Brownian Motion Stochastic Calculus

Chapter 4
Brownian Motion and Stochastic Calculus
Brownian motion is a continuous-time stochastic process having stationary

and independent Gaussian distributed increments, and continuous paths.
This chapter presents the constructions of Brownian motion and its asso-
ciated Itô stochastic integral, which will be used for the random modeling of
asset and portfolio prices in continuous time.
4.1 Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

4.2 Three Constructions of Brownian Motion . . . . . . . . 135
4.3 Wiener Stochastic Integral . . . . . . . . . . . . . . . . . . . . . . 139
4.4 Itô Stochastic Integral . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.5 Stochastic Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
4.1 Brownian Motion
We start by recalling the definition of Brownian motion, which is a funda-

mental example of a stochastic process. The underlying probability space
(Ω, F, P) of Brownian motion can be constructed on the space Ω = C0 (R+ )
of continuous real-valued functions on R+ started at 0.
Definition 4.1. The standard Brownian motion is a stochastic process
(Bt )t∈R+ such that
1. B0 = 0 almost surely,
2. The sample trajectories t 7→ Bt are continuous, with probability 1.
3. For any finite sequence of times t0 < t1 < · · · < tn , the increments
Bt1 − Bt0 , Bt2 − Bt1 , . . . , Btn − Btn−1
" 131
This version: December 23, 2020
https://personal.ntu.edu.sg/nprivault/indext.html
N. Privault
are mutually independent random variables.
4. For any given times 0 6 s < t, Bt − Bs has the Gaussian distribution

N (0, t − s) with mean zero and variance t − s.
In particular, for t ∈ R+ , the random variable Bt ' N (0, t) has a Gaussian
distribution with mean zero and variance t > 0. Existence of a stochastic pro-
cess satisfying the conditions of Definition 4.1 will be covered in Section 4.2.
In Figure 4.1 we draw three sample paths of a standard Brownian motion

obtained by computer simulation using (4.3). Note that there is no point in
“computing” the value of Bt as it is a random variable for all t > 0. How-
ever, we can generate samples of Bt , which are distributed according to the
centered Gaussian distribution with variance t.
Bt 3
Bt 2
Bt 1
0 t1 t2 t3
-1
0 0.2 0.4 0.6 0.8 1
Fig. 4.1: Sample paths of a one-dimensional Brownian motion.
In particular, Property 4 in Definition 4.1 implies
IE[Bt − Bs ] = 0 and Var[Bt − Bs ] = t − s, 0 6 s 6 t,
and we have
Cov(Bs , Bt ) = IE[Bs Bt ]
= IE[Bs (Bt − Bs + Bs )]
= IE Bs (Bt − Bs ) + (Bs )2

= IE[Bs (Bt − Bs )] + IE (Bs )2

= IE[Bs ] IE[Bt − Bs ] + IE (Bs )2

= Var[Bs ]
= s, 0 6 s 6 t,
hence
132 "
Cov(Bs , Bt ) = IE[Bs Bt ] = min(s, t), s, t ∈ R+ , (4.1)

cf. also Exercise 4.2-(4.1). The following graphs present two examples of
possible modeling of random data using Brownian motion.
Fig. 4.2: Evolution of the fortune of a poker player vs number of games played.
Fig. 4.3: Web traffic ranking.
In the sequel, we denote by (Ft )t∈R+ the filtration generated by the Brownian
paths up to time t, defined as
F t : = σ ( Bs : 0 6 s 6 t ) , t > 0. (4.2)
Property 3 in Definition 4.1 shows that Bt − Bs is independent of all Brownian

increments taken before time s, i.e.
⊥ (Bt1 − Bt0 , Bt2 − Bt1 , . . . , Btn − Btn−1 ),

( Bt − Bs ) ⊥
0 6 t0 6 t1 6 · · · 6 tn 6 s 6 t, hence Bt − Bs is also independent of the

whole Brownian history up to time s, hence Bt − Bs is in fact independent
of Fs , s > 0. As in Example 2 page 2 we have the following result.
" 133
N. Privault
Proposition 4.2. Brownian motion (Bt )t∈R+ is a continuous-time martin-

gale.
Proof. We have
IE[Bt | Fs ] = IE[Bt − Bs + Bs | Fs ]
= IE[Bt − Bs | Fs ] + IE[Bs | Fs ]
= IE[Bt − Bs ] + Bs
= Bs , 0 6 s 6 t,
because it has centered and independent increments, cf. Section 7.1.

The n-dimensional Brownian motion can be constructed as (Bt1 , Bt2 , . . . , Btn )t∈R+
where (Bt1 )t∈R+ , (Bt2 )t∈R+ , . . .,(Btn )t∈R+ are independent copies of (Bt )t∈R+ .
Next, we turn to simulations of 2 dimensional and 3 dimensional Brownian
motions in Figures 4.4 and 4.5. Recall that the movement of pollen particles
originally observed by R. Brown in 1827 was indeed 2-dimensional.
2
1.5
0.5
-0.5
-1
-1.5
-2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
Fig. 4.4: Two sample paths of a two-dimensional Brownian motion.
-1
-2
-1
0
1 2
1
0
2 -1
-2
Fig. 4.5: Sample path of a three-dimensional Brownian motion.
134 "
Figure 4.6 presents an illustration of the scaling property of Brownian motion.
Fig. 4.6: Scaling property of Brownian motion.∗
4.2 Three Constructions of Brownian Motion

We refer to Chapter 1 of Revuz and Yor (1994) and to Theorem 10.28 in
Folland (1999) for the proof of existence of Brownian motion as a stochastic
process (Bt )t∈R+ satisfying the above Conditions 1-4.
Brownian motion as a random walk
For convenience we will informally regard Brownian motion as a random walk

over infinitesimal time intervals of length ∆t, whose increments
∆Bt := Bt+∆t − Bt ' N (0, ∆t)
over the time interval [t, t + ∆t] will be approximated by the Bernoulli random
variable √
∆Bt = ± ∆t (4.3)
with equal probabilities (1/2, 1/2). Figure 4.7 presents a simulation of Brow-
nian motion as a random walk with ∆t = 0.1.
∗
The animation works in Acrobat Reader on the entire pdf file.
" 135
N. Privault
2.5
1.5
Bt
1
0.5
-0.5
-1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
t
Fig. 4.7: Construction of Brownian motion as a random walk.∗
Remark 4.3. The choice of the square root in (4.3) is in fact not fortuitous.
Indeed, any choice of ±(∆t)α with a power α > 1/2 would lead to explosion
of the process as dt tends to zero, whereas a power α ∈ (0, 1/2) would lead
to a vanishing process.
Note that we have
1√ 1√
IE[∆Bt ] = ∆t − ∆t = 0,
2 2
and
1 1
Var[∆Bt ] = IE (∆Bt )2 = ∆t + ∆t = ∆t.

2 2
According to this representation, the paths of Brownian motion are not dif-
ferentiable, although they are continuous by Property 2, as we have
√
dBt ± dt 1
' = ± √ ' ±∞. (4.4)
dt dt dt
In order to recover the Gaussian distribution property of the random variable
BT , we can split the time interval [0, T ] into N subintervals
k−1

k
T, T , k = 1, 2, . . . , N ,
N N
of same length ∆t = T /N , with N “large”.
0 T 2T T
N N
∗
136 "
Defining the Bernoulli random variable Xk as

√
Xk := ± T
with equal probabilities (1/2, 1/2), we have Var(Xk ) = T and
X √
∆Bt := √ k = ± ∆t
N
is the increment of Bt over ((k − 1)∆t, k∆t], and we get
X X1 + X2 + · · · + XN
BT ' ∆Bt ' √ .
0<t<T
N
Hence by the central limit theorem we recover the fact that BT has the cen-
tered Gaussian distribution N (0, T ) with variance T , cf. point 4 of the above
Definition 4.1 of Brownian motion, and the illustration given in Figure 4.8.
N=1000; t <- 0:N; dt <- 1.0/N; nsim <- 10 # using Bernoulli samples
X <- matrix((dt)^0.5*(rbinom( nsim * N, 1, 0.5)-0.5)*2, nsim, N)
X <- cbind(rep(0, nsim), t(apply(X, 1, cumsum)))
plot(t, X[1, ], xlab = "time", ylab = "", type = "l", ylim = c(-2, 2), col = 0)
for (i in 1:nsim){lines(t, X[i, ], xlab = "time", type = "l", ylim = c(-2, 2), col = i)}
# using Gaussian samples
nsim=100;X <- matrix(rnorm(nsim*N,mean=0,sd=sqrt(dt)), nsim, N)
X <- cbind(rep(0, nsim), t(apply(X, 1, cumsum)));H<-hist(X[,N])
layout(matrix(c(1,2), nrow =1, byrow = TRUE));par(mar=c(2,2,2,0), oma = c(2, 2, 2, 2))
plot(t*dt, X[1, ], xlab = "time", ylab = "", type = "l", ylim = c(-2, 2), col = 0)
for (i in 1:nsim){lines(t*dt, X[i, ], xlab = "time", type = "l", ylim = c(-2, 2), col = i)}
for (i in 1:nsim){points(N, X[i,N], pch=1, lwd = 5, col = i)}
x <- seq(-2,2, length=100);px <- dnorm(x);par(mar = c(2,2,2,2))
plot(NULL , xlab="", ylab="", xlim = c(0, max(px,H$density)), ylim = c(-2,2),axes=F)
rect(0, H$breaks[1:(length(H$breaks) - 1)], col=rainbow(20,start=0.08,end=0.6), H$density,
H$breaks[2:length(H$breaks)])
lines(px,x, type="l", lty=1, col="black",lwd=2,xlab="",ylab="", main="")
2
1
0
−1
−2
0.0 0.2 0.4 0.6 0.8 1.0
time
Fig. 4.8: Statistics of one-dimensional Brownian paths vs Gaussian distribution.
" 137
N. Privault
Indeed, the central limit theorem states that given any sequence (Xk )k>1 of
independent identically distributed centered random variables with variance
σ 2 = Var[Xk ] = T , the normalized sum
X1 + X2 + · · · + XN
√
N
converges (in distribution) to the centered Gaussian random variable N (0, σ 2 )

with variance σ 2 as N goes to infinity. As a consequence, ∆Bt could in fact
be replaced by any centered random variable with variance ∆t in the above
description.
Lévy’s construction of Brownian motion
Figure 4.9 represents the construction of Brownian motion by successive lin-

ear interpolations, see Problem 4.20 for a proof of existence of Brownian
motion based on this construction.
n= 12
1.5
1.0
0.5
0.0
0.0 0.2 0.4 0.6 0.8 1.0
Fig. 4.9: Lévy’s construction of Brownian motion.∗
The following R code is used to generate Figure 4.9.†
∗
†
Download the corresponding R code or the IPython notebook that can be run
here.
138 "
alpha=1/2;t <- 0:1;dt <- 1;z=rnorm(1,mean=0,sd=dtâlpha)

plot(t*dt, c(0, z), xlab = "t", ylab = "", main = "", type = "l", xaxs="i")
k=0;while (k<10) {readline("Press <return> to continue")
m <- (z+c(0,head(z,-1)))/2;y <- rnorm(length(t)-1,mean=0,sd=(dt/4)âlpha)
x <- m+y;x <- c(matrix(c(x,z), 2, byrow = T));n=2*length(t)-2;t <- 0:n
plot(t*dt/2, c(0, x), xlab = "t", ylab = "", main = "", type = "l", xaxs="i");z=x;dt=dt/2}
Construction by series expansions
Brownian motion on [0, T ] can also be constructed by Fourier synthesis via

the Paley-Wiener series expansion
√
2T X sin n − 1/2 πt/T

X
Bt = Xn fn (t) = Xn , t ∈ [0, T ],
π (n − 1/2)
n>1 n>1
where (Xn )n>1 is a sequence of independent N (0, 1) standard Gaussian ran-

dom variables, as illustrated in Figure 4.10.∗
2
n=500
4
10
12
0 200 400 600 800
Fig. 4.10: Construction of Brownian motion by series expansions.†
4.3 Wiener Stochastic Integral
In this section, we construct the Wiener stochastic integral of square-

integrable deterministic functions of time with respect to Brownian motion.
Recall that the price St of risky assets was originally modeled in Bachelier
(1900) as St := σBt , where σ is a volatility parameter. The stochastic integral
∗
Download the corresponding IPython notebook that can be run here.
†
" 139
N. Privault
wT wT
f (t)dSt = σ f (t)dBt
0 0
can be used to represent the value of a portfolio as a sum of profits and
losses f (t)dSt where dSt represents the stock price variation and f (t) is the
quantity invested in the asset St over the short time interval [t, t + dt].
A naive definition of the stochastic integral with respect to Brownian mo-

tion would consist in letting
wT wT dBt
f (t)dBt := f (t) dt,
0 0 dt
and evaluating the above integral with respect to dt. However this definition
fails because the paths of Brownian motion are not differentiable, cf. (4.4).
Next we present Itô’s construction of the stochastic integral with respect to
Brownian motion. Stochastic integrals will be first constructed as integrals
of simple step functions of the form
n
ai 1(ti−1 ,ti ] (t),
X
f (t) = t ∈ [0, T ], (4.5)
i=1
i.e. the function f takes the value ai on the interval (ti−1 , ti ], i = 1, 2, . . . , n,

with 0 6 t0 < · · · < tn 6 T , as illustrated in Figure 4.11.
f (t)
a2
a1
a4
t0 t1 t2 t3 t4 t
Fig. 4.11: Step function t 7→ f (t).
ti<-c(0,2,4.5,7,9)
ai<-c(0,3,1,2,1,0)
plot(stepfun(ti,ai),xlim = c(0,10),do.points = F,main="")
Recall that the classical integral of f given in (4.5) is interpreted as the area
under the curve f , and computed as
wT n
X
f (t)dt = ai (ti − ti−1 ).
0
i=1
140 "
f (t)
6
a2 b r
b r b r
a1
a4 b r
-
t0 t1 t2 t3 t4 t
Fig. 4.12: Area under the step function t 7→ f (t).
In the next Definition 4.4 we use such step functions for the construction
of the stochastic integral with respect to Brownian motion. The stochastic
integral (4.6) for step functions will be interpreted as the sum of profits and
losses ai (Bti − Bti−1 ), i = 1, 2, . . . , n, in a portfolio holding a quantity ai of
a risky asset whose price variation is Bti − Bti−1 at time i = 1, 2, . . . , n.
Definition 4.4. The stochastic integral with respect to Brownian motion
(Bt )t∈[0,T ] of the simple step function f of the form (4.5) is defined by
wT n
X
f (t)dBt := ai (Bti − Bti−1 ). (4.6)
0
i=1
wT
In the next Lemma 4.5 we determine the probability distribution of f (t)dBt
0
and we show that it is independent of the particular representation (4.5) cho-
sen for f (t).
Lemma 4.5. Let f be a simple step function f of the form (4.5). The stochas-
wT
tic integral f (t)dBt defined in (4.6) has the centered Gaussian distribution
0
wT wT
f (t)dBt ' N 0, |f (t)|2 dt
0 0
hw T i
with mean IE f (t)dBt = 0 and variance given by the Itô isometry
0
hw T i h w T 2 i w T
Var f (t)dBt = IE f (t)dBt = |f (t)|2 dt. (4.7)
0 0 0
Proof. Recall that if X1 , X2 , . . . , Xn are independent Gaussian random vari-
ables with probability distributions N (m1 , σ12 ), . . . , N (mn , σn2 ) then the sum
X1 + · · · + Xn is a Gaussian random variable with distribution
N m1 + · · · + mn , σ12 + · · · + σn2 .

As a consequence, the stochastic integral
" 141
N. Privault
wT n
X
f (t)dBt = ak (Btk − Btk−1 )
0
k =1
of the step function

n
ak 1(tk−1 ,tk ] (t),
X
f (t) = t ∈ [0, T ],
k =1
has the centered Gaussian distribution with mean 0 and variance

hw T
" n #
i X
Var f (t)dBt = Var ak (Btk − Btk−1 )
0
k =1
n
X
= Var[ak (Btk − Btk−1 )]
k =1
X n
= |ak |2 Var[Btk − Btk−1 ]
k =1
X n
= (tk − tk−1 )|ak |2
k =1
X n w tk
= |ak |2 dt
tk−1
k =1
wT X
n
= |ak |2 1(tk−1 ,tk ] (t)dt
0
k =1
wT
= |f (t)|2 dt,
0
since the simple function

n
a2i 1(ti−1 ,ti ] (t),
X
f 2 (t) = t ∈ [0, T ],
i=1
takes the value a2i on the interval (ti−1 , ti ], i = 1, 2, . . . , n, as can be checked

from the following Figure 4.13.
142 "
f2
6
a22 b r
b r b r
a21
a24 b r
-
t0 t1 t2 t3 t4 t
Fig. 4.13: Squared step function t 7→ f 2 (t).

In the sequel, we will make a repeated use of the space L2 ([0, T ]) of square-
integrable functions.
Definition 4.6. Let L2 ([0, T ]) denote the space of (measurable) functions
f : [0, T ] −→ R such that
wT
r
kf kL2 ([0,T ]) := |f (t)|2 dt < ∞, f ∈ L2 ([0, T ]). (4.8)
0
For example, the function f (t) := tα , t ∈ (0, T ], belongs to L2 ([0, T ]) if and

only if α > −1/2, as we have
+∞ if α 6 −1/2,

wT wT



2 2α
f (t)dt = t dt = 1+2α t=T
0 0 t T 1+2α
= < ∞ if α > −1/2.


1 + 2α t=0 1 + 2α

The norm k · kL2 ([0,T ]) on L2 ([0, T ]) induces a distance between two functions
f and g in L2 ([0, T ]), defined as
wT
r
kf − gkL2 ([0,T ]) := |f (t) − g (t)|2 dt < ∞,
0
cf. e.g. Chapter 3 of Rudin (1974) for details.

Definition 4.7. Convergence in L2 ([0, T ]). We say that a sequence (fn )n∈N
of functions in L2 ([0, T ]) converges in L2 ([0, T ]) to another function f ∈
L2 ([0, T ]) if
wT
r
lim kf − fn kL2 ([0,T ]) = lim |f (t) − fn (t)|2 dt = 0.
n→∞ n→∞ 0
" 143
N. Privault
f = function(x){exp(sin(x*1.8*pi))}
for (i in 3:9){n=2î;x<-cumsum(c(0,rep(1,n)))/n;
z<-c(NA,head(x,-1))
y<-c(f(x)-pmax(f(x)-f(z),0),f(1))
t=seq(0,1,0.01);
plot(f,from=0,to=1,ylim=c(0.3,2.9),type="l",lwd=3,col="red",main="",xaxs="i",yaxs="i")
lines(stepfun(x,y),do.points=F,lwd=2,col="blue",main="");
readline("Press <return> to continue");}
2.5
2.0
f
1.5
1.0
0.5
0.0 0.2 0.4 0.6 0.8 1.0
Fig. 4.14: Step function approximation.∗
By e.g. Theorem 3.13 in Rudin (1974) or Proposition 2.4 page 63 of Hirsch

and Lacombe (1999), we have the following result which states that the set
of simple step functions f of the form (4.5) is a linear space which is dense
in L2 ([0, T ]) for the norm (4.8), as stated in the next proposition.
Proposition 4.8. For any function f ∈ L2 ([0, T ]) satisfying (4.8) there
exists a sequence (fn )n∈N of simple functions converging to f in L2 ([0, T ])
in the sense that
wT
r
lim kf − fn kL2 ([0,T ]) = lim |f (t) − fn (t)|2 dt = 0.
n→∞ n→∞ 0
wT
In order to extend the definition (4.6) of the stochastic integral f (t)dBt
0
to any function f ∈ L2 ([0, T ]), i.e. to f : [0, T ] −→ R measurable such that
wT
|f (t)|2 dt < ∞, (4.9)
0
we will make use of the space L2 (Ω) of square-integrable random variables.

Definition 4.9. Let L2 (Ω) denote the space of random variables F : Ω −→
R such that
∗
144 "
q
IE F 2 < ∞.

kF kL2 (Ω×[0,T ]) :=
The norm k · kL2 (Ω) on L2 (Ω) induces the distance

q
IE (F − G)2 < ∞,

kF − GkL2 (Ω) :=
between the square-integrable random variables F and G in L2 (Ω).

Definition 4.10. Convergence in L2 (Ω). We say that a sequence (Fn )n∈N
of random variables in L2 (Ω) converges in L2 (Ω) to another random variable
F ∈ L2 (Ω) if
q
lim kF − Fn kL2 (Ω) = lim IE (F − Fn )2 = 0.

n→∞ n→∞
The next proposition allows us to extend Lemma 4.5 from simple step func-
tions to square-integrable functions in L2 ([0, T ]).
wT
Proposition 4.11. The definition (4.6) of the stochastic integral f (t)dBt
wT 0
can be extended to any function f ∈ L2 ([0, T ]). In this case, f (t)dBt has
0
the centered Gaussian distribution
wT wT
f (t)dBt ' N 0, |f (t)|2 dt
0 0
w
T
with mean IE f (t)dBt = 0 and variance given by the Itô isometry
0
w "
wT 2 # w
T T
Var f (t)dBt = IE f (t)dBt = |f (t)|2 dt. (4.10)
0 0 0
Proof. The extension of the stochastic integral to all functions satisfying

(4.9) is obtained by a denseness and Cauchy∗ sequence argument, based on
the isometry relation (4.10).
i) Given f a function satisfying (4.9), consider a sequence (fn )n∈N of sim-
ple functions converging to f in L2 ([0, T ]), i.e.
wT
r
lim kf − fn kL2 ([0,T ]) = lim |f (t) − fn (t)|2 dt = 0
n→∞ n→∞ 0
as in Proposition 4.8.
ii) By the isometry (4.10) and the triangle inequality† we have
∗
See MH3100 Real Analysis I.
†
The triangle inequality kfk − fn kL2 ([0,T ]) 6 kfk − f kL2 ([0,T ]) + kf − fn kL2 ([0,T ])
follows from the Minkowski inequality.
" 145
N. Privault
w wT
T
0 fk (t)dBt − 0 fn (t)dBt 2

L (Ω)
v "
u
u wT wT 2 #
= tIE fk (t)dBt − fn (t)dBt
0 0
v "
u w 2 #
u T
= IE
t (fk (t) − fn (t))dBt
0
= kfk − fn kL2 ([0,T ])

6 kfk − f kL2 ([0,T ]) + kf − fn kL2 ([0,T ]) ,
r
T
which tends to 0 as k and n tend to infinity, hence 0 fn (t)dBt
n∈N
is a Cauchy sequence
r in L (Ω) by
2
for the L (Ω)-norm.
2
T
iii) Since the sequence 0 fn (t)dBt is Cauchy and the space L2 (Ω) is
n∈N
complete, cf. e.g. Theorem 3.11 in Rudin (1974) or Chapter 4 of Dudley
wT
(2002), we conclude that fn (t)dBt converges for the L2 -norm
0 n∈N
to a limit in L2 (Ω). In this case we let
wT wT
f (t)dBt := lim fn (t)dBt ,
0 n→∞ 0
which also satisfies (4.10) from (4.7) From (4.10) we can check that the
limit is independent of the approximating sequence (fn )n∈N .
iv) Finally, from the convergence of Gaussian characteristic functions
w w
T T
IE exp iα f (t)dBt = IE lim exp iα fn (t)dBt
0 n→∞ 0
w
T
= lim IE exp iα fn (t)dBt
n→∞ 0
α2 w T

= lim exp − |fn (t)|2 dt
n→∞ 2 0
α2 w T

= exp − |f (t)|2 dt ,
2 0
wT
f ∈ L2 ([0, T ]), α ∈ R, we check that f (t)dBt has the centered Gaus-
0
sian distribution
wT w
T

f (t)dBt ' N 0, |f (t)|2 dt ,
0 0
see Theorem 23.11.

146 "

The next corollary is obtained by bilinearity from the Itô isometry (4.10).
Corollary 4.12. The stochastic integral with respect to Brownian motion
(Bt )t∈R+ satisfies the isometry
w wT w
T T
IE f (t)dBt g (t)dBt = f (t)g (t)dt,
0 0 0
for all square-integrable deterministic functions f , g ∈ L2 ([0, T ]).

Proof. Applying the Itô isometry (4.10) to the processes f + g and f − g, we
have
w wT
T
IE f (t)dBt g (t)dBt
0 0
wT wT 2 w wT
" 2 #
1 T
= IE f (t)dBt + g (t)dBt − f (t)dBt − g (t)dBt
4 0 0 0 0
w w
" 2 # " 2 #
1 T 1 T
= IE (f (t) − g (t))dBt − IE (f (t) − g (t))dBt
4 0 4 0
1wT 1 wT
= (f (t) + g (t))2 dt − (f (t) − g (t))2 dt
4 0 4 0
1 Tw
(f (t) + g (t))2 − (f (t) − g (t))2 dt

=
4 0
wT
= f (t)g (t)dt.
0

wT
For example, e −t dBt has the centered Gaussian distribution with variance
0
wT
1
t=T
1
e −2t dt = − e −2t = 1 − e −2T .
0 2 t=0 2
wT
Remark 4.13. The Wiener stochastic integral f (s)dBs is a Gaussian
0
random variable which cannot be “computed” in the way standard integral
are computed via the use of primitives. However, when f ∈ L2 ([0, T ]) is in
C 1 ([0, T ]),∗ we have the integration by parts relation
wT wT
f (t)dBt = f (T )BT − Bt f 0 (t)dt. (4.11)
0 0
When f ∈ L2 (R+ ) is in C 1 (R+ ) we also have following formula

∗
This means that f is continuously differentiable on [0, T ].
" 147
N. Privault
w∞ w∞
f (t)dBt = − Bt f 0 (t)dt, (4.12)
0 0
provided that limt→∞ t|f (t)|2 = 0 and f ∈ L2 (R+ ), cf. e.g. Remark 2.5.9 in
Privault (2009).
4.4 Itô Stochastic Integral
In this section we extend the Wiener stochastic integral from deterministic

functions in L2 ([0, T ]) to random square-integrable (random) adapted pro-
cesses. For this, we will need the notion of measurability.
The extension of the stochastic integral to adapted random processes is

actually necessary in order to compute a portfolio value when the portfolio
process is no longer deterministic. This happens in particular when one needs
to update the portfolio allocation based on random events occurring on the
market.
A random variable F is said to be Ft -measurable if the knowledge of F

depends only on the information known up to time t. As an example, if
t =today,
• the date of the past course exam is Ft -measurable, because it belongs to
the past.
• the date of the next Chinese new year, although it refers to a future event,
is also Ft -measurable because it is known at time t.
• the date of the next typhoon is not Ft -measurable since it is not known
at time t.
• the maturity date T of the European option is Ft -measurable for all

t ∈ [0, T ], because it has been determined at time 0.
• the exercise date τ of an American option after time t (see Section 15.1)
is not Ft -measurable because it refers to a future random event.
In the next definition, (Ft )t∈[0,T ] denotes the information flow defined in
(4.2), i.e.
F t : = σ ( Bs : 0 6 s 6 t ) , t > 0.
Definition 4.14. A stochastic process (Xt )t∈[0,T ] is said to be (Ft )t∈[0,T ] -
adapted if Xt is Ft -measurable for all t ∈ [0, T ].
For example,
148 "
- (Bt )t∈R+ is an (Ft )t∈R+ -adapted process,
- (Bt+1 )t∈R+ is not an (Ft )t∈R+ -adapted process,
- (Bt/2 )t∈R+ is an (Ft )t∈R+ -adapted process,
- B √t is not an (Ft )t∈R+ -adapted process,

t∈R+
- Maxs∈[0,t] Bs )t∈R+ is an (Ft )t∈R+ -adapted process,

w
t
- Bs ds is an (Ft )t∈R+ -adapted process,
0 t∈R
w +
t
- f (s)dBs is an (Ft )t∈[0,T ] -adapted process when f ∈ L2 ([0, T ]).
0 t∈[0,T ]
In other words, a stochastic process (Xt )t∈R+ is (Ft )t∈[0,T ] -adapted if the
value of Xt at time t depends only on information known up to time t. Note
that the value of Xt may still depend on “known” future data, for example
a fixed future date in the calendar, such as a maturity time T > t, as long as
its value is known at time t.
The next Figure 4.15 shows an adapted portfolio strategy on two assets,
constructed from a sign-switching signal based on spread data, see § 1.5 in
Privault (2020) and this R code.
0.2
Pair trading
200
0.1
Performance
0.0
Spread
150
−0.1
−0.2 100
Benchmark (0.5,0.5)
−0.3
2017 2018 2019 2020 2017 2018 2019 2020
Fig. 4.15: Adapted pair trading portfolio strategy.
The stochastic integral of adapted processes is first constructed as integrals

of simple predictable processes.
Definition 4.15. A simple predictable processes is a stochastic process
(ut )t∈R+ of the form
n
Fi 1(ti−1 ,ti ] (t),
X
ut := t ∈ R+ , (4.13)
i=1
" 149
N. Privault
where Fi is an Fti−1 -measurable random variable for i = 1, 2, . . . , n, and

0 = t0 < t1 < · · · < tn−1 < tn = T .
For example, a natural approximation of (Bt )t∈R+ by a simple predictable
process can be constructed as
n n
Fi 1(ti−1 ,ti ] (t) := Bti−1 1(ti−1 ,ti ] (t),
X X
ut = t ∈ R+ , (4.14)
i=1 i=1
since Fi := Bti−1 is Fti−1 -measurable for i = 1, 2, . . . , n. The notion of simple

predictable process makes full sense in the context of portfolio investment,
in which Fi will represent an investment allocation decided at time ti−1 and
to remain unchanged over the time interval (ti−1 , ti ].
By convention, u : Ω × R+ −→ R is denoted in the sequel by ut (ω ), t ∈ R+ ,

ω ∈ Ω, and the random outcome ω is often dropped for convenience of
notation.
Definition 4.16. The stochastic integral with respect to Brownian motion
(Bt )t∈R+ of any simple predictable process (ut )t∈R+ of the form (4.13) is
defined by
wT Xn
ut dBt := Fi (Bti − Bti−1 ), (4.15)
0
i=1
with 0 = t0 < t1 < · · · < tn−1 < tn = T .
The use of predictability in the definition (4.15) is essential from a financial
point of view, as Fi will represent a portfolio allocation made at time ti−1
and kept constant over the trading interval [ti−1 , ti ], while Bti − Bti−1 repre-
sents a change in the underlying asset price over [ti−1 , ti ]. See also the related
discussion on self-financing portfolios in Section 5.3 and Lemma 5.15 on the
use of stochastic integrals to represent the value of a portfolio.
Definition 4.17. Let L2 (Ω × [0, T ]) denote the space of stochastic processes
u : Ω × [0, T ] −→ R
(ω, t) 7−→ ut (ω )
such that
s w
T
kukL2 (Ω×[0,T ]) := IE |ut |2 dt < ∞, u ∈ L2 (Ω × [0, T ]).
0
The norm k · kL2 (Ω×[0,T ]) on L2 (Ω × [0, T ]) induces a distance between two

stochastic processes u and v in L2 (Ω × [0, T ]), defined as
150 "
s w
T
ku − vkL2 (Ω×[0,T ]) = IE |ut − vt |2 dt .
0
Definition 4.18. Convergence in L2 (Ω × [0, T ]). We say that a sequence

u(n) n∈N of processes in L2 (Ω × [0, T ]) converges in L2 (Ω × [0, T ]) to an-

other process u ∈ L2 (Ω × [0, T ]) if

s
wT
( ) (n)
lim u − u L2 (Ω×[0,T ]) = lim
n
IE |ut − ut |2 dt = 0.

n→∞ n→∞ 0
By Lemma 1.1 of Ikeda and Watanabe (1989), pages 22 and 46, or Propo-
sition 2.5.3 in Privault (2009), the set of simple predictable processes forms
a linear space which is dense in the subspace L2ad (Ω × R+ ) made of square-
integrable adapted processes in L2 (Ω × R+ ), as stated in the next proposi-
tion.
Proposition 4.19. Given u a square-integrable adapted process there exists a
sequence (u(n) )n∈N of simple predictable processes converging to u in L2 (Ω ×
R+ ), i.e.
s
wT
lim u − u (n)
= lim IE ut − u(n) 2 dt = 0.

L2 (Ω×[0,T ])

n→∞ n→∞ t 0
The next Proposition 4.20 extends the construction of the stochastic integral
from simple predictable processes to square-integrable (Ft )t∈[0,T ] -adapted
processes (ut )t∈R+ for which the value of ut at time t can only depend on
information contained in the Brownian path up to time t.
This restriction means that the Itô integrand ut cannot depend on future
information, for example a portfolio strategy that would allow the trader
to “buy at the lowest” and “sell at the highest” is excluded as it would
require knowledge of future market data. Note that the difference between
Relation (4.16) below and Relation (4.10) is the presence of an expectation
on the right hand side.
Proposition 4.20. The stochastic integral with respect to Brownian motion
(Bt )t∈R+ extends to all adapted processes (ut )t∈R+ such that
w
T
kuk2L2 (Ω×[0,T ]) := IE |ut |2 dt < ∞,
0
with the Itô isometry

w 2 "
wT 2 # w
T T
:= IE = IE |ut |2 dt . (4.16)

0 ut dBt 2 ut dBt

L (Ω) 0 0
" 151
N. Privault
In addition, the Itô integral of an adapted process (ut )t∈R+ is always a cen-
tered random variable: w
T
IE ut dBt = 0. (4.17)
0
Proof. We start by showing that the Itô isometry (4.16) holds for the simple
predictable process u of the form (4.13). We have
 !2 
wT n
" 2 # X
IE ut dBt = IE  Fi (Bti − Bti−1 ) 
0
i=1
 ! n 
n
X X
= IE  Fi (Bti − Bti−1 )  Fj (Btj − Btj−1 )
i=1 j =1
 
n
X
= IE  Fi Fj (Bti − Bti−1 )(Btj − Btj−1 )
i,j =1
n
" #
X
= IE |Fi |2 (Bti − Bti−1 )2
i=1
 
X
+2 IE  Fi Fj (Bti − Bti−1 )(Btj − Btj−1 )
16i<j 6n
n
" #
X
= IE 2
|Fi | (Bti − Bti−1 ) 2
i=1
X
+2 IE Fi Fj (Bti − Bti−1 )(Btj − Btj−1 )

16i<j 6n
n
X
= IE[IE[|Fi |2 (Bti − Bti−1 )2 |Fti−1 ]]
i=1
X
+2 IE[IE[Fi Fj (Bti − Bti−1 )(Btj − Btj−1 )|Ftj−1 ]]
16i<j 6n
n
X
= IE[|Fi |2 IE[(Bti − Bti−1 )2 |Fti−1 ]]
i=1
X
+2 IE Fi Fj (Bti − Bti−1 ) IE[(Btj − Btj−1 )|Ftj−1 ]

16i<j 6n
| {z }
=0
n
X
IE |Fi |2 IE (Bti − Bti−1 )2

=
i=1
X
+2 IE[Fi Fj (Bti − Bti−1 ) IE[Btj − Btj−1 ]]
16i<j 6n
| {z }
=0
152 "
n
X
= IE[|Fi |2 (ti − ti−1 )]
i=1
n
" #
X
= IE 2
|Fi | (ti − ti−1 )
i=1
hw T i
= IE |ut |2 dt ,
0
where we applied the “tower property” (23.38) of conditional expectations

and the facts that Bti − Bti−1 is independent of Fti−1 with
IE[Bti − Bti−1 ] = 0, IE (Bti − Bti−1 )2 = ti − ti−1 , i = 1, 2, . . . , n.

u2
6
F22 b r
b r b r
F12
F42 b r
-
t0 t1 t2 t3 t4 t
Fig. 4.16: Squared simple predictable process t 7→ u2 (t).
The extension of the stochastic integral to square-integrable adapted pro-

cesses (ut )t∈R+ is obtained by a denseness and Cauchy sequence argument
using the isometry (4.16), in the same way as in the proof of Proposition 4.11.
i) By Proposition 4.19 given u ∈ L2 (Ω × [0, T ]) a square-integrable adapted
process there exists a sequence (u(n) )n∈N of simple predictable processes
such that
r hw
T i
lim ku − u(n) k 2 = lim IE ut − u(n) 2 dt = 0.
n→∞ L (Ω×[0,T ]) n→∞ 0 t
ii) Since the sequence (u(n) )n∈N converges it is a Cauchy sequence in
r T (n)
L2 (Ω × R+ ), hence by the Itô isometry (4.16), the sequence 0 ut dBt
n∈N
is a Cauchy sequence in L2 (Ω), therefore it admits a limit in the com-
plete space L2 (Ω). In this case we let
wT wT
(n)
ut dBt := lim ut dBt
0 n→∞ 0
and the limit is unique from (4.16) and satisfies (4.16).

wT
iii) The fact that the random variable ut dBt is centered can be proved
0
first on simple predictable process u of the form (4.13) as
" 153
N. Privault
w " n #
T X
IE ut dBt = IE Fi (Bti − Bti−1 )
0
i=1
n
X
= IE[IE[Fi (Bti − Bti−1 ) | Fti−1 ]]
i=1
Xn
= IE[Fi IE[Bti − Bti−1 | Fti−1 ]]
i=1
Xn
= IE[Fi IE[Bti − Bti−1 ]]
i=1
= 0,
and this identity extends as above from simple predictable processes to

adapted processes (ut )t∈R+ in L2 (Ω × R+ ).

As an application of the Itô isometry (4.16), we note in particular the identity
wT w w wT
" 2 #
T T T2
IE = IE |Bt |2 dt = IE |Bt |2 dt = ,

Bt dBt tdt =
0 0 0 0 2
with
wT L2 (Ω)
n
X
lim

Bt dBt = Bti−1 Bti − Bti−1
0 n→∞
i=1
from (4.14).
The next corollary is obtained by bilinearity from the Itô isometry (4.16) by
the same argument as in Corollary 4.12.
Corollary 4.21. The stochastic integral with respect to Brownian motion
(Bt )t∈R+ satisfies the isometry
w wT w
T T
IE ut dBt vt dBt = IE ut vt dt ,
0 0 0
for all square-integrable adapted processes (ut )t∈R+ , (vt )t∈R+ .

Proof. Applying the Itô isometry (4.16) to the processes u + v and u − v we
have
w wT
T
IE ut dBt vt dBt
0 0
wT wT 2 w wT
" 2 #!
1 T
= IE ut dBt + vt dBt − ut dBt − vt dBt
4 0 0 0 0
154 "
w wT
" 2 # " 2 #!
1 T
= IE (ut − vt )dBt − IE (ut − vt )dBt
4 0 0
w w
1

T T
= IE (ut + vt )2 dt − IE (ut − vt )2 dt
4 0 0
w
1

T
= IE (ut + vt )2 − (ut − vt )2 dt

4 0
w
T
= IE ut vt dt .
0

In addition, when the integrand (ut )t∈R+ is not a deterministic function of
wT
time, the random variable ut dBt no longer has a Gaussian distribution,
0
except in some exceptional cases.
Definite stochastic integral
The definite stochastic integral of an adapted process u ∈ L2ad (Ω × R+ ) over

an interval [a, b] ⊂ [0, T ] is defined as
wb wT
ut dBt := 1[a,b] (t)ut dBt ,
a 0
with in particular
wb wT
dBt = 1[a,b] (t)dBt = Bb − Ba , 0 6 a 6 b,
a 0
We also have the Chasles relation

wc wb wc
ut dBt = ut dBt + ut dBt , 0 6 a 6 b 6 c,
a a b
and the stochastic integral has the following linearity property:

wT wT wT
(ut + vt )dBt = ut dBt + vt dBt , u, v ∈ L2 (R+ ).
0 0 0
" 155
N. Privault
4.5 Stochastic Calculus
Fig. 4.17: NGram Viewer output for the term "stochastic calculus".
Stochastic modeling of asset returns
In the sequel we will define the return at time t ∈ R+ of the risky asset
(St )t∈R+ as
dSt
dSt = µSt dt + σSt dBt , or = µdt + σdBt . (4.18)
St
with µ ∈ R and σ > 0. Using the relation
wT
XT = X0 + dXt , T > 0,
0
which holds for any process (Xt )t∈R+ , Equation (4.18) can be rewritten in
integral form as
wT wT wT
ST = S0 + dSt = S0 + µ St dt + σ St dBt , (4.19)
0 0 0
hence the need to define an integral with respect to dBt , in addition to the
usual integral with respect to dt. Note that in view of the definition (4.15),
this is a continuous-time extension of the notion portfolio value based on a
predictable portfolio strategy.
In Proposition 4.20 we have defined the stochastic integral of square-

integrable processes with respect to Brownian motion, thus we have made
sense of the equation (4.19) where (St )t∈R+ is an (Ft )t∈[0,T ] -adapted pro-
cess, which can be rewritten in differential notation as in (4.18).
This model will be used to represent the random price St of a risky asset
at time t. Here the return dSt /St of the asset is made of two components: a
156 "
constant return µdt and a random return σdBt parametrized by the coeffi-
cient σ, called the volatility.
Our goal is now to solve Equation (4.18) and for this we will need to introduce
Itô’s calculus in Section 4.5 after a review of classical deterministic calculus.
Deterministic calculus
The fundamental theorem of calculus states that for any continuously differ-
entiable (deterministic) function f we have the integral relation
wx
f (x) = f (0) + f 0 (y )dy.
0
In differential notation this relation is written as the first order expansion
df (x) = f 0 (x)dx, (4.20)
where dx is “infinitesimally small”. Higher-order expansions can be obtained

from Taylor’s formula, which, letting
∆f (x) := f (x + ∆x) − f (x),
states that
1 1 1
∆f (x) = f 0 (x)∆x + f 00 (x)(∆x)2 + f 000 (x)(∆x)3 + f (4) (x)(∆x)4 + · · · .
2 3! 4!
Note that Relation (4.20), i.e. df (x) = f 0 (x)dx, can be obtained by neglecting
all terms of order higher than one in Taylor’s formula, since (∆x)n << ∆x,
n > 2, as ∆x becomes “infinitesimally small”.
Stochastic calculus
Let us now apply Taylor’s formula to Brownian motion, taking

√
∆Bt = Bt+∆t − Bt ' ± ∆t,
and letting
∆f (Bt ) := f (Bt+∆t ) − f (Bt ),
we have
∆f (Bt )
1 1 1
= f 0 (Bt )∆Bt + f 00 (Bt )(∆Bt )2 + f 000 (Bt )(∆Bt )3 + f (4) (Bt )(∆Bt )4 + · · · .
2 3! 4!
" 157
N. Privault
From
√ the construction of Brownian motion by its small increments ∆Bt =
± ∆t, it turns out that the terms in (∆t)2 and ∆t∆Bt ' ±(∆t)3/2 can
be neglected in Taylor’s formula at the first order of approximation in ∆t.
However, the term of order two
√
(∆Bt )2 = (± ∆t)2 = ∆t
can no longer be neglected in front of ∆t itself.
Basic Itô formula
For f ∈ C 2 (R), Taylor’s formula written at the second order for Brownian
motion reads
1
df (Bt ) = f 0 (Bt )dBt + f 00 (Bt )dt, (4.21)
2
for “infinitesimally small” dt. Note that writing this formula as
df (Bt ) dBt 1
= f 0 ( Bt ) + f 00 (Bt )
dt dt 2
does not make sense because the pathwise derivative
√
dBt dt 1
'± ' ± √ ' ±∞
dt dt dt
of Bt with respect to t does not exist. Integrating (4.21) on both sides and
using the relation wt
f (Bt ) − f (B0 ) = df (Bs )
0
we get the integral form of Itô’s formula for Brownian motion, i.e.
wt 1 w t 00
f (Bt ) = f (B0 ) + f 0 (Bs )dBs + f (Bs )ds.
0 2 0
Itô processes
We now turn to the general expression of Itô’s formula, which is stated for
Itô processes.
Definition 4.22. An Itô process is a stochastic process (Xt )t∈R+ that can
be written as
wt wt
Xt = X0 + vs ds + us dBs , t ∈ R+ , (4.22)
0 0
158 "
or in differential notation
dXt = vt dt + ut dBt ,
where (ut )t∈R+ and (vt )t∈R+ are square-integrable adapted processes.
Given (t, x) 7→ f (t, x) a smooth function of two variables on R+ × R, from
∂f
now on we let denote partial differentiation with respect to the first (time)
∂t
∂f
variable in f (t, x), while denotes partial differentiation with respect to
∂x
the second (price) variable in f (t, x).
Theorem 4.23. (Itô formula for Itô processes). For any Itô process (Xt )t∈R+
of the form (4.22) and any f ∈ C 1,2 (R+ × R) we have
f (t, Xt )
w t ∂f w t ∂f w t ∂f
= f (0, X0 ) + (s, Xs )ds + vs (s, Xs )ds + us (s, Xs )dBs
0 ∂s 0 ∂x 0 ∂x
1wt 2
2∂ f
+ |us | (s, Xs )ds. (4.23)
2 0 ∂x2
Proof. The proof of the Itô formula can be outlined as follows in the case
where (Xt )t∈R+ = (Bt )t∈R+ is a standard Brownian motion. We refer to
Theorem II-32, page 79 of Protter (2004) for the general case.
Let {0 = tn0 6 tn1 6 · · · 6 tnn = t}, n > 1, be a refining sequence of

partitions of [0, t] tending to the identity. We have the telescoping identity
n
X
f (Btni ) − f (Btni−1 ) ,

f ( Bt ) − f ( B0 ) =
k =1
and from Taylor’s formula
∂f 1 ∂2f
f (y ) − f (x) = (y − x) (x) + (y − x)2 2 (x) + R(x, y ),
∂x 2 ∂x
where the remainder R(x, y ) satisfies R(x, y ) 6 o(|y − x|2 ), we get
n
X ∂f 1 ∂2f
f ( Bt ) − f ( B0 ) = (Btni − Btni−1 ) (Btni−1 ) + |Btni − Btni−1 |2 2 (Btni−1 )
∂x 2 ∂x
k =1
n
X
+ R(Btni , Btni−1 ).
k =1
" 159
N. Privault
It remains to show that as n tends to infinity the above converges to

w t ∂f 1 w t ∂2f
f ( Bt ) − f ( B0 ) = (Bs )dBs + (Bs )ds.
0 ∂x 2 0 ∂x2

From the relation
wt
df (s, Xs ) = f (t, Xt ) − f (0, X0 ),
0
we can rewrite (4.23) as

wt wt ∂f w t ∂f
df (s, Xs ) = vs (s, Xs )ds + us (s, Xs )dBs
0 0 ∂x 0 ∂x
w t ∂f 1wt ∂2f
+ (s, Xs )ds + |us |2 2 (s, Xs )ds,
0 ∂s 2 0 ∂x
which allows us to rewrite (4.23) in differential notation, as
df (t, Xt ) (4.24)
∂f ∂f ∂f 1 ∂2f
= (t, Xt )dt + vt (t, Xt )dt + ut (t, Xt )dBt + |ut |2 2 (t, Xt )dt.
∂t ∂x ∂x 2 ∂x
In case the function x 7→ f (x) does not depend on the time variable t we get
∂f ∂f 1 ∂2f
df (Xt ) = ut (Xt )dBt + vt (Xt )dt + |ut |2 2 (Xt )dt.
∂x ∂x 2 ∂x
Taking ut = 1, vt = 0 and X0 = 0 in (4.22) yields Xt = Bt , in which case

the Itô formula (4.23) reads
w t ∂f w t ∂f 1 w t ∂2f
f (t, Bt ) = f (0, B0 ) + (s, Bs )ds + (s, Bs )dBs + (s, Bs )ds,
0 ∂s 0 ∂x 2 0 ∂x2
i.e. in differential notation:
∂f ∂f 1 ∂2f
df (t, Bt ) = (t, Bt )dt + (t, Bt )dBt + (t, Bt )dt. (4.25)
∂t ∂x 2 ∂x2
Itô multiplication table
Next, consider two Itô processes (Xt )t∈R+ and (Yt )t∈R+ written in integral
form as wt wt
Xt = X0 + vs ds + us dBs , t ∈ R+ ,
0 0
160 "
and wt wt
Yt = Y0 + bs ds + as dBs , t ∈ R+ ,
0 0
or in differential notation as
dXt = vt dt + ut dBt , and dYt = bt dt + at dBt , t ∈ R+ .
Bivariate Itô formula
The Itô formula can also be written for functions f ∈ C 1,2,2 (R+ × R2 ) we
have of two state variables as
∂f ∂f 1 ∂2f
df (t, Xt , Yt ) = (t, Xt , Yt )dt + (t, Xt , Yt )dXt + |ut |2 2 (t, Xt , Yt )dt
∂t ∂x 2 ∂x
∂f 1 ∂2f ∂2f
+ (t, Xt , Yt )dYt + |at |2 2 (t, Xt , Yt )dt + ut at (t, Xt , Yt )dt,
∂y 2 ∂y ∂x∂y
(4.26)
which can be used to show that
d(Xt Yt ) = Xt dYt + Yt dXt + dXt • dYt
where the product dXt • dYt is computed according to the Itô rule
dt • dt = 0, dt • dBt = 0, dBt • dBt = dt, (4.27)
which can be encoded in the following Itô multiplication table:
• dt dBt
dt 0 0
dBt 0 dt
Table 4.1: Itô multiplication table.
It follows from the Itô Table 4.1 that
dXt • dYt = (vt dt + ut dBt ) • (bt dt + at dBt )

= bt vt (dt)2 + bt ut dtdBt + at vt dtdBt + at ut (dBt )2
= at ut dt.
Hence we also have
(dXt )2 = (vt dt + ut dBt )2

" 161
N. Privault
= (vt )2 (dt)2 + (ut )2 (dBt )2 + 2ut vt (dt • dBt )

= (ut )2 dt,
according to the Itô Table 4.1. Consequently, (4.24) can be rewritten as
∂f ∂f 1 ∂2f
df (t, Xt ) = (t, Xt )dt + (t, Xt )dXt + (t, Xt )(dXt )2 ,
∂t ∂x 2 ∂x2
and the Itô formula for functions f ∈ C 1,2,2 (R+ × R2 ) of two state variables
can be rewritten as
∂f ∂f 1 ∂2f
df (t, Xt , Yt ) = (t, Xt , Yt )dt + (t, Xt , Yt )dXt + (t, Xt , Yt )(dXt )2
∂t ∂x 2 ∂x2
∂f 1 ∂2f ∂2f
+ (t, Xt , Yt )dYt + (t, Xt , Yt )(dYt )2 + (t, Xt , Yt )(dXt • dYt ).
∂y 2 ∂y 2 ∂x∂y
Examples
Applying Itô’s formula (4.25) to Bt2 with
Bt2 = f (t, Bt ) and f (t, x) = x2 ,
and
∂f ∂f 1 ∂2f
(t, x) = 0, (t, x) = 2x, (t, x) = 1,
∂t ∂x 2 ∂x2
we find
d(Bt2 ) = df (Bt )
∂f ∂f 1 ∂2f
= (t, Bt )dt + (t, Bt )dBt + (t, Bt )dt
∂t ∂x 2 ∂x2
= 2Bt dBt + dt.
Note that from the Itô Table 4.1 we could also write directly
d(Bt2 ) = Bt dBt + Bt dBt + (dBt )2 = 2Bt dBt + dt.
Next, by integration in t ∈ [0, T ] we find

wT wT wT
BT2 = B0 + 2 Bs dBs + dt = 2 Bs dBs + T , (4.28)
0 0 0
hence the relation
162 "
wT 1
BT2 − T .

Bs dBs =
0 2
Similarly, we have
d(Bt3 ) = 3Bt2 dBt + 3Bt dt,






1



d sin Bt = cos(Bt )dBt − sin(Bt )dt,


2






1


d e t = e Bt dBt + e Bt dt,
 B

2


1 1


d log Bt =

 dBt − 2 dt,
2B

B

t



 t


t2


de = Bt e tBt dt + e tBt dt + t e tBt dBt ,

 tBt

2
etc.
Notation
We close this section with some comments on the practice of Itô’s calculus. In
certain finance textbooks, Itô’s formula for e.g. geometric Brownian motion
(St )t∈R+ given by
dSt = µSt dt + σSt dBt
can be found written in the notation
wT ∂f wT ∂f
f (T , ST ) = f (0, X0 ) + σ St (t, St )dBt + µ St (t, St )dt
0 ∂St 0 ∂St
w T ∂f 1 w T 2
∂ f
+ (t, St )dt + σ 2 St2 2 (t, St )dt,
0 ∂t 2 0 ∂St
or
∂f ∂f 1 ∂2f
df (St ) = σSt (St )dBt + µSt (St )dt + σ 2 St2 2 (St )dt.
∂St ∂St 2 ∂St
∂f
The notation (St ) can in fact be easily misused in combination with the
∂St
fundamental theorem of classical calculus, and potentially leads to the wrong
identity
∂f

df (S = (St )dSt .
t)
∂St
Similarly, writing
" 163
N. Privault
∂f 1 ∂2f
df (Bt ) = (Bt )dBt + (Bt )dt
∂x 2 ∂x2
is consistent, while writing
∂f (Bt ) 1 ∂ 2 f ( Bt )
df (Bt ) = dBt + dt
∂Bt 2 ∂Bt2
is potentially a source of confusion. Note also that the right hand side of the
Itô formula uses partial derivatives while its left hand side is a total derivative.
Stochastic differential equations
In addition to geometric Brownian motion there exists a large family of

stochastic differential equations that can be studied, although most of the
time they cannot be explicitly solved. Let now
σ : R+ × Rn −→ Rd ⊗ Rn
where Rd ⊗ Rn denotes the space of d × n matrices, and
b : R+ × Rn −→ R
satisfy the global Lipschitz condition
kσ (t, x) − σ (t, y )k2 + kb(t, x) − b(t, y )k2 6 K 2 kx − yk2 ,
t ∈ R+ , x, y ∈ Rn . Then there exists a unique strong solution to the stochastic

differential equation
wt wt
Xt = X0 + b(s, Xs )ds + σ (s, Xs )dBs , t ∈ R+ , (4.29)
0 0
i.e., in differential notation
dXt = b(t, Xt )dt + σ (t, Xt )dBt , t ∈ R+ ,
where (Bt )t∈R+ is a d-dimensional Brownian motion, see e.g. Protter (2004),
Theorem V-7. In addition, the solution process (Xt )t∈R+ of (4.29) has the
Markov property, see § V-6 of Protter (2004).
The term σ (s, Xs ) in (4.29) will be interpreted later on in Chapters 8-9 as a

local volatility component.
Stochastic differential equations can be used to model the behaviour of a

variety of quantities, such as
• stock prices,
164 "
• interest rates,
• exchange rates,
• weather factors,
• electricity/energy demand,
• commodity (e.g. oil) prices, etc.
Next, we consider several examples of stochastic differential equations that
can be solved explicitly using Itô’s calculus, in addition to geometric Brown-
ian motion. See e.g. § II-4.4 of Kloeden and Platen (1999) for more examples
of explicitly solvable stochastic differential equations.
Examples of stochastic differential equations
1. Consider the stochastic differential equation
dXt = −αXt dt + σdBt , X0 = x0 , (4.30)
with α > 0 and σ > 0.

N=10000; t <- 0:(N-1); dt <- 1.0/N;alpha=5; sigma=0.4;
Z <- rnorm(N,mean=0,sd=sqrt(dt));X <- rep(0,N);X[1]=0.5
for (j in 2:N){X[j]=X[j-1]-alpha*X[j-1]*dt+sigma*Z[j]}
plot(t, X, xlab = "t", ylab = "", type = "l", ylim = c(-0.5,1), col = "blue")
abline(h=0)
We look for a solution of the form

wt
Xt = a(t)Yt = a(t) x0 + b(s)dBs
0
where a(·) and b(·) are deterministic functions of time. After applying
rt
Theorem 4.23 to the Itô process x0 + 0 b(s)dBs of the form (4.22) with
ut = b(t) and v (t) = 0, and to the function f (t, x) = a(t)x, we find
dXt = d(a(t)Yt )
= Yt a0 (t)dt + a(t)dYt
= Yt a0 (t)dt + a(t)b(t)dBt . (4.31)
By identification of (4.30) with (4.31), we get

 0
 a (t) = −αa(t)
a(t)b(t) = σ,

hence a(t) = a(0) e −αt

= e −αt
and b(t) = σ/a(t) = σ e αt , which shows
that wt
Xt = x0 e −αt + σ e −(t−s)α dBs , t ∈ R+ , (4.32)
0
Using integration by parts, we can also write
" 165
N. Privault
wt
Xt = x0 e −αt + σBt − σα e −(t−s)α Bs ds, t ∈ R+ , (4.33)
0
Remark: the solution of the equation (4.30) cannot be written as a func-

tion f (t, Bt ) of t and Bt as in the proof of Proposition 5.16.
0.6
0.5
0.4
0.3
Xt
0.2
0.1
-0.1
-0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t
Fig. 4.18: Simulated path of (4.30) with α = 10 and σ = 0.2.

2 /2
dXt = tXt dt + e t dBt , X0 = x0 . (4.34)
N=10000; T<-2.0; t <- 0:(N-1); dt <- T/N;

Z <- rnorm(N,mean=0,sd= sqrt(dt));X <- rep(0,N);X[1]=0.5
for (j in 2:N){X[j]=X[j-1]+j*X[j-1]*dt*dt+exp(j*dt*j*dt/2)*Z[j]}
plot(t, X, xlab = "t", ylab = "", type = "l", ylim = c(-0.5,10), col = "blue")
abline(h=0)
rt
Looking for a solution of the form Xt = a(t) X0 + 0 b(s)dBs , where
a(·) and b(·) are deterministic functions of time, we get a0 (t)/a(t) = t
2 2
and a(t)b(t) = e t /2 , hence a(t) = e t /2 and b(t) = 1, which yields
2 /2
Xt = e t (X0 + Bt ), t ∈ R+ .
7
4
Xt
3
-1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
t
Fig. 4.19: Simulated path of (4.34).
166 "

p
dYt = (2µYt + σ 2 )dt + 2σ Yt dBt , (4.35)
where µ ∈ R and σ > 0.
N=10000; t <- 0:(N-1); dt <- 1.0/N;mu=-5;sigma=1;

Z <- rnorm(N,mean=0,sd=sqrt(dt));Y <- rep(0,N);Y[1]=0.5
for (j in 2:N){
Y[j]=max(0,Y[j-1]+(2*mu*Y[j-1]+sigma*sigma)*dt+2*sigma*sqrt(Y[j-1])*Z[j])}
plot(t, Y, xlab = "t", ylab = "", type = "l", ylim = c(-0.1,1), col = "blue")
abline(h=0)
√
Letting Xt = Yt , we find that dXt = µXt dt + σdBt , hence
wt 2
e µ(t−s) dBs
p
Yt = (Xt )2 = e µt Y0 + σ .
0
0.7
0.6
0.5
0.4
Yt
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t
Fig. 4.20: Simulated path of (4.35) with µ = 5 and σ = 1.
Exercises
Exercise 4.1 Compute IE[Bt Bs ] in terms of s, t ∈ R+ .
Exercise 4.2 Let (Bt )t∈R+ denote a standard Brownian motion. Let c > 0.
Among the following processes, tell which is a standard Brownian motion
and which is not. Justify your answer.
a) (Xt )t∈R+ := Bc+t − Bc t∈R ,

+
b) (Xt )t∈R+ := cBt/c2 t∈R ,

+
c) (Xt )t∈R+ := Bct2 t∈R ,

+
d) (Xt )t∈R+ := Bt + Bt/2 t∈R .
+
" 167
N. Privault
Exercise 4.3 Let (Bt )t∈R+ denote a standard Brownian motion. Compute
the stochastic integrals
wT wT
2dBt and 2 × 1[0,T /2] (t) + 1(T /2,T ] (t) dBt

0 0
and determine their probability distributions (including mean and variance).
Exercise 4.4 Determine the probability distribution (including mean and

w 2π
variance) of the stochastic integral sin(t) dBt .
0
Exercise 4.5 Let T > 0. Show that for f : [0, T ] 7→ R a differentiable function
such that f (T ) = 0, we have
wT wT
f (t)dBt = − f 0 (t)Bt dt.
0 0
Hint: Apply Itô’s calculus to t 7→ f (t)Bt .
Exercise 4.6
r1
a) Find the probability distribution of the stochastic integral 0 t2 dBt with
respect to Brownian motion (Bt )t∈R+ .
r1
b) Find the probability distribution of the stochastic integral 0 t−1/2 dBt
with respect to Brownian motion (Bt )t∈R+ .
Exercise 4.7 Given (Bt )t∈R+ a standard Brownian motion and n > 1, let
the random variable Xn be defined as
w 2π
Xn := sin(nt)dBt , n > 1.
0
a) Give the probability distribution of Xn for all n > 1.

b) Show that (Xn )n>1 is a sequence of pairwise independent and identically
distributed random variables.
1
Hint: We have sin a sin b = cos(a − b) − cos(a + b) , a, b ∈ R.

2
Exercise 4.8 Apply the Itô formula to the process Xt := sin2 (Bt ), t ∈ R+ .
Exercise 4.9 Let (Bt )t∈R+ denote a standard Brownian motion.

a) Using the Itô isometry and the known relations
wT wT
BT = dBt and BT2 = T + 2 Bt dBt ,
0 0
compute the third and fourth moments IE[BT3 ] and IE[BT4 ].
168 "
b) Give the third and fourth moments of the centered normal distribution
with variance σ 2 .
Exercise 4.10 Consider an asset price (St )t∈R+ given by the stochastic
differential equation
dSt = rSt dt + σSt dBt , (4.36)
where (Bt )t∈R+ is a standard Brownian motion, with r ∈ R and σ > 0.
a) Find the stochastic differential equation satisfied by the power (Stp )t∈R+
of order p ∈ R of (St )t∈R+ .
b) Construct a probability measure under which the discounted process
e −rt Stp t∈R is a martingale.
+
Exercise 4.11 Given T > 0, find the stochastic integral decomposition of

(BT )3 as
wT
( BT ) 3 = C + ζt,T dBt (4.37)
0
where C ∈ R is a constant and (ζt,T )t∈[0,T ] is an adapted process to be de-
termined.
Exercise 4.12 Let f ∈ L2 ([0, T ]), and consider a standard Brownian motion
(Bt )t∈[0,T ] .
a) Compute the conditional expectation
h rT i
IE e 0 f (s)dBs Ft , 0 6 t 6 T,
where (Ft )t∈[0,T ] denotes the filtration generated by (Bt )t∈[0,T ] .

b) Using the result of Question (a), show that the process
w
1wt 2

t
t 7−→ exp f (s)dBs − f (s)ds , t ∈ [0, T ],
0 2 0
is an (Ft )-martingale, where (Ft )t∈[0,T ] denotes the filtration generated

by (Bt )t∈[0,T ] .
2
c) Show that the process ( e σBt −σt )t∈R+ is an (Ft )-martingale for any
σ ∈ R.
Exercise 4.13 Compute the expected value

w
T
IE exp β Bt dBt
0
for all β < 1/T . Hint: Expand (BT )2 using the Itô formula as in (4.28).
" 169
N. Privault
Exercise 4.14
a) Solve the stochastic differential equation
dXt = −bXt dt + σ e −bt dBt , t ∈ R+ ,
where (Bt )t∈R+ is a standard Brownian motion and σ, b ∈ R.

b) Solve the stochastic differential equation
dXt = −bXt dt + σ e −at dBt , t ∈ R+ ,
where (Bt )t∈R+ is a standard Brownian motion and a, b, σ ∈ R are posi-

tive constants.
Exercise 4.15 Given T > 0, let (XtT )t∈[0,T ) denote the solution of the
stochastic differential equation
XtT
dXtT = σdBt − dt, t ∈ [0, T ), (4.38)
T −t
under the initial condition X0T = 0 and σ > 0.

a) Show that
wt 1
XtT = (T − t)σ dBs , t ∈ [0, T ).
0 T −s
Hint: Start by computing d(XtT /(T − t)) using the Itô formula.
b) Show that IE[XtT ] = 0 for all t ∈ [0, T ).
c) Show that Var[XtT ] = σ 2 t(T − t)/T for all t ∈ [0, T ).
d) Show that limt→T XtT = 0 in L2 (Ω). The process (XtT )t∈[0,T ] is called a
Brownian bridge.
Exercise 4.16 Exponential Vašíček (1977) model (1). Consider a Vasicek

process (rt )t∈R+ solution of the stochastic differential equation
drt = (a − brt )dt + σdBt , t ∈ R+ ,
where (Bt )t∈R+ is a standard Brownian motion and σ, a, b > 0 are positive
constants. Show that the exponential Xt := e rt satisfies a stochastic differ-
ential equation of the form
dXt = Xt (ã − b̃f (Xt ))dt + σg (Xt )dBt ,
where the coefficients ã and b̃ and the functions f (x) and g (x) are to be
determined.
170 "
Exercise 4.17 Exponential Vasicek model (2). Consider a short-term rate

interest rate process (rt )t∈R+ in the exponential Vasicek model:
drt = rt (η − a log rt )dt + σrt dBt , (4.39)
where η, a, σ are positive parameters and (Bt )t∈R+ is a standard Brownian

motion.
a) Find the solution (Zt )t∈R+ of the stochastic differential equation
dZt = −aZt dt + σdBt
as a function of the initial condition Z0 , where a and σ are positive pa-

rameters.
b) Find the solution (Yt )t∈R+ of the stochastic differential equation
dYt = (θ − aYt )dt + σdBt (4.40)
as a function of the initial condition Y0 . Hint: Let Zt := Yt − θ/a.

c) Let Xt = e Yt , t ∈ R+ . Determine the stochastic differential equation
satisfied by (Xt )t∈R+ .
d) Find the solution (rt )t∈R+ of (4.39) in terms of the initial condition r0 .
e) Compute the conditional mean∗ IE[rt |Fu ].
f) Compute the conditional variance
Var[rt |Fu ] := IE[rt2 |Fu ] − (IE[rt |Fu ])2
of rt , 0 6 u 6 t, where (Fu )u∈R+ denotes the filtration generated by the

Brownian motion (Bt )t∈R+ .
g) Compute the asymptotic mean and variance limt→∞ IE[rt ] and limt→∞ Var[rt ].
Exercise 4.18 Cox-Ingersoll-Ross (CIR) model. Consider the equation

√
drt = (α − βrt )dt + σ rt dBt (4.41)
modeling the variations of a short-term interest rate process rt , where α, β, σ

and r0 are positive parameters and (Bt )t∈R+ is a standard Brownian motion.
a) Write down the equation (4.41) in integral form.
b) Let u(t) = IE[rt ]. Show, using the integral form of (4.41), that u(t) satis-
fies the differential equation
u0 (t) = α − βu(t),
and compute IE[rt ] for all t ∈ R+ .

2
/2
∗
One may use the Gaussian moment generating function IE[ e X ] = e α for X '
N (0, α2 ).
" 171
N. Privault
c) By an application of Itô’s formula to rt2 , show that
drt2 = rt (2α + σ 2 − 2βrt )dt + 2σrt3/2 dBt . (4.42)
d) Using the integral form of (4.42), find a differential equation satisfied by

v (t) := IE[rt2 ] and compute IE[rt2 ] for all t ∈ R+ .
e) Show that
σ 2 −βt ασ 2 2
Var[rt ] = r0 e − e −2βt + 1 − e −βt , t ∈ R+ .
β 2β 2
Problem 4.19 Itô-Tanaka formula. Let (Bt )t∈R+ be a standard Brownian

motion started at B0 ∈ R.
a) Does the Itô formula apply to the European call option payoff function
f (x) := (x − K )+ ? Why?
b) For every ε > 0, consider the approximation fε (x) of f (x) := (x − K )+
defined by
if x > K + ε,


 x−K


1


fε (x) := (x − K + ε)2 if K − ε < x < K + ε,

 4ε



0 if x < K − ε.

Plot the graph of the function x 7→ fε (x) for ε = 1 and K = 10.

c) Using the Itô formula, show that we have
wT 1
fε0 (Bt )dBt +
` t ∈ [0, T ] : K − ε < Bt < K + ε ,

fε (BT ) = fε (B0 ) +
0 4ε
(4.43)
where ` denotes the measure of time length (Lebesgue measure) in R.
d) Show that limε→0 k1[K,∞) (·) − fε0 (·)kL2 (R+ ) = 0.
e) Show, using the Itô isometry,∗ that the limit
1
[0,T ] : = lim
LK
ε→0 2ε
`({t ∈ [0, T ] : K − ε < Bt < K + ε})
exists in L2 (Ω), and prove the Itô-Tanaka formula

wT 1
(BT − K )+ = (B0 − K )+ + 1[K,∞) (Bt )dBt + LK
[0,T ] . (4.44)
0 2
hwT 2 i
∗
Hint: Show that lim IE 1[K,∞) (Bt ) − fε0 (Bt ) dt = 0.
ε→0 0
172 "
The quantity LK
[0,T ] is called the local time spent by Brownian motion at
the level K.
Problem 4.20 Lévy’s construction of Brownian motion. The goal of this

problem is to prove the existence of standard Brownian motion (Bt )t∈[0,1] as
a stochastic process satisfying the four properties of Definition 4.1, i.e.:
1. B0 = 0 almost surely,
2. The sample trajectories t 7→ Bt are continuous, with probability 1.
3. For any finite sequence of times t0 < t1 < · · · < tn , the increments
Bt1 − Bt0 , Bt2 − Bt1 , . . . , Btn − Btn−1
are independent.
4. For any given times 0 6 s < t, Bt − Bs has the Gaussian distribution

N (0, t − s) with mean zero and variance t − s.
The construction will proceed by the linear interpolation scheme illustrated
in Figure 4.9. We work on the space C0 ([0, 1]) of continuous functions on
[0, 1] started at 0, with the norm
kf k∞ := Max |f (t)|
t∈[0,1]
and the distance

kf − gk∞ := Max |f (t) − g (t)|.
t∈[0,1]
The following ten questions are interdependent.

a) Show that for any Gaussian random variable X ' N (0, σ 2 ) we have
σ 2 2
P(|X| > ε) 6 √ e −ε /(2σ ) , ε > 0.
ε π/2
Hint: Start from the inequality IE[(X − ε)+ ] > 0 and compute the left-
hand side.
b) Let X and Y be two independent centered Gaussian random variables

with variances α2 and β 2 . Show that the conditional distribution
P(X ∈ dx | X + Y = z )
of X given X + Y = z is Gaussian with mean α2 z/(α2 + β 2 ) and variance

α2 β 2 / ( α2 + β 2 ) .
" 173
N. Privault
Hint: Use the definition

P(X ∈ dx and X + Y ∈ dz )
P(X ∈ dx | X + Y = z ) :=
P(X + Y ∈ dz )
and the formulas

1 2 / (2α2 ) 1 2 2
dP(X 6 x) := √ e −x dx, dP(Y 6 x) := p e −x /(2β ) dx,
2πα2 2πβ 2
where dx (resp. dy) represents a “small” interval [x, x + dx] (resp. [y, y +
dy ]).
c) Let (Bt )t∈R+ denote a standard Brownian motion and let 0 < u < v. Give
the distribution of B(u+v )/2 given that Bu = x and Bv = y.
Hint: Note that given that Bu = x, the random variable Bv can be written
as
Bv = (Bv − B(u+v )/2 ) + (B(u+v )/2 − Bu ) + x, (4.45)
and apply the result of Question (b) after identifying X and Y in the
above decomposition (4.45).
d) Consider the random sequences

(0)
 Z (0) = 0, Z1






 Z (1) = 0, Z (1) , Z (0)




 1/2 1



 Z (2) = 0, Z (2) , Z (1) , Z (2) , Z (0)




 1/4 1/2 3/4 1



(3) (2) (3) (1) (3) (2) (3) (0)

Z (3) = 0, Z1/8 , Z1/4 , Z3/8 , Z1/2 , Z5/8 , Z3/4 , Z7/8 , Z1



.. ..


. .









(n) (n) (n) (n) (n)
Z (n) = 0, Z1/2n , Z2/2n , Z3/2n , Z4/2n , . . . , Z1









 Z (n+1)= 0, Z (n+1) , Z (n) , Z (n+1) , Z (n+1) , Z (n+1) , Z (n+1) , . . . , Z (n+1)


1/2n+1 1/2n 3/2n+1 2/2n 5/2n+1 3/2n 1
(n)
with Z0 = 0, n > 0, defined recursively as
(0)
i) Z1 ' N (0, 1),
(0) (0)
(1) Z0 + Z1
ii) Z1/2 ' + N (0, 1/4),
2
(1) (1) (1) (0)
(2) Z0 + Z1/2 (2) Z + Z1
iii) Z1/4 ' + N (0, 1/8), Z3/4 ' 1/2 + N (0, 1/8),
2 2
174 "
and more generally

(n) (n)
(n+1)
Zk/2n + Z(k+1)/2n
Z(2k+1)/2n+1 = + N (0, 1/2n+2 ), k = 0, 1, . . . , 2n − 1,
2
where N (0, 1/2n+2 ) is an independent centered Gaussian sample with
(n+1) (n)
variance 1/2n+2 , and Zk/2n := Zk/2n , k = 0, 1, . . . , 2n .
(n)
In the sequel we denote by Zt the continuous-time random path
t∈[0,1]
(n)
obtained by linear interpolation of the sequence points in Zk/2n k=0,1,...,2n .
(0) (1)
Draw a sample of the first four linear interpolations Zt t∈[0,1] , Zt t∈[0,1] ,
(2) (3) (n)
Zt t∈[0,1] , Zt t∈[0,1] , and label the values of Zk/2n on the graphs for

k = 0, 1, . . . , 2 and n = 0, 1, 2, 3.
n
e) Using an induction argument, explain why for all n > 0 the sequence
(n) (n) (n) (n) (n)
Z (n) = 0, Z1/2n , Z2/2n , Z3/2n , Z4/2n , . . . , Z1
has same distribution as the sequence
B (n) := B0 , B1/2n , B2/2n , B3/2n , B4/2n , . . . , B1 .

Hint: Compare the constructions of Questions (c) and (d) and note that
under the above linear interpolation, we have
(n) (n)
(n)
Zk/2n + Z(k+1)/2n
Z(2k+1)/2n+1 = , k = 0, 1, . . . , 2n − 1.
2
f) Show that for any εn > 0 we have
(n+1) (n)
P Z (n+1) − Z (n) ∞ > εn 6 2n P |Z1/2n+1 − Z1/2n+1 | > εn .

Hint: Use the inequality

2n −1 n −1
2X
!
[
P Ak 6 P ( Ak )
k =0 k =0
for a suitable choice of events (Ak )k=0,1,...,2n −1 .

g) Use the results of Questions (a) and (f) to show that for any εn > 0 we
have
2n/2 2 n+1
P Z (n+1) − Z (n) ∞ > εn 6 √ e −εn 2 .

εn 2π
h) Taking εn = 2−n/4 , show that
" 175
N. Privault
 
Z (n+1) − Z (n) < ∞ = 1.
X
P

∞
n>0
Hint: Show first that

X
P Z (n+1) − Z (n) ∞ > 2−n/4 < ∞,

n>0
and apply the Borel-Cantelli lemma. n o

(n)
i) Show that with probability one, the sequence Zt t∈[0,1] , n > 1 con-
verges uniformly on [0, 1] to a continuous (random) function (Zt )t∈[0,1] .
Hint: Use the fact that C0 ([0, 1]) is a complete space for the k · k∞ norm.
j) Argue that the limit (Zt )t∈[0,1] is a standard Brownian motion on [0, 1]
by checking the four relevant properties.
Problem 4.21 Consider (Bt )t∈R+ a standard Brownian motion, and for any
n > 1 and T > 0, define the discretized quadratic variation
n
(n)
X
QT := (BkT /n − B(k−1)T /n )2 , n > 1.
k =1
h i
(n)
a) Compute IE QT , n > 1.
(n)
b) Compute Var[QT ], n > 1.
c) Show that
(n)
lim QT = T,
n→∞
where the limit is taken in L2 (Ω), that is, show that

(n)
lim kQT − T kL2 (Ω) = 0,
n→∞
where q
(n) (n) 2
IE QT − T , n > 1.

Q − T 2
T L (Ω)
: =
d) By the result of Question (c), show that the limit
wT n
X
Bt dBt := lim (BkT /n − B(k−1)T /n )B(k−1)T /n
0 n→∞
k =1
exists in L2 (Ω), and compute it.
176 "
Hint: Use the identity

1 2
(x − y )y = (x − y 2 − (x − y )2 ), x, y ∈ R.
2
e) Consider the modified quadratic variation defined by
n
e (n) : =
X
Q T (B(k−1/2)T /n − B(k−1)T /n )2 , n > 1.
k =1
(n)
Compute the limit limn→∞ Q e
T in L ( Ω ) by repeating the steps of Ques-
2
tions (a)-(c).
f) By the result of Question (e), show that the limit
wT n
X
Bt ◦ dBt := lim (BkT /n − B(k−1)T /n )B(k−1/2)T /n
0 n→∞
k =1
Hint: Use the identities

1 2
(x − y )y = (x − y 2 − (x − y )2 ),
2
and
1 2
(x − y )x =(x − y 2 + (x − y )2 ), x, y ∈ R.
2
g) More generally, by repeating the steps of Questions (e) and (f), show that
for any α ∈ [0, 1] the limit
wT n
X
Bt ◦ dα Bt := lim (BkT /n − B(k−1)T /n )B(k−α)T /n
0 n→∞
k =1

h) Comparison with deterministic calculus. Compute the limit
n
X T T T
lim (k − α ) k − (k − 1)
n→∞ n n n
k =1
for all values of α in [0, 1].
Exercise 4.22 Let (Bt )t∈R+ be a standard Brownian motion generating the
information flow (Ft )t∈R+ .
a) Let 0 6 t 6 T . What is the probability distribution of BT − Bt ?
" 177
N. Privault
b) From the answer to Exercise A.4-(c), show that

r
T − t −B 2 /(2(T −t)) Bt
IE[(BT )+ | Ft ] = e t + Bt Φ √ ,
2π T −t
0 6 t 6 T . Hint: Use the time splitting decomposition BT = BT − Bt +
Bt .
c) Let σ > 0, ν ∈ R, and Xt := σBt + νt, t ∈ R+ . Compute e Xt by applying
the Itô formula
wt ∂f w t ∂f 1 w t 2 ∂2f
f (Xt ) = f (X0 ) + us(Xs )dBs + vs (Xs )ds + u (Xs )ds
0 ∂x 0 ∂x 2 0 s ∂x2
wt wt
to f (x) = e x , where Xt is written as Xt = X0 + us dBs + vs ds,
0 0
t ∈ R+ .
d) Let St = e Xt , t ∈ R+ , and r > 0. For which value of ν does (St )t∈R+
satisfy the stochastic differential equation
dSt = rSt dt + σSt dBt ?
Exercise 4.23 From the answer to Exercise A.4-(c), show that for any β ∈ R
we have
r
T − t −(β−Bt )2 /(2(T −t)) β − Bt
IE[(β − BT )+ | Ft ] = e + ( β − Bt ) Φ √ ,
2π T −t
0 6 t 6 T.
Hint: Use the time splitting decomposition BT = BT − Bt + Bt .
178 "

Brownian Motion Stochastic Calculus

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Brownian Motion Stochastic Calculus

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Brownian Motion Stochastic Calculus

Uploaded by

Copyright:

Available Formats

Chapter 4

Brownian Motion and Stochastic Calculus

Brownian motion is a continuous-time stochastic process having stationary

4.1 Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

4.1 Brownian Motion

We start by recalling the definition of Brownian motion, which is a funda-

2. The sample trajectories t 7→ Bt are continuous, with probability 1.

Bt1 − Bt0 , Bt2 − Bt1 , . . . , Btn − Btn−1

are mutually independent random variables.

4. For any given times 0 6 s < t, Bt − Bs has the Gaussian distribution

In Figure 4.1 we draw three sample paths of a standard Brownian motion

Fig. 4.1: Sample paths of a one-dimensional Brownian motion.

In particular, Property 4 in Definition 4.1 implies

IE[Bt − Bs ] = 0 and Var[Bt − Bs ] = t − s, 0 6 s 6 t,

= IE[Bs (Bt − Bs )] + IE (Bs )2

= IE[Bs ] IE[Bt − Bs ] + IE (Bs )2

Cov(Bs , Bt ) = IE[Bs Bt ] = min(s, t), s, t ∈ R+ , (4.1)

Fig. 4.3: Web traffic ranking.

Property 3 in Definition 4.1 shows that Bt − Bs is independent of all Brownian

⊥ (Bt1 − Bt0 , Bt2 − Bt1 , . . . , Btn − Btn−1 ),

0 6 t0 6 t1 6 · · · 6 tn 6 s 6 t, hence Bt − Bs is also independent of the

Proposition 4.2. Brownian motion (Bt )t∈R+ is a continuous-time martin-

because it has centered and independent increments, cf. Section 7.1.

Fig. 4.4: Two sample paths of a two-dimensional Brownian motion.

Fig. 4.5: Sample path of a three-dimensional Brownian motion.

Figure 4.6 presents an illustration of the scaling property of Brownian motion.

Fig. 4.6: Scaling property of Brownian motion.∗

4.2 Three Constructions of Brownian Motion

Brownian motion as a random walk

For convenience we will informally regard Brownian motion as a random walk

∆Bt := Bt+∆t − Bt ' N (0, ∆t)

Fig. 4.7: Construction of Brownian motion as a random walk.∗

of same length ∆t = T /N , with N “large”.

Defining the Bernoulli random variable Xk as

with equal probabilities (1/2, 1/2), we have Var(Xk ) = T and

0.0 0.2 0.4 0.6 0.8 1.0

Fig. 4.8: Statistics of one-dimensional Brownian paths vs Gaussian distribution.

converges (in distribution) to the centered Gaussian random variable N (0, σ 2 )

Lévy’s construction of Brownian motion

Figure 4.9 represents the construction of Brownian motion by successive lin-

0.0 0.2 0.4 0.6 0.8 1.0

Fig. 4.9: Lévy’s construction of Brownian motion.∗

The following R code is used to generate Figure 4.9.†

alpha=1/2;t <- 0:1;dt <- 1;z=rnorm(1,mean=0,sd=dt^alpha)

Construction by series expansions

Brownian motion on [0, T ] can also be constructed by Fourier synthesis via

where (Xn )n>1 is a sequence of independent N (0, 1) standard Gaussian ran-

Fig. 4.10: Construction of Brownian motion by series expansions.†

4.3 Wiener Stochastic Integral

In this section, we construct the Wiener stochastic integral of square-

A naive definition of the stochastic integral with respect to Brownian mo-

i.e. the function f takes the value ai on the interval (ti−1 , ti ], i = 1, 2, . . . , n,

Fig. 4.11: Step function t 7→ f (t).

Fig. 4.12: Area under the step function t 7→ f (t).

As a consequence, the stochastic integral

of the step function

has the centered Gaussian distribution with mean 0 and variance

since the simple function

takes the value a2i on the interval (ti−1 , ti ], i = 1, 2, . . . , n, as can be checked