Lectures on the

Foundations of Mathematical Finance

M. R. Grasselli
Fields Institute, Winter 2010
Notes by E. Ha and W. Mnif
The aim of these lectures is to give an introduction to the mathematical foundations of nance,
rather than to mathematical nance per se. The reader is assumed to know the basics of stochastic
dierential equations and mathematical nance (at the level of Shreves textbooks [5], [6]).
1 No-arbitrage for Finite Probability Spaces
The notion of arbitrage will be one of the main themes of the course.
We will start the course by examining models based on nite probability spaces with discrete
time. By studying this toy model we can introduce the necessary ideas and language of func-
tional analysis at a relatively non-technical level. Later we will study models based on arbitrary
probability spaces with continuous time.
In this lecture we shall follow sections 2.12 of Delbaen and Schachermayer [2] very closely.
1.1 Finite Model of a Financial Market ([2], 2.1)
The nancial market model that we shall consider here is based on a nite probability space
(, T, P), where =
1
, . . . ,
N
is a nite set, P is a probability measure with P() > 0 for
all , and T is the -algebra of all subsets of . In addition, we consider a ltration
T
0
T
T
= T of sub--algebras of T with T
0
= , .
Denition 1.1.1. Asset prices are given by an R
d+1
valued adapted process

S = (

S
0
t
, . . . ,

S
d
t
)
T
t=0
.
We will assume that

S
0
0
= 1 and

S
0
t
> 0 for all t = 1, , T.
The requirement of being an adapted process, that is

S
t
is T
t
-measurable, simply means that
the prices at t are known at time t, despite being uncertain at any earlier time. The rst component
S
0
t
will play the role of a numeraire, which is a fancy way to call the units used to express the
value of the other assets. In the simplest case we have S
0
t
1 for all t, so we can think of it as a
xed currency amount, say one Canadian dollar. More generally, S
0
t
can represent the value of a
bank account accumulating interest as time goes by.
Denition 1.1.2. A trading strategy is an R
d+1
valued predictable process

H = (

H
0
t
, . . . ,

H
d
t
)
T
t=1
.
The components of a trading strategy H
t
are the number of units of an asset being held from
time t 1 until time t. The holding decision is made at t 1, which explains why H
t
needs to be
predictable, that is

H
t
is T
t1
-measurable.
Given an asset

S
t
and a trading strategy

H
t
, the inner product

V
t
=

H
t

S
t
:=
d

i=0

H
i
t

S
i
t
(1)
is called the portfolio value at time t.
1
Denition 1.1.3. The trading strategy

H is said to be selfnancing if

H
t+1

S
t
=

H
t

S
t
, (2)
for every t = 1, . . . , T 1.
The lefthand side of (2) corresponds the amount of money necessary to form a portfolio to be
held from time t until time t +1 at the market prices prevailing at time t, whereas the righthand
side is the amount of money obtained from the portfolio held from time t 1 until time t. In other
words, a trading strategy is selfnancing provided there is no injection or withdraw of funds at
any given time.
Given asset prices

S
t
= (

S
0
t
, . . . ,

S
d
t
), let S
t
= (S
1
t
, . . . , S
d
t
) be the R
d
valued adapted process
of discounted prices with components
S
j
t
=

S
j
t

S
0
t
, j = 1, . . . , d.
Accordingly, given an R
d+1
valued trading strategy

H
t
= (

H
0
t
, . . . ,

H
d
t
), let H
t
= (H
1
t
, . . . , H
d
t
) be
the R
d
-valued trading strategy with components
H
j
t
=

H
j
t
, j = 1, . . . , d.
Notice that, for every predictable R
d
valued process H
t
we can construct a unique R
d+1
valued
selfnancing trading strategy

H
t
such that

H
j
t
= H
j
t
for all j = 1, . . . , d and all t = 1, . . . , T
simply by setting H
0
1
= 0 and nding H
0
t
inductively for j = 2, . . . , T using (2). Together with the
next proposition, this construction shows that, as long as we are interested in discounted portfolio
values only, there is no loss of information when we consider the restricted R
d
valued strategy H
t
instead of the R
d+1
value strategy

H
t
.
Proposition 1.1.4. Let

H
t
be the unique R
d+1
value selfnancing trading strategy associated
with an arbitrary R
d
value strategy H
t
through the construction above. Then the discounted
portfolio value V
t
=

V
t
/S
0
t
is independent of the scalar process

H
0
t
.
Proof. Since

S
0
0
= 1 and H
0
1
= 0, we have that
V
0
=

V
0
=

H
1
1

S
1
1
+ +

H
d
1

S
d
1
.
Using the self-nancing condition

H
t

S
t
=

H
t+1

S
t
, we nd that the change in V
t
is
V
t+1
:= V
t+1
V
t
=

V
t+1
/

S
0
t+1

V
t
/

S
0
t
=

H
t+1

S
t+1
/

S
0
t+1

H
t+1

S
t
/

S
0
t
=

H
0
t+1
(1 1) +
d

j=1

H
j
t+1
(

S
j
t+1
/

S
0
t+1

S
j
t
/

S
0
t
)
=
d

j=1
H
j
t+1
S
j
t+1
= H
t+1
S
t+1
.
Thus V
t
= V
0
+H
1
S
1
+ +H
t
S
t
, which is manifestly independent of

H
0
t
.
In particular, using the following standard notation for stochastic integrals (see [3])
(H S)
T
= (H S)
T
t=1
:=
T

t=1
H
t
S
t
we have that
V
T
= V
0
+ (H S)
T
. (3)
2
1.2 No Arbitrage and the FTAP
Let 1 be the space of predictable R
d
-valued processes H
t
for a nancial market with discounted
asset prices S
t
. Let L
0
(, T, P) denote the space of all measurable functions on , which for nite
sample spaces is canonically isomorphic to R
N
. Similarly, let L

(, T, P) = R
N
denote the space
of bounded measurable functions on .
Denition 1.2.1. The subspace K L
0
(, T, P) dened by
K = (H S)
T
[ H 1,
is called the set of attainable claims at price 0, whereas the convex cone C L

(, T, P) dened
by
C = g L

(, T, P) [ g f for some f K
is called the set of super-replicable claims at price 0. (We say super-replicable since such claims
are dominated by attainable claims the terminology is not ideal!) For a R, the sets K
a
= a+K
and C
a
= a +C are respectively the sets of attainable and superreplicable claims at price a.
Remark 1.2.2. Observe that we can write C = K + L
0

. It then follows that C is a closed set,

since K is closed (being a linear subspace of R
N
) and L
0

is a closed polyhedral cone. This is one

of the many instances where working with nitedimensional vector spaces simplies the analysis
tremendously.
Denition 1.2.3. A nancial market S satises the no arbitrage condition (NA) if
K L
0
+
(, T, P) = 0.
or equivalently
C L

+
(, T, P) = 0.
Because this is the central concept in this notes, it deserves further explanation. In view of
(3), the set K consists of random variables that coincide with the discounted terminal values of
selfnancing trading strategies starting at zero initial value. On the other hand, L
0
+
is the set of
nonnegative vectors in R
N
. Thus, an arbitrage is a selfnancing trading strategy starting with
zero initial value and with terminal value given by a random variable that is nonnegative and not
identically equal to zero. That, an arbitrage is a strategy that starts at zero, never loses money,
and has a strictly positive probability of making money.
Proposition 1.2.4. The condition (NA) implies that C (C) = K.
Proof. Clearly K C(C). For the reverse inclusion, consider an element g C(C). It then
follows from Remark 1.2.2 that we can write g = f
1
h
1
= f
2
+ h
2
for some elements f
1
, f
2
K
and h
1
, h
2
L
0
+
. But then f
1
f
2
= h
1
+h
2
is in K L
0
+
, which is 0 by (NA). Thus h
1
= h
2
= 0
and so g K.
Denition 1.2.5. A probability measure Q on (, T) is an equivalent martingale measure (EMM)
for S if Q P (that is, Q[
n
] > 0 for all n) and S is a Q-martingale, that is
E
Q
[S
t+1
[ T
t
] = S
t
, t = 0, 1, . . . , T 1.
The set of EMMs for S is denoted /
e
(S).
Lemma 1.2.6. For a probability measure Q on (, T), the following are equivalent:
(i) S is a Qmartingale.
(ii) E
Q
[f] = 0 for all f K.
(iii) E
Q
[g] 0 for all g C.
3
Proof. We shall prove only that (i) (ii), since (ii) (iii) is obvious. First observe that for a
Qmartingale S and a predictable trading strategy H
t
we have
E
Q
[H
t
S
t
[T
t1
] = E
Q
[
d

j=1
H
j
t
(S
j
t
S
j
t1
[T
t1
]
=
d

j=1
H
j
t
E
Q
[S
j
t
S
j
t1
[T
t1
] = 0.
But this shows that (H S)
t
is also a Qmartingale, since
E
Q
[(H S)
t
[T
t1
] = E
Q
[
t1

s=1
H
s
S
s
+H
t
S
t
[T
t1
] = (H S)
t1
.
In particular, E
Q
[(H S)
T
] = (H S)
0
= 0, which shows that (1) (2). Conversely, let A be
an arbitrary T
t1
measurable set and consider the strategy H(, s) = 1
A
()1
(t1,t]
(s). Then
(H S)
T
= 1
A
(S
t
S
t1
) and (ii) implies that
E
Q
[1
A
(S
t
S
t1
)] = 0
which is equivalent to
E
Q
[S
t
[T
t1
] = S
t1
,
which in turn means that S is a Qmartingale.
We are now in a position to prove the following theorem, known as the Fundamental Theorem
of Asset Pricing.
Theorem 1.2.7 (FTAP). For a nancial market modeled on a nite probability space (, T, P; T
t
),
the following are equivalent:
1. S satises (NA).
2. /
e
(S) ,= .
Proof. (2) (1) (easy part): Suppose Q /
e
(S). By Lemma 1.2.6, E
Q
[g] 0 for all g C. On
the other hand, if there were a non-zero element g C L

+
, then we would have E
Q
[g] > 0, since
Q P. So necessarily S must satisfy (NA).
(1) (2) (interesting part): By the condition (NA), K L

+
= 0, and so K and L

+
are
disjoint convex sets. Let B =

n
1
n
[
n
0,

n
= 1. Then B L

+
is a convex
compact set which is disjoint from K. Now, by the separating hyperplane theorem (take the
proof in [4, Theorem V.4] and eliminate the use of Hahn-Banach), there is a linear functional
Q (L

= L
1
separating B and K. This means that we can nd numbers < such that
Q[f] < Q[g], for all f K, g B.
Since K is linear, we have 0, and without loss of generality we can take it to be 0, which
implies that > 0. Let e
n
be the n-th canonical basis vector of R
N
. Since e
n
B we have that
Q(e
n
) > 0. Moreover, let I = (1, . . . , 1). Then by linearity Q[I] > 0. Normalizing so that Q[I] = 1,
we can associate Q with a probability measure equivalent to P satisfying property (ii) of Lemma
1.2.6. Then Q /
e
(S).
Corollary 1.2.8. Let S satisfy (NA) and let f = a +(H S)
T
for some H 1 and a R. Then
a and H are uniquely determined by this expression and, moreover, a = E
Q
[f] and a +(H S)
t
=
E
Q
[f [ T
t
].
4
Proof. For uniqueness, suppose that f = a
1
+ (H
1
S)
T
= a
2
+ (H
2
S)
T
, and say a
1
> a
2
. Then
_
(H
2
H
1
) S
_
T
= a
1
a
2
> 0 is an arbitrage. But since we are assuming that S satisfy (NA),
we must have a
1
= a
2
. Next suppose that H
1
,= H
2
and dene
A = [ (H
1
S)
t
(H
2
S)
t
> 0,
for some t. Then H := (H
1,t
H
2,t
)1
A
1
(0,t]
is an arbitrage trading strategy because (H S)
T
= 0
outside A, while (H S)
T
= (H
1
S)
t
(H
2
S)
t
> 0 on A. But again, since we are assuming that
S satisfy (NA), we must have H
1
= H
2
.
The last part follows from the fact (already established) the stochastic integral (H S)
t
is a
Q-martingale.
1.2.1 Convex cones and polar sets
Using a standard denition in convex analysis, let the polar set of our cone C of super-replicable
claims be given by
C

= f L
1
(, T, P) [ E[fg] 0, f C.
According to the bipolar theorem (for a very general version, see [1]), the bipolar set C

:= (C

coincides with the closed convex hull of C. But by virtue of Remark 1.2.2, we know that C is
already a closed set, from which we conclude that C

= C.
Now denote by /
a
S the set absolutely continuous martingale measures for S, that is probability
measures Q which are absolutely continuous with respect to P and such that S is a Qmartingale.
Consider the cone generated by /
a
S, that is,
cone(/
a
S) :=
_
f =
dQ
dP
, 0, Q /
a
S
_
The next proposition establishes a perfect polar relation between C and /
a
S.
Proposition 1.2.9. Suppose S satises (NA). Then C

= cone /
a
(S), and /
e
(S) is dense
in /
a
(S).
Proof. Since S satises (NA), /
e
(S) ,= by the FTAP. Pick any Q

/
e
(S). Then for all
Q /
a
(S) and 0 < 1, we have Q

+ (1 )Q /
e
(S), since (1 )Q is absolutely
continuous with respect to /
e
(S). In particular, /
a
(S) is arbitrarily close to /
e
(S), which
proves the density statement.
Now let Q /
a
(S) and let > 0. Then by Lemma 1.2.6, we have
E
_

Q
dP
g
_
= E
Q
[g] 0, g C.
This shows that cone(/
a
(S)) C

. For the converse, observe rst that L

C because 0 is
achievable, which implies that C

L
1
+
. This means that f C

L
1
+
can be written as f =
dQ
dP
for some 0 and some probability measure Q. But then
0 E[fg] = E
Q
[g], g C
which means that Q /
a
(S), by Lemma 1.2.6. Thus C

cone(/
a
(S)).
In view of Lemma 1.2.6, another way to formulate the proposition is as follows.
Proposition 1.2.10. For all g L

(, T, P), the following are equivalent:

1. g C.
2. E
Q
[g] 0 for all Q /
a
(S).
3. E
Q
[g] 0 for all Q /
e
(S).
5
2 Utility Maximization in Finite Probability Spaces
Consider
U : R R ,
satisfying
(i) U is increasing on R;
(ii) U is continuous on dom(U) = x/U(x) > ;
(iii) U is strictly concave on the interior of dom(U);
(iv) U

() = 0.
Regarding negative wealth, we assume:
Case 1 U(x) = , x < 0;
U(x) > , x > 0;
U

(0) = +;
Examples: U(x) = log(x), U(x) =
x
p
p
, p (, 1)0.
Case 2 U(x) > , x R;
U

() = +;
Example: U(x) =
e
x

, > 0.
The utility maximization we are interested in is:
u(x) := sup
HH
E[U(x + (H S)
T
)], x dom(U) (4)
Before tackling this problem, it is convenient to dene the conjugate function V by
V (y) = sup
xR
(U(x) xy), y > 0,
which can be seen as the Legendre transform of U(x). It follows that V satises the following
properties:
(i) V : R R is nite valued;
(ii) V is dierentiable and strictly convex on (0, +);
(iii) V

(0) = .
Moreover,
Case 1 V () = lim
x0
U(x) and V

() = 0;
Case 2 V () = + and V

() = +.
In addition, U(x) = inf
y>0
[V (y) +yx] and V

(U

(x)) = x. In other words, I := (U

)
1
= V

.
6
2.1 The complete market case
Assume that /
e
(S) = Q and consider the Arrow-Debreux securities 1
{n}
so that E
Q
[1
{n}
] =
Q(
n
) := q
n
and because the market is complete, 1
{n}
= q
n
+ (H
n
S) for some H
n
1. It
follows from the previous lemmas that a random variable f L

satises f x + (H S)
T
for
some H 1 i E
Q
(f) x. Therefore, in this nite, complete case, we can rewrite (4) as the
following concave optimization with a linear constraint:
u(x) = sup
f R
N
E
Q
[f] x
E[U(f)]
= sup
(f
1
, ..., fn)

N
n=1
fnqn x
N

n=1
p
n
U(f
n
). (5)
To solve this problem, let us introduce the Lagrangian:
L(f
1
, ..., f
n
, y) =
N

n=1
p
n
U(f
n
) y(
N

n=1
f
n
q
n
x)
=
N

n=1
p
n
(U(f
n
) y
q
n
p
n
f
n
) +xy.
It follows from the saddle point theorem that a solution to (5) is given by a saddle point (

f
1
, ...,

f
N
, y)
of L, that is,
L(f
1
, ..., f
N
, y) L(

f
1
, ...,

f
N
, y) (

f
1
, ...,

f
N
, y), f R
N
, y > 0.
To see this, dene
(f
1
, .., f
N
) = inf
y0
L(f
1
, ..., f
N
, y), f
n
dom(U)
and
(y) = inf
fR
N
L(f
1
, ..., f
N
, y), y 0.
Now notice that if f = (f
1
, .., f
N
) satises E
Q
[f] x, then (f
1
, .., f
N
) = L(f
1
, ..., f
N
, 0) =

N
n=1
p
n
U(f
n
). Conversely, if E
Q
(f) > x, then (f
1
, .., f
N
) = . Therefore,
sup
fR
N
(f
1
, .., f
N
) = sup
f R
N
E
Q
[f] x
N

n=1
p
n
U(f
n
) = u(x).
Moreover, observe that for xed y > 0, the optimization over R
N
appearing in the denition of
(y) splits into N separate one dimensional optimization problems. Explicitly, using the denition
of V , we see that:
(y) =
N

n=1
p
n
V (y
q
n
p
n
) +xy
= E[V (y
dQ
dP
)] +xy
:= v(y) +xy.
Observe that v(y) inherits all the proprieties of V . In particular, for x dom(U), there exists a
unique y = y(x) > 0 such that v

( y(x)) = x, which is therefore the unique optimizer for (y).

7
Fixing y(x), we see that the function (f
1
, .., f
N
) L(f
1
, .., f
N
, y) achieves its maximum at
(

f
1
, ...,

f
N
) satisfying U

f
n
) = y(x)
qn
pn

f
n
= I( y(x)
qn
pn
) which implies that
inf
y>0
(y) = inf
y>0
(v(y) +xy)
= v( y(x)) +x y(x)
=
N

n=1
p
n
V ( y(x)) +x y(x)
=
N

n=1
p
n
(U(

f
n
) x y(x)
q
n
p
n
) +x y(x)
= L(

f
1
, ...,

f
N
, y).
Notice that

f
n
is in the interior of dom(U), which means that L is continuously dierentiable at
(

f
1
, ...,

f
N
, y) and
L
y
[

f, y
= 0 so that the constraint is binding, that is

N
n=1
q
n

f
n
= x.
Finally, it is clear that

N
n=1
p
n
U(

f
n
) u(x). Conversely, for all (f
1
, .., f
n
) satisfying E
Q
[f]
x, we have
N

n=1
p
n
U(f
n
) L(f
1
, ..., f
N
, y) L(

f
1
, ...,

f
N
, y) =
N

n=1
p
n
U(

f
n
).
Therefore, u(x) = v( y(x)) +x y(x) u

= y(x). So that u inherits all the properties from U.

Theorem 2.1.1. For a nite complete case, dene u(x) = sup
HH
E[U(X
T
)], x dom(u),
X
T
= x + (H S)
T
, and v(x) = E[V (y
dQ
dP
)], y > 0. Then,
(i) u and v are conjugates and inherit the proprieties of U and V ;
(ii)

X
T
(
n
) = I(y
dQ
dP
) (or equivalent U

(

X
T
(
n
)) = y
dQ
dP
(
n
)) is optimal wealth, where y satises
u

(x) = y (or equivalently v

(y) = x).
Notes:
(1) U

(

X
T
(
n
)) = y
qn
pn
, where U

represents the marginal utility,

X
T
(
n
) is the optimal wealth,
and q
n
is the price of the Arrow-Debrew security 1
{n}
with a probability of its success p
n
.
(2) Observe that u

(x) = y and U

(

X
T
) = y
dQ
dP
implies that u

(x) = E[U

(

X
T
)].
Consider an agent with x + as an initial endowment who uses x to nance

X
x
T
= x + (

H S)
T
for some

H 1 and to buy the numeraire. Thereby ending with

X
T
+ at T. Comparing this
by the optimal wealth

X
x+
T
gives:
lim
0
+
u(x +) u(x)

= lim
0
+
E[U

(

X
x+
T
) U

(

X
x
T
)]

lim
0
+
E[U

(

X
x
T
+) U

(

X
x
T
)]

= E[U

(

X
x
T
)].
The argument with < 0 implies that u

(x) = E[U

(

X
x
T
)] the agent is indierent. Similarly, we
can prove that xu

(x) = E[

X
x
T
U

(

X
x
T
)].
8
3 The DalangMortonWillinger Theorem
3.1 The closedness of C
Let X = S
1
= S
1
S
0
and its corresponding subspaces:
E
X
= H : R
d
, F
0
-measurable and X H = 0 a.s and
1
X
= f : R
d
, F
0
-measurable and Pf = f,
where P = IP

and P

being the projection associated to E

X
. Observe that HX = H(S
1
S
0
) =
(H.S)
1
. Dene:
I : /
0
(, F
0
, P; R
d
) /
0
(, F
0
, P)
H (H S)
1
Denition 3.1.1. We say H is a canonical form for S if H 1
X
, where X = S.
Lemma 3.1.2. The kernel of I is E
X
. The restriction of I to 1
X
is injective, linear and has full
range.
Proof. The rst statement follows from the denition of E
X
. For the second, let H and H

1
X
with I(H) = I(H

). Then X (H H

) = 0 a.s (H H

) E
X
. But (H H

) 1
X
, then
(H H

) E
X
1
X
= 0 H = H

a.s.
Proposition 3.1.3. Let (S
t
)
1
t=0
be adapted to (, (F
t
)
1
t=0
, P) and let (H
n
)

n=1
be a sequence in
/
0
(, F
0
, P; R
d
) in canonical form. Then:
(i) (H
n
)

n=1
is bounded i (H
n
S)

n=1
is.
(ii) (H
n
)

n=1
converges a.s i (H
n
S)

n=1
does.
If, in addition, S satises (NA), then
(i) (H
n
)

n=1
is bounded i ((H
n
S)

n=1
is.
(ii) (H
n
)

n=1
converges to zero a.s i ((H
n
S)

n=1
does.
Proof. We consider just the if part of each statement.
(i) and (i): Suppose that (H
n
)

n=1
is not bounded. Let K = R
d
and take x
0
= K.
Since (H
n
)

n=1
diverges to on a set B of positive measure, there exists a subsequence (L
k
)

k=1
=
(H

k
)

k=1
such that (L
k
())

k=1
diverges to on B.
Now put

L
k
=
L
k
|L
k
|
I
B{|L
k
|1}
so that [

L
k
()[ = 1 for B and k suciently large.
By passing to a subsequence again, we may suppose that (

L
k
)

k=1
converges to

L, which is in
canonical form and satises [

L[ = 1 on B. Therefore we assume that (H

n
S)

n=1
is bounded.
Then (

L
n
S)

n=1
necessarily goes to zero a.s. But then

L S = lim
k

L S = 0 a.s and
since

L is in canonical form, this implies that

L = 0 a.s (contradiction).
In addition, suppose that S satises (NA). Then if we assume that ((H
n
S)

n=1
is bounded.
So,

L S

= lim
k

L
k
S

= 0 a.s. (NA) implies

L S

= 0 a.s

L = 0 a.s (contradiction).
(ii) and (ii): Suppose that (H
n
)

n=1
does not converge a.s, but (H
n
S)

n=1
does. Then,
(H
n
S)

n=1
is bounded a.s. Therefore (H
n
)
n=1
is also bounded a.s from the previous item.
Using again the compact set K = R
d
; we can nd a subsequence (H

k
)

k=1
comparing
to some H
0
1
X
. Since (H
n
)

n=1
itself does not converge, we can nd another subsequence
(H

k
)

k=1
converging to some

H
0
with P[H
0
,=

H
0
] > 0. But since (H
n
S)

n=1
converge, we
must have (H
0

H
0
) S = lim
k
H

k
S lim
k
H

k
S = 0 a.s. Therefore, H
0
=

H
0
.
Assume (NA) and also that (H
n
)

n=1
does not converge to zero but ((H
n
S)

n=1
does. We
can again nd a convergent subsequence (H

k
)

k=1
converging to some

H
0
such that P(

H
0
,= 0) >
0. But

H
0
S

= lim
k
H

k
S

= 0 a.s. which, together with (NA), means that

H
0
= 0
a.s.
9
Theorem 3.1.4. Let (S
t
)
1
t=0
be a one step process.
(i) K = H S/H L
0
(, F
0
, P; R
d
) is closed in L
0
(, F
1
, P).
(ii) If S satises (NA), then C=K-L
0
+
(, F
1
, P) is also closed.
Proof. (i) let (f
n
) = (H
n
S)

n=1
be a sequence in K converging to f
0
L
0
(, F
1
, P), with each
H
n
in canonical form. By passing to a subsequence, we can suppose that (f
n
)

n=1
converges to f
0
a.s. Then, (H
n
)

n=1
converges a.s to some H
0
L
0
(, F
0
, P; R
d
) so that f
0
= H
0
S K.
(ii) Let f
n
= g
n
h
n
be a sequence in C converging to f
0
L
0
(, F
1
, P), where g
n
= H
n
S
for H
n
in canonical form and h
n
L
0
+
(, F
1
, P).
Again by passing to a subsequence, we can assume that (f
n
) converges to f
0
a.s. Since g
n
f
n
we have that g
n
is bounded. Because of (NA), we then have that (H
n
)

n=1
is also bounded a.s.
By passing to a convergent subsequence (H

k
)

k=1
, we may suppose that g

k
= H

k
S converges
a.s to g
0
= H
0
S, H
0
= L
0
(, F
0
, P; R
d
). Since (f

k
) still converge a.s to f
0
, we have that
h

k
= g

k
f

k
converges a.s to h
0
0. Thus f
0
= g
0
h
0
C.
3.2 The DMW Theorem for T = 1
Theorem 3.2.1. Let (S
t
)
1
t=0
be a one-step price process adapted to (, (F
t
)
1
t=0
, P) satisfying the
(NA) condition. Then, an equivalent probability measure Q such that:
(i) S
0
, S
1
L
1
(, F
1
, Q; R
d
);
(ii) E
Q
[S
1
[ F
0
] = S
0
;
(iii)
dQ
dP
is bounded.
Proof. First construct P
1
given by:
dP
1
dP
= ce
S1S0
,
so that P
1
P,
dP1
dP
is bounded, S
0
, S
1
L
1
(, F
1
, P
1
; R
d
). Next, take C
1
= CL
1
(, F
1
, P
1
; R
d
).
Then it is easy to showthat C
1
is closed in L
1
(, F
1
, P
1
; R
d
), because C is closed in L
0
(, F
1
, P; R
d
) =
L
0
(, F
1
, P
1
; R
d
). Moreover, C
1
is a convex cone because C is a convex cone and by (NA),
C
1
L
1
+
(, F
1
, P
1
; R
d
) = 0. It then follows from the Hahn-Banach Theorem (see the gen-
eral version next lecture!) that we can nd an equivalent probability measure Q such that
dQ
dP1
is
bounded and E
Q
[f] 0 for all f C
1
. We then have that S
0
, S
1
L
1
(, F
1
, Q; R
d
) and that
dQ
dP
=
dQ
dP1
dP1
dP
is bounded.
For the martingale property, observe that for each component j = i, .., d and each A F
0
, we
have that I
A
(S
j
1
S
j
0
) C
1
and I
A
(S
j
1
S
j
0
) C
1
. Therefore, E
Q
[I
A
(S
j
1
S
j
0
) [ F
0
] = 0 and so
E
Q
[I
A
(S
1
S
0
) [ F
0
] = 0.
3.3 Proof of the DMW theorem for T 1
Let us use induction on the number of intervals necessary to reach T.
For T = 1, the result holds.
Suppose that it holds for n = T 1, that is, consider t = 1, ..., T and the process (S
t
)
T
t=1
adapted to (, (F
t
)
T
t=1
, P; R
d
) for which there exists on equivalence probability measure Q
1
on F
T
such that:
(i)
dQ
1
dP
is bounded;
(ii) S
1
, ..., S
T
L
1
(, F
T
, Q
1
; R
d
);
10
(iii) (S
t
)
T
t=1
is a Q
1
-martingale.
Using the one-step DMW Theorem for (S
t
)
1
t=1
and (, (F
t
)
1
t=0
, Q
1
; R
d
), we can nd a bounded
function f
1
such that f
1
is F
1
-measurable, f
1
> 0, E
Q
1 [f
1
] = 1, E
Q
1 [[ S
1
[ f
1
] < , E
Q
1 [[ S
0
[
f
1
] < and for all A F
0
:
_
A
S
0
f
1
dQ
1
=
_
A
S
1
f
1
dQ
1
.
Now dene Q on F
T
by Q[A] =
_
A
f
1
dQ
1
, A F
T
. That is,
dQ
dP
= f
1
dQ
1
dP
is bounded and
dQ
dP
> 0.
So Q P.
Moreover, for t = 0, ..., T, we have
_

[S
t
[dQ =
_

[S
t
[f
1
dQ
1
< .
Finally, for the martingale property, observe that for all A F
0
, we have:
_
A
S
0
dQ =
_
A
S
0
f
1
dQ
1
=
_
A
S
1
f
1
dQ
1
=
_
A
S
1
dQ.
So that E
Q
[S
1
[ F
0
] = S
0
, for t 1, let A F
t
, then
_
A
S
t
dQ =
_
A
S
t
f
1
dQ
1
=
_
A
S
t+1
f
1
dQ
1
=
_
A
S
t+1
dQ.
So, E
Q
[S
t+1
[ F
t
] = S
t
.
4 No-arbitrage in continuous time
4.1 Stochastic integrals for semimartingales
Recall that we have dened the stochastic integral H (H.W)
t
pathwise for bounded simples
strategies and used the isometry
| H |
L
2
(R+,P,P)
=| (H.S)

|
L
2
(,F,P)
to extend it by continuity to the entire space L
2
(R
+
, T, P ), in such way that (H.W)
t
is an
L
2
-bounded martingale, that is, sup
t
| (H.W)
t
|
L
2
(,F,P)
< . When H is locally in L
2
(P )
(which is equivalent to
_
t
0
H
2
s
ds < ), the same construction yields a local martingale (H.W)
t
which is locally L
2
bounded.
Recall also that for an L
2
bounded martingale S
t
, we dene the quadratic variation measure
on T as:
d[S](][, ][) := E([S

[
2
)
and the following isometry holds for bounded simple integrand H:
| H |
L
2
(R+,P,d[S])
=| (H.S)

|
L
2
(,F,P)
we can then extend H (H.S)
t
to the entire space L
2
( R
+
, T, d[S]) by continuity in such a
way that (H.S)
t
is also an L
2
bounded martingale. We can use again localization to extend this
to a locally L
2
integrand H and locally L
2
bounded local martingale. Since every continuous local
martingale S is automatically locally L
2
-bounded, this is the right degree of generality for this
class. To include integrators with jumps, we will extend the theory even beyond local martingales.
Suppose rst that S is a c` adl` ag adapted process of bounded variation, that is,
[S[
t
:= sup
0t0t1...tnt
n

i=0
[S
ti+1
S
ti
[ < a.s for all t < .
11
Then for almost all , the path (S
t
())
0t<
is of bounded variation on compact subsets of
R
+
by dS()(]a, b]) = S
b
() S
a
(). So that the stochastic integral (H.S)
t
() =
_
t
0
H
u
()dS
u
()
is well dened as a Lebesgue-Sieltjes integral for each process H such that (H
u
())
0ut
is dS()-
integrable. We have then led to investigate process of the form
S = M +A (6)
where M is a bounded martingale and A is locally of bounded variation.
For that, let us dene o as the class of bounded simple integrands H with the topology of
uniform convergence which is given by the norm:
| H |

= sup| H
t
|
L

(,Ft,P)
t R+ (7)
For this class, we can dene the stochastic integral as before:
I(H) = (H.S)

=
n

i=1
f
i1
(S
i
S
i1
) (8)
for any c` adl` ag process S.
Denition 4.1.1. S is a strict semi-martingale if the map:
I : o L
0
(, F

, P) (9)
H I(H) = (H S)

(10)
is continuous for the topologies of ||

on o and convergence in probability on L

0
.
(i) S is a semi-martingale if it is locally a strict semi-martingale.
Theorem 4.1.2 (Bicheteler-Dellacherie). S is a semi-martingale in the sense of the denition
above if and only if it can be decomposed as S = M +A as in (6).
We say that S is a special semi-martingale if in addition the process A is predictable.
It is relatively easy to show that (H.S)
t
is a semi-martingale, even when H is only locally in
o. To extend for H beyond L

-bounded simple, consider the semi-martingale topology induced

on the set of one-dimensional semi-martingale by the distance:
D[S] =

n=1
2
n
supE[[ (K S)
n
[] 1/K 1
where K is a predictable process. This means that S
n
0 i (K S
n
)
t
0 uniformly in t and K.
One can show that this space is complete. We then say that an R
d
-valued process H is S-integrable
(L(S)) w.r.t a semi-martingale S if (H1
Hn
S)

n=1
is a cauchy sequence. We dene (H S)
t
as
the limit of the sequence.
Notice (H S)
t
is a semi-martingale and can be decomposed to a local martingale

M
t
+

A
t
that
are dierent from (H M)
t
and (H .A)
t
resp.
Remark 4.1.3. One can construct examples where S = M +A is a special semi-martingale, H is
Sintegral, so that (H S)
t
exists, but (H A) does not exist.
Lemma 4.1.4. Let S be a special semi-martingale with decomposition S = M + A and H be
an R
d
valued predictable process. If the stochastic integral (H S) is itself special, then (H A)
exists as a Lebesgue-Stieltjes integral.
One can nd examples of a martingale M
t
and an Mintegrable process H such that (H M)
t
is not a local martingale.
12
Lemma 4.1.5. Let M be an R
d
valued local martingale and let H be an R
d
valued Mintegarble
process. Then (H M)
t
is a local martingale if there exists a sequence of stopping times
n

and integrable function
n
L
1
with
n
0 s.t < H, M >

m

n
.
Theorem 4.1.6. If S is a special semi-martingale with canonical decomposition S = M +A and
if H is Sintegrable then (H.S) is special martingale i:
i (H M) is dened as an integral in local martingale sense
ii (H A) is dened as a Lesbegue-Stieltjes integral.
Proof. Let H be Sintegrable. If (H S) is special then (H A) exists as a L-S integrable by
Lemma 1, which gives (ii). Moreover, (H S) is a special, it must be locally integrable, that is,
there is a sequence
n
and
n
L
1
s.t (H S)
n

n
. Now let
n
be stopping times such
that (
_
n
0
[ H
s
[ dA
s
) L
1
(which must exist since (H A) is a regular L-S integral). Then for each
n, (H M)
nn

n

_
n
0
[ H
s
[ dA
s
and Lemma 2 shows that (H M) is a local martingale.
Conversely, if (i) and (ii) hold, then (H S) is the sum of a local martingale and a predictable
bounded variation process (H A) and therefore special.
References
[1] W. Brannath and W. Schachermayer. A bipolar theorem for L
0
+
(, T, P). In Seminaire de
Probabilites, XXXIII, volume 1709 of Lecture Notes in Math., pages 349354. Springer, Berlin,
1999.
[2] F. Delbaen and W. Schachermayer. The Mathematics of Arbitrage. Springer Finance. Springer-
Verlag, Berlin, 2006.
[3] Philip Protter. Stochastic integration and dierential equations, volume 21 of Applications of
Mathematics. Springer-Verlag, Berlin, 1990. A new approach.
[4] M. Reed and B. Simon. Methods of Modern Mathematical Physics. I. Academic Press Inc.
[Harcourt Brace Jovanovich Publishers], New York, second edition, 1980. Functional analysis.
[5] S. E. Shreve. Stochastic calculus for nance. I. Springer Finance. Springer-Verlag, New York,
2004. The binomial asset pricing model.
[6] S. E. Shreve. Stochastic calculus for nance. II. Springer Finance. Springer-Verlag, New York,
2004. Continuous-time models.
13