Introduction to Levy processes

Graduate lecture 22 January 2004

Matthias Winkel
Departmental lecturer
(Institute of Actuaries and Aon lecturer in Statistics)
1. Random walks and continuous-time limits
2. Examples
3. Classication and construction of Levy processes
4. Examples
5. Poisson point processes and simulation
1
1. Random walks and continuous-time limits
Denition 1 Let Y
k
, k 1, be i.i.d. Then
S
n
=
n

k=1
Y
k
, n N,
is called a random walk.
16 8 0
4
0
-4
Page 1
Random walks have stationary and independent increments
Y
k
= S
k
S
k1
, k 1.
Stationarity means the Y
k
have identical distribution.
Denition 2 A right-continuous process X
t
, t R
+
, with
stationary independent increments is called Levy process.
2
What are S
n
, n 0, and X
t
, t 0?
Stochastic processes; mathematical objects, well-dened,
with many nice properties that can be studied.
If you dont like this, think of a model for a stock price
evolving with time. There are also many other applications.
If you worry about negative values, think of logs of prices.
What does Denition 2 mean?
Increments X
t
k
X
t
k1
, k = 1, . . . , n, are independent and
X
t
k
X
t
k1
X
t
k
t
k1
, k = 1, . . . , n for all 0 = t
0
< . . . < t
n
.
Right-continuity refers to the sample paths (realisations).
3
Can we obtain Levy processes from random walks?
What happens e.g. if we let the time unit tend to zero, i.e.
take a more and more remote look at our random walk?
If we focus at a xed time, 1 say, and speed up the process
so as to make n steps per time unit, we know what happens,
the answer is given by the Central Limit Theorem:
Theorem 1 (Lindeberg-Levy) If
2
= V ar(Y
1
) < , then
S
n
E(S
n
)

n
Z N(0,
2
) in distribution, as n .
4
Theorem 2 (Donsker) If
2
=V ar(Y
1
)<, then for t 0
X
(n)
t
=
S
[nt]
E(S
[nt]
)

n
X
t
N(0,
2
t) in distribution.
Furthermore, X
(n)
X where X is Brownian motion with
diusion coecient .
16 8 0
4
0
-4
Page 1
64 32 0
8
0
-8
Page 1
256 128 0
16
0
-16
Page 1
It can be shown that X has continuous sample paths.
5
We had to put the condition V ar(Y
1
) < . If it fails:
Theorem 3 (Doeblin) If = sup{ 0 : E(|Y
1
|

) < }
(0, 2] and E(Y
1
) = 0 for > 1, then (under a weak regu-
larity condition)
S
n
n
1/
(n)
Z in distribution, as n .
for a slowly varying . Z has a so-called stable distribution.
Think 1. The family of stable distributions has three
parameters, (0, 2], c

R
+
, E(|Z|

) < <
(or = 2).
6
Theorem 4 If = sup{ 0 : E(|Y
1
|

) < } (0, 2] and

E(Y
1
) = 0 for > 1, then (under a weak regularity cond.)
X
(n)
t
=
S
[nt]
n
1/
(n)
X
t
in distribution, as n .
Also, X
(n)
X where X is a so-called stable Levy process.
16 8 0
4
0
-4
Page 1
64 32 0
64
0
-64
Page 1
256 128 0
1024
0
-1024
Page 1
One can show that (X
ct
)
t0
(c
1/
X
t
)
t0
(scaling).
7
To get to full generality, we need triangular arrays.
Theorem 5 (Khintchine) Let Y
(n)
k
, k = 1, . . . , n, be i.i.d.
with distribution changing with n 1, and such that
Y
(n)
1
0, in probability, as n .
If
S
(n)
n
Z in distribution,
then Z has a so-called innitely divisible distribution.
Theorem 6 (Skorohod) In Theorem 5, k 1, n 1,
X
(n)
t
= S
(n)
[nt]
X
t
in distribution, as n .
Furthermore, X
(n)
X where X is a Levy process.
8
Do we really want to study Levy processes as limits?
No! Not usually. But a few observations can be made:
Brownian motion is a very important Levy process.
Jumps seem to be arising naturally (in the stable case).
We seem to be restricting the increment distributions to:
Denition 3 A r.v. Z has an innitely divisible distribution
if Z=

S
(n)
n
for all n 1 and suitable random walks

S
(n)
.
Example for Theorems 5 and 6: B(n, p
n
) Poi() for
np
n
is a special case of Theorem 5 (Y
(n)
k
B(1, p
n
)),
and Theorem 6 turns out to give an approximation of the
Poisson process by Bernoulli random walks.
9
2. Examples
Example 1 (Brownian motion) X
t
N(0, t).
Example 2 (Stable process) X
t
stable.
Example 3 (Poisson process) X
t
Poi(t/2)
To emphasise the presence of jumps, we remove the verti-
cal lines.
256 128 0
16
0
-16
Page 1
256 128 0
1024
0
-1024
Page 1
16 8 0
8
4
0
Page 1
10
3. Classication and construction of Levy proc.
Why restrict to innitely divisible increment distrib.?
You have restrictions for random walks: nstep increments
S
n
must be divisible into n iid increments, for all n 2.
For a Levy process, X
t
, t > 0, must be divisible into n 2
iid random variables
X
t
=
n

k=1
(X
tk/n
X
t(k1)/n
),
since these are successive increments (hence independent)
of equal length (hence identically distrib., by stationarity).
11
Approximations are cumbersome. Often nicer direct argu-
ments exist in continuous time, e.g., because the class of
innitely divisible distributions can be parametrized nicely.
Theorem 7 (Levy-Khintchine) A r.v. Z is innitely divis-
ible i its characteristic function E(e
iZ
) is of the form
exp
_
i
1
2

2
+
_
R

_
e
ix
1 ix1
{|x|1}
_
(dx)
_
where R,
2
0 and is a measure on R

such that
_
R

(1 x
2
)(dx) < , [(dx) f(x)dx.]
Only if is hard to prove. We give an indication of if by
constructing the associated Levy processes.
12
The hidden beauty of this abstract parametrization can
only be appreciated when interpreted for the Levy process
with given innitely divisible increment distribution.
Theorem 8 (Levy-Ito) Let X be a Levy process, the dis-
tribution of X
1
parametrized by (,
2
, ). Then X decom-
poses
X
t
= t +B
t
+J
t
+M
t
where B is a Brownian motion, and X
t
= X
t
X
t
, t 0,
an independent Poisson point process with intensity mea-
sure ,
J
t
=

st
X
s
1
{|X
s
|>1}
and M is a martingale with jumps M
t
= X
t
1
{|X
s
|1}
.
13
4. Examples
Z 0 E(e
Z
) = exp
_

_
(0,)
_
1 e
x
_
(dx)
_
Example 4 (Poisson proc.)

= 0, (dx) =
1
(dx).
Example 5 (Gamma process)

= 0, (x) = f(x)dx with

f(x) = ax
1
exp{bx}, x > 0. Then X
t
Gamma(at, b).
Example 6 (stable subordinator)

= 0, f(x) = cx
3/2
Page 1 Page 1 Page 1
14
Example 7 (Compound Poisson process) =
2
= 0.
Choose jump distribution, e.g. density g(x), intensity > 0,
(dx) = g(x)dx. J in Theorem 8 is compound Poisson.
Example 8 (Brownian motion) = 0,
2
> 0, 0.
Example 9 (Cauchy process) =
2
= 0, (dx) = f(x)dx
with f(x) = x
2
, x > 0, f(x) = |x|
2
, x < 0.
Page 1 Page 1 Page 1
15
Page 1
16
Page 1
17
Page 1
Page 1 Page 1
Page 1 Page 1 Page 1
Page 1 Page 1 Page 1
= 5 = 0 = 2
18
5. Poisson point processes and simulation
What does it mean that (X
t
)
t0
is a Poisson point pro-
cess with intensity measure ?
Denition 4 A stochastic process (H
t
)
t0
in a measurable
space E = E

{0} is called a Poisson point process with

intensity measure on E

if
N
t
(A) = #{s t : H
s
A} , t 0, A E

measurable
satises
N
t
(A), t 0, is a Poisson process with intensity (A).
For A
1
, . . . , A
n
disjoint, the processes N
t
(A
1
), . . . , N
t
(A
n
),
t 0, are independent.
19
N
t
(A), t 0, counts the number of points in A, but does
not tell where in A they are. Their distribution on A is :
Theorem 9 (Ito) For all measurable A E

with M =
(A) < , denote the jump times of N
t
(A), t 0, by
T
n
(A) = inf{t 0 : N
t
(A) = n}, n 1.
Then
Z
n
(A) = H
T
n
(A)
, n 1,
are independent of N
t
(A), t 0, and iid with common
distribution M
1
( A).
This is useful to simulate H
t
, t 0.
20
Simulation of (X
t
)
t0
Time interval [0, 5]
> 0 small, M = ((, )
c
) <
N Poi(5((, )
c
))

T
1
, . . . ,

T
N
U(0, 5)
X

T
j
M
1
( (, )
c
)
5 4 3 2 1 0
40
20
0
-20
-40
-60
Page 1
5 4 3 2 1 0
1
0
-1
Page 1

zoom

5 4 3 2 1 0
40
20
0
-20
-40
-60
Page 1
21
Corrections to improve approximation
First, add independent Brownian motion t +B
t
.
If X symmetric, small means small error.
If X not symmetric, we need a drift correction t
=
_
A[1,1]
x(dx), where A = (, )
c
.
Also, one can often add a Normal variance correction C
t

2
=
_
(,)
x
2
(dx), C independent Brownian motion
to account for the many small jumps thrown away. This
can be justied by a version of Donskers theorem and
E(X
1
) = +
_
[1,1]
c
x(dx), V ar(X
1
) =
2
+
_
R

x
2
(dx).
22
Decomposing a general Levy process X
Start with intervals of jump sizes in sets A
n
with
_
n1
A
n
= (0, ), A
n
= A
n
,
_
n1
A
n
= (, 0)
so that (A
n
+A
n
) 1, say. We construct X
(n)
, n Z

,
independent Levy processes with jumps in A
n
according to
( A
n
) (and drift correction
n
t). Now
X
t
=

nZ
X
(n)
t
, where X
(0)
t
= t +B
t
.
In practice, you may cut o the series when X
(n)
t
is small,
and possibly estimate the remainder by a Brownian motion.
23
A spectrally negative Levy process: ((0, )) = 0

2
= 0,
(dx) = f(x)dx,
f(x) = |x|
5/2
,
x [3, 0),
= 0.845

1
= 0,

0.3
= 1.651,

0.1
= 4.325,

0.01
= 18,

as 0.
eps=1
15 10 5 0
8
4
0
-4
Page 1
eps=0.01
15 10 5 0
4
0
-4
-8
Page 1
eps=0.3
15 10 5 0
4
0
-4
-8
Page 1
eps=0.1
15 10 5 0
4
0
-4
-8
Page 1
24
Another look at the Levy-Khintchine formula
E(e
iX
1
) = E
_
_
_exp
_

_
i

nZ
X
(n)
1
_

_
_
_
_ =

nZ
E(e
iX
(n)
1
)
can be calculated explicitly, for n Z

(n = 0 obvious)
E(e
iX
(n)
1
) = E
_
_
exp
_
_
_
i
_
_

n
+
N

k=1
Z
k
_
_
_
_
_
_
_
= exp
_
i
_
A
n
[1,1]
x(dx)
_

m=0
P(N = m)
_
E(e
iZ
1
)
_
m
= exp
__
A
n
_
e
ix
1ix1
{|x|1}
_
(dx)
_
to give for E(e
iX
1
) the Levy-Khintchine formula
exp
_
i
1
2

2
+
_
R

_
e
ix
1 ix1
{|x|1}
_
(dx)
_
25
Summary
Weve dened Levy processes via stationary independent
increments.
Weve seen how Brownian motion, stable processes and
Poisson processes arise as limits of random walks, indi-
cated more general results.
Weve analysed the structure of general Levy processes and
given representations in terms of compound Poisson pro-
cesses and Brownian motion with drift.
Weve simulated Levy processes from their marginal distri-
butions and from their Levy measure.
26