Documente Academic
Documente Profesional
Documente Cultură
Kazuhisa Matsuda
Department of Economics
The Graduate Center, The City University of New York,
365 Fifth Avenue, New York, NY 10016-4309
Email: maxmatsuda@maxmatsuda.com
http://www.maxmatsuda.com/
January 2005
Abstract
This paper presents the basic knowledge of a standard Brownian motion which is a
building block of all stochastic processes. A standard Brownian motion is a subclass of 1)
continuous martingales, 2) Markov processes, 3) Gaussian processes, and 4) Itô diffusion
processes. It is also a subclass of Lévy processes although we will not discuss this in this
sequel.
1
© 2005 Kazuhisa Matsuda All rights reserved.
(1) Its increments are independent. In other words, for 0 ≤ t1 < t2 < ... < tn < ∞ :
(2) Its increments are stationary (time homogeneous): i.e. for h ≥ 0 , Bt + h − Bt has the
same distribution as Bh . In other words, the distribution of increments does not depend
on t .
(3) P ( B0 = 0) = 1 . The process starts from 0 almost surely (with probability 1).
(4) Bt ∼ Normal (0, t ) . Its increments follow a Gaussian distribution with the mean 0 and
the variance t .
Definition 1.2 Standard Brownian motion with starting point c Let c be a real
valued constant or a random variable independent of a standard Brownian motion
( Bt∈[0,∞ ) ) . Then, a standard Brownian motion with starting point c is a real valued
stochastic process defined on a filtered probability space (Ω, Ft∈[0,∞ ) , P ) :
(c + Bt∈[0,∞ ) ) .
(2) Its sample path (trajectory) is continuous in t (i.e. continuous ∈ rcll) almost surely.
Proof
Consult Karlin (1975). We have to remind you that this proof is not that simple.
2
© 2005 Kazuhisa Matsuda All rights reserved.
Definition 1.3 Brownian motion with drift Let ( Bt∈[0,∞ ) ) be a standard Brownian
motion process defined on a filtered probability space (Ω, Ft∈[0,∞ ) , P ) . Then, a Brownian
motion with drift is a real valued stochastic process defined on a filtered probability
space (Ω, Ft∈[0,∞ ) , P ) as:
( X t∈[0,∞ ) ) ≡ ( µ t + σ Bt∈[0,∞ ) ) ,
(1) Its increments are independent. In other words, for 0 ≤ t1 < t2 < ... < tn < ∞ :
P ( X t0 ∩ X t1 − X t0 ∩ X t2 − X t1 ∩ ... ∩ X tn − X tn−1 )
= P ( X t0 ) P ( X t1 − X t0 ) P ( X t2 − X t1 )...P ( X tn − X tn−1 ) .
(2) Its increments are stationary (time homogeneous): i.e. for h ≥ 0 , X t + h − X t has the
same distribution as X h . In other words, the distribution of increments does not depend
on t .
(3) X t ≡ µ t + σ Bt ∼ Normal ( µ t , σ 2t ) . Its increments follow a Gaussian distribution with
the mean µt and the variance σ 2t .
(4) Its sample path (trajectory) is continuous in t (i.e. continuous ∈ rcll) almost surely.
Before discussing the sample paths properties of Brownian motion, take a look at
simulated sample paths of a standard Brownian motion on Panel (A) in Figure 1.1 and
those of a Brownian motion with drift on Panel (B) and (C).
20
10
-10
-20
0 100 200 300 400 500
Time
3
© 2005 Kazuhisa Matsuda All rights reserved.
50
0 µ=0.2,σ=1
-50 µ=0,σ=1
-100 µ=−0.2,σ=1
B) Sample Paths of a Brownian Motion with Drift. Different drifts and same diffusion
parameters.
250
200
µ=0,σ=1
150
100 µ=0,σ=5
50
µ=0,σ=10
0
C) Sample Paths of a Brownian Motion with Drift. Zero drifts and different diffusion
parameters.
Theorem 1.2 Sample paths properties of Brownian motion with drift Consider a
real valued Brownian motion with drift ( X t∈[0,T ] ) ≡ ( µ t + σ Bt∈[0,∞ ) ) defined on a filtered
probability space (Ω, Ft∈[0,∞ ) , P ) . Then, the sample paths of ( X t∈[0,T ] ) possess following
properties:
4
© 2005 Kazuhisa Matsuda All rights reserved.
(2) Sample paths are of infinite variation on any finite interval [0, t ] . In other words, the
total variation on any finite interval [0, t ] of a sample path of a Brownian motion with
drift is infinite with probability 1 in the limit n → ∞ (as the partition becomes finer and
finer):
⎛ n
⎞
P ⎜ lim T ( X ) = lim sup ∑ X (ti ) − X (ti −1 ) = ∞ ⎟ = 1 .
⎝ n→∞ n →∞
i =1 ⎠
Intuitively speaking, the infinite variation property means highly oscillatory sample
paths.
(3) The quadratic variations of sample paths of Brownian motions with drift ( X t∈[0,T ] ) are
finite on any finite interval [0, t ] and converge to σ 2t with probability 1 in the limit
n → ∞ (as the partition becomes finer and finer):
⎛ n
⎞
P ⎜ lim T 2 ( X ) = lim sup ∑ X (ti ) − X (ti −1 ) = σ 2t < ∞ ⎟ = 1 .
2
⎝ n→∞ n →∞
i =1 ⎠
For more details and proofs about theorem 1.2, consult Sato (1999) page 22 – 28 and
Karatzas and Shreve (1991) section 1.5 and 2.9. We also recommend Rogers and
Williams (2000) chapter 1.
(1) A standard Brownian motion ( Bt∈[0,∞ ) ) is symmetric. In other words, the process
( − Bt∈[0,∞ ) ) is also a standard Brownian motion:
( Bt∈[0,∞ ) ) d (− Bt∈[0,∞ ) ) .
(2) A standard Brownian motion ( Bt∈[0,∞ ) ) has a time shifting property. In other words,
the process ( Bt + A − BA ) is also a standard Brownian motion for ∀A ∈ R + :
( Bt + A − BA ) d ( Bt∈[0,∞ ) ) .
(3) Time scaling property of a standard Brownian motion. For any nonzero c ∈ R , the
1
process ( cBt / c ) or ( Bct ) is also a standard Brownian motion:
c
5
© 2005 Kazuhisa Matsuda All rights reserved.
1
( Bct ) d ( cBt / c ) d ( Bt∈[0,∞ ) ) .
c
(4) Time inversion property of a standard Brownian motion (i.e. a variant of (3)). The
process defined as:
⎧0 if t = 0
( Bt∈[0,∞ ) ) = ⎨ ,
⎩(tB1/ t ) if 0 < t < ∞
( Bt∈[0,∞ ) ) d ( Bt∈[0,∞ ) ) .
Proof
These are easy exercises for readers. For the proof of the continuity of ( Bt∈[0,∞ ) ) at 0,
consult Rogers and Williams (2000) page 4.
Consider a real valued Brownian motion with drift process ( X t∈[0,∞ ) ) ≡ ( µ t + σ Bt∈[0,∞ ) )
defined on a filtered probability space (Ω, Ft∈[0,∞ ) , P ) . Its characteristic function can be
obtained by the direct use of the definition of a characteristic function (i.e. Fourier
transform of the probability density function with Fourier transform parameters (1,1) ):
∞
φ X (ω ) ≡ F [ P( x)] ≡ ∫ eiω x P( x)dx
t −∞
∞ 1 ⎧⎪ ( x − µ t )2 ⎫⎪
φ X t (ω ) = ∫ e iω x
exp ⎨− ⎬ dx
−∞
2πσ 2t ⎪⎩ 2σ 2t ⎪
⎭
σ tω
2 2
φ X t (ω ) = exp(i µ tω − ).
2
(1) X t is nonanticipating.
6
© 2005 Kazuhisa Matsuda All rights reserved.
In other words, if a stochastic process is a martingale, then, the best prediction of its
future value is its present value. Note that the definition of martingale makes sense only
when the underlying probability measure P and the filtration Ft have been specified.
The fundamental property of a martingale process is that its future variations are
completely unpredictable with the filtration Ft :
∀u > 0, E[ xt +u − xt Ft ] = E[ xt +u Ft ] − E[ xt Ft ] = xt − xt = 0 .
Finite mean condition (2) is necessary to ensure the existence of the conditional
expectation.
Proof
u
Bu = Bt + ∫ dBv . (1)
t
Using the equation (1) and the fact that a Brownian motion is a nonanticipating process,
i.e. E[ Bt Ft ] = Bt :
u
E[ Bu − Bt Ft ] = E[ Bu Ft ] − E[ Bt Ft ] = E[ Bt + ∫ dBv Ft ] − Bt
t
u
E[ Bu − Bt Ft ] = E[ Bt Ft ] + E[ ∫ dBv Ft ] − Bt
t
E[ Bu − Bt Ft ] = Bt + 0 − Bt = 0 ,
or in other words:
7
© 2005 Kazuhisa Matsuda All rights reserved.
u u
E[ Bu Ft ] = E[ Bt + ∫ dBv Ft ] = E[ Bt Ft ] + E[ ∫ dBv Ft ] = Bt + 0
t t
E[ Bu Ft ] = Bt ,
Proof
Using the equation (1) and independent increments condition, for 0 ≤ t ≤ u < ∞ :
u
E[ Bu 2 Ft ] = E[( Bt + ∫ dBv )2 Ft ]
t
u u
E[ Bu Ft ] = E[ Bt + 2 Bt ∫ dBv + ∫ dBv 2 Ft ]
2 2
t t
u u
E[ Bu 2 Ft ] = E[ Bt 2 Ft ] + E[2 Bt ∫ dBv Ft ] + E[ ∫ dBv 2 Ft ]
t t
u
E[ Bu Ft ] = E[ Bt Ft ] + E[ ∫ dBv Ft ]
2 2 2
t
E[ Bu Ft ] = Bt + (u − t ) ,
2 2
Proof
Using the equation (1) and independent increments condition, for 0 ≤ t ≤ u < ∞ :
u
E[ Bu 2 − u Ft ] = E[( Bt 2 − t ) + {( ∫ dBv ) 2 − (u − t )} Ft ]
t
8
© 2005 Kazuhisa Matsuda All rights reserved.
u
E[ Bu 2 − u Ft ] = E[( Bt 2 − t ) Ft ] + E[( ∫ dBv )2 − (u − t ) Ft ]
t
u
E[ Bu 2 − u Ft ] = Bt 2 − t + E[( ∫ dBv )2 Ft ] − (u − t )
t
E[ Bu − u Ft ] = Bt − t + (u − t ) − (u − t ) = Bt 2 − t ,
2 2
Theorem 2.5 Brownian motion with drift is not a continuous martingale Let
( Bt∈[0,∞ ) ) be a standard Brownian motion process defined on a filtered probability space
(Ω, Ft∈[0,∞ ) , P ) . Then, a Brownian motion with drift ( X t∈[0,∞ ) ) ≡ ( µ t + σ Bt∈[0,∞ ) ) is not a
continuous martingale with respect to the filtration Ft∈[0,∞ ) and the probability measure P .
Proof
u
X u = X t + ∫ dX v . (2)
t
Using the equation (2) and the fact that a Brownian motion with drift is a nonanticipating
process, i.e. E[ X t Ft ] = X t :
u u
E[ X u Ft ] = E[ X t + ∫ dX v Ft ] = E[ X t Ft ] + E[ ∫ dX v Ft ]
t t
E[ X u Ft ] = X t + µ (u − t ) ,
9
© 2005 Kazuhisa Matsuda All rights reserved.
is a continuous martingale with respect to the filtration Ft∈[0,∞ ) and the probability
measure P .
Proof
For ∀0 ≤ t ≤ u < ∞ :
u u
E[ X u − µu Ft ] = E[( X t − µt ) + ( ∫ dX v − µ ∫ dv) Ft ]
t t
u u
E[ X u − µu Ft ] = E[( X t − µt ) Ft ] + E[( ∫ dX v − µ ∫ dv) Ft ]
t t
E[ X u − µu Ft ] = X t − µt + µ (u − t ) − µ (u − t )
E[ X u − µu Ft ] = X t − µt ,
1
Z t = exp(θ Bt − θ 2t ) , (3)
2
is a continuous martingale with respect to the filtration Ft∈[0,∞ ) and the probability
measure P .
Proof
10
© 2005 Kazuhisa Matsuda All rights reserved.
1
E[exp(θ X )] = exp(θµ t + θ 2σ 2t ) . (4)
2
Proof
∞ 1 ( X − µt )2
E[exp(θ X )] = ∫ exp(θ X ) exp{− }dX
−∞
2πσ 2t 2σ 2t
∞ 1 −θ X 2σ 2t + X 2 − 2 X µ t + µ 2t 2
=∫ exp{− }dX
−∞
2πσ 2t 2σ 2t
∞ 1 X 2 − 2(θσ 2t + µ t ) X + µ 2t 2
=∫ exp{− }dX
−∞
2πσ 2t 2σ 2t
∞ 1 ( X − (θσ 2t + µ t )) 2 − (θσ 2t + µ t ) 2 + µ 2t 2
=∫ exp{− }dX
−∞
2πσ 2t 2σ 2t
∞ 1 ( X − (θσ 2t + µ t )) 2 (θσ 2t + µ t ) 2 − µ 2t 2
=∫ exp{− }exp{ }dX
−∞
2πσ 2t 2σ 2t 2σ 2t
(θσ 2t + µ t ) 2 − µ 2t 2 ∞ 1 ( X − (θσ 2t + µ t )) 2
= exp{ }∫ exp{− }dX
2σ 2t −∞
2πσ 2t 2σ 2t
(θσ 2t + µt ) 2 − µ 2t 2 θ 2σ 4t 2 + 2θσ 2t µ t
= exp{ } = exp{ }
2σ 2t 2σ 2t
1
= exp{θµ t + θ 2σ 2t}
2
Now we are ready to prove the Brownian exponential defined by the equation (3) is a
martingale.
1
E[ Z t ] = E[exp(θ Bt − θ 2t )]
2
1
E[ Z t ] = E[exp(θ Bt ) exp(− θ 2t )]
2
1
E[ Z t ] = exp( − θ 2t ) E[exp(θ Bt )] ,
2
1
using the proposition 2.1, E[exp(θ Bt )] = exp( θ 2t ) :
2
11
© 2005 Kazuhisa Matsuda All rights reserved.
1 1
E[ Z t ] = exp(− θ 2t ) exp( θ 2t ) = 1 .
2 2
1
E[ Z t + h Ft ] = E[exp{θ Bt + h − θ 2 (t + h)} Ft ] .
2
1
E[ Z t + h Ft ] = E[exp(θ Bt − θ Bt ) exp{θ Bt + h − θ 2 (t + h)} Ft ]
2
1 1
E[ Z t + h Ft ] = E[exp(θ Bt ) exp(−θ Bt ) exp(θ Bt + h ) exp(− θ 2t ) exp(− θ 2 h) Ft ]
2 2
1 1
E[ Z t + h Ft ] = E[exp(θ Bt − θ 2t ) exp{θ ( Bt + h − Bt ) − θ 2 h} Ft ] .
2 2
1 1
E[ Z t + h Ft ] = E[exp(θ Bt − θ 2t ) Ft ]E[exp{θ ( Bt + h − Bt ) − θ 2 h} Ft ] ,
2 2
1 1
E[ Z t + h Ft ] = exp(θ Bt − θ 2t ) E[exp{θ ( Bt + h − Bt ) − θ 2 h}] .
2 2
By the definition of Z t :
1
E[ Z t + h Ft ] = Z t E[exp{θ ( Bt + h − Bt )}exp(− θ 2 h)] ,
2
1
and since exp(− θ 2 h) is a constant:
2
1
E[ Z t + h Ft ] = Z t exp(− θ 2 h) E[exp{θ ( Bt + h − Bt )}] .
2
12
© 2005 Kazuhisa Matsuda All rights reserved.
1 1
E[ Z t + h Ft ] = Z t exp(− θ 2 h) exp( θ 2 h)
2 2
E[ Z t + h Ft ] = Z t .
(1) Means E[ X t ] .
13
© 2005 Kazuhisa Matsuda All rights reserved.
Proof
By the definition of a standard Brownian motion. For more details, we recommend Karlin
and Taylor (1975) page 376-377.
Theorem 3.2 A standard Brownian motion A real valued one dimensional Gaussian
stochastic process ( X t∈[0,∞ ) ) defined on a filtered probability space (Ω, Ft∈[0,∞ ) , P ) is a
standard Brownian motion with drift, if its mean and covariance function satisfy:
(1) E[ X t ] = 0 .
(2) Cov( X t , X u ) = t ∧ u = min{t , u}
Proof
14
© 2005 Kazuhisa Matsuda All rights reserved.
We first introduce the definitions and terminologies used in the study of Markov
processes.
Pt ,v ( x, B ) = P0,v −t ( x, B ) = Pv −t ( x, B ) ,
which indicates that the transition function Pt ,v ( x, B ) of moving from a state x at time t
to a state B at time v is equivalent to the transition function P0,v −t ( x, B ) of moving from
a state x at time 0 to a state B at time v − t . In other words, the transition function is
independent of the time t and depends only on the interval of time v − t .
∫
R
P0,t ( x, dy )P0,u ( y, B) = ∫ Pt ( x, dy )Pu ( y, B) = Pt +u ( x, B) .
R
15
© 2005 Kazuhisa Matsuda All rights reserved.
P ( X u Ft ) = P ( X u X 0 , X t1 , X t2 ,..., X tn , X t ) = P ( X u X t ) ,
Informally, Markov property means that the probability of a random variable X u at time
u ≥ t (tomorrow) conditional on the entire history of the stochastic process F[0,t ] ≡ X [0,t ]
is equal to the probability of a random variable X u at time u ≥ t (tomorrow) conditional
only on the value of a random variable at time t (today). In other words, the history
(sample path) of the stochastic process F[0,t ] is of no importance in that the way this
stochastic process evolved or the dynamics does not mean a thing in terms of the
conditional probability of the process. This property is sometimes called a memoryless
property.
E[ X v Ft ] = E[ X v X t ] ,
Pt ,v ( x, B) = ∫ Pt ,u ( x, dy )Pu ,v ( y, B) .
R
(1) It is a time homogeneous Markov process. In other words, for any bounded Borel
function f : R → R and for ∀0 ≤ t ≤ u < ∞ :
E[ f ( Bu ) Ft ] = P0,u −t f ( Bt ) = Pu −t f ( Bt ) .
16
© 2005 Kazuhisa Matsuda All rights reserved.
1 ⎧ ( x − y)2 ⎫
Ph ( x, y ) = exp ⎨− ⎬.
2π t ⎩ 2t ⎭
⎧⎪ f ( x) if t = 0
(3) Ph f ( x) = ⎨ ∞ .
⎪⎩ ∫−∞ h
P ( x , y ) f ( y ) dy if t > 0
Proof
P ( X u − X t X t1 − X 0 , X t2 − X t1 ,..., X t − X tn )
P( X u − X t ∩ X t1 − X 0 , X t2 − X t1 ,..., X t − X tn )
=
P( X t1 − X 0 , X t2 − X t1 ,..., X t − X tn )
P( X u − X t )P( X t1 − X 0 , X t2 − X t1 ,..., X t − X tn )
=
P( X t1 − X 0 , X t2 − X t1 ,..., X t − X tn )
= P( X u − X t ) ,
P ( X u X 0 , X t1 , X t2 ,..., X tn , X t ) = P (( X u − X t ) + X t X 0 , X t1 , X t2 ,..., X tn , X t )
= P( X u X t ) ,
17
© 2005 Kazuhisa Matsuda All rights reserved.
Definition 5.1 Itô diffusion processes An Itô diffusion process is a real valued
stochastic process ( X t∈[0,∞ ) ) defined on a filtered probability space (Ω, Ft∈[0,∞ ) , P ) whose
dynamics (or motion) is governed by a stochastic differential equation of the form:
X t = X 0 + ∫ b( s, X s )ds + ∫ σ ( s, X s )dBs .
t t
0 0
dX t = b( X t )dt + σ ( X t )dBt ,
Theorem 5.1 Time homogeneous Itô diffusion processes are a subclass of time
homogeneous Markov processes A time homogeneous Itô diffusion process ( X t∈[0,∞ ) )
18
© 2005 Kazuhisa Matsuda All rights reserved.
E[ f ( Bu ) Ft ] = P0,u −t f ( Bt ) = Pu −t f ( Bt ) ,
⎧⎪ f ( x) if t = 0
Ph f ( x) = ⎨ ∞ .
⎪⎩ ∫−∞ Ph ( x , y ) f ( y ) dy if t > 0
Proof
dX t = dBt .
Proof
For more details about Itô diffusion processes, consult an excellent book Oksendal (2003)
chapter 3 and 7.
19
© 2005 Kazuhisa Matsuda All rights reserved.
References
Karlin, S., and Taylor, H., 1975, A First Course in Stochastic Processes, Academic Press.
Karatzas, Ioannis., and Shreve, S. E., 1991, Brownian Motion and Stochastic Calculus,
Springer-Verlag.
Rogers, L.C.G., and Williams, D., 2000, Diffusions, Markov Processes and Martingales,
Cambridge University Press.
Sato, K., 1999, Lévy process and Infinitely Divisible Distributions, Cambridge
University Press.
20