitely divisible if for each positive integer
n there exists of a set of IID random variable Y
j
for j = 1: n such that the sum
S
n
=
n

j=1
Y
j
dist
= X,
where X has distribution
X
.
Innitely Divisibility can be related to the central limit theorem and is closely
connected to Levy processes via compound Poisson processes as follows [59].
D
R
A
F
T
bk13abstract
2006/5/24
page 514
i
i
i
i
i
i
i
i
514 Chapter 13. Applied Guide to Abstract Stochastic Processes
Proposition 13.18. Levy processes and Innitely Divisibility:
Let X(t) be a Levy process for t 0 on R
nx
, then for every t, X(t) has an innitely
divisible distribution. Conversely, if is an innitely divisible distribution, then
there exists a Levy process X(t) with the distribution .
The compound Poisson process is included rmly as a Levy process by the
following result proved in Cont and Tankov [59],
Proposition 13.19. Compound Poisson Processes as Levy Processes:
The process CP(t) for t 0 is a compound Poisson process, i.e.,
CP(t) =
P(t)

j=1
Q
i
, (13.51)
where P(t) is a simple Poisson process with constant rate
0
and the Q
j
are
IID random jump-amplitudes with common distribution
Y
(y) such that
0
are
independent of the Q
i
,
if and only if
CP(t) is a Levy process and its sample paths are piecewise constant functions.
Characteristic Functions and Levy Characteristic Exponents
Denition 13.20. Characteristic Function:
The characteristic function of a random vector X on R
nx
is the complex-valued
function,
C
X
(z) E
X
_
exp
_
iz

X
_
(13.52)
for all z R
nx
and i is the imaginary unit.
Clearly, the characteristic function of a continuous random variable X is
the Fourier transform of the density of X, i.e.,
C
X
(z) =
_
R
nx
e
iz

X
(x)dx,
while if X is a discrete scalar random variable with distribution given by the
countable sequence of probabilities
k
= Prob[X = k], then the characteristic
function is the discrete Fourier transform,
C
X
(z) =

k=0
e
izk

k
.
This is the basic random vector denition, but here the interest will be the
same denition when the random vector is a function of time t, i.e., a stochastic
process X(t),
C
X(t)
(z) E
X(t)
_
exp
_
iz

X(t)
_
.
D
R
A
F
T
bk13abstr
2006/5/24
page 515
i
i
i
i
i
i
i
i
13.3. It o, Levy and Jump-Diusion Comparisons 515
One of the most important features of a Levy process is that the characteristic
function has relatively simple form [59, 12]
Proposition 13.21. Levy Characteristic Functions and Exponents:
If X(t) is a Levy process for t 0 on R
nx
, then there exist a real-valued continuous
function
X(t)
(z) of the characteristic vector z R
nx
, called the characteristic
exponent, such that
C
X(t)
(z) = E
X(t)
_
exp
_
iz

X(t)
_
= exp
_
it
X(t)
(z)
_
. (13.53)
However, for nonstationary problems without the Levy stationarity condition
(13.50), then it would be expected that in general the exponent will not be linear
in t,
C
X(t)
(z) = exp
_
i
X(t)
(z; t)
_
.
It is well-known that Fourier transforms, and the characteristic function, is mainly
useful for constant coecients, with few exceptions.
Examples 13.22. Characteristic Functions and Exponents of Levy Pro-
cesses:
Standard Wiener Process W(t) on R:
C
W(t)
(z) = E
W(t)
_
e
izW(t)
_
=
1

2t
_

e
izw
e
w
2
/(2t)
dw
= e
(tz)
2
/(2t)
1

2t
_

e
(w itz)
2
/(2t)
dw = e
tz
2
/2
,
using the completing the square technique, so the Levy characteristic exponent
is

W(t)
(z) =
1
2
z
2
. (13.54)
IID Wiener Vector Process W(t) on R
nw
with Cov[W(t), W

(t)] = tI
nw
:
C
W(t)
(z) = E
W(t)
_
e
iz

W(t)
_
=
nw

j=1
1

2t
_

e
iz
j
w
j
e
w
2
j
/(2t)
dw
j
=
nw

j=1
C
Wj (t)
(z
j
) = exp
_
_
t
nx

j=1
z
2
j
/2
_
_
= exp
_
t|z|
2
/2
_
,
so the Levy characteristic exponent is

W(t)
(z) =
1
2
|z|
2
. (13.55)
D
R
A
F
T
bk13abstr
2006/5/24
page 516
i
i
i
i
i
i
i
i
516 Chapter 13. Applied Guide to Abstract Stochastic Processes
IID Gaussian Vector Process G(t) =
0
t +
0
W(t) on R
nx
with Cov[W(t), W

(t)] = tI
nw
, Constant
0
R
nx
and Constant
0
R
nxnw
:
C
G(t)
(z) = E
W(t)
_
e
iz

(
0
t +
0
W(t))
_
= e
iz

0
t
nw

k=1
1

2t
_

exp
_
_
i
nw

j=1
z
j

0,j,k
w
k
_
_
e
w
2
k
/(2t)
dw
k
= exp
_
_
itz

nx

j=1
z
j
nx

=1
z

nw

k=1

0,j,k

0,,k
/2
_
_
= exp
_
itz

0
tz

(
0

0
)z
_
,
so the Levy characteristic exponent is

G(t)
(z) = iz

1
2
z

(
0

0
)z/2. (13.56)
Simple Poisson Process P(t) on R with Constant Jump-Rate
0
:
C
P(t)
(z) = E
P(t)
_
e
izP(t)
_
= e

0
t

k=0
(
0
t)
k
k!
e
izk
= e

0
t

k=0
_

0
te
iz
_
k
k!
= e

0
t +
0
te
iz
= e

0
t (e
iz
1)
,
so the Levy characteristic exponent is

P(t)
(z) =
0
(e
iz
1). (13.57)
Centered or Martingale Form of Poisson Process

P(t) P(t)
0
t on R with Constant Jump-Rate
0
:
C
e
P(t)
(z) = E
P(t)
_
e
iz(P(t)
0
t)
_
= e

0
t

k=0
(
0
t)
k
k!
e
iz(k
0
t)
= e

0
t(1 + iz)
C
P(t)
(z) = e

0
t(e
iz
1 iz)
,
so the Levy characteristic exponent is

e
P(t)
(z) =
0
(e
iz
1 iz). (13.58)
D
R
A
F
T
bk13abstr
2006/5/24
page 517
i
i
i
i
i
i
i
i
13.3. It o, Levy and Jump-Diusion Comparisons 517
Simple Poisson Vector Process P(t) on R
np
with Independent Com-
ponents and Constant Jump-Rate Vector
0
= [
0,j
]
np1
:
C
P(t)
(z) = E
P(t)
_
exp
_
iz

P(t)
_
=
np

j=1
e

0,j
t

kj=0
(
0,j
t)
kj
k
j
!
e
iz
j
k
j
=
np

j=1
exp(
0,j
t (exp(iz
j
) 1)) = exp
_
_
t
np

j=1

0,j
(exp(iz
j
) 1)
_
_
= exp
_
tn
p
_

0
exp(iz)
0
__
,
where
0

np
j=1

0,j
/n
p
and
0
exp(iz)

np
j=1

0,j
exp(iz
j
)/n
p
, so the
Levy characteristic exponent is

P(t)
(z) = n
p
_

0
exp(iz)
0
_
. (13.59)
Simple Compound Poisson Process CP(t) =

P(t)
=1
Q

on R with Con-
stant Jump-Rate
0
and IID Jump-Amplitudes Q

with Distribution

Q
(q):
C
CP(t)
(z) = E
P(t),Q
_
e
izX(t)
_
= e

0
t

k=0
(
0
t)
k
k!
E
Q
_
exp
_
iz
k

=1
Q

__
= e

0
t

k=0
(
0
t)
k
k!
k

=1
E
Q
[exp(izQ

)]
= e

0
t

k=0
(
0
t)
k
k!
E
k
Q
[exp(izQ)] = exp (
0
t (E
Q
[exp(izQ)] 1)) ,
using the iterated conditional expectation technique and IID, so the Levy char-
acteristic exponent, substituting E
Q
[exp(izQ)] = C
Q
(z), is

CP(t)
(z) =
0
(C
Q
(z) 1) (13.60)
and the simple Poisson process exponent is recovered if Q

w.p.o.
= 1 1.
Vector Compound Poisson Process CP(t) =

P(t)
=1
Q

on R
nx
with
Constant Jump-Rate
0
and IID Vector Jump-Amplitudes Q

with
Distribution
Q
(q): Note that the Q

are IID as vectors not necessarily as


D
R
A
F
T
bk13abstract
2006/5/24
page 518
i
i
i
i
i
i
i
i
518 Chapter 13. Applied Guide to Abstract Stochastic Processes
components, thus,
C
CP(t)
(z) = E
P(t),Q
_
e
iz

CP(t)
_
= e

0
t

k=0
(
0
t)
k
k!
E
Q
_
exp
_
iz

=1
Q

__
= e

0
t

k=0
(
0
t)
k
k!
k

=1
E
Q
_
exp
_
iz

_
= e

0
t

k=0
(
0
t)
k
k!
E
k
Q
_
exp
_
iz

_
= exp
_

0
t
_
E
Q
_
exp
_
iz

_
1
__
,
using the iterated conditional expectation technique and IID again, so the Levy
characteristic exponent, substituting E
Q
_
exp(iz

Q)

= C
Q
(z), is

CP(t)
(z) =
0
(C
Q
(z) 1). (13.61)
Levy-Klintchine Jump-Diusion Formula
In these examples, the ingredients for the fundamental theorem of the Levy-
Klintchine representation formula specialized to jump-diusion processes has
been derived, based on the vector Gaussian process exponent result in (13.56) and
the vector compound Poisson process exponent result in (13.61).
Theorem 13.23. Levy-Klintchine Formula for Jump-Diusion Processes:
Let X(t) be the jump-diusion process on R
nx
for t 0,
X(t) = X(0) +
0
t +
0
W(t) +
P(t)

=1
Q

, (13.62)
with Levy characteristic triplet (
0

0
,
0

Q
(dq)dt,
0
), where
0
R
nx
is a
constant,
0
R
nxnw
is a constant, W(t) R
nw
is a vector Wiener process,
P(t) R is a simple Poisson process with constant and nite jump-rate
0
R and
compounded with IID vector jump-amplitudes Q

R
nx
with distribution
Q
(q).
The random triplet (W(t), P(t), Q) are independent variables, except that the jump-
amplitude Q requires the existence of a jump of the Poisson process.
Then, the characteristic function with z R
nx
for the initial condition trans-
lated process
Y(t) X(t) X(0) (13.63)
is
C
Y(t)
(z) = E
Y(t)
_
exp(iz

Y(t)

= exp
_
t
Y(t)
(z)
_
,
D
R
A
F
T
bk13abstr
2006/5/24
page 519
i
i
i
i
i
i
i
i
13.3. It o, Levy and Jump-Diusion Comparisons 519
where the Levy characteristic exponent is

Y(t)
(z) = i
0
t
1
2
z

0
z +
0
_
R
nx
_
exp(iz

q) 1
_

Q
(q)dq. (13.64)
Except for the technical details, the Levy characteristic exponent result (13.64)
follows from (13.56) for G(t) and from (13.61) for CP(t) by the independence
properties between G(t) and (P(t), Q) and by iterative conditional expectation
between P(t) and Q that is conditioned on the existence of a jump as for (13.61).
Thus,
C
Y(t)
(z) = E
Y(t)
_
exp
_
iz

Y(t)
_
= E
W(t)
_
exp
_
iz

G(t)
_
E
P(t),Q
_
exp
_
iz

CP(t)
_
= C
G(t)
(z) C
P(t)
(z)
= exp
_
t
G(t)
(z)
_
exp
_
t
CP(t)
(z)
_
= exp
_
t
_

G(t)
(z) +
CP(t)
(z)
__
so substituting (13.56) and (13.61) and expanding the expectations leads directly to
the main result (13.64). It should be noted that embedded in this derivation is the
semi-group property [12, 59] of the characteristic function in the case of constant
coecients.
In the case to the geometric or linear jump-diusion process (5.35) with con-
stant rate coecients for X(t) R with SDE,
dX(t) = X(t)
_
_

0
dt +
0
dW(t) +
dP(t)

k=1
_
e
Q
k
1
_
_
_
,
the solution is exponential via a logarithmic change of variable technique,
X(t) = X(0) exp
_
_
(
0

2
0
/2)t +
0
W(t) +
P(t)

k=1
Q
k
_
_
, (13.65)
with X(0) > 0, is obviously not a Levy process due to the exponential time-
dependence, without further transformation:
Corollary 13.24. Levy-Klintchine Transformed Geometric
Jump-Diusion Formula:
Assuming the hypotheses of Th. 13.23, except that, n
x
= 1, n
w
= 1 and that
the Levy characteristic triplet is (
2
0
,
0

Q
(dq)dt,
0

2
0
), then the characteristic
function with z R of the the logarithmic-translated process Y (t).,
Y (t) ln(X(t)/X(0)), (13.66)
D
R
A
F
T
bk13abstr
2006/5/24
page 520
i
i
i
i
i
i
i
i
520 Chapter 13. Applied Guide to Abstract Stochastic Processes
corresponding to the geometric process (13.65), is
C
Y (t)
(z) = E
Y (t)
[exp(izY (t)] = exp
_
t
Y (t)
(z)
_
,
where the Levy characteristic exponent is

Y (t)
(z) = i(
0

2
0
/2)t
1
2

2
0
z
2
+
0
_
R
(exp(izq) 1)
Q
(q)dq. (13.67)
Levy-Klintchine Levy Process Formula including Innite Rate Processes
So far the jump-rate
0
has been assumed to be constant and either explicitly or
implicitly nite in this Subsection on Levy processes. However, the innite jump-
rates is a distinguishing feature of Levy processes, so that, in general, it is not valid
to write the jump-rate symbol
0
in Levy process formulas. Instead, it is necessary
to refer to the number of jumps rather than to the jump-rate.
Recall the denition (0.176) on page 64 of the jump function of a process:
[X](t) X(t) X(t),
written here for RCLL vector processes (caution: in some of the literature X(t) is
used, but can be confused with the analytic or numerical time increment, X(t)
X(t + t) X(t)). At points where X(t) is continuous, then [X](t) = 0.
Denition 13.25. Number of Jumps of a Process, Poisson Random Mea-
sure and Levy Measure: The number of jumps in the open set S, assuming a
bounded number of jumps and excluding zero jumps (0 / S) on the interval (0, t],
is
P((0, t], S) =

s(0,t]
1
[X](s)S
. (13.68)
Here, P((0, t], S) is the Poisson random or jump measure [219]. The dierential
form is denoted by P(dt, dq) = P((t, t+dt, (q, q+dq]), as previous used in Chapt. 5.
An alternate form [228] uses a sequence of stopping or jump times,
T
k+1
(S) = inf{t | t > T
k
(S), [X](t) S}; T
0
(S) 0,
such that
P((0, t], S) =

k=1
1
T
k
(S)t
.
The zero-mean (centered or Martingale) form is denoted by

P(dt, q) = P(dt, dq)


(L)
(dq)dt, (13.69)
where now
(L)
(dq)dt = E[P(dt, dq)] and
(L)
is called the Levy measure in
general .
D
R
A
F
T
bk13abstract
2006/5/24
page 521
i
i
i
i
i
i
i
i
13.3. It o, Levy and Jump-Diusion Comparisons 521
A fundamental tool for separating out the large jumps in the presence of
innite jump-rates is the following decomposition after the concise form of ksendal
and Sulem [219]:
Theorem 13.26. Levy-It o Decompsition: Let 0 R < be a jump-amplitude
cuto, then a Levy process X
(L)
(t) on R
nx
has the decomposition,
X
(L)
(t) =
0,R
t +
0
W(t) +
_
|q|<R
q

P(t, dq) +
_
|q|R
qP(t, dq), (13.70)
where W(t) R
nw
is an independent vector Wiener process,
0,R
R
nx
is a
constant adjusted with R from the original drift
0
R
nx
,
0
R
nxnw
is a constant
.
In particular, the Levy-Ito decomposition states that the Levy process is, as
is the jump-diusion, decomposable into a continuous process and a discontinuous
process,
X
(L)
(t) = X
(cont)
(t) +X
(discont )
(t);
X
(cont)
(t)) =
0,R
t +
0
W(t);
X
(discont)
(t)) = X
(L)
(t) X
(cont)
(t).
One consequence of this Levy-Ito decomposition is another fundamental re-
sult [219, 228]:
Theorem 13.27. Levy-Klintchine Representation Formula for Levy Pro-
cesses: Let X
(L)
(t) be a Levy process for t 0 with Levy measure
(L)
on R
nx
,
given constants
0,R
R
nx
and
0
R
nxnw
, then the jump-count satises
_
R
nx
min(|q|
2
, R)
(L)
(dq) <
and the characteristic function on z R
nx
for X(t) = X
(L)
(t) is
C
X(t)
(z) = E
X(t)
_
exp
_
iz

X(t)
_
= exp
_
t
X(t)
(z)
_
,
where the Levy characteristic exponent is

X(t)
(z) = i
0,R
t
1
2
z

0
z
+
_
|q|<R
_
exp
_
iz

q
_
1 iz

q
_

(L)
(dq)
+
_
|q|R
_
exp
_
iz

q
_
1
_

(L)
(dq).
(13.71)
Conversely, given constants
0,R
R
nx
and
0
R
nxnw
, along with the Levy
measure
(L)
on R
nx
such that the jump-count satises
_
R
nx
min(|q|
2
, R)
(L)
(dq) < ,
D
R
A
F
T
bk13abstract
2006/5/24
page 522
i
i
i
i
i
i
i
i
522 Chapter 13. Applied Guide to Abstract Stochastic Processes
then there exists a Levy process X
(L)
that is unique in distribution such that the
Levy characteristic is (13.71) for z R
nx
.
Note that the extra linear term iz

q in the rst or inner integral of (13.71) is


related to the zero-mean Poisson

P(t) form iz found in (13.58) but not in (13.57)
for P(t).
Although jump processes time-dependent coecients, like drift and volatility
coecients, do not strictly satisfy the stationary increment condition (13.49) for
a Levy process, ksendal and Sulem [219] dene Levy-driven processes which
satisfy the Levy-Ito decomposition formula (13.70), but not the constant coecient
condition. For example, analogous to the Wiener-driven Ito process (13.44),
there is the Levy-driven Ito-Levy process [219, Th. 1.14, p. 6] on R with time-
random coecients,
dX(t) =
0,R
(t; )dt +
0
(t; )dW(t)
+
_
|q|<R
h(t, q; )

P(dt, dq) +
_
|q|R
h(t, q; )P(dt, dq),
(13.72)
for some R [0, ), (
0,R
(t; ),
0
(t; ), h(t, q) are integrable function and is
some background random variable.
The Levy-driven geometric Levy process [219, Example 1.15, p. 7] is
similarly dened,
dX(t) = X(t)
_

0,R
(t; )dt +
0
(t; )dW(t)
+
_
|q|<R
h(t, q; )

P(dt, dq) +
_
|q|R
h(t, q; )P(dt, dq)
_
,
(13.73)
where, in addition, the jump-amplitude h(t, q; ) 1 to preserve positivity as-
suming X(0) > 0, with more potential uses in nancial applications.
In general, these processes are special cases of what ksendal and Sulem [219,
Th. 1.19,p. 10] call Levy diusions governed by Levy stochastic dierential
equations,
dX(t) = (t, X(t))dt + (t, X(t))dW(t)
+
_
|q|<R
h(t, q; )

P(dt, dq) +
_
|q|R
h(t, X(t), q)P(dt, dq),
(13.74)
where 0 t T, X R
nx
, R
nx
, W R
nw
, R
nxnw
, P R
np
, Q R
nx
and
h R
nxnp
, subject to the usual linear growth and Lipschitz continuity conditions.
For many other Levy process models, including models which push the limits
of the assumptions here, see Applebaum [12, Subsect. 5.4.7, p. 286]
Concluding this subsection like the last, the size of the Levy processes is
compared to that of jump-diusions. According to the strict Levy process denition
leading to a restriction to constant coecients,
_
constant coecient
jump-diusion processes
_
Levy processes, (13.75)
D
R
A
F
T
bk13abstract
2006/5/24
page 523
i
i
i
i
i
i
i
i
13.3. It o, Levy and Jump-Diusion Comparisons 523
since ordinarily jump-diusions based upon Poisson processes do not allow for in-
nite jump-rates on [0, t]. However, if the innite jump activity is controlled for,
then
_
nite jump-rate
Levy processes
_
jump-diusion processes, (13.76)
since jump-diusions in general include variable coecients and nonlinear terms.
If the comparison is made to the Levy-driven processes discussed by ksendal
and Sulem [219] and summarized here, then
{jump-diusion processes} Levy-driven processes, (13.77)
due to the inclusion of innite jump-rates with nonlinear and time-dependent coef-
cients in Levy-driven processes.
Suggested References for Further Reading
Applebaum, 2004 [12].
Bain, 2006 [17].
Baxter and Rennie, 1996 [22].
Billingsley, 1986 Billingsley86.
Bingham and Kiesel, 2004 [33]
Bossaerts, 2002 [41].
Bremaud, 1981 [43].
Cont and Tankov, 2004 [59].
Cyganowski, Kloeden and Ombach, 2002 [66].
Doob, 1953 [69].
Due, 1992 [74].
Gihman and Skorohod, 1972 [94].
Glasserman, 2003 [96].
Heath and Schweizer, 2000 [132].
Hull, 2000 [144].
It o, 1951 [146].
Karlin and Taylor, 1981 [158].
Klebaner, 1998 [160].
D
R
A
F
T
bk13abstract
2006/5/24
page 524
i
i
i
i
i
i
i
i
524 Chapter 13. Applied Guide to Abstract Stochastic Processes
Mikosch, 1998 [205].
Neftci, 2000 [213].
ksendal, 1998 [218].
ksendal and Sulem, 2005 [219].
Pliska, 1997 [221].
Protter, 2004 [228].
Runggaldier, 2003 [235].
Rogers and Williams, 2000 [232].
Steele, 2001 [251].
Yong and Zhou, 1999 [283].