c 2007 by Karl Sigman

Copyright

1
1.1

IEOR 4700: Introduction to stochastic integration

Riemann-Stieltjes integration

Recall from calculus how the Riemann integral ab h(t)dt is defined for a continuous function
h over the bounded interval [a, b]. We partition the interval [a, b] into n small subintervals
a = t0 < t1 < < tn = b, and sum up the area of the corresponding rectangles to obtain a
Riemann sum;
R

h(t)dt

n
X

h(tj1 )(tj tj1 ).

j=1

(h(tj1 ) can be replaced by h(sj ), for any sj [tj1 , tj ].) As n gets larger and larger while
the partition gets finer and finer, the approximating sum to the true area under the function
becomes exact. This dt integration can be generalized to increments dG(t) of (say) any
monotone increasing function G(t) by using G(tj ) G(tj1 ) in place of tj tj1 yielding the
so-called Riemann-Stieltjes integral
Z

h(t)dG(t)

n
X

h(tj1 )(G(tj ) G(tj1 )).

j=1

G need not be monotone to define such integration. The needed condition on a function G
ensuring the existence of such integrals (via a limit of Riemann-Stieltjes sums) is that its
variation be bounded (finite) over [a, b]:
def

V (G)[a, b] = limn

n
X

|G(tj ) Gtj1 )| < ,

j=1

where by taking the limit we mean taking a finer and finer partition of [a, b]. (And the answer can
not depend on how we partition, as long as the subinterval lengths tend to 0.) Examples include
tj = a + j(b a)/n, j = 0, 1, 2, . . . , n, in which case each subinterval is of length (b a)/n. If we
imagine a particle whose yaxis position at time t is given by G(t), then this variation simply
measures the total yaxis distance (e.g., up-down distance) travelled during the time interval
[a, b]. Any differentiable function with continuous derivative g(t) = G0 (t) has finite variation
R
R
R
and in fact then V (G)[a, b] = ab |g(t)|dt, and integration reduces to ab h(t)dG(t) = ab h(t)g(t)dt;
we then write dG(t) = g(t)dt.
In this differentiable case, the variation is related to the arclength over [a, b] given by
Rbp
1 + (G0 (t))2 dt.
a
(It can be proved that a function G has finite variation if and only if G(t) = G1 (t) G2 (t)
where both G1 and G2 are monotone.)
Assume that G is of bounded variation and continuous (so it has no jumps). Let us focus for
simplicity on the interval [a, b] = [0, t]. If f = f (x) is a differentiable function with a continuous
derivative f 0 (x), then the differential of f (G(t)) can be dealt with by calculus yielding
df (G(t)) = f 0 (G(t))dG(t), differential form
Z

f (G(t)) = f (G(0)) +

f 0 (G(s))dG(s), integral form.

(1)
(2)

More generally, if f = f (t, x) is a real-valued function of two variables (t for time, x for space)
f
with both partial derivatives continuous, f
t (t, x), x (t, x), then the differential of f (t, G(t)) can
be dealt with by calculus yielding
f
f
(t, G(t))dG(t) +
(t, G(t))dt, differential form
(3)
x
t
Z t
Z t
f
f
(s, G(s))dG(s) +
(s, G(s))ds, integral form. (4)
f (t, G(t)) = f (0, G(0)) +
x
0 t
0

df (t, G(t)) =

1.2

Brownian motion has paths of unbounded variation

It should be somewhat intuitive that a typical Brownian motion path cant possibly be of
bounded variation; although continuous the paths seem to exhibit rapid infinitesimal movement
up and down in any interval of time. This is most apparant when recalling how BM can be
obtained bya limiting (in n) procedure by scaling a simple symmetric random walk (scaling
space by 1/ n and scaling time by n): In any time interval, no matter how small, the number
of moves of the random walk tends to infinity, while the size of each move tends to 0. This
intuition turns out to be correct (we state without proof):
Proposition 1.1 With probability 1, the paths of Brownian motion {B(t)} are not of bounded
variation, in fact they are differentiable nowhere: P (V (B)[0, t] = ) = 1 for all fixed t > 0,
and P (B 0 (t) does not exist at any value of t) = 1.
R t As a consequence, we can not naively define sample-path by sample-path an integral,
0 h(s)dB(s), in the Riemann-Stieltjes sense.
Rt

Of course, since BM has continuous sample paths, we can use it as an integrand;

exits in the Riemann sense for example, it is called integrated BM.

1.3

B(s)ds

Ito integration, Ito Calculus

To see what we can do to remedy the problem, let us try to make sense of
Z

B(s)dB(s),
0

using approximating sums of the form

n
X

B(tk1 )(B(tk ) B(tk1 )),

k=1

where 0 = t0 < t1 < < tn = t denotes a partition that becomes finer and finer as n .
Using the identity
1
1
B(tk1 )(B(tk ) B(tk1 )) = (B 2 (tk ) B 2 (tk1 )) (B(tk ) B(tk1 ))2
2
2
yields
n
X
k=1

B(tk1 )(B(tk ) B(tk1 ) =

n
1 2
1X
B (t)
(B(tk ) B(tk1 )2 .
2
2 k=1

The sum of squares nk=1 (B(tk ) B(tk1 ))2 is the key to understanding what to do here.
As the partition gets finer and finer, it turns out that the limiting so-called squared variation
exists and equals t, that is
P

def

Q(B)[0, t] = lim

n
X

(B(tk ) B(tk1 ))2 = t.1 .

k=1

To get some intuition: Since B(tk ) B(tk1 ) N (0, tk tk1 ), it follows that E(B(tk )
B(tk1 ))2 = V ar(B(tk ) B(tk1 )) = tk tk1 and thus
E

n
hX

(B(tk ) B(tk1 ))

n
X

(tk tk1 ) = t.

k=1

k=1

Since this is true for any partition, it holds in the limit too. Heuristically, we write (dB(t))2 =
dt. (It is important to realize that for continuous functions G of bounded variation, a quadratic
variation term does not arise, it is 0; (dG(t))2 = 0. For
G(t) = t,
P
P example, consider the function
and use the standard partition tk = kt/n. Then nk=1 (G(tk ) G(tk1 ))2 = nk=1 (t/n)2 =
t2 /n 0, as n ; in other words (dt)2 = 0.)
Thus we conclude that we have derived an interesting formula:
t

Z
0

1
1
B(s)dB(s) = B 2 (t) t.
2
2

(5)

The standard calculus in (1) applied to f (G(t)) with f (x) = x2 (and assuming G(0) = 0)
yields d(G2 (t)) = 2G(t)dG(t) or
t

Z
0

1
G(s)dG(s) = G2 (t).
2

For example, if G is a differentiable function with a continuous derivative and G(0) = 0 then
Rt
Rt
1 2
0
0 G(s)dG(s) = 0 G(s)G (s)ds = 2 G (t). (5) is different due to the non-zero quadratic variation; a second term gets included.
Unlike Riemann-Stieltjes integration, however, the above derivation of (5) fails if we choose
a different value for B(tk1 ) in our approximating sums, for example if we use
n
X

B(tk )(B(tk ) B(tk1 )),

k=1

or

n
X

B(tM
k )(B(tk ) B(tk1 )),

k=1

where tM
k = tk1 + (1/2)(tk tk1 ) denotes the midpoint of the interval [tk1 , tk ]. Each of these
other schemes leads to different answers. It is the one we derived using the values B(tk1 ) for
the increments B(tk ) B(tk1 ), that is called the Ito integral. It generalizes to integrals of the
R
form 0t X(s)dB(s) for appropriate stochastic processes {X(t) : t 0}. For example, we can
1

The convergence here, in general, is in probability or in L2

Rt

take any continuous function f = f (x) and define

R
did for deriving 0t B(s)dB(s), using sums
n
X

f (B(s))dB(s) using the same method we

f (B(tk1 ))(B(tk ) B(tk1 )).

k=1

The Ito integral leads to a nice Ito calculus so as to generalize (1) and (3); it is summarized
by Itos Rule:

Itos Rule
Proposition 1.2 If f = f (x) is a twice differentiable function with a continuous second derivative f 00 (x), then
1
df (B(t)) = f 0 (B(t))dB(t) + f 00 (B(t))dt, differential form
2
Z t
Z
1 t 00
0
f (B(s))dB(s) +
f (B(t)) = f (B(0)) +
f (B(s))ds, integral form.
2 0
0
f
t (t, x)

Proposition 1.3 If f = f (t, x) is a function such that both

uous, then
df (t, B(t)) =

2f
(t, x)
x2

and

(6)
(7)

are contin-

f
f
1 2f
dB(t) + (
+
)dt,
x
t
2 x2

which in integral form is given by

Z

f (t, B(t)) = f (0, 0) +

0

f
(s, B(s))dB(s) +
x

Z
0

f
1
(s, B(s))ds +
t
2

Z
0

2f
(s, B(s))ds.
x2

Examples
1. Take f (x) = x2 . Then f 0 (x) = 2x and f 00 (x) = 2 yielding d(B 2 (t)) = 2B(t)dB(t) + dt
which in integral form is precisely (5).
2. Consider geometric BM of the form
S(t) = eB(t) = f (B(t)), where f (x) = ex . Then f 0 (x) = f 00 (x) = ex and thus Itos formula
yields
1
d(eB(t) ) = eB(t) dB(t) + eB(t) dt,
2
or, as a stochastic differential equation (SDE)
1
dS(t) = S(t)dB(t) + S(t)dt.
2
In integral form we have
Z

B(s)

B(t)

dB(s) = e

1
1
2

Z
0

eB(s) ds.

Notice how this generalizes the standard

t

es ds = et 1.

3. BM with drift and variance term: As our next example, take f (t, x) = x + t, and let
f
2f
X(t) = B(t) + t. Then f
x = , t = and x2 = 0 yielding (in differential form)
dX(t) = dB(t) + dt
4. Geometric BM : As our last example, consider geometric BM of the form
S(t) = S0 eB(t)+t , where S0 > 0 is a constant initial value.
We use f (t, x) = S0 ex+t and observing that

f
x

= f ,

f
t

= f and

2f
x2

= 2 f , obtain

1
dS(t) = S(t)dB(t) + ( + 2 )S(t)dt.
2
This is the classic stochastic differential equation (SDE) for geometric BM.

Sketch of a proof of Itos formula f = f (x) case:

The proof of Itos formula relies on using the Taylors series expansion
1
f (x + h) f (x) = f 0 (x)h + h2 f 00 (x) + R(x, h),
2
where R(x, h) is the remainder.We partition [0, t] into n subintervals and use Taylors formula
on each subinterval;
f (B(t)) f (0)) =
=

n
X
k=1
n
X

(f (B(tk ) f (B(tk1 ))
f 0 (B(tk1 ))(B(tk ) B(tk1 )) +

k=1

n
1X
f 00 (B(tk1 ))(B(tk ) B(tk1 ))2 + ERROR.
2 k=1

As n , the first sum converges to 0t f 0 (B(s))dB(s), the second to 12 0t f 00 (B(s))ds (recall

(dB(s))2 = ds ) while the ERROR tends to 0; we have derived Itos formula.
R

1.4

Ito processes

If we replace BM with another process Y (t) and try to define yet again such integration as
Z

Y (s)dY (s),
0

we once again can use the identity

1
1
Y (tk1 )(Y (tk ) Y (tk1 )) = (Y 2 (tk ) Y 2 (tk1 )) (Y (tk ) Y (tk1 ))2
2
2

yielding
n
X

Y (tk1 )(Y (tk ) Y (tk1 )) =

k=1

n
1 2
1X
Y (t)
(Y (tk ) Y (tk1 ))2 .
2
2 k=1

As with BM, we thus need to be able to compute the squared variation:

def

Q(Y )[0, t] = lim

n
X

(Y (tk ) Y (tk1 ))2 .

k=1

Then we would have a formula for (dY (t))2 , and obtain a calculus
1
df (Y (t)) = f 0 (Y (t))dY (t) + f 00 (Y (t))(dY (t))2 .
2
If Y (t) itself is related to BM via dY (t) = H(t)dB(t) + K(t)dt, we call it an Ito process, in
which case it can be shown that (dY (t))2 = H 2 (t)dt.
A heuristic way to see this: squaring both sides of dY (t) = H(t)dB(t)+K(t)dt and recalling
that (dB(t))2 = dt, while (dt)2 = 0, yields
(dY (t))2 = H 2 (t)dt + K 2 (t)(dt)2 + 2K(t)H(t)dB(t)dt
= H 2 (t)dt + 2K(t)H(t)dB(t)dt. But it can be shown that dB(t)dt = 0, and the result follows.
Here, the point is that the cross variation
lim

n
X

(B(tk ) B(tk1 ))(tk tk1 ) = 0.

k=1

We summarize with a general result:

Theorem 1.1 Suppose dY (t) = H(t)dB(t) + K(t)dt is an Ito process. If f = f (t, x) is a
2f
function such that both f
t (t, x) and x2 (t, x) are continuous, then for Z(t) = f (t, Y (t)),
f
dY (t) +
x
h f
K(t) +
=
x

dZ(t) =

f
1 2f 2
dt +
H (t)dt
t
2 x2
f
1 2f 2 i
f
+
H (t) dt +
H(t)dB(t).
2
t
2 x
x

Application to geometric BM
As an application of Theorem 1.1, recall that geometric BM is indeed an Ito process: dS(t) =
S(t)dB(t)+(+ 21 2 )S(t)dt; H(t) = S(t) and K(t) = (+ 21 2 )S(t); thus for Z(t) = f (t, S(t)),
we have
f
f
1 2f 2 2
dS(t) +
dt +
S (t)dt
x
t
2 x2
h f
1
f
1 2f 2 2 i
f
=
( + 2 )S(t) +
+
S (t) dt +
S(t)dB(t).
2
x
2
t
2 x
x

dZ(t) =

(8)
(9)

1.5

Derivation of the Black-Scholes PDE using Ito calculus

Theorem 1.2 (Black-Scholes Partial Differential Equation (PDE)) Let f (t, x) denote
the price at time t of a European style derivative of stock (such as a European call option) with
expiration date T , when S(t) = x, where S(t) = S0 eB(t)+t is geometric Brownian motion,
modeling the price per share of the stock. Then f must satisfy the partial differential equation:
f
f
1 2f 2 2
+
rx +
x = rf.
t
x
2 x2
Proof :
As in our proof under the binomial lattice model, we will replicate our option payoff by
creating a portfolio of stock and risk-free asset, readjusting the portfolio over time. By doing
this we will see that f must satisfy the PDE in question.
We have two continuous time processes.
1. The stock price per share: S(t) = S0 eB(t)+t ; the SDE for this geometric BM is dS(t) =
S(t)dt + S(t)dB(t), where = + 2 /2.
2. The risk-free asset (bond (say)) per share: b(t) = ert , a deterministic function of t where
r > 0 is the interest rate; the differential equation for this asset is db(t) = rb(t)dt.
We wish to construct a portfolio ((t), (t)), denoting the number of shares of stock and riskfree asset we have at any time t [0, T ], the value of which matches the option payoff. At any
given time t [0, T ] the value of our portfolio is V (t) = (t)S(t) + (t)b(t). Our objective is
to buy some initial shares at time t = 0, (0 , 0 ), and then continuously readjust our portfolio
in such a way that V (t) = f (t, S(t)) for all t [0, T ], in particular the price of the option (at
time t = 0) will be C0 = V (0) = 0 S0 + 0 .
We are not allowed to either insert or remove funds along the way, that is, we assume a
self-financing strategy, mathematically enforced by assuming that
dV (t) = (t)dS(t) + (t)db(t), self-financing condition.

(10)

In integral form, this condition is given by

Z

(s)dS(s) +

V (t) = 0 S0 + 0 +
0

(s)db(s).
0

It simply ensures that any change in V (t) must equal the profit or loss due to changes in the
price of the two assets themselves. For example, if at time t we have (t)S(t) = $500, then we
can exchange (t)/2 shares of stock for $250 worth of risk-free asset (so V (t) does not change),
but we cant just sell the (t)/2 shares and remove this money from the portfolio.
Using our differentials dS(t) and db(t) together with the self-financing condition yields
dV (t) = (t)[S(t)dt + S(t)dB(t)] + (t)rb(t)dt
= ((t)S(t) + (t)rb(t))dt + (t)S(t)dB(t).

(11)
(12)

We are looking for factors (t) and (t) such that V (t) = (t)S(t)+(t)b(t)) = f (t, S(t)), t
[0, T ]. We will thus equate the two differentials dV (t) = df (t, S(t)) so as to figure out what
these factors must be, and also see what properties f must satisfy.
To this end, using Itos formula on f (t, S(t)) exactly as in (8) yields
7

f
f
1 2f
dt +
dS(t) +
(dS(t))2
t
x
2 x2
f
f
1 2f 2 2
df (t, S(t)) =
dt +
dS(t) +
S (t)dt
t
x
2 x2
h f
i
f
1 2f 2 2
f

S
(t)
+
=
+
S(t)dB(t).
S(t)
dt +
t
2 x2
x
x
df (t, S(t)) =

(13)
(14)
(15)

Equating the dB(t) coefficients from (12) and (15) yields

f
(t, S(t)).
x
Similarly equating the two dt coefficients yields
(t) =

(t) =


1 2f
1  f
2 2
(t, S(t)) +
(t,
S(t))
S
(t)
.
rb(t) t
2 x2

Finally, plugging in these values for (t) and (t) while equating (t)S(t) + (t)b(t) =
f (t, S(t)) then replacing S(t) with x yields the Black-Scholes PDE as was to be shown.