Blatt 11

Lecture Notes
FINANCIAL MATHEMATICS I:
Stochastic Calculus, Option Pricing,
Portfolio Optimization
Holger Kraft
University of Kaiserslautern
Department of Mathematics
Mathematical Finance Group
Summer Term 2005
This Version: August 4, 2005
CONTENTS 1
Contents
1 Introduction 3
1.1 What is an Option? . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Fundamental Relations . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Discrete-time State Pricing with Finite State Space 6
2.1 Single-period Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.1 Description of the Model . . . . . . . . . . . . . . . . . . . . 6
2.1.2 Arrow-Debreu Securities . . . . . . . . . . . . . . . . . . . . . 7
2.1.3 Deator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.4 Risk-neutral Measure . . . . . . . . . . . . . . . . . . . . . . 10
2.1.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Multi-period Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.1 Modeling Information Flow . . . . . . . . . . . . . . . . . . . 14
2.2.2 Description of the Multi-period Model . . . . . . . . . . . . . 16
2.2.3 Some Basic Denitions Revisited . . . . . . . . . . . . . . . . 19
2.2.4 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.5 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3 Introduction to Stochastic Calculus 23
3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.1 Why Stochastic Calculus? . . . . . . . . . . . . . . . . . . . . 23
3.1.2 What is a Reasonable Model for Stock Dynamics? . . . . . . 23
3.2 On Stochastic Processes . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2.1 Cadlag and Caglad . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2.2 Modication, Indistinguishability, Measurability . . . . . . . 25
3.2.3 Stopping Times . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Prominent Examples of Stochastic Processes . . . . . . . . . . . . . . 29
3.4 Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.5 Completition, Augmentation, Usual Conditions . . . . . . . . . . . . 34
3.6 Ito-Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.7 Itos Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4 Continuous-time Market Model and Option Pricing 47
CONTENTS 2
4.1 Description of the Model . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2 Black-Scholes Approach: Pricing by Replication . . . . . . . . . . . . 49
4.3 Pricing with Deators . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.4 Pricing with Risk-neutral Measures . . . . . . . . . . . . . . . . . . . 57
5 Continuous-time Portfolio Optimization 68
5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.2 Martingale Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.3 Stochastic Optimal Control Approach . . . . . . . . . . . . . . . . . 73
5.3.1 Heuristic Derivation of the HJB . . . . . . . . . . . . . . . . . 73
5.3.2 Solving the HJB . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.3.3 Verication . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.4 Intermediate Consumption . . . . . . . . . . . . . . . . . . . . . . . . 76
5.5 Innite Horizon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
References 86
1 INTRODUCTION 3
1 Introduction
1.1 What is an Option?
There exist two main variants: call and put.
A European call gives the owner the right to buy some asset (e.g. stock, bond,
electricity, corn . . . ) at a future date (maturity) for a xed price (strike price).
Payo: maxS(T) K; 0, where S(T) is the asset price at maturity T and
K is the strike price.
An American call gives the owner the right to buy some asset during the whole
life-time of the call.
Put: replace buy with sell
Since the owner has the right but not the obligation, this right has a positive
value.
1.2 Fundamental Relations
Consider an option written on an underlying which pays no dividends etc. during
the life-time of the option. In the following we assume that the underlying is a stock.
The time-t call and put prices are denoted by Call(t) and Put(t). The time-t price
of a zero-coupon bond with maturity T is denoted by p(t, T).
Proposition 1.1 (Bounds)
(i) For a European call we get: S(t) Call(t) maxS(t) Kp(t, T); 0
(ii) For a European put we have: Kp(t, T) Put(t) maxKp(t, T) S(t); 0
Proof. (i) The relation S(t) Call(t) is obvious because the stock can be in-
terpreted as a call with strike 0. We now assume that Call(t) < maxS(t)
Kp(t, T); 0. W.l.o.g. S(t) Kp(t, T) 0. Consider the strategy
buy one call,
sell one stock,
buy K zeros with maturity T,
leading to an initial inow of S(t)Call(t)Kp(t, T) > 0. At time T, we distinguish
two scenarios:
1. S(T) > K: Then the value of the strategy is S(T) +S(T) K +K = 0.
2. S(T) K: Then the value of the strategy is S(T) + 0 +K 0.
Hence, the strategy would be an arbitrage opportunity.
(ii) The relation Kp(t, T) Put(t) is valid because otherwise the strategy
sell one put,
buy K zeros with maturity T
1 INTRODUCTION 4
would be an arbitrage strategy. The second relation can be proved similarly to the
relation for the call in (i). 2
Proposition 1.2 (Put-Call Parity)
The following relation holds: Put(t) = Call(t) S(t) +Kp(t, T)
Proof. Consider the strategy
buy one put,
sell one call,
buy one stock,
sell K zeros with maturity T.
It is easy to check that the time-T value of this strategy is zero. Since no additional
payments need to be made, no arbitrage dictates that the time-t value is zero as
well implying the desired result. 2
Remark. This is a very important relationship because it ties together European
call and put prices.
Proposition 1.3 (American Call)
Assume that interest rates are positive. Early exercise of an American call is not
optimal, i.e. Call
am
(t) = Call(t).
Proof. The following relations hold
Call
am
(t) Call(t)
()
maxS(t) Kp(t, T); 0 S(t) K
where () holds due to Proposition 1.1 (i). Hence, the American call is worth more
alive than dead. 2
Remark.
Unfortunately, this result does not hold for American puts.
If dividends are paid, then the result for the American call breaks down as well.
Proposition 1.4 (American Put)
Assume that interest rates are positive. Then the following relations hold:
(i) Call
am
(t) S(t) +Kp(t, T) Put
am
(t) Call
am
(t) S(t) +K
(ii) Put(t) Put
am
(t) Put(t) +K(1 p(t, T)).
Proof. (i) The rst inequality follows from the put-call parity and Proposition 1.3.
For the second one assume that Put
am
(t) > Call
am
(t) S(t) +K and consider the
following strategy:
sell one put,
buy one call,
sell one stock,
invest K Euros in the money market account.
1
1
The time-t value of the money market account is denoted by M(t). By convention, M(0) = 1.
1 INTRODUCTION 5
By assumption, the initial value of this strategy is strictly positive. For t [0, T]
we distinguish to scenarios:
1. S(t) > K: Then the value of the strategy is
0 + (S(t) K) S(t) +KM(t) = K(M(t) 1) 0
since interest rates are positive and thus M(t) 1.
2. S(t) K: Then the value of the strategy is
(K S(t)) + 0 S(t) +KM(t) = K(M(t) 1) 0.
Hence, the strategy would be an arbitrage opportunity.
(ii) We have
Put(t) Put
am
(t) Call
am
(t) S(t) +K = Call(t) S(t) +K
= Put(t) +S(t) Kp(t, T) S(t) +K = Put(t) +K(1 p(t, T)).
2
Remark. If interest rates are zero, then Put(t) = Put
am
(t).
2 DISCRETE-TIME STATE PRICING WITH FINITE STATE SPACE 6
2 Discrete-time State Pricing with Finite State
Space
2.1 Single-period Model
2.1.1 Description of the Model
We consider a single-period model starting today (t = 0) and ending at time
t = T. Without loss of generality T = 1.
Uncertainty at time t = 1 is represented by a nite set =
1
, . . . ,
I
of
states, one of which will be revealed as true. is said to be the state space.
Each state
i
can occur with probability P(
i
) > 0. This denes a
probability measure P : (0, 1).
At time t = 0 one can trade in J + 1 assets. The time-0 prices of these
assets are described by the vector S
= (S
0
, S
1
, . . . , S
J
). The state-dependent
payos of the assets at time t = 1 are described by the payo matrix
X =
X
0
(
1
) X
J
(
1
)
X
0
(
2
) X
J
(
2
)
.
.
.
.
.
.
.
.
.
X
0
(
I
) X
J
(
I
)
.
To shorten notations, X
ij
:= X
j
(
i
).
The rst asset is assumed to be riskless, i.e. X
i0
= X
k0
= const. for all i,
k 1, . . . , I.
Therefore, at time t = 0 there exists a riskless interest rate given by
r =
X
0
S
0
1.
An investor buys and sells assets at time t = 0. This is modeled via a trading
strategy = (
0
, . . . ,
J
)
, where
j
denotes the number of asset j which the
investor holds in his portfolio (combination of assets).
Denition 2.1 (Arbitrage)
We distinguish two kinds of arbitrage opportunities:
(i) A free lunch is a trading strategy with
S
< 0 and X 0.
(ii) A free lottery is a trading strategy with
S
= 0 and X 0 and X = 0.
Assumptions:
(a) All investors agree upon which states cannot occur at time t = 1.
(b) All investors estimate the same payo matrix (homogeneous beliefs).
(c) At time t = 0 there is only one market price for each security and each investor
can buy and sell as many securities as desired without changing the market
price (price taker).
(d) Frictionless markets, i.e.
assets are divisible,
no restrictions on short sales,
2
no transaction costs,
no taxes.
(e) The asset market contains no arbitrage opportunities, i.e. there are neither free
lunches nor free lotteries.
2.1.2 Arrow-Debreu Securities
To characterize arbitrage-free markets, the notion of an Arrow-Debreu security is
crucial:
Denition 2.2 (Arrow-Debreu Security)
An asset paying one Euro in exactly one state and zero else is said to be an Arrow-
Debreu security (ADS). Its price is an Arrow-Debreu price (ADP).
Remarks.
In our model, we have as many ADS as states at time t = 1, i.e. we have I
Arrow-Debreu securities.
The i-th ADS has the payo e
i
= (0, . . . , 0, 1, 0, . . . , 0)
, where the i-th entry is

1. The corresponding Arrow-Debreu price is denoted by
i
.
Unfortunately, in practice ADSs are not traded.
We can now prove the following theorem characterizing markets which do not con-
tain arbitrage opportunities.
Theorem 2.1 (Characterization of No Arbitrage)
Under assumptions (a)-(d) the following statements are equivalent:
The market does not contain arbitrage opportunities (Assumption (e)).
2
A short sale of an assets is equivalent to selling an asset one does not own. An investor can do
this by borrowing the asset from a third party and selling the borrowed security. The net position
is a cash inow equal to the price of the security and a liability (borrowing) of the security to
the third party. In abstract mathematical terms, a short sale corresponds to buying a negative
amount of a security.
There exist strictly positive Arrow-Debreu prices, i.e.
i
> 0, i = 1, . . . , I, such
that
S
j
=
I
i=1
i
X
ij
.
Proof. Suppose (ii) to hold. We consider a strategy with X 0 and X = 0.
Then
S
(ii)
=
X
()
> 0,
where () holds because > 0, X 0, and X = 0. For this reason, there are
no free lunches. The argument to exclude free lotteries is similar. The implication
(i)(ii) follows from the so-called Farkas-Stiemke lemma (See Mangasarian (1969,
p. 32) or Dax (1997)). 2
Remarks.
The ADPs are the solution of the following system of linear equations:
S = X
.
The hard part of the proof is actually the proof of implication (i)(ii). For
brevity, this proof is omitted here.
Given our market with prices S and payo matrix X, it is now of interest how to
price other assets.
Denition 2.3 (Contingent Claim)
In the single-period model, a contingent claim (derivative, option) is some additional
asset with payo C IR
I
.
Denition 2.4 (Replication)
Consider a market with payo matrix X. A contingent claim with payo C can be
replicated (hedged) in this market if there exists a trading strategy such that
C = X.
Remark. If an asset with payo C is replicable via a trading strategy , by the
no arbitrage assumption its price S
C
has to be (law of one price)
S
C
=
J
j=1
j
S
j
.
Denition 2.5 (Completeness)
An arbitrage-free market is said to be complete if every contingent claim is replica-
ble.
Theorem 2.2 (Characterization of a Complete Market)
Assume that the nancial market contains no arbitrage opportunities. Then the
market is complete if and only if the Arrow-Debreu prices are uniquely determined,
i.e.
Rank(X) = I.
Proof.
is unique.
X
= S has a unique solution.

Rank (X) = I
Range (X) = IR
I
Every contingent claim is replicable.
2
Remarks.
The market is complete if the number of assets with linearly independent payos
is equal to the number of states.
If the number of assets is smaller than the number of states (J < I), the market
cannot be complete.
2.1.3 Deator
Main result from the previous subsection: In an arbitrage-free market, there
exist strictly positive ADPs such that
S
j
=
I
i=1
i
X
ij
.
In this subsection: Using the physical probabilities P(
i
), i = 1, . . . , I, arbitrage-
free asset prices can be represented as discounted expected values of the correspond-
ing payos.
Denition 2.6 (Deator)
In an arbitrage-free market, a deator H is dened by
H =

P
.
Remarks.
Since the market is arbitrage-free, there exists at least one vector of ADSs , but
and thus the deator H are only unique if the market is complete.
The ADS and the probability measure P can be interpreted as random variables
with realizations
i
= (
i
) and P(
i
).
Proposition 2.3 (Pricing with a Deator)
In an arbitrage-free market, the asset prices are given by
S
j
= E
P
[H X
j
].
Proof.
S
j
=
I
i=1
i
X
ij
=
I
i=1
P(
i
)
(
i
)
P(
i
)
X
j
(
i
) = E
P
P
X
j
= E
P
[H X
j
]
2
Remarks.
The deator H is the appropriate discount factor for payos at time t = 1.
The deator H ist a random variable with realizations (
i
)/P(
i
).
On average, one needs to discount with the riskless rate r because
E
P
[H] =
I
i=1
P(
i
)
(
i
)
P(
i
)
=
I
i=1
(
i
) =
1
1 +r
.
Interpretation of a Deator:
In an equilibrium model one can show that the deator equals the marginal
rate of substitution between consumption at time t = 0 and t = 1 of a repre-
sentative investor, i.e. (roughly speaking) we have
H =
u/c
1
u/c
0
,
where u = u(c
0
, c
1
) is the utility function of the representative investor.
Setting u
0
:= u/c
0
and u
1
:= u/c
1
leads to
r =
1
E
P
[H]
1 =
u
0
E
P
(u
1
)
1,
which gives the following result:
(i) u
0
> E(u
1
) (time preference):
3
r > 0,
(ii) u
0
< E(u
1
): r < 0.
2.1.4 Risk-neutral Measure
Two important observations:
(i) If an investor buys all ADSs, he gets one Euro at time t = 1 for sure, i.e. we get
I
i=1
i
=
1
1 +r

I
i=1
i
(1 +r) = 1.
(ii) In an arbitrage-free market, it holds that
i
> 0, i = 1, . . . , I.
3
Intuitively, time preference describes the preference for present consumption over future
consumption.
Proposition 2.4 (Risk-neutral Measure)
In an arbitrage-free market, there exists a probability measure Q dened via
Q(
i
) = q
i
:=
i
(1 +r)
This measure is said to be the risk-neutral measure or equivalent martingale mea-
sure.
Proof. follows from (i) and (ii). 2
Using the risk-neutral measure, there is a third representation of an asset price:
Proposition 2.5 (Risk-neutral Pricing)
In an arbitrage-free market, asset prices possess the following representation
S
j
= E
Q
X
j
R
,
where R := 1 +r.
Proof.
S
j
=
I
i=1
i
X
ij
=
I
i=1
i
R
X
ij
R
=
I
i=1
q
i
X
ij
R
= E
Q
X
j
R
.
2
Remarks.
Warning: The risk-neutral measure Q is dierent from the physical measure P.
Roughly speaking, risk-neutral probabilities are compounded Arrow-Debreu se-
curities.
The notion risk-neutral measure comes from the fact that under this measure
prices are discounted expected values where one needs to discount using the
riskless interest rate.
To motivate the notion equivalent martingale measure, let us recall some basic
denitions:
Denition 2.7 (Equivalent Probability Measures)
A Probability Measure Q is said to be continuous w.r.t. another probability measure
P if for any event A
P(A) = 0 =Q(A) = 0.
If Q is continuous w.r.t. P and vice versa, then P and Q are said to be equivalent,
i.e. both measures possess the same null sets.
Theorem 2.6 (Radon-Nikodym)
If Q is continuous w.r.t. P, then Q possesses a density w.r.t. Q, i.e. there exists a
positive random variable D =
dQ
dP
such that
E
Q
[Y ] = E
P
[D Y ]
for any random variable Y . This density is uniquely determined (P-a.s.).
Denition 2.8 (Stochastic Process in a Single-period Model)
Consider a single-period model with time points t = 0, 1. The nite sequence
Y (t)
t=0,1
forms a stochastic process.
Denition 2.9 (Martingale in a Single-period Model)
Consider a single-period model with nite state space. Uncertainty at time t = 1
is modeled by some probability measure Q. A real-valued stochastic process is said
to be a martingale w.r.t. Q (short: Q-martingale) if
E
Q
[Y (1)] = Y (0),
where it is assumed that E
Q
[[Y (1)[] < .
We can now characterize risk-neutral measures as follows:
Proposition 2.7 (Properties of a Risk-neutral Measure)
(i) The discounted price processes S
j
, X
j
/R are martingales under a risk-neutral
measure Q.
(ii) The physical measure P and a risk-neutral measure Q are equivalent.
Proof. (i) follows from Proposition 2.5.
(ii) Since ADPs are strictly positive, we have Q(
i
) > 0 and P(
i
) > 0 which gives
the desired result. 2
Remarks.
Proposition 2.7 is the reason why Q is said to be an equivalent martingale
measure.
From Propositions 2.3 and 2.5, we get
E
Q
X
j
R
= S
j
= E
P
[H X
j
].
Therefore, we obtain
E
Q
[X
j
] = E
P
[ H R
. .. .
=dQ/dP
X
j
],
i.e. H = (dQ/dP)/R. Analogously, one can show dP/dQ = 1/(H R).
2.1.5 Summary
Theorem 2.8 (No Arbitrage)
Consider a nancial market with price vector S and payo matrix X. If the as-
sumptions (a)-(d) are satised, then the following statements are equivalent:
(i) The nancial market contains no arbitrage opportunities (assumption (e)).
(ii) There exist strictly positive Arrow-Debreu prices
i
, i = 1, . . . , I, such that
S
j
=
I
i=1
i
X
ij
.
(iii) There exists a state-dependent discount factor H, the so-called deator, such
that for all j = 0, . . . , J
S
j
= E
P
[H X
j
].
(iv) There exists an equivalent martingale measure Q such that the discounted
price processes S
j
, X
j
/R are martingales under Q, i.e. for all j = 0, . . . , J
S
j
= E
Q
X
j
R
Theorem 2.9 (Completeness)

Consider a nancial market with price vector S and payo matrix X. If the as-
sumptions (a)-(e) are satised, then the following statements are equivalent:
(i) The nancial market is complete, i.e. each additional asset can be replicated.
(ii) The Arrow-Debreu prices
i
, i = 1, . . . , I, are uniquely determined.
(iii) There exists only one deator H.
(iv) There exists only one equivalent martingale Q.
Remarks.
In a complete market, the price of any additional asset is given by the price of
its replicating portfolio.
The value of this portfolio can be calculated applying the following theorem.
Theorem 2.10 (Pricing of Contingent Claims)
Consider an arbitrage-free and complete market with price vector S and payo
matrix X. Let C IR
I
be the payo of a contingent claim. Its todays price S
C
can be calculated by using one of the following formulas:
(i) S
C
= C
=
I
i=1
i
C(
i
) (Pricing with ADSs).
(ii) S
C
= E
P
[H C] (Pricing with the deator).
(iii) S
C
= E
Q
C
R
(Pricing with the equivalent martingale measure).

Proof. In an arbitrage-free and complete market, there exist ADPs which are
uniquely determined. The no-arbitrage requirement together with the denition of
an ADP gives formula (i). Formulas (ii) and (iii) follow from (i) using H = /P
and Q(
i
) =
i
R. 2
2.2 Multi-period Model
In the previous section we have modeled only two points in time:
t = 0: The investor chooses the trading strategy.
t = T: The investor liquidates his portfolio and gets the payos.
This has the following consequences:
We have disregarded the information ow between 0 and T.
Every trading strategy was a static strategy (buy and hold), i.e. intermediate
portfolio adjustments was not allowed.
In this section
we will model the information ow between 0 and T.
Filtration
and the investor can adjust his portfolio according to the arrival of new informa-
tion about prices.
dynamic (self-nancing) trading strategy.
2.2.1 Modeling Information Flow
Motivation
Consider the dynamics of a stock in a two-period binomial model:
121
110
100
90
81
99
1
P
P
P
P
P
PPq
1
P
P
P
P
P
PPq
1
P
P
P
P
P
PPq
t = 0 t = 0.5 t = 1
Due to the additional trading date at time t = 0.5, the investor gets more infor-
mation about the stock price at time t = 1.
At t = 0 it is possible that the stock price at time t = 1 equals 81, 99 or 121.
If, however, the investor already knows that at time t = 0.5 the stock price equals
110 (90), then at time t = 1 the stock price can only take the values 121 or 99
(99 or 81).
It is one of the goals of this section to model this additional piece of information!
Reference Example
To demonstrate how one can model the information ow, we consider the following
example:
1
,
2
1
,
2
,
3
,
4
3
,
4
2
,
3
1
P
P
P
P
P
PPq
1
P
P
P
P
P
PPq
1
P
P
P
P
P
PPq
t = 0 t = 1 t = 2
The state space thus equals: =
1
,
2
,
3
,
4
.
Example for the interpretation of the states.
1
: DAX stands at 5000 at t = 1 and at 6000 at time t = 2.
2
3
4
First Approach to Model the Flow of Information
Denition 2.10 (Partition)
A partition of the state space is a set A
1
, . . . , A
K
of subsets A
k
of a state
space such that
the subsets A
k
are disjoint, i.e. A
k
A
l
= ,
= A
1
. . . A
K
.
Example.
1
,
2
,
3
,
4
.
Denition 2.11 (Information Structure)
An information structure is a sequence F
t
t
of partitions F
0
, . . ., F
T
such that
F
0
=
1
, . . . ,
I
( At the beginning each state can occur.),
F
T
=
1
, . . . ,
I
( At the end, one knows for sure which
state has occurred. ),
the partitions become ner over time.
Example.
F
0
=
1
,
2
,
3
,
4
.
F
1
=
1
,
2
,
3
,
4
.
F
2
=
1
,
2
,
3
,
4
.
Remarks.
Problem: This intuitive denition cannot be generalized to the continuous-time
case.
However, it provides a good intuition of what information ow actually means.
Second Approach to Model the Flow of Information: Filtration
Denition 2.12 (-Field)
A system T of subsets of the state space is said to be a -eld if
(i) T,
(ii) A T A
C
T,
(iii) for every sequence A
n
nIN
with A
n
T we have
n=1
A
n
T.
Remark. For a nite -eld, it is sucient if nite (instead of innite) unions of
events belong to the -eld (see requirement (iii) of Denition 2.12).
Example. ,
1
,
2
,
3
,
4
, .
Denition 2.13 (Filtration)
Consider a state space T. A family T
t
t
of sub--elds T
t
, t 1, is said to be a
ltration if T
s
T
t
, s t and s, t 1. Here 1 denotes some ordered index set.
Remark. In our discrete time model, we can always use ltrations with the fol-
lowing properties ({() denotes the power-set of ):
T
0
= , (At the beginning, one has no information.),
T
T
= {() (At the end one is fully informed.),
T
0
T
1
. . . T
T
(The state of information increases over time.).
Example. T
0
= , , T
1
= ,
1
,
2
,
3
,
4
, , T
2
= {().
Some facts about ltrations:
Filtrations are used to model the ow of information over time.
At time t, one can decide if the event A T
t
has occurred or not.
Taking conditional expectation E[Y [T
t
] of a random variable Y means that one
takes expectation on the basis of all information available at time t.
2.2.2 Description of the Multi-period Model
Properties of the Model:
Trading takes place at T + 1 dates, t = 0, . . . , T.
Investors can trade in J + 1 assets with price processes (S
0
(t), . . . , S
J
(t)), t =
0, . . . , T.
S
j
(t) denotes the price of the j-th asset at time t.
The range of S
j
(t) is nite.
Trading of an investor is modeled via his trading strategy (t) = (
0
(t), . . . ,
J
(t))
,
where
j
(t) denotes the number of asset j at time t in the investors portfolio.
A trading strategy needs to be predictable, i.e. on the basis of all information
about prices S(t 1) at time t 1 the investor selects his portfolio which he holds
until time t.
The value of a trading strategy at time t is given by:
V
(t) =
S(t)
(t) =

J
j=0
S
j
(t)
j
(t) f ur t = 1, . . . , T
S(0)
(1) =

J
j=0
S
j
(0)
j
(1) f ur t = 0
Denition 2.14 (Contingent Claim)
A contingent claim leads to a state-dependent payo C(t) at time t, which is mea-
surable w.r.t. T
t
.
Remarks.
A typical contingent claim is a European call with payo C(T) = maxS(T)
K; 0, where S(T) denotes the value of the underlying at time T and K the strike
price.
American options are no contingent claims in the sense of Denition 2.14, but
one can generalize this denition accordingly.
Recall the following paradigm: Arbitrage-free pricing of a contingent claim
means nding a hedging (replication) strategy of the corresponding payo.
In the single-period model: The price of the replication strategy is then the price
of the contingent claim.
In the multi-period model we need an additional requirement: The replication
strategy should not require intermediate inows or outows of cash.
Only in this case, the initial costs are equal to the fair price of the contingent
claim.
This leads to the notion of a self-nancing trading strategy.
Denition 2.15 (Self-nancing Trading Strategy)
A trading strategy (t) = (
1
(t), . . . ,
J
(t)) is said to be self-nancing if for t =
1, . . . , T 1
S(t)
(t) = S(t)
(t + 1)
J
j=0
S
j
(t)
j
(t) =
J
j=0
S
j
(t)
j
(t + 1).
Remarks.
To buy assets, the investor has to sell other ones.
For this reason, the portfolio value changes only if prices change and not if the
numbers of assets in the portfolio change.
Notation:
Change of the value of strategy at time t:
V
(t) := V
(t) V
(t 1)
Price changes at time t
S(t) = (S
0
(t), . . . , S
J
(t)) := (S
0
(t) S
0
(t 1), . . . , S
J
(t) S
J
(t 1))
Gains process of the trading strategy at time t:
(t)
S(t) =
J
j=0
j
(t)S
j
(t)
Cumulative gains process of the trading strategy at time t:
G
(t) :=
t
=1
()
S() =
t
=1
J
j=0
j
()S
j
()
Proposition 2.11 (Characterization of a Self-nancing Trading Strategy)
A self-nancing trading strategy (t) has the following properties:
(i) The value of changes only due to price changes, i.e.
V
(t) = (t)
S(t).
(ii) The value of is equal to its initial value plus the cumulative gains, i.e.
V
(t) = V
(0) +G
(t) = V
(0) +
t
=1
()
S().
Proof. (i) For a self-nancing strategy , we obtain
V
(t) = V
(t) V
(t 1)
=
J
j=0
S
j
(t)
j
(t)
J
j=0
S
j
(t 1)
j
(t 1)
()
=
J
j=0
S
j
(t)
j
(t)
J
j=0
S
j
(t 1)
j
(t)
=
J
j=0
j
(t)
S
j
(t) S
j
(t 1)
=
J
j=0
j
(t)S
j
(t)
= (t)
S(t).
The equality () is valid because is self-nancing.
(ii) Any (not necessary self-nancing) trading strategy satises
V
(t) = V
(0) +
t
=1
V
().
Since is assumed to be self-nancing, by (i), we obtain
V
(t) = V
(0) +
t
=1
V
().
(i)
= V
(0) +
t
=1
()
S()
= V
(0) +G
(t).
2
Remarks.
In continuous-time models, the relation (i) is still valid if is replace by d.
Unfortunately, the intuitive relation
S(t)
(t) = S(t)
(t + 1)
has no continuous-time equivalent.
2.2.3 Some Basic Denitions Revisited
The notion of a self-nancing trading strategy is also important to generalize the
following denitions to the multi-period setting:
Denition 2.16 (Arbitrage)
We distinguish two kinds of arbitrage opportunities:
(i) A free lunch is a self-nancing trading strategy with
V
(0) < 0 und V
(T) 0.
(ii) A free lottery is a self-nancing trading strategy with
V
(0) = 0 und V
(T) 0 und V
(T) = 0.
Denition 2.17 (Replication of a Contingent Claim)
A contingent claim with payo C(T) is said to be replicable if there exists a self-
nancing trading strategy such that
C(T) = V
(T).
The trading strategy is said to be a replication (hedging) strategy for the claim
C(T).
Remarks.
By denition, a replication strategy of a claim leads to the same payo as the
claim.
In general, replication strategies are dynamic strategies, i.e. at every trading
date the hedge portfolio needs to be adjusted according to the arrival of new
information about prices.
The denition of a complete market remains the same:
Denition 2.18 (Completeness)
An arbitrage-free market is said to be complete if each contingent claim can be
replicated.
2.2.4 Assumptions
(a) All investors agree upon which states cannot occur at times t > 0.
(b) All investors agree upon the range of the price processes S(t).
(c) At time t = 0 there is only one market price for each security and each investor
can buy and sell as many securities as desired without changing the market
price (price taker).
(d) Frictionless markets, i.e.
assets are divisible,
no restrictions on short sales,
4
no transaction costs,
no taxes.
(e) The asset market contains no arbitrage opportunities, i.e. there are neither free
lunches nor free lotteries.
2.2.5 Main Results
Conventions:
The 0-th asset is a money market account, i.e. it is locally risk-free with
S
0
(0) = 1 und S
0
(t) =
t
u=1
(1 +r
u
), t = 1, . . . , T,
where r
u
denotes the interest rate for the period [ 1, ].
The discounted (normalized) price processes are dened by
Z
j
(t) :=
S
j
(t)
S
0
(t)
.
The money market account is thus used as a numeraire.
Denition 2.19 (Martingal)
A real-valued adapted process Y (t)
t
, t = 0, . . . , T, with E[[Y (t)[] < is a mar-
tingale w.r.t. a probability measure Q if for s t
E
Q
[Y (t)[T
s
] = Y (s).
Remark. A martingale is thus a process being on average stationary.
Important Facts:
4
A short sale of an assets is equivalent to selling an asset one does not own. An investor can do
this by borrowing the asset from a third party and selling the borrowed security. The net position
is a cash inow equal to the price of the security and a liability (borrowing) of the security to
the third party. In abstract mathematical terms, a short sale corresponds to buying a negative
amount of a security.
A multi-period model consists of a sequence of single-period models.
Therefore, a multi-period model is arbitrage-free if all single-period models are
arbitrage-free (See, e.g., Bingham/Kiesel (2004), Chapter 4).
For these reasons, the following three theorems hold:
Theorem 2.12 (Characterization of No Arbitrage)
If assumptions (a)-(d) are satised, the following statements are equivalent:
(i) The nancial market contains no arbitrage opportunities (assumption (e)).
(ii) There exist a state-dependent discount factor H, the so-called deator, such
that for all time points s, t 0, . . . , T with s t and all assets j = 0, . . . , J
we have
S
j
(s) =
E
P
[H(t)S
j
(t)[T
s
]
H(s)
,
i.e. all prices can be represented as discounted expectations under the physical
measure.
(iii) There exists a risk-neutral probability measure Q such that all discounted
price processes Z
j
(t) are martingales under Q, i.e. for all time points s, t
0, . . . , T with s t and all assets j = 1, . . . , J we have
Z
j
(s) = E
Q
[Z
j
(t)[T
s
]
S
j
(s)
S
0
(s)
= E
Q
S
j
(t)
S
0
(t)
T
s
Remarks.
The equivalence No Arbitrage Existence of Q is known as Fundamental
Theorem of Asset Pricing.
Unfortunately, in continuous-time models this equivalence breaks down.
However, the implication Existence of Q No Arbitrage is still valid,
Theorem 2.13 (Pricing of Contingent Claims)
Consider a nancial market satisfying assumptions (a)-(e) and a contingent claim
with payo C(T). Then we obtain for s, t 0, . . . , T with s t:
(i) The replication strategy for the claim is unique, i.e. the arbitrage-free price
of the claim is uniquely determined by
C(s) = V
(s).
(ii) Using the deator, we can calculate the claims time-s price
C(s) =
E
P
[H(t)C(t)[T
s
]
H(s)
.
(iii) The following risk-neutral pricing formula holds:
C(s) = S
0
(s) E
Q
C(t)
S
0
(t)
T
s
.
Remarks.
Result (i) says that the todays price of a claim is uniquely determined by its
replication costs, i.e. C(0) = V
(0).
From (iii) we get
C(s)
S
0
(s)
= E
Q
C(t)
S
0
(t)
T
s
,
i.e. the discounted price processes of contingent claims are Q-martingales as well.
Note that one needs to discount with the money market account.
If interest rates are constant, then from (iii)
C(0) = E
Q
C(T)
(1 +r)
T
=
E
Q
[C(T)]
(1 +r)
T
.
Pricing of contingent claims boils down to computing expectations under the
risk-neutral measure.
Theorem 2.14 (Completeness)
Consider a nancial market satisfying assumptions (a)-(e). Then the following
statements are equivalent
(i) The market is complete.
(ii) There exists a unique deator H.
(iii) There exists a unique martingale measure Q.
2.2.6 Summary
The following results are important:
It is always valid that a nancial market contains no arbitrage opportunities if
an equivalent martingale measure exists.
In our discrete model, the converse is also valid: If the market contains no arbi-
trage opportunities, there exists a martingale measure.
If a martingale measure Qexists, the discounted price processes are Q-martingales.
If a unique martingale measure exists, the market contains is arbitrage-free and
complete.
If a unique martingale measure exists, the todays price of a contingent claim is
given by the following Q-expectation
C(0) = E
Q
C(T)
S
0
(T)
.
3 INTRODUCTION TO STOCHASTIC CALCULUS 23
3 Introduction to Stochastic Calculus
3.1 Motivation
3.1.1 Why Stochastic Calculus?
In the previous section, the investor was able to trade at nitely many dates
t = 0, . . . , T only.
In a continuous-time model, trading takes place continuously, i.e. at any date
in the time period [0, T].
Consequently, price dynamics are modeled continuously.
In continuous-time models asset dynamics are modeled via stochastic dier-
ential equations.
Example Black-Scholes Model. Two investment opportunities:
money market account: riskless asset with constant interest rate r modeled via
an ordinary dierential equation (ODE):
dM(t) = M(t)rdt, M(0) = 1.
stock: risky asset modeled via a stochastic dierential equation (SDE)
dS(t) = S(t)
dt +dW(t)
, S(0) = s
0
,
where W is a Brownian motion.
In this section we answer the following questions:
Q1 What is a Brownian motion?
Q2 What is the meaning of dW(t)? Stochastic Integral
Q3 What are stochastic dierential equations and do explicit solutions exist?
Variation of Constant
Q4 How does the dynamics of contingent claims look like? Ito Formula
3.1.2 What is a Reasonable Model for Stock Dynamics?
Model for the Money Market Account
The solution of the ODE B
(t) = r B(t) with B(0) = 1 is

B(t) = b
0
exp(rt)
and thus
ln(B(t)) = ln(b
0
) +r t.
Model for the Stock
The result for the money market account motivates the following model for the
stock dynamics:
ln(S(t)) = ln(s
0
) + t + randomness
Notation: Y (t) = randomness, i.e. Y (t) = ln(S(t)) ln(s
0
) t
Reasonable properties of Y :
1) no tendency, i.e. E[Y (t)] = 0,
2) increasing with time t, i.e. Var(Y (t)) increases with t,
3) no memory, i.e. for Y (t) = Y (s) +[Y (t) Y (s)] the dierence [Y (t) Y (s)]
depends only on t s and
is independent of Y (s),
Properties which would make life easier:
4) normally distributed, i.e. Y (t) ^(0,
2
t), > 0,
5) continuous.
3.2 On Stochastic Processes
Denition 3.1 (Stochastic Process)
A stochastic process X dened on (, T) is a collection of random variables X
t
)
tI
,
i.e. X
t
T.
Remark. As otherwise stated, our stochastic processes take values in the mea-
surable space (IR
n
, B(IR
n
)), i.e. the n-dimensional Euclidian space equipped with
-eld of Borel sets. Therefore, every X
t
is (T, B(IR
n
))-measurable.
3.2.1 Cadlag and Caglad
Denition 3.2 (Cadlag and Caglad Functions)
Let f : [0, T] IR
n
be a function with existing left and right limits, i.e. for each
t [0, T] the limits
f(t) = lim
st
f(s), f(t+) = lim
st
f(s)
exist. The function f is cadlag if f(t) = f(t+) and caglad if f(t) = f(t).
A cadlag function cannot jump around too widly as the next proposition shows.
Proposition 3.1
(i) A cadlag function f can have at most a countable number of discontinuities, i.e.
t [0, T] : f(t) = f(t) is countable.
(ii) For any > 0, the number of discontinuities on [0, T] larger than is nite.
Proof. See Fristedt/Gray (1997).
3.2.2 Modication, Indistinguishability, Measurability
Recall the following denition:
Denition 3.3 (Filtration)
Consider a -eld T. A family T
t
of sub--elds T
t
, t 1, is said to be a
ltration if T
s
T
t
, s t and s, t 1. Here 1 denotes some ordered index set.
Denition 3.4 (Adapted Process)
A process X is said to be adapted to the ltration T
t
tI
if every X
t
is measurable
w.r.t. T
t
, i.e. X
t
T
t
.
Remarks.
The notation X
t
, T
t
tI
indicates that the process X is adapted to T
t
tI
.
For xed , the realization X
t
()
t
is said to be a path of X. For this reason,
a stochastic process can be interpreted as function-valued random variable.
Denition 3.5 (Modication, Indistinguishability)
Let X and Y be stochastic processes.
(i) Y is a modication of X if
P( : X
t
() = Y
t
()) = 1 for all t 0.
(ii) X and Y are indistinguishable if
P( : X
t
() = Y
t
() for all t 0) = 1,
i.e. X and Y have the same paths P-a.s.
Example. Consider a positive random variable with a continuous distribution.
Set X
t
0 and Y
t
= 1
{t=}
. Then Y is a modication of X, but P(X
t
=
Y
t
for all t 0) = 0.
Proposition 3.2 (Modication, Indistinguishability)
Let X and Y be stochastic processes.
(i) If X and Y are indistinguishable, then Y is a modication of X and vice versa.
(ii) If X and Y have right-continuous paths and Y is a modication of X, then X
and Y are indistinguishable.
(iii) Let X be adapted to T
t
. If Y is a modication of X and T
0
contains all
P-null sets of T, then Y is adapted to T
t
t
.
Proof. (i) is obvious.
(ii) Since Y is a modication of X, we have
P([X
t
() = Y
t
() for all t [0, ) Q) = 1.
By the right-continuity assumption, the claim follows.
(iii) Exercise. 2
If X is an (adapted) stochastic process, then every X
t
is (T, B(IR
n
))-measurable
((T
t
, B(IR
n
))-measurable). However, X is actually a function-valued random vari-
able, and so, for technical reasons, it is often convenient to have some joint measur-
ability properties.
Denition 3.6 (Measurable Stochastic Process)
A stochastic process is said to measurable, if for each A B(IR
n
) the set (t, ) :
X
t
() A belongs to the product -eld B(IR
n
) T, i.e. if the mapping
(t, ) X
t
() : ([0, ) , B([0, )) T) (IR
n
, B(IR
n
))
is measurable.
Denition 3.7 (Progressively Measurable Stochastic Process)
A stochastic process is said to progressively measurable w.r.t. the ltration T
t
, if
for each t 0 and A B(IR
n
) the set (s, ) [0, t] : X
s
() A belongs to
the product -eld B([0, t]) T
t
, i.e. if the mapping
(s, ) X
s
() : ([0, t] , B([0, t]) T
t
) (IR
n
, B(IR
n
))
is measurable for each t 0.
Remark. Any progressively measurable process is measurable and adapted (Exer-
cise). The following proposition provides the extent to which the converse is true.
Proposition 3.3
If a stochastic process is measurable and adapted to the ltration T
t
, then it has
a modication which is progressively measurable w.r.t. T
t
.
Proof. See Karatzas/Shreve (1991, p. 5).
Fortunately, if a stochastic process is left- or right-continuous, a stronger result can
be proved:
Proposition 3.4
If the stochastic process X is adapted to the ltration T
t
and right- or left-
continuous, then X is also progressively measurable w.r.t. T
t
.
Proof. See Karatzas/Shreve (1991, p. 5).
Remark. If X is right- or left-continuous, but not necessarily adapted to T
t
,
then X is at least measurable.
3.2.3 Stopping Times
Consider a measurable space (, T). An T- measurable random variable :
1 where 1 = [0, ) or 1 = IN is said to be a random time. Two special
classes of random times are of particular importance.
Denition 3.8 (Stopping Time, Optional Time)
(i) A random time is a stopping time w.r.t. the ltration T
t
tI
if the event
t T
t
for every t 1.
(ii) A random time is an optional time w.r.t. the ltration T
t
tI
if the event
< t T
t
for every t 1.
Propositions 3.5 and 3.6 show how both concepts are connected.
Proposition 3.5
If is optional and c > 0 is a positive constant, then +c is a stopping time.
Proof. Exercise.
Denition 3.9 (Continuity of Filtrations)
(i) A ltration (
t
is said to be right-continuous if
(
t
= (
t+
:=
>0
(
t+
(ii) A ltration (
t
is said to be left-continuous if
(
t
= (
t
:=
s<t
(
s
Remark. Let X be stochastic process.

Loosely speaking, left-continuity of F
X
t
means that X
t
can be guessed by ob-
serving X
s
, 0 s < t.
Right-continuity means intuitively that if X
s
has been observed for 0 s t,
then one cannot learn more by peeking innitesimally far into the future.
Proposition 3.6
Every stopping time is optional, and both concepts coincide if the ltration is right-
continuous.
Proof. Consider a stopping time . Then < t =

t>>0
T t and
T t T
t
T
t
, i.e. is optional. On the other hand, consider an optional
time w.r.t. a right-continuous ltration T
t
. Since t =

t+>u>t
T < u,
we have t T
t+
for every > 0 implying t =

>0
T
t+
= T
t+
= T
t
.
2
Remark. In our applications, we will always work with right-continuous ltrations.
5
5
See Denition 3.18.
Proposition 3.7
(i) Let
1
and
2
be stopping times. Then
1
2
= min
1
,
2
,
1
2
= max
1
,
2
,
1
+
2
, and
1
with 1 are stopping times.
(ii) Let
n
nIN
be a sequence of optional times. Then the random times
sup
n1
n
, inf
n1
n
, lim
n
n
, lim
n
n
are all optional. Furthermore, if the
n
s are stopping times, then so is sup
n1
n
.
Proof. Exercise.
Proposition and Denition 3.8 (Stopping Time -Field)
Let be a stopping time of the ltration T
t
. Then the set
A T : A t T
t
for all t 0
is a -eld, the so-called stopping time -eld T
.
Proof. Exercise.
Proposition 3.9 (Properties of a Stopped -Field)
Let
1
and
2
be stopping times.
(i) If
1
= t [0, ), then T
1
= T
t
.
(ii) If
1

2
, then T
1
T
2
.
(iii) T
2
= T
1
T
2
and the events
1
<
2
,
2
<
1
,
1

2
,
2

1
,
1
=
2
belong to T
1
T
2
.
Proof. Exercise.
Denition 3.10 (Stopped Process)
(i) If X is a stochastic process and is random time, we dene the function X
on
the event < by
X
() := X
()
()
If X
() is dened for all , then X
can also be dened on , by setting

X
() := X
() on = .
(ii) The process stopped at time is the process X
t
t[0,)
.
Proposition 3.10 (Measurability of a Stopped Process)
(i) If the process X is measurable and the random time is nite, then the function
X
is a random variable.
(ii) If the process X is progressively measurable w.r.t. T
t
t[0,)
and is a stopping
time of this ltration, then the random variable X
, dened on the set <

T
, is T
-measurable, and the stopped process X

t
, T
t
t[0,)
is progressively
measurable.
Proof. Karatzas/Shreve (1991, p. 5 and p. 9).
Remark. The following result is obvious: If X is adapted and cadlag and is a
stopping time, then
X
t
= X
t
1
{t<}
+X
1
{t}
.
is also adapted and cadlag.
Denition 3.11 (Hitting Time)
Let X be a stochastic process and let A B(IR
n
) be a Borel set in IR
n
. Dene
A
() = inft > 0 : X
t
() A.
Then is called a hitting time of A for X.
Proposition 3.11
Let X be a right-continuous process adapted to the ltration T
t
.
(i) If A
o
is an open set, then the hitting time of A
o
is an optional time.
(ii) If A
c
is a closed set, then the random variable

A
c
() = inft > 0 : X
t
() A
c
or X
t
() A
c
.
is a stopping time.
Proof. (i) For simplicity n = 1. We have
A
o
< t =

sQ[0,t)
X
s
A
o
. The
inclusion is obvious. Suppose that the inclusion does not hold. Then we have
X
i
A
o
for some i IR ` Q and i < t, but X
q
A
o
for all q Q with q < t.
Since A
o
is open, there exists a > 0 such that (X
i
, X
i
+ ) A
o
. Therefore,
X
q
(X
i
, X
i
+). However, then there exists a sequence (q
n
)
nIN
with q
n
`s
0
and q
n
Q [0, t), but X
q
n
X
i
which is a contradiction to the right-continuity
of X.
(ii) See Karatzas/Shreve (1991, p. 39) and Protter (2005, p. 5). 2
Remarks.
If the ltration is right-continuous, then the hitting time
A
o
is a stopping time.
If X is continuous , then the stopping time
A
c
is a hitting time of A
c
.
Even if X is continuous, in general the hitting time of A
o
is not a stopping time.
For example, suppose that X is real-valued, that for some , X
t
() 1 for t 1,
and that X
1
() = 1. Let A
o
= (1, ). Without looking slightly ahead of time 1,
we cannot tell wheter or not
A
o
= 1.
3.3 Prominent Examples of Stochastic Processes
Theorem and Denition 3.12 (Brownian Motion)
(i) A Brownian motion is an adapted, real-valued process W
t
, T
t
t[0,)
such that
W
0
= 0 a.s. and
W
t
W
s
^(0, t s) for 0 s < t stationary increments
W
t
W
s
independent of T
s
for 0 s < t independent increments
exists and is said to be a one-dimensional Brownian motion.
(ii) An n-dimensional Brownian motion W := (W
1
, . . . , W
n
) is an n-dimensional
vector of independent one-dimensional Brownian motions.
Proof. Applying Kolmogorovs extension theorem,
6
it is straightforward to prove
the existence of a Brownian motion. See, e.g., Karatzas/Shreve (1991, pp. 47). 2
Theorem 3.13 (Kolmogorovs Continuity Theorem)
Suppose that the process X = X
t
t[0,)
satises the following condition: For all
T > 0 there exist positive constants , , C such that
E[[X
t
X
s
[
] C [t s[
1+
, 0 s, t T.
Then there exists a continuous modication of X.
Proof. See Karatzas/Shreve (1991, pp.53).
Corollary 3.14 (Continuous Version of the Brownian Motion)
An n-dimensional Brownian motion possesses a continuous modication.
Proof. One can show that
E[[W
t
W
s
[
4
] = n(n + 2)[t s[
2
.
Hence, we can apply Kolmogorovs continuity theorem with = 4, C = n(n + 2),
and = 1. 2
Remark. We will always work with this continuous version.
Denition 3.12 (Poisson Process)
A Poisson process with intensity > 0 is an adapted, integer-valued cadlag process
N = N
t
, T
t
t[0,)
such that N
0
= 0, and for 0 t < , the dierence N
t
N
s
is
independent of T
s
and is Poisson distributed with mean (t s).
7
3.4 Martingales
Denition 3.13 (Martingale, Supermartingale, Submartingale)
A process X = X
t
t[0,)
adapted to the the ltration T
t
t[0,)
is called a
martingale (resp. supermartingale, submartingale) with respect to T
t
t[0,)
if
(i) X
t
L
1
(dP), i.e. E[[X
t
[] < .
6
Kolmogorovs extension theorem basically says that for given (nite-dimensional) distributions
satisfying a certain consistency condition there exists a probability space and a stochastic process
such that its (nite-dimensional) distributions coincide with the given distributions. See, e.g.,
Karatzas/Shreve (1991, p. 50).
7
A random variable Y taking values in IN {0} is Poisson distributed with parameter > 0 if
P(Y = n) = e
n
/(n!), n IN.
(ii) if s t, then E[X
t
[T
s
] = X
s
, a.s. (resp. E[X
t
[T
s
] X
s
, resp. E[X
t
[T
s
] X
s
).
Examples.
Let W be a one-dimensional Brownian motion. Then W is a martingale and W
2
is a submartingale.
The arithmetic Brownian motion X
t
:= t +W
t
, , IR, is a martingale for
= 0, a submartingale for > 0, and a supermartingale for < 0.
Let N be a Poisson process with intensity . Then N
t
t
t
is a martingale.
Denition 3.14 (Square-Integrable)
Let X = X
t
, T
t
t[0,)
be a right-continuous martingale. We say that X is square-
integrable if E[X
2
t
] < for every t 0. If, in addition, X
0
= 0 a.s. we write
X
2
(or X
c
2
if X is also continuous).
Proposition 3.15 (Functions of Submartingales)
(i) Let X = (X
1
, . . . , X
n
) with real-valued components X
k
be a martingale w.r.t.
T
t
t[0,)
and let : IR
n
IR be a convex function such that E[[(X
t
)[] <
for all t [0, ). Then (X) is a submartingale w.r.t. T
t
t[0,)
(ii) Let X = (X
1
, . . . , X
n
) with real-valued components X
k
be a submartingale w.r.t.
T
t
t[0,)
and let : IR
n
IR be a convex, non-decreasing function such that
E[[(X
t
)[] < for all t [0, ). Then (X) is a submartingale w.r.t. T
t
t[0,)
Proof. follows from Jensens inequality (Exercise). 2
For any X
2
and 0 t < , we dene
[[X[[
t
:=
E[X
2
t
] and [[X[[ :=
n=1
[[X[[
n
1
2
n
.
Note that [[XY [[ denes a the pseudo-metric on
2
. If we identify indistinguish-
able processes, the pseudo-metric [[X Y [[ becomes a metric on
2
.
8
Proposition 3.16
If we identify indistinguishable processes, the pair (
2
, [[) is a complete metric
space, and
c
2
is a closed subset of
2
.
Proof. See, e.g., Karatzas/Shreve (1991, pp. 37f).
An important question now arises: What happens to the martingale property of a
stopped martingale? For bounded stopping times, nothing changes.
Theorem 3.17 (Optional Sampling Theorem, 1st Version)
Let X = X
t
, T
t
t[0,)
be a right-continuous submartingale (supermartingale) and
let
1
and
2
be bounded stopping times of the ltration T
t
with
1

2
c IR
a.s. Then
E[X
2
[T
1
] ()X
1
8
This means: (i) ||X Y || 0 and ||X Y || = 0 i X = Y . (ii) ||X Y || = ||Y X||. (iii)
||X Y || = ||X Z|| +||Z Y ||.
If the stopping times are unbounded, in general, this result does only hold if X can
be extended to such that

X := X
t
, T
t
t[0,]
is still a sub- or supermartingale.
Theorem 3.18 (Sub- and Supermartingale Convergence)
Let X = X
t
, T
t
t[0,)
be a right-continuous submartingale (supermartingale)
and assume sup
t0
E[X
+
t
] < (sup
t0
E[X
t
] < ). Then the limit X
() =
lim
t
X
t
() exists for a.e. and E[X
[ < .
Proof. See Karatzas/Shreve (1991, pp. 17f).
This proposition ensures that a limit exists, but it is not clear at all if

X is a sub- or
supermartingale. To solve this problem, we need to require the following condition:
Denition 3.15 (Uniform Integrability)
A familiy of random variables (U)
aA
is uniformly integrable (UI) if
lim
n
sup
a
{|U
|n}
[U
[dP = 0.
Using the denition, it is sometimes not easy to check if a family of random variables
is uniformly integrable.
Theorem 3.19 (Characterization of UI)
Let (U)
aA
be a subset of L
1
. The following are equivalent:
(i) (U)
aA
is uniformly integrable.
(ii) There exists a positive, increasing, convex function g dened on [0, ) such that
lim
x
g(x)
x
= + and sup
aA
E[g [U
[] < .
Remark. A good choice is often g(x) = x
1+
with > 0.
Proposition 3.20
The following conditions are equivalent for a nonnegative, right-continuous sub-
martingale X = X
t
, T
t
t[0,)
:
(i) It is a uniformly integrable family of random variables.
(ii) It converges in L
1
, as t .
(iii) It converges P-a.s. (t ) to an integrable random variable X
, such that
X
t
, T
t
t[0,]
is a submartingale.
For the implication (i)(ii)(iii), the assumption of nonnegativity can be dropped.
Theorem 3.21 (UI Martingales can be closed)
The following conditions are equivalent for a right-continuous martingale X =
X
t
, T
t
t[0,)
:
(i) It is a uniformly integrable family of random variables.
(ii) It converges in L
1
, as t .
(iii) It converges P-a.s. (t ) to an integrable random variable X
, such that
X
t
, T
t
t[0,]
is a martingale, i.e. the martingale is closed by Y .
(iv) There exists an integrable random variable Y , such that X
t
= E[Y [T
t
] a.s. for
all t 0.
If (iv) holds and X
is the random variable in (c), then E[Y [T
] = X
a.s.
We now formulate a more general version of Doobs optional sampling theorem:
Theorem 3.22 (Optional Sampling Theorem, 2nd Version)
Let X = X
t
, T
t
t[0,]
be a right-continuous submartingale (supermartingale)
with a last element X
and let
1
and
2
be stopping times of the ltration T
t
with
1

2
a.s. Then
E[X
2
[T
1
] ()X
1
Proof. See, e.g., Karatzas/Shreve (1991, pp. 19f).
Proposition 3.23 (Stopped (Sub)-Martingales)
Let X = X
t
, T
t
t[0,)
be an adapted right-continuous (sub)-martingale and let
be a bounded stopping time. The stopped process X
t
, T
t
t[0,)
is also an
adapted right-continuous (sub)-martingale.
Proof. Exercise.
Remarks.
The diculty in this proposition is to show that the stopped process is a (sub)-
martingale for the ltration T
t
. It is obvious that the stopped process is a
(sub)-martingale for T
t
.
If X
t
, T
t
t[0,]
is a UI right-continuous martingale, then the stopping time
may be unbounded (See Protter (2005, p. 10)).
Corollary 3.24 (Conditional Expectation and Stopped -Fields)
Let
1
and
2
be stopping times and Z be an integrable random variable. Then
E[E[Z[T
1
][T
2
] = E[E[Z[T
2
][T
1
] = E[Z[T
2
], P a.s.
Proof. Exercise.
For technical reasons, we sometimes need to weaken the martingale concept.
Denition 3.16 (Local Martingale)
Let X = X
t
, T
t
t[0,)
be a (continuous) process. If there exists a non-decreasing
sequence
n
nIN
of stopping times of T
t
, such that X
(n)
t
:= X
t
n
, T
t
t[0,)
is a martingale for each n 1 and P(lim
n
n
= ) = 1, then we say that X is
a (continuous) local martingale.
Remark.
Every martingale is a local martingale (Choose
n
= n).
However, there exist local martingales which are no martingales. Especially,
E[X
t
] need not to exist for local martingales.
Proposition 3.25
A non-negative local martingale is a supermartingale.
Proof. follows from Fatous lemma (Exercise). 2
This concept can be applied to other situations (e.g. locally bounded, locally square
integrable):
Denition 3.17 (Local Property)
Let X be a stochastic process. A property is said to hold locally if there exists a
sequence of stopping times
n
nIN
with P(lim
n
n
= ) = 1 such that X
(n)
has property for each n 1.
3.5 Completition, Augmentation, Usual Conditions
The following denition is very important:
Denition 3.18 (Usual Conditions)
A ltered probability space (, (, P, (
t
) satises the usual conditions if
(i) it is complete,
9
(ii) (
0
contains all the P-null sets of (, and
(iii) (
t
is right-continuous.
In the sequel, we impose the following
Standing assumption: We always consider a ltered probability space (, T, P, T
t
)
satisfying the usual conditions.
One of the main reasons for this assumption is that the following proposition holds:
Proposition 3.26
Let (, (, P, (
t
) satisfy the usual conditions. Then every martingale for (
t
possesses a unique cadlag modication.

Proof. See, e.g. Karatzas/Shreve (1991, pp. 16f).
Unfortunately, we are confronted with the following result.
Proposition 3.27
The ltration T
W
t
is left-continuous, but fails to be right-continuous.
9
A probability space is complete if for any N such that there exists an

N G with N

N
and P(

N) = 0, we have N G.
Fortunately, there is a way out. Let T
:=
t0
T
t
and consider the space
(, T
W
, P). Dene the collection of P-null sets by

^ := A : B T
W
with A B, P(B) = 0.
The P-augmentation of T
W
t
is dened by T
t
:= T
W
t
^ and is called the
Brownian ltration.
Theorem 3.28 (Brownian Filtration)
The Brownian ltration, is left- and right-continuous. Besides, W is still a Brownian
motion w.r.t. the Brownian ltration.
This is the reason why we work with the Brownian ltration instead of T
W
t
. For
a Poisson process, a similar result holds.
Proposition 3.29 (Augmented Filtration of a Poisson Process)
The P-augmentation of the natural ltration induced by a Poisson process N is
right-continuous. Besides, N is still a Poisson process w.r.t. this P-augmentation.
The proofs of the previous results can be found in Karatzas/Shreve (1991), pp. 89.
Remark. More generally, one can prove the following results:
(i) If a stochastic process is left-continuous, then the ltration T
X
t
is left-continuous.
(ii) Even if X is is continuous, T
X
t
need not to be right-continuous.
(iii) For an n-dimensional strong Markov process X the augmented ltration is
right-continuous.
3.6 Ito-Integral
Our goal is to dene an integral of the form:
t
0
X(s)dW(s),
where X is an appropriate stochastic process. From the Lebesgue-Stieltjes theory
we know that for a measurable function f and a processes A of nite variation on
[0, t] one can dene
t
0
f(s)dA(s) = lim
n
n1
k=0
f
kt
n
(k+1)t
n
kt
n
.
However, the variation of a Brownian motion is unbounded on any interval. There-
fore, naive stochastic integration is impossible (See Protter (2005, pp. 43)).
Good News: We can dene a stochastic integral in terms of L
2
-convergence!
Agenda. (disregarding measurability for the moment)
1st Step. Dene a stochastic integral for simple processes (constant process
except for nitely many jumps).
I
t
() :=
t
0
s
dW
s
:=?
2nd Step. Every L
2
[0, T]-process X can be approximated by a sequence
(n)
n
of simple processes such that
[[X
(n)
[[
2
T
:= E
T
0
(X(s)
(n)
(s))
2
ds
0.
3rd Step. For each n the stochastic integral
I
t
(
(n)
) =
t
0
(n)
(s) dW(s)
is a well-dened random variable and the sequence I
t
(
(n)
)
n
converges (in some
sense) to a random variable I
4th Step. Dene the stochastic integral for a L
2
[0, T]-process X as follows:
t
0
X(s) dW(s) := lim
n
I
t
(
(n)
) = lim
n
t
0
(n)
(s) dW(s).
5th Step. Extend the denition by localization.
Standing Assumption. T
t
is the Brownian ltration of a Brownian motion W.
1st Step.
Denition 3.19 (Simple Process)
A stochastic process
t
t[0,T]
is a simple process if there exist real numbers 0 =
t
0
< t
1
< . . . < t
p
= T, p IN, and bounded random variables
i
: IR,
i = 0, 1, . . . , p such that
0
is T
0
-measurable,
i
is T
t
i1
-measurable, and for all
we have
t
() =
0
()1
{0}
(t) +
p
i=1
i
()1
(t
i1
,t
i
]
(t)
Remarks.

t
is T
t
i
1
-measurable for t (t
i1
, t
i
] (measurable w.r.t. the left-hand side of
the interval).
is a left-continuous step function.
Denition 3.20 (Stochastic Integral for Simple Processes)
For a simple process and t [0, T] the stochastic integral I() is dened by
I
t
() :=
t
0
s
dW
s
:=
1ip
i
(W
t
i
t
W
t
i1
t
)
Proposition 3.30 (Stochastic Integral for Simple Processes)
For a simple process we have:
(i) The integral I() = I
t
()
t[0,T]
is a continuous martingale for T
t
t[0,T]
.
(ii) E
(I
t
())
2
= E
t
0

2
s
ds
for all t [0, T].

(iii) E
sup
0tT
[I
t
()[
4E
T
0

2
s
ds

Proof. (i) Exercise.
(ii) For simplicity, t = t
k+1
. Then
E[I
t
()
2
] =
k
i=1
k
j=1
E
j
(W
t
i
W
t
i1
)(W
t
j
W
t
j1
)
.
1st case: i > j.
E
j
(W
t
i
W
t
i1
)(W
t
j
W
t
j1
)
= E
j
(W
t
i
W
t
i1
)(W
t
j
W
t
j1
)[T
t
i1
= E[
i
j
(W
t
j
W
t
j1
)(E[W
t
i
[T
t
i1
]
. .. .
=W
t
i1
W
t
i1
)]
= 0.
2nd case i = j.
E
2
i
(W
t
i
W
t
i1
)
2
= E
2
i
E[(W
t
i
W
t
i1
)
2
[T
t
i1
]
= E[
2
i
(t
i
t
i1
)]
Hence,
E[I
t
()
2
] = E
i=1
2
i
(t
i
t
i1
)
= E
t
0
2
s
ds
.
(iii) Apply (i), (ii), and Doobs inequality (see Kartzas/Shreve (1991, p. 14): For a
right-continuous martingale M
t
with E[M
2
t
] < for all t 0 we have
E
sup
0T
[M
t
[
4E[M
2
T
].
It remains to check:
E[I
t
()
2
] = E
i=1
2
i
(t
i
t
i1
)
. .. .
T
T
k
i=1
E
2
i
< ,
since, by Denition 3.19, is bounded. 2
Remarks.
Stochastic integrals over [t, T] are dened by
T
t
s
dW
s
:=
T
0
s
dW
s
t
0
s
dW
s
.
Let
1
and
2
be simple processes and a, b IR. The stochastic integral for
simple processes is linear, i.e.
I
t
(a
1
+b
2
) = aI
t
(
1
) +bI
t
(
2
).
This follows directly from Denition 3.20. Note that linear combinations of simple
functions are simple.
For 1,

t
0
1 dW
s
= W
t
. Hence,
E
t
0
dW
s
= E[W
2
t
] = t =
t
0
ds.
This is the reason why people sometimes write dW
t
=

dt, which is wrong from
a mathematical point of view.
2nd Step. Every L
2
[0, T]-process X can be approximated by a sequence
(n)
n
of simple processes.
Dene the vector space
L
2
[0, T] := L
2
[0, T], , T, P, T
t
t[0,T]
:=
X
t
, T
t
t[0,T]
: progressively measurable, real-valued with E
T
0
X
2
s
ds
<
with (semi-)norm
[[X[[
2
T
:= E
T
0
X
2
s
ds
,
where we identify processes for which X = Y only Lebesgue P-a.s.
For simple processes the mapping I() induces a norm on the vector space of
the corresponding stochastic integrals via
[[I()[[
2
L
T
:= E
T
0
s
dW
s
Prop. 3.30 (ii)

= E
T
0
2
s
ds
= [[[[
2
T
.
Since the mapping I(X) is linear and norm preserving, it is an isometry, called the
Ito isometry.
Proposition 3.31 (Simple Processes Approximate L
2
[0, T]-processes)
The linear subspace of simple processes is dense in X L
2
([0, T]), i.e. for all
X L
2
[0, T] there exists a sequence
(n)
n
of simple processes with
[[X
(n)
[[
T
= E
T
0
(X
s
(n)
s
)
2
ds
0 ( as n )
Proof. Assume that X L
2
[0, T] is continuous and bounded. Choose
(n)
t
() := X
0
()1
{0}
(t) +
2
n
1
k=0
X
kT/2
n()1
(kT/2
n
,(k+1)T/2
n
]
(t).
Since X is bounded,
(n)
n
is uniformly bounded and we can apply the dominated
convergence theorem to obtain the desired result.
For the general case, see Karatzas/Shreve (1991, p. 133). 2
3rd and 4th Step.
Dene
c
2
([0, T]) := M : M continuous martingale on [0, T] w.r.t. T
t
,
E[M
2
t
] < for every t [0, T]
Theorem and Denition 3.32 (Ito Integral for X L

2
[0, T])
There exists a unique linear mapping J : L
2
[0, T]
c
2
([0, T]) with the following
properties:
(i) If is simple, then P(J
t
() = I
t
(X) for all t [0, T]) = 1.
(ii) E[J
t
(X)
2
] = E
t
0
X
2
s
ds
Ito isometry
J is unique in the following sense: If there exists a second mapping J
satisfying (i)
and (ii) for all X L
2
[0, T], then J(X) and J
(X) are indistinguishable.

For X L
2
[0, T], we dene
t
0
X
s
dW
s
:= J
t
(X)
for every t [0, T] and call it the Ito integral or stochastic integral of X (w.r.t. W).
Proof. Let X L
2
[0, T].
(a) Let
(n)
n
be an approximating sequence, i.e. [[X
(n)
[[
T
0 as n .
Since
(n)
n
is also a Cauchy sequence w.r.t. [[ [[
T
, we obtain for every t [0, T]
(as n, m ):
E[(I
t
(
(n)
)I
t
(
(m)
))
2
]
I linear
= E[I
t
(
(n)
(m)
)
2
]
Ito Isom.
= E
t
0
(
(n)
(m)
)
2
ds
0.
Hence, I
t
(
(n)
)
n
is a Cauchy sequence in the ordinary L
2
(, T
t
, P) for random
variables (t xed!) which is a complete space and thus
I
t
(
(n)
)
L
2
Z
t
,
where Z
t
is T
t
-measurable and E[Z
2
t
] < . Since I
t
(
(n)
)
P
Z
t
as well, there
exists a subsequence n
k
k
such that
I
t
(
(n
k
)
)
a.s.
Z
t
(1)
as k , i.e. Z
t
is only P-a.s. unique (may depend on sequence!).
(b) Z is a martingale: Note that I
t
(
(n)
) is a martingale for all n IN. By the
denition of the conditional expected value, we thus get
A
I
t
(
(n)
) dP =
A
I
s
(
(n)
) dP
for s < t and for all A T
s
and n IN. The L
2
-convergence (1) leads to (as
n )
10
A
Z
t
dP
A
I
t
(
(n)
) dP =
A
I
s
(
(n)
) dP
A
Z
s
dP.
Since Z is adapted, (ii) is proved.
(c) By Proposition 3.26, Z possesses a right-continuous modication J
t
(X), T
t
t[0,T]
still being a square-integrable martingale. Applying Doobs inequality and the
Borel-Cantelli Lemma, one can prove that J(X) is even continuous (Exercise).
(d) J
t
(X) is independent of the approximating sequence: Let
(n)
n
and

(n)
n
be two approximating sequences for X with I
t
(
(n)
)
L
2
Z
t
and I
t
(
(n)
)
L
2
10
See Bauer (1992, MI, p. 92 and p. 99).
Z
t
. Then
(n)
n
with
(2n)
=
(2n)
(even) and
(2n+1)
=
2n+1
(odd) is an
approximating sequence for X with I
t
(
(n)
)
n
converges in L
2
to both Z
t
and

Z
t
.
Thus they need to coincide with J
t
(X) P-a.s.
11
(e) Ito isometry:
E
J
t
(X)
2
()
= lim
n
E
I
t
(
(n)
)
2
= lim
n
E
t
0
(
(n)
s
)
2
ds
()
= E
t
0
X
2
s
ds
() holds because I
t
(
(n)
)
L
2
J
t
(X) (as n ). This is the L
2
-convergence in
the ordinaryL
2
(, T
t
, P).
() holds because
(n)
L
2
X (as n ). This is the L
2
-convergence in ([0, T]
, B[0, T] T, Lebesgue P).
12
(f) Linearity of J follows from the linearity of I and from the fact that if
(n)
X

n
approximates X and
(n)
Y

n
approximates Y , then a
(n)
X
+b
(n)
Y

n
approximates
aX +bY (a, b IR). Therefore,
aJ
t
(X) +bJ
t
(Y ) = a lim
n
I
t
(
(n)
X
) +b lim
n
I
t
(
(n)
Y
) = lim
n
I
t
a
(n)
X
+b
(n)
Y
= J
t
(aX +bY ).
(g) Uniqueness of J: Let J
be a linear mapping with properties (i) and (ii).

Then
J
t
(X) = J
t
( lim
n
(n)
)
()
= lim
n
J
t
(
(n)
)
(i)
= lim
n
I
t
(
(n)
) = J
t
(X) P a.s.
for all t [0, T]. Since J(X) and J
(X) are continuous processes, by Proposition

3.2 (ii), they are indistinguishable, i.e. P(J
t
(X) = J
t
(X) for all t [0, T]) = 1. The
equality () holds because linear mappings are continuous. 2
5th Step. The class of admissible integrands can be extended. Dene
H
2
[0, T] := H
2
[0, T], , T, P, T
t
t[0,T]
:=
X
t
, T
t
t[0,T]
: progressively measurable, real-valued with
P
T
0
X
2
s
ds <
= 1
In general, for X H
2
[0, T] it can happen that
[[X[[
2
T
= E
T
0
X
2
s
ds
= .
Therefore, the Ito isometry needs not to hold and the approximation argument
breaks down. Fortunately, one can use a localization argument:
13
n
() := T inf
t [0, T] :
t
0
X
2
s
() ds n
,
X
(n)
t
() := X
t
()1
{
n
()t}
.
11
See Bauer (1992, MI, p. 130).
12
Both () and () are a consequence of Bauer (1992, MI, p. 92).
13
Every
n
is a stopping time because it is the rst-hitting time of [n, ) for the continuous
process {
t
0
X
2
s
ds}
t
. See Proposition 3.11.
Hence, X
(n)
t
() L
2
[0, T] and J(X
(n)
) is dened and a martingale for xed n IN.
Now dene
J
t
(X) := J
t
(X
(n)
) for X H
2
[0, T] and 0 t
n
.
Note that
J
t
(X
(n)
) = J
t
(X
(m)
) for all 0 t
n

m
, i.e. J is well-dened,
lim
n
n
= T a.s. (since P
T
0
X
2
s
ds <
= 1, i.e. there exists an event

A with P(A) = 1 such that for all A there exists an c() > 0 with
T
0
X
2
s
() ds c().)
Important Results.
By construction, for X H
2
[0, T] the process J
t
(X)
t[0,T]
is a continuous local
martingale with localizing sequence
n
n
.
J is still linear, i.e. J
t
(aX + bY ) = aJ
t
(X) + bJ
t
(Y ) for a, b IR and X,
Y H
2
[0, T]. (Choose
n
:=
X
n

Y
n
as localizing sequence of aX + bY , where
X
n

n
is the localizing sequence of X.)
E[J
t
(X)] needs not to exist.
Denition 3.21 (Multi-dimensional Ito Integral)
Let W(t), T
t
be an m-dimensional Brownian motion and X(t), T
t
an IR
n,m
-
valued progressively measurable process with all components X
ij
H
2
[0, T]. Then
we dene the Ito integral of X w.r.t. W as
t
0
X(s) dW(s) :=
m
j=1
t
0
X
1j
(s)dW
j
(s)
.
.
.
m
j=1
t
0
X
nj
(s)dW
j
(s)
, t [0, T],
where
t
0
X
ij
(s)dW
j
(s) is an one-dimensional Ito integral.
3.7 Itos Formula
Recall our
Standing assumption: (, T, P, T
t
) is a ltered probability space satisfying
the usual conditions.
Additional Assumption: W
t
, T
t
is a (multi-dimensional) Brownian motion of
appropriate dimension.
Denition 3.22 (Ito Process)
Let W be an m-dimensional Brownian motion.
(i) X(t), T
t
is said to be a real-valued Ito process if for all t 0 it possesses the
representation
X(t) = X(0) +
t
0
(s) ds +
t
0
(s)
dW(s)
= X(0) +
t
0
(s) ds +
m
j=1
t
0
j
(s)
dW
j
(s)
for an T
0
- measurable random variable X(0), and progressively measurable pro-
cesses , with
t
0
[(s)[ ds < ,
m
j=1
t
0
2
j
(s) ds < , P a.s. for all t 0. (2)
(ii) An n-dimensional Ito process X = (X
(1)
, . . . , X
(n)
) is a vector of one-dimensional
Ito processes X
(i)
.
Remarks.
To shorten notations, one sometimes uses the symbolic dierential notation
dX(t) = (t)dt +(t)
dW(t)
It can be shown that the processes and are unique up to indistinguishability.
By (2), we have
j
H
2
[0, T].
For a real-valued Ito process X and a real-valued progressively measurable process
Y , we dene
t
0
Y
s
dX
s
:=
t
0
Y
s
s
ds +
t
0
Y
s
s
dW
s
Denition 3.23 (Quadratic Variation and Covariation)
Let X
(1)
and X
(2)
two real-valued Ito processes with
X
(i)
(t) = X(0) +
t
0
(i)
(s) ds +
t
0
(i)
(s)
dW(s),
where i 1, 2. Then,
< X
(1)
, X
(2)
>
t
:=
m
j=1
t
0
(1)
j
(s)
(2)
j
(s) ds
is said to be the quadratic covariation of X
(1)
and X
(2)
. In particular, < X >
t
:=<
X, X >
t
is said to be the quadratic variation of an Ito process X.
Remarks. The denition may appear a bit unfounded. However, one can show
the following (see Karatzas/Shreve (1991, pp. 32)): For xed t 0, let =
t
0
, . . . , t
n
with 0 = t
0
< t
1
< . . . < t
n
= t be a partition of [0, t]. Then the p-th
variation of X over the partition is dened by (p > 0)
V
(p)
t
() =
n
k=1
[X
t
k
X
t
k1
[
p
.
The mesh of is dened by [[[[ = max
1kn
[t
k
t
k1
[. For X
c
2
, it holds
that
lim
||||0
V
(2)
t
() =< X >
t
in probability.
The main tool for working with Ito processes is Itos formula:
Theorem 3.33 (Itos Formula)
Let X be an n-dimensional Ito process with
X
i
(t) = X
i
(0) +
t
0
i
(s) ds +
m
j=1
t
0
ij
(s) dW
j
(s).
Let f : [0, ) IR
n
IR be a C
1,2
-function. Then
f(t, X
1
(t), . . . , X
n
(t)) = f(0, X
1
(0), . . . , X
n
(0)) +
t
0
f
t
(s, X
1
(s), . . . , X
n
(s)) ds
+
n
i=1
t
0
f
x
i
(s, X
1
(s), . . . , X
n
(s)) dX
i
(s)
+0.5
n
i,j=1
t
0
f
x
i
x
j
(s, X
1
(s), . . . , X
n
(s)) d < X
i
, X
j
>
s
,
where all integrals are dened.
Proof. For simplicity, X(0) = const., n = m = 1, and f : IR IR.
1st step. By applying a localization argument, we can assume that
M
t
:=
t
0
s
dW
s
, B
t
:=
t
0
s
ds,

B
t
:=
t
0
[
s
[ ds,
t
0

2
s
ds, and thus X are uniformly bounded by a constant C. Hence, the relevant
support of f, f
, and f
is compact implying that these functions are bounded as

well. Besides, M is a martingale (since L
2
[0, T]).
2nd step. Let = t
0
, . . . , t
p
be a partition of [0, t]. Taylor expension leads to
(Note: f(X
t
k
) = f(X
t
k1
) +f
(X
t
k1
)(X
t
k
X
t
k1
) +. . ..)
f(X
t
) f(X
0
) =
p
k=1
(f(X
t
k
) f(X
t
k1
))
=
p
k=1
f
(X
t
k1
)(X
t
k
X
t
k1
)
. .. .
()
+0.5
p
k=1
f
(
k
)(X
t
k
X
t
k1
)
2
. .. .
()
with
k
= X
t
k1
+
k
(X
t
k
X
t
k1
),
k
(0, 1).
3rd step. Convergence of (*). We get
() =
p
k=1
f
(X
t
k1
)(B
t
k
B
t
k1
)
. .. .
A
1
()
+
p
k=1
f
(X
t
k1
)(M
t
k
M
t
k1
)
. .. .
A
2
()
.
(i) For [[[[ 0, we obtain (Lebesgue-Stieltjes theory)
A
1
()
a.s.,L
1
t
0
f
(X
s
)dB
s
=
t
0
f
(X
s
)
s
ds
(ii) Approximate f
(X
s
) by a simple process
Y
s
() := f
(X
0
)1
{0}
(s) +
p
k=1
f
(X
t
k1
())1
(t
k1
,t
k
]
(s).
Note that A
2
() =

t
0
Y
s
dM
s
.
14
By the Ito isometry,
E
t
0
f
(X
s
) dM
s
A
2
()
2
= E
t
0
(f
(X
s
) Y
s
)
s
dW
s
2
= E
t
0
(f
(X
s
) Y
s
)
2
2
s
ds
L
2
0,
where the limit follows from the dominated convergence theorem (as [[[[ 0).
Therefore,
15
(as [[[[ 0)
() = A
1
() +A
2
()
L
1
t
0
f
(X
s
)
s
ds +
t
0
f
(X
s
) dM
s
4th step. Convergence of (**). It can be shown that (See Korn/Korn (1999,
pp. 53))
()
L
1
t
0
f
(X
s
)
2
s
ds.
5th step. By steps 2-4, there exists a subsequence
k
k
, such that the integrals
converges a.s. towards the corresponding limits. Therefore, for xed t both sides
of Itos formula are equal a.s. Since both sides are continuous processes in t, by
Proposition 3.2 (ii), they are indistinguishable. 2
Remark. Often a short-hand symbolic dierential notation is used (f : [0, )
IR IR)
df(t, X
t
) = f
t
(t, X
t
)dt +f
x
(t, X
t
)dX
t
+ 0.5f
xx
(t, X
t
)d < X >
t
Examples.
X
t
= t and f C
2
leads to the fundamental theorem of integration.
X
t
= h(t) and f C
2
leads to the chain rule of ordinary calculus.
X
t
= W
t
and f(x) = x
2
:
W
2
t
= 2
t
0
W
s
dW
s
+t
X
t
= x
0
+t +W
t
, x
0
, , IR, and f(x) = e
x
.
df(X
t
) = f
(X
t
)dX
t
+ 0.5f
(X
t
)d < X >
t
= f(X
t
)
dt +dW
t
+ 0.5
2
dt
= f(X
t
)
( + 0.5
2
)dt +dW
t
.
Therefore, the solution of the stock price dynamics in the Black-Scholes model,
dS
t
= S
t
[dt +dW
t
], reads
S
t
= s
0
e
(0.5
2
)t+W
t
. (3)
14
This is actually a consequence of the more general denition of Ito integrals in Karatzas/Shreve
(1991, pp. 129).
15
L
2
-convergence implies L
1
-convergence
Itos product rule: Let X and Y be Ito processes and f(x, y) = x y Then
(omitting dependencies)
d(XY ) = f
x
dX +f
y
dY + 0.5 f
xx
....
=0
d < X > +0.5 f
yy
....
=0
d < Y > + f
xy
....
=1
d < X, Y >
= XdY +Y dX +d < X, Y >
Denition 3.24 (Stochastic Dierential Equation)
If for a given ltered probability space (, T, P, T
t
) there exists an adapted pro-
cess X with X(0) = x IR
n
xed and
X
i
(t) = x
i
+
t
0
i
(s, X(s)) ds +
m
j=0
t
0
ij
(s, X(s)) dW
j
(s)
P-a.s. for every t 0 and i 1, . . . , n, such that
t
0
[
i
(s, X(s))[ +
m
j=1
2
ij
(s, X(s))
ds <
P-a.s. for every t 0 and i 1, . . . , n, then X is a strong solution of the stochastic
dierential equation (SDE)
dX(t) = (t, X(t))dt +(t, X(t))dW(t) (4)
X(0) = x,
where : [0, ) IR
n
IR
n
and : [0, ) IR
n
IR
n,m
are given functions.
Theorem 3.34 (Existence and Uniqueness)
If the coecients and are continuous functions satisfying the following global
Lipschitz and linear growth conditions for xed K > 0
[[(t, z) (t, y)[[ +[[(t, z) (t, y)[[ K[[z y[[,
[[(t, y)[[
2
+[[(t, y)[[
2
K
2
(1 +[[y[[
2
),
z, y IR, then there exists an adapted continuous process X being a strong solution
of (4) and satisfying
E[[[X
t
[[
2
] C(1 +[[x[[
2
)e
CT
for all t [0, T],
where C = C(K, T) > 0. X is unique up to indistinguishability.
Proof. See Korn/Korn (1999, pp. 128).
In general, there do not exist explicit solutions to SDEs. However, the important
family of linear SDEs admits explicit solutions:
Proposition 3.35 (Variation of Constants)
Fix x IR and let A, a, B, b be progressively measurable real-valued processes with
(P-a.s. for every t 0)
t
0
([A(s)[ +[a(s)[) ds < ,
m
j=1
t
0
B
j
(s)
2
+b
2
j
(s)
ds < .
Then the linear SDE
dX(t) = (A(t)X(t) +a(t))dt +
m
j=1
(B
j
(t)X(t) +b
j
(t))dW
j
(t),
X(0) = x,
possesses an adapted Lebesgue P-unique solution
X(t) = Z(t)
x +
t
0
1
Z(u)
a(u)
m
j=1
B
j
(u)b
j
(u)
du +
m
j=1
t
0
b
j
(u)
Z(u)
dW
j
(u)
,
where
Z(t) = exp
t
0
A(u) 0.5[[B(u)[[
2
du +
t
0
B(u)
dW(u)
. (5)
is the solution of the homogeneous SDE
dZ(t) = Z(t) [A(t)dt +B(t)
dW(t)] ,
Z(0) = 1.
Proof. For simplicity, n = m = 1. As in (3), one can show that Z solves the
homogeneous SDE. To prove uniqueness, note that Itos formula leads to
d
1
Z
=
1
Z
(A+B
2
)dt BdW
.
Let

Z be another solution. Then, by Itos product rule,
d
Z
1
Z
=

Zd
1
Z
+
1
Z
d
Z+ <
1
Z
,

Z >
=

Z
Z
(A+B
2
)dt BdW
+

Z
Z
[Adt +BdW]
Z
Z
B
2
dt = 0
implying

Z
1
Z
= const. Due to the initial condition, we obtain

Z = Z.
To show that X satises the linear SDE, dene Y := X/Z. Itos product rule
applied to Y Z gives the desired result. To prove uniqueness, let X and

X be two
solutions to the linear SDE. Then

X :=

X X solves the homogeneous SDE
d

X =

X[Adt +BdW]
with

X
0
= 0. Applying Itos product rule to 0 Z shows

X
t
= 0 a.s. for all t 0. 2
4 CONTINUOUS-TIME MARKET MODEL AND OPTION PRICING 47
Example. In Brownian models the money market account and stocks are modeled
via (i = 1, . . . , I)
dS
0
(t) = S
0
(t)r(t)dt, S(0) = 1, (6)
dS
i
(t) = S
i
(t)
i
(t)dt +
m
j=1
ij
(t)dW
j
(t)
, S(0) = s
i0
> 0.
Applying variation of constants leads to
S
0
(t) = e
t
0
r(s) ds
,
S
i
(t) = s
i0
e
t
0

i
(s)0.5
m
j=1

2
ij
(s) ds+
m
j=1
t
0

ij
(s)dW
j
(s)
.
4 Continuous-time Market Model and Option Pric-
ing
4.1 Description of the Model
We consider a continuous-time security market as given by (6).
Mathematical Assumptions.
Recall the standing assumption: (, T, P, T
t
) is a ltrered probability space
satisfying the usual conditions.
W
t
, T
t
is a (multi-dimensional) Brownian motion of appropriate dimension.
T
t
is the Brownian ltration.
The coecients r, , and are progressively measurable processes being uni-
formly bounded.
Notation.

i
(t): number of stock i at time t

0
(t)S
0
(t): money invested in the money market account at time t
X
(t) :=

I
i=0
i
(t)S
i
(t): wealth of the investor
Assumption: x
0
:= X(0) =

I
i=0
i
(0)S
i
(0) 0 (non-negative initial wealth)

i
(t) :=
i
(t)S
i
(t)/X(t): proportion invested in stock i
1
i
: proportion invested in the money market account
Denition 4.1 (Trading Strategy, Self-nancing, Admissible)
(i) Let = (
0
, . . . ,
I
)
be an IR
I+1
-valued process which is progressively measur-
able w.r.t. T
t
and satises (P-a.s.)
T
0
[
0
(t)[dt < ,
I
i=1
T
0
(
i
(t)S
i
(t))
2
dt < . (7)
Then is said to be a trading strategy.
(ii) If
dX(t) = (t)
dS(t)
holds, then is said to be self-nancing.
(iii) A self-nancing is said to be admissible if X(t) 0 for all t 0.
Remarks.
The corresponding portfolio process is said to be self-nancing or admissible
if is self-nancing or admissible (Recall
i
=
i
S
i
/X). Note that if we work
with , then we need to require that X(t) > 0 for all t 0. Otherwise, is not
well-dened.
(7) ensures that all integrals are dened.
Compare the denition (ii) with the result of Proposition 2.11.
Every trading strategy satises
dX(t) = d((t)
S(t)).
Hence, in abstract mathematical terms, self-nancing means that can be put
before the d operator.
In the following, we always consider admissible trading strategies.
Since admissible strategies are progressive, investors cannot see into the future
(no clairvoyants).
Denition 4.2 (Arbitrage, Option, Replication, Fair Value)
(i) An admissible trading strategy is an arbitrage opportunity if (P-a.s.)
X
(0) = 0, X
(T) 0, P(X
(T) > 0) > 0.

(ii) An T
T
-measurable random variable C(T) 0 is said to be a contingent claim,
T 0.
(iii) An admissible strategy is a replication strategy for the claim C(T) if
X
(T) = C(T) P a.s.

(iv) The fair time-0 value of a claim is dened as
C(0) = infp : D(p) = ,
where
D(x) := : admissible and replicates C(T) and X
(0) = x
Remark. The strategy in (i) is also known as free lottery.
Economical Assumptions.
All investors model the dynamics of the securities as above.
All investors agree upon , but may disagree upon .
Investors are price takers.
Investors use admissible strategies.
Frictionless markets
No arbitrage (see Denition 4.2 (i))
Proposition 4.1 (Wealth Equation)
The wealth dynamics of a self-nancing strategy are described by the SDE
dX(t) = X(t)
(r(t) +(t)
((t) r(t)1))dt +(t)
(t)dW(t)
, X(0) = x
0
,
the so-called wealth equation.
Proof. By denition,
dX =
dS
=
0
S
0
rdt +
I
i=1
i
S
i
i
dt +
m
j=1
ij
dW
j
1
I
i=1
Xrdt +
I
i=1
i
X
i
dt +
m
j=1
ij
dW
j
= X
r +
I
i=1
i
(
i
r)
dt +
m
j=1
I
i=1
ij
dW
j
2
Remarks.
Due to the boundedness of the coecients, by variation of constants, the con-
dition (P a.s.)
t
0
i
(s)
2
ds <
is sucient for the wealth equation to possess a unique solution.
One can directly start with (instead of ) and require that for the wealth
equation
t
0
X(s)
2
i
(s)
2
ds <
4.2 Black-Scholes Approach: Pricing by Replication
We wish to price a European call written on a stock S with strike K, maturity T,
and payo
C(T) = maxS(T) K; 0,
where
dS(t) = S(t)
dt +dW(t)
with constants and . The interest rate r is constant as well.

Assumptions.
No payout are made to the stocks during the life-time of the call.
There exists a C
1,2
-function f = f(t, s) such that the time-t price of the call is
given by C(t) = f(t, S(t)).
Then
dC(t) = f
t
(t, S(t))dt +f
s
(t, S(t))dS(t) + 0.5f
ss
(t, (S(t))d < S >
t
= f
t
dt +f
s
S
dt +dW
+ 0.5f
ss
S
2
2
dt
=
f
t
+f
s
S + 0.5f
ss
S
2
dt +f
s
SdW
Consider a self-nancing trading strategy = (
M
,
S
,
C
) and choose
C
(t)
1. Then
dX
(t) =
M
(t)dM(t) +
S
(t)dS(t) dC(t)
=
0
rMdt +
S
S
dt +dW
f
t
+f
s
S + 0.5f
ss
S
2
dt f
s
SdW
=
M
rM +
S
S f
t
f
s
S 0.5f
ss
S
2
2
. .. .
()
dt +
1
S f
s
S
. .. .
()
dW
Choosing
S
(t) = f
s
(t) leads to () = 0. Hence, the strategy is locally risk-free
implying () = X
(t)r (Otherwise there is an arbitrage opportunity with the money

market account). Rewriting (*):
() =
M
rM f
t
0.5f
ss
S
2
2
= r
M
M +
S
S C
. .. .
=X
rf
s
S +rC f
t
0.5f
ss
S
2
2
. .. .
()
.
Therefore, ( ) = 0, i.e.
rf
s
(t, S(t))S(t) rf(t, S(t)) +f
t
(t, S(t)) + 0.5f
ss
(t, S(t))S
2
(t)
2
= 0.
It follows:
Theorem 4.2 (PDE of Black-Scholes (1973))
In an arbitrage-free market the price function f(t, s) of a European call satises the
Black-Scholes PDE
f
t
(t, s) +rsf
s
(t, s) + 0.5s
2
2
f
ss
(t, s) rf(t, s) = 0,
(t, s) [0, T] IR
+
, with terminal condition f(T, s) = maxs K, 0.
Remark. We have not used the terminal condition to derive the PDE. There-
fore, every European (path-independent) option satises this PDE. Of course, the
terminal condition will be dierent.
Question: How should we solve this PDE?
1st approach: Merton (1973). Solve the PDE by transforming it to the heat
equation u
t
= u
xx
which has a well-known solution.
16
2nd approach. Apply the Feynman-Kac theorem (see below) which states that
the time-0 solution of the Black-Scholes PDE is given by
C(0) = E
maxY (T) K, 0
Y (0) = S(0)
e
rT
,
where Y satises the SDE
dY (t) = Y (t)
rdt +dW(t)
, Y (0) = S(0).
Theorem 4.3 (Black-Scholes Formula)
(i) The price of the European call is given by
C(t) = S(t) N(d
1
(t)) K N(d
2
(t)) e
r(Tt)
with
d
1
(t) =
ln(S(t)/K) + (r + 0.5
2
)(T t)
T t
,
d
2
(t) = d
1
(t)
T t.
(ii) The trading strategy := (
M
,
S
) with
M
(t) =
C(t) f
s
(t, S(t)) S(t)
M(t)
,
S
(t) = f
s
(t, S(t)),
is self-nancing and a replication strategy for the call.
Proof. (i) Compute (exercise)
C(t) = E
maxY (T) K, 0
Y (t) = S(t)
e
r(Tt)
.
(ii) We have
X
(t) =
M
(t) M(t) +
S
(t) S(t) = C(t). (8)
Hence, (
M
,
S
) replicates the call and is self-nancing because
dX
(8)
= dC
Ito
=
f
t
+f
s
S + 0.5f
ss
S
2
dt +f
s
SdW
(+)
=
r[C f
s
S] +f
s
S
dt +f
s
SdW
Def. (
M
,
S
)
=
rM
M
+
S
S
dt +
S
SdW
=
M
rMdt +
S
S
dt +SdW
=
M
dM +
S
dS.
16
See, e.g., Korn/Korn (1999, p. 124).
Note: (+) holds due to the Black-Scholes PDE
f
t
+ 0.5f
ss
S
2
2
= r[C f
s
S]
2
Remark. Since (
M
,
S
) is self-nancing, the strategy (
M
,
S
,
C
) with
C
= 1
is self-nancing as well because
d(
C
(t)C(t)) = d(C(t)) = dC(t) =
C
(t)dC(t).
Theorem 4.4 (Feynman-Kac Representation)
Let and be functions of (t, x) [0, ) IR. Under technical conditions (see
Korn/Korn (1999, p. 136)), there exists a unique C
1,2
-solution F = F(t, x) of the
PDE
F
t
(t, x) +(t, x)F
x
(t, x) + 0.5
2
(t, x)F
xx
(t, x) rF(t, x) = 0 (9)
with terminal condition F(T, x) = h(x) possessing the representation
F(t
0
, x) = e
r(Tt
0
)
E[h(X(T))[X(t
0
) = x],
where X satises the SDE
dX(s) = (s, X(s))ds +(s, X(s))dW(s)
with initial condition X(t
0
) = x.
Sketch of Proof. (i) Firstly assume r = 0. If F is a C
1,2
-function, we can apply
Itos formula
dF = F
t
dt +F
x
dX + 0.5F
xx
d < X > .
Since d < X >=
2
dt, we get
dF = F
t
dt +F
x
dt +dW
+ 0.5
2
F
xx
dt
=
F
t
+F
x
+ 0.5
2
F
xx
dt +F
x
dW.
Suppose that there exists a function F satisfying the PDE, i.e.
F
t
+F
x
+ 0.5
2
F
xx
= 0
with F(T, x) = h(x). Then
dF = F
x
dW
or in integral notation
F(s, X(s)) = F(t
0
, X(t
0
)) +
s
t
0
(u, X(u))F
x
(u, X(u))dW(u).
If E

s
t
0
(u, X
u
)
2
F
x
(u, X
u
)
2
du
< , the stochastic integral has zero expectation

and thus
F(t
0
, x) = E[F(s, X(s)[X(t
0
) = x].
Choose s = T and note that F(T, X(T)) = h(X(T)).
(ii) For r = 0, multiply the PDE (9) by e
rt
:
F
t
(t, x)e
rt
+(t, x)F
x
(t, x)e
rt
+ 0.5
2
(t, x)F
xx
(t, x)e
rt
rF(t, x)e
rt
= 0.
Set G(t, x) := F(t, x)e
rt
. Since
G
t
(t, x) = F
t
(t, x)e
rt
rF(t, x)e
rt
,
we get
G
t
(t, x) +(t, x)G
x
(t, x) + 0.5
2
(t, x)G
xx
(t, x) = 0
with terminal condition G(T, x) = e
rT
h(x). Now we can apply (i):
G(t
0
, x) = E[e
rT
h(X(T))[X(t
0
) = x].
Substituting G(t
0
, x) = e
rt
0
F(t
0
, x) gives the desired result. 2
Remarks.
Two crucial points are not proved: Firstly, we have not checked under which
conditions a C
1,2
-function F exists. Secondly, we have not checked under which
conditions (, X)F
x
(, X) L
2
[0, T].
To solve the Black-Scholes PDE, choose (t, x) = rx and (t, x) = x.
The Feynman-Kac representation can be generalized to a multidimensional set-
ting, i.e. X is a multidimensional Ito process, with stochastic short rate r = r(X).
4.3 Pricing with Deators
In a single-period model the deator is given by (See remark to Proposition 2.7)
H =
1
1 +r
dQ
dP
This motivates the denition of the deator in continuous-time:
H(t) :=
1
M(t)
Z(t),
where
M(t) := exp
t
0
r(s) ds
,
Z(t) := exp
0.5
t
0
[[(s)[[
2
ds
t
0
(s)
dW(s)
,
The process is implicitly dened as the solution to the following system of linear
equations:
(t)(t) = (t) r(t)1. (10)
Applying Itos product rule yields the following SDE for H:
dH(t) = d(Z(t)M(t)
1
) = H(t)[r(t)dt +(t)
dW(t)]
Remarks.
In this section we are going to work under the physical measure. Nevertheless
we will see later on that Z can be interpreted as a density inducing a change of
measure (Girsanovs Theorem).
The system of equations (10) needs not to have a (unique) solution.
The process is said to be the market price of risk and can be interpreted as a
risk premium (excess return per units of volatility).
The following theorem shows that H has indeed the desired properties:
Theorem 4.5 (Arbitrage-free Pricing with Deator and Completeness)
(i) Assume that (10) has a bounded solution for . Let be an admissible portfolio
strategy. Then H(t)X
(t)
t[0,T]
is both a local martingale and a supermartingale
implying
E[H(t)X
(t)] X
(0) for all t [0, T]. (11)

Besides, the market is arbitrage-free.
(ii) Assume that I = m (# stocks = # sources of risk) and that is uniformly
positive denite.
17
Then (10) has a unique solution which is uniformly bounded.
Furthermore, let C(T) be an T
T
-measurable contingent claim with
x := E[H(T)C(T)] < (12)
Then there exists a P Lebesgue-unique portfolio process such that is a repli-
cation strategy for C(T), i.e. (P-a.s.)
X
(T) = C(T),
and X
(0) = x. Hence, the market is complete. The time-t price of the claim
equals
C(t) =
E[H(T)C(T)[T
t
]
H(t)
,
i.e. the process H(t)C(t)
t[0,T]
is a martingale.
Remarks.
The boundedness assumption in (i) can be weakened.
The price in (ii) is thus equal to the expected discounted payo under the phys-
ical measure where we discount with the state-dependent discount factor H, the
deator.
Compare (ii) with Theorem 2.13.
Proof. For simplicity, I = 1. (i) Applying Itos product rule, one gets for an
admissible = (
0
,
1
):
18
d(HX
) = HdX
+X
dH +d < H, X
>
17
The process is uniformly positive denite if there exists a constant K > 0 such that
x
(t)(t)
x Kx
x for all x IR
m
and all t [0, T], P-a.s. This requirement implies that
(t) is regular.
18
Note = r.
= HX
rdt +H
1
S( r)dt +H
1
SdW X
Hrdt
X
HdW H
1
Sdt
= H(
1
S X
)dW. (13)
Since is admissible, we have X
(t) 0. By denition, H(t) 0 and thus

H(t)X(t) 0. By Proposition 3.25, HX is a supermartingale which proves (11).
Note that an arbitrage strategy would violate this supermartingale property.
(ii) Clearly, = ( r)/ is unique and uniformly bounded by some K
> 0.
19
Dene
X(t) :=
E[H(T)C(T)[T
t
]
H(t)
By denition, X(t) is adapted, X(0) = x, X(t) 0, and X(T) = C(T) (P-a.s.).
20
Set M(t) := H(t)X(t) = E[H(T)C(T)[T
t
]. Due to (12), M is a martingale with
M(0) = x. Applying the martingale representation theorem, there exists an T
t
-
progressively measurable real-valued process with

T
0
(s)
2
ds < such that
(P a.s.)
H(t)X(t) = M(t) = x +
t
0
(s) dW(s). (14)
Since H and

0
(s) dW(s) are continuous, we can choose X to be continuous.
Comparing (13) and (14) yields
21
H(
1
S X) = .
Since there exists a K
> 0 such that (t) K
for all t [0, T],

22
we get
1
=
/H +X
S
.
Since X
=
0
M +
1
S for any strategy , we dene
0
:=
X
1
S
M
which implies that is self-nancing and X = X
. Since M > 0, H > 0, and

X 0 are continuous, M as well as H are pathwise bounded away from zero and X
is pathwise bounded from above, i.e. M() K
M
() > 0, H() K
H
() > 0, and
X() K
X
() < for a.a. . Therefore, the process is a trading strategy
because
23
T
0
(
1
(s)S(s))
2
ds =
T
0
(s)/H(s) +X(s)(s)
(s)
2
ds
2
K
2
1
K
2
H
T
0
(s)
2
ds +K
2
X
K
2
< ,
19
Note that all coecients are bounded and is uniformly positive denite.
20
Since F
0
is the trivial -algebra augemented by null sets, is every F
0
-measurable random
variable a.s. constant and the conditioned expected value is equal to the unconditioned expected
value.
21
Note that the representation of an Ito process is unique.
22
By assumption, is uniformly positive denite.
23
(a +b)
2
2a
2
+ 2b
2
.
T
0
[
0
(s)[ ds =
T
0
[X(s)
1
(s)S(s)[
M(s)
ds
1
K
M
K
X
T +
T
0
1 + (
1
(s)S(s))
2
ds
<
for P-a.s. 2
Theorem 4.6 (Martingale Representation Theorem)
Let M
t
t[0,T]
be a real-valued process being adapted to the Brownian ltration.
(i) If M is a square integrable martingale, i.e. E[M
2
t
] < for all t [0, T], then
there exists a progressively measurable IR
m
-valued process
t
t[0,T]
with
E
T
0
[[
s
[[
2
ds
< (15)
and
M
t
= M
0
+
t
0
s
dW
s
, P-a.s.
(ii) If M is a local martingale, then the result from (i) holds with
T
0
[[
s
[[
2
ds <
instead of (15).
In both cases, is P Lebesgue-unique.
Proof. See Korn/Korn (1999, pp. 81).
Remarks.
All Brownian (local) martingales are stochastic integrals and vice versa!
The quadratic (co)variation is dened for all (local) Brownian martingales.
Proposition and Denition 4.7 (Quadratic (Co)variation)
(i) Let X, Y be Brownian local martingales. The quadratic covariation < X, Y >
is the unique adapted, continuous process of bounded variation with < X, Y >
0
= 0
such that XY < X, Y > is a Brownian local martingale. If X = Y , this process
is non-decreasing.
(ii) Let (, (, P, (
t
) be a ltered probability space satisfying the usual conditions
and let X, Y be continuous local martingales for the ltration (
t
. Then the result
of (i) holds with XY < X, Y > being a local martingale for (
t
.
Proof. (i) By the martingale representation theorem, dX = dW and dY = dW
with progressively measurable processes and . Itos product rule yields
d(XY ) = Y dX +XdY +d < X, Y >= Y dW +XdW +dt.
Since Ito integrals are local Brownian martingales, XY < X, Y > is a Brownian
martingale. The rest is obvious.
(ii) See Kartzas/Shreve (1991, p. 36). 2
Remark. One can show that if X and Y are square-integrable, then XY <
X, Y > is a martingale.
4.4 Pricing with Risk-neutral Measures
The deator is dened as
H(t) =
1
M(t)
Z(t),
where
dZ(t) = Z(t)(t)
dW(t)
is a local martingale. If it is even a martingale, we have E[Z(T)] = 1 and we can
dene a probability measure on T
T
by
Q(A) := E[1
A
Z(T)], A T
T
.
Hence, Z(T) is the density of Q and we can rewrite the result from Theorem 4.5 as
C(0) = E[H(T)C(T)] = E
Z(T)
C(T)
M(T)
= E
Q
C(T)
M(T)
,
i.e. the price of a claim equals its expected discounted payo where we discount with
the money market account (instead of the deator) and expectation is calculated
w.r.t. the measure Q (instead of the physical measure).
Two important questions arise:
Q1. Under which conditions is Z a martingale? (Novikov condition)
Q2. What happens to the dynamics of the assets if we change to the measure Q?
(Girsanovs theorem)
Proposition 4.8
Let be progressive and K > 0. If
T
0
[[(s)[[
2
ds K, then the process Z(t)
t[0,T]
dened by
dZ(t) = Z(t)(t)
dW(t), Z(0) = 1, (16)

is a martingale.
Proof. For simplicity, m = 1. The process Z is a local martingale. By variations
of constants, we get
Z(t) = exp
0.5
t
0
(s)
2
ds
t
0
(s) dW(s)
0. (17)
By Proposition 3.25, it is thus a supermartingale. Hence, it is sucient to show
that E[Z(T)] = 1. Dene
n
:= inf
t 0 :
t
0
(s) dW(s)
.
Due to (17), for xed n IN, we get Z(T
n
)
2
e
2n
and thus
T
n
0
Z(s)
2
(s)
2
ds
K e
2n
. From (16), we obtain E[Z(T
n
)] = 1. Therefore, the claim is proved if
Z(T
n
)
nIN
is UI because in this case
lim
n
E[Z(T
n
)] = E[Z(T)].
We obtain the estimate
[Z(T
n
)[
1+
exp
0.5( +
2
)
T
0
(s)
2
ds
. .. .
independent of n and bounded by assumption
Y (T
n
),
where
dY (t) = Y (t)(1 +)(s)dW(s), Y (0) = 1.
With the same arguments leading to E[Z(T
n
)] = 1 we conclude E[Y (T
n
)] = 1.
Therefore,
sup
nIN
E[Z(T
n
)[
1+
< ,
which, by Proposition 3.19, proves UI of Z(T
n
)
nIN
. 2
Remark. The requirement in the proposition is a strong one. It can be replaced
by the Novikov condition (See Karatzas/Shreve (1991, pp. 198)):
E
exp
0.5
T
0
[[(s)[[
2
ds
< .
We turn to the second question and dene (Q := Q
T
)
Q
t
(A) := E[1
A
Z(t)], A T
t
, t [0, T]. (18)
The following lemma shows that Q
t
is well-dened.
Lemma 4.9
For a bounded stopping time [0, T] and A T
, we get Q
T
(A) = Q
(A).
Proof. The optional sampling theorem (OS) yields (See Theorem 3.17)
Q
T
(A) = E[1
A
Z(T)] = E[E[1
A
Z(T)[T
]] = E[1
A
E[Z(T)[T
]]
OS
= E[1
A
Z()] = Q
(A).
2
To prove Girsanovs theorem, we need some additional tools.
Theorem 4.10 (Levys Characterization of the Brownian Motion)
An m-dimensional stochastic process X with X(0) = 0 is an m-dimensional Brow-
nian motion if and only if it is a continuous local martingale with
24
< X
j
, X
k
>
t
=
jk
t.
24
Note that
ij
= 1
{i=j}
is the Kronecker delta.
Proof. For simplicity m = 1. Clearly, the Brownian motion is a continuous local
martingale with < W >
t
= t. To prove the converse, x u IR and set f(t, x) =
exp(iux + 0.5u
2
t) and Z
t
:= f(t, X
t
), where i =

1. Since f C
1,2
is analytic, a
complex-valued version of Itos formula yields
dZ = f
t
dt +f
x
dX + 0.5f
xx
d < X >
= 0.5u
2
fdt +iufdX 0.5u
2
fd < X >
<X>
t
=t
= iufdX.
Since an integral w.r.t. a local martingale is a local martingale as well,
25
Z is a
(complex-valued) local martingale which is bounded on [0, T] for every T IR.
Hence, Z is a martingale implying E[exp(iuX
t
+ 0.5u
2
t[T
s
] = exp(iuX
s
+ 0.5u
2
s),
i.e.
E[exp(iu(X
t
X
s
))[T
s
] = exp(0.5u
2
(t s)).
Since this hold for any u IR, the dierence X
t
X
s
^(0, t s) and X
t
X
s
is
independent of T
s
. 2
Proposition 4.11 (Bayes Rule)
Let and be two probability measures on a measurable space (, () such that
d/d = f for some f L
1
(). Let X be a random variable on (, () such that
E
[[X[] =
[X[fd < .
Let H be a -eld with H (. Then
E
[X[H] E
[f[H] = E
[fX[H] a.s.
Proof. Let H H. We start with the left-hand side:
H
E
[fX[H] d =
H
Xf d =
H
Xd =
H
E
[X[H] d =
H
E
[X[H]f d
= E
[X[H] f1
H
= E
[E
[E
[X[H] f1
H
[H]]
= E
[1
H
E
[X[H] E
[f[H]] =
H
E
[X[H] E
[f[H] d.
2
Lemma 4.12
Let Y
t
, T
t
t[0,T]
be a right-continuous adapted process. The following statements
are equivalent:
(i) Z
t
Y
t
, T
t
t[0,T]
is a local P-martingale.
(ii) Y
t
, T
t
t[0,T]
is a local Q
T
-martingale.
25
For instants, a Brownian local martingale can be represented as Ito integral.
Proof. Assume (i) to hold and let
n
n
be a localizing sequence of ZY . Then
26
E
Q
T
[Y
t
n
[T
s
]
Lem 4.9
= E
Q
t
n
[Y
t
n
[T
s
]
Bayes
=
E
P
[Z
t
n
Y
t
n
[T
s
]
E
P
[Z
t
n
[T
s
]
=
Z
s
n
Y
s
n
Z
s
n
= Y
s
n
.
The converse implication can be proved similarly (exercise). 2
Theorem 4.13 (Girsanovs Theorem)
We assume that Z is a martingale and dene the process W
Q
(t), T
t
by
W
Q
j
(t) := W
j
(t) +
t
0
j
(s) ds, t 0,
j 1, . . . , m. For xed T [0, ), the process W
Q
(t), T
t
t[0,T]
is an m-
dimensional Brownian motion on (, T
T
, Q
T
), where Q
T
is dened by (18).
Proof. For simplicity, m = 1. By Levys characterization, we need to show that
(i) W
Q
(t), T
t
t[0,T]
is a continuous local martingale on (, T
T
, Q
T
).
(ii) < W
Q
>
t
= t on (, T
T
, Q
T
).
ad (i). Itos product rule yields
d(ZW
Q
) = ZdW
Q
+W
Q
dZ +d < W
Q
, Z >
= ZdW +Zdt W
Q
ZdW Zdt
= (Z W
Q
Z)dW,
i.e. ZW
Q
is a local P-martingale. By Lemma 4.12, the process W
Q
is a local
Q
T
-martingale as well.
ad (ii). Itos product rule leads to
d
Z ((W
Q
)
2
t)
= d
W
Q
(ZW
Q
)
d(Z t)
= W
Q
d(ZW
Q
) +ZW
Q
dW
Q
+d < ZW
Q
, W
Q
> Zdt tdZ
= W
Q
d(ZW
Q
) +ZW
Q
(dt +dW) + (Z W
Q
Z)dt Zdt +tZdW
= W
Q
d(ZW
Q
) +Z(W
Q
+t)dW,
i.e. Z
t
((W
Q
t
)
2
t) is a local P-martingale. Applying Lemma 4.12 again yields
that (W
Q
t
)
2
t is a local Q
T
-martingale. By Proposition 4.7 (ii), the claim (ii)
follows. 2
Corollary 4.14 (Girsanov II)
If M
t
t[0,T]
is a Brownian local P-martingale and
D
t
:=
t
0
1
Z
s
d < Z, M >
s
,
26
Note that Y
t
n
is F
t
n
-measurable, i.e. the conditional expected value can be computed
w.r.t. Q
t
n
instead of Q
T
. Lemma 4.9 ensures that both measures coincide on F
t
n
. Besides,
F
t
n
F
T
, i.e. Bayes rule can be applied. Furthermore, by Proposition 3.23, E
P
[Z
t
n
|F
s
] =
Z
s
n
.
then M
t
D
t
t[0,T]
is a continuous local Q
T
-martingale and < M D >
Q
=<
M >, i.e. the quadratic variations of M D under Q
T
and M under P coincide.
Proof. By the martingale representation theorem, we have dM = dW for some
progressive . Hence,
dD =
1
Z
d < Z, M >=
1
Z
Zdt = dt
and thus
d(M D) = dW +dt = dW
Q
.
Therefore, M D is a local Q-martingale. Besides,
d < M D >
Q
t
=
2
dt = d < M >
t
,
which completes the proof. 2
Theorem 4.15 (Converse Girsanov)
Let Q be an equivalent probability measure to P on T
T
. Then the density process
Z(t), T
t
t[0,T]
dened by
Z(t) :=
dQ
dP
F
t
is a positive Brownian P-martingale possessing the representation
dZ(t) = Z(t)(t)dW(t), Z(0) = 1,
for a progressively measurable m-dimensional process with
T
0
[[(s)[[
2
ds < P a.s. (19)
Proof. Since Q and P are equivalent on T
T
, both measures are equivalent on T
t
,
t [0, T] , i.e. Z(t) > 0 for all t [0, T]. Besides, for A T
t
Q
Z(T)dP =
A
dQ/dP[
F
T
Lem 4.9
=
A
dQ/dP[
F
t
=
Q
Z(t)dP,
i.e. Z(t) = E[Z(T)[T
t
] implying that Z is a Brownian martingale. By the martingale
representation theorem,
Z(t) = 1 +
t
0
(s)
dW(s)
with a progressive process satisfying

T
0
[[(s)[[
2
ds < , P a.s.. Since Z > 0,
we have Z() K
Z
() > 0 for a.a. . Dening
(t) =
(t)
Z(t)
thus implies (19) which completes the proof. 2
Denition 4.3 (Doleans-Dade Exponential)
For an Ito process X with X
0
= 0 the Doleans-Dade exponential (stochastic ex-
ponential) of X, written c(X), is the unique Ito process Z that is the solution
of
Y
t
= 1 +
t
0
Y
s
dX
s
.
Remarks.
Therefore, the density Z is sometimes written as c(
0

s
dW
s
).
One can show that c(X)c(Y ) = c(X + Y + < X, Y >) and that (c(X))
1
=
c(X+ < X, X >).
We now come back to our pricing problem. We set Q := Q
T
. Recall our denitions
H(t) :=
1
M(t)
Z(t),
where
M(t) := exp
t
0
r(s) ds
,
Z(t) := exp
0.5
t
0
[[(s)[[
2
ds
t
0
(s)
dW(s)
,
and is implicitly dened by:
(t)(t) = (t) r(t)1. (20)
For bounded , by Girsanovs theorem, Z(t) = dQ/dP[
F
t
is a density process for
the measure Q
t
(A) = E[Z(t)1
A
] and W
Q
(t), T
t
dened by
W
Q
j
(t) := W
j
(t) +
t
0
j
(s) ds, t 0,
is a Q-Brownian motion.
Denition 4.4 (Equivalent Martingale Measure)
A probability measure Q is a (weak) equivalent martingale measure if
(i) Q P and
(ii) all discounted price processes S
i
/M, i = 1, . . . , I, are (local) Q-martingales.
Remark. An equivalent martingale measure (EMM) is also said to be a risk-neutral
measure.
Proposition 4.16
If Q is a weak EMM and
E
exp
0.5
T
0
[[
i
(s)[[
2
ds
< (21)
for every i 1, . . . , I, then Q is already an EMM.
Proof. Dene Y
i
:= S
i
/M. Since Q is a weak martingale measure, Y
i
is a local
Q-martingale. By the martingale representation theorem and the uniqueness of the
representation of an Ito process,
dY
i
= Y
i
i
dW
Q
.
By the Novikov condition (21), this is a Q-martingale. 2
Remark. Since we have assumed that is uniformly bounded, we need not to
distinguish between EMM and weak EMM.
Theorem 4.17 (Characterization of EMM)
In the Brownian market (6), the following are equivalent:
(i) The probability measure Q is an EMM.
(ii) There exists a solution to the system of equations (20) such that the process
Z = c(
0

s
dW
s
) is a Girsanov density inducing the probability measure Q.
Proof. Dene Y
i
:= S
i
/M. Assume (i) to hold. Since Q is a martingale measure,
dY
i
= Y
i
i
dW
Q
. (22)
Besides, since Q is equivalent to P, by the converse Girsanov, there exists a process
such that dQ/dP = c(
s
dW
s
) and dW
Q
= dW +dt is a Brownian increment
under Q. Therefore,
dS
i
= S
i
[
i
dt +
i
dW] = S
i
[(
i
i
)dt +
i
dW
Q
]
implying
dY
i
= Y
i
[(
i
i
r)dt +
i
dW
Q
] (23)
Comparing (22) and (23) gives (ii).
On the other hand, assume (ii) to hold. By denition, Q is equivalent to P and
dW
Q
= dW +dt is a Q-Brownian increment. Itos formula leads to
dY
i
= Y
i
[(
i
r)dt +
i
dW] = Y
i
[(
i
r
i
)dt +
i
dW
Q
] = Y
i
i
dW
Q
,
which is Q-martingale increment because is assumed to be uniformly bounded. 2
Corollary 4.18 (Q-Dynamics of the Assets)
Under an EMM Q, the dynamicss of asset i are given by
dS
i
(t) = S
i
(t)[r(t)dt +
i
(t)
dW
Q
(t)],
i.e. the drift does not matter under Q.
Theorem 4.19 (Arbitrage-free Pricing under Q and Completeness)
(i) Let Qbe an EMM and let be an admissible portfolio strategy. Then X
(t)/M(t)
t[0,T]
is both a local Q-martingale and a Q-supermartingale implying that
E
Q
(t)
M(t)
(0) for all t [0, T].

Besides, the market is arbitrage-free. If additionally
E
Q
T
0
i
(s)
S
i
(s)
M(s)
2
ds
< (24)
for every i = 1, . . . , I, then X
(t)/M(t)
t[0,T]
is a Q-martingale.
(ii) Assume that I = m (# stocks = # sources of risk) and that is uniformly
positive denite. Then (10) has a unique solution which is uniformly bounded
implying that there exists a unique EMM Q. Besides, the market is complete.
Furthermore, let C(T) be an T
T
-measurable contingent claim with
x := E
Q
[C(T)/M(T)] <
Then the time-t price of the claim equals
C(t) = M(t) E
Q
C(T)
M(T)
T
t
,
i.e. the discounted price process C(t)/M(t)
t[0,T]
is a Q-martingale.
Proof. (i) Consider
dX
dS = Xrdt +
I
i=1
i
S
i
i
dW
Q
.
Since d(1/M) = (1/M)rdt,
d
X
M
= Xd
1
M
+
1
M
dX =
1
M
I
i=1
S
i
i
dW
Q
which is a positive local Q-martingale and thus a Q-supermartingale. If (24) is
satised, then it is a Q-martingale.
(ii) Since is unique, by Theorem 4.17, the EMM Q is unique. The claim price
follows from Theorem 4.5 because H = Z/M. 2
Remark.
An EMM exists if, for instants, is bounded. This follows from Theorem 4.17.
If is uniquely bounded, then (24) holds because , , and r are assumed to be
uniformly bounded.
If the market is incomplete, then there exist more than one market price of risk
. Let Q and

Q be two corresponding EMM. If (24) holds, then for an admissible
strategy we get
M(t) E
Q
(T)
M(T)
T
t
= X
(t) = M(t) E
(T)
M(T)
T
t
.
This implies that any EMM assigns the same value to a replicable claim, namly
its replication costs. However, the prices of non-replicable claims may dier
under dierent EMM. This is the reason why pricing in incomplete markets is so
complicated and in some sense arbitrary.
Corollary 4.20 (Black-Scholes Again)
The price of the European call is given by
C(t) = S(t) N(d
1
(t)) K N(d
2
(t)) e
r(Tt)
with
d
1
(t) =
ln(S(t)/K) + (r + 0.5
2
)(T t)
T t
,
d
2
(t) = d
1
(t)
T t.
Proof. By Theorem 4.19,
C(t) = E
Q
[maxS(T) K; 0] e
r(Tt)
,
where due to Corollary 4.18
dS(t) = S(t)[rdt +dW
Q
(t)].
Thus S(T) has the same distribution as Y (T) in the proof of Proposition 4.3 and
the result follows. 2
Corollary 4.21 (Q-Dynamics of Claims)
Under the assumptions of Theorem 4.19, the dynamics of a replicable claim C(T)
are given by
dC(t) = C(t)[r(t)dt +(t)
dW
Q
(t)] (25)
with a suitable progressively measurable process .
Proof. By Theorem 4.19, C/M is a (local) Q-martingale and thus (25) follows
from the martingale representation theorem. 2
Proposition 4.22 (Black-Scholes PDE Again and Hedging)
Under the assumptions of Theorem 4.19, let C(T) be a replicable claim with C(t) =
f(t, S(t)) and C(T) = h(S(T)), where f is a C
1,2
-function.
(i) The pricing function f satises the following multi-dimensional Black-Scholes
PDE:
f
t
+
I
i=1
f
i
s
i
r + 0.5
I
i,k=1
m
j=1
s
i
s
k
ij
kj
f
ik
rf = 0, f(T, s) = h(s), s IR
I
+
,
where f
i
denotes the partial derivative w.r.t. the i-th asset price.
(ii) A replication strategy for C(T) is given by
i
(t) = f
i
(t, S(t)), i = 1, . . . , I,
where the remaining funds C(t)
I
i=1
i
(t)S
i
(t) are invested in the money market
account, i.e.
M
(t) = (C(t)
I
i=1
i
(t)S
i
(t))/M(t).
Proof. (i) Itos formula yields
dC =
f
t
+
I
i=1
f
i
S
i
r + 0.5
I
i,k=1
m
j=1
f
ik
S
i
S
k
ij
kj
dt +
I
i=1
m
j=1
f
i
S
i
ij
dW
Q
j
. (26)
Since the representation of an Ito process is unique, comparing the drift with the
drift of (25) yields the Black-Scholes PDE.
(ii) Since C(T) is attainable, there exists a replication strategy with X
(T) =
C(T) and Q-dynamics
dX
= X
rdt +
I
i=1
m
j=1
i
S
i
ij
dW
Q
j
,
where due to Theorem 4.19 (ii) the drift equals r.
27
Comparing the diusion part
with the diusion part of (26) shows that one can choose
i
= f
i
. 2
Remarks.
The strategy in (ii) is said to be a delta hedging strategy.
Under the conditions of Theorem 4.19 (ii), the delta hedging strategy is the
unique replication strategy.
In the Black-Scholes model, we get for the European call (see Exercise 29):
S
=
N(d
1
). Due to the put-call parity, the delta hedge for the put reads
S
=
N(d
1
).
If I > m, i.e. (# stocks > # sources of risk), then in general there exist further
replication strategies, i.e. the choice
i
= f
i
in the proof of (ii) is not the only
possible choice.
In Theorem 4.19, we have proved that the following implications hold:
1) Existence of EMM = No arbitrage
2) Uniqueness of EMM = Completeness
The following theorems show to which extend the converse is true.
Theorem 4.23 (Existence of a Market Price of Risk)
Assume that the market (6) is arbitrage-free. Then there exists a market price of
risk, i.e. there exists a progressively measurable process solving the system (10).
Proof. Dene
i
:=
i
S
i
(amount of money invested in asset i) leading to the
following version of the wealth equation:
dX(t) = X(t)r(t)dt +(t)
((t) r(t)1)dt +(t)
(t)dW(t).
27
The representation of the diusion part follows from rewriting X
dW
Q
in terms of =
X/S.
Assume that there exists a self-nancing strategy with
(t)
(t) = 0 and (t)
((t) r(t)1) = 0.
This strategy would be an arbitrage since it does not involve risk, but its wealth
equation has a drift dierent from r. Thus in an arbitrage-free market every vector
in the kernel Ker((t)
) should be orthogonal to the vector (t) r(t)1. But the

orthogonal complement of Ker((t)
) is Range((t)). Hence, (t) r(t)1 should be

in the range of (t), i.e. there exists a (t) such that
(t) r(t)1 = (t)(t)
The progressive measurability of is proved in Karatzas/Shreve (1999, pp. 12).2
Remark. In this theorem, nothing is said about whether the corresponding process
Z is a martingale. Although a market price of risk exists, it is not clear if an EMM
exists.
Theorem 4.24 (Uniqueness of EMM)
Assume that the market is complete and let Qbe an EMM such that X
(t)/M(t)
t[0,T]
is a Q-martingale for all admissible strategies with E[X
(T)/M(T)] < . If

Q
is another EMM, then Q =

Q, i.e. Q is the unique EMM.
Proof. Fix j 1, . . . , m. Since the market is complete, there exists an admissible
trading strategy
j
such that X
j
(T) = Y
j
(T), where
Y
j
(T) = exp
T
0
r(s) ds 0.5T +W
Q
j
(T)
.
From Theorem 4.19, we know that the process X
j
/M is a local martingale under
Q and

Q. By assumption, X
j
/M is even a Q-martingale implying
X
j
(t) = M(t)E
Q
j
(T)
M(T)
T
t
= M(t)E
Q
Y
j
(T)
M(T)
T
t
= Y
j
(t) P a.s.,
i.e. Y
j
is a modication of X
j
. By Proposition 3.2, X
j
and Y
j
are indistinguish-
able. Hence, Y
j
/M is a local martingale under Q and

Q. By Theorem 4.17, both
measures are characterized by market prices of risk and

such that
dW
Q
j
(t) = dW
j
(t) +(t)dt,
dW

Q
j
(t) = dW
j
(t) +

(t)dt.
Therefore,
Y
j
(t)
M(t)
= 1 +
t
0
Y
j
(s)
M(s)
dW
Q
j
(s) = 1 +
t
0
Y
j
(s)
M(s)
dW

Q
j
(s) +
t
0
Y
j
(s)
M(s)
(
j
(s)

j
(s)) ds.
Since Y
j
/M is a Brownian local

Q-martingale, the process Y
j
(
j

j
)/M needs to
be indistinguishable from zero.
28
Since Y
j
/M > 0, the market prices of risk
j
and
j
are indistinguishable. Since this holds for every j, we get Q =

Q. 2
Remarks.
28
Recall that the representation of Ito processes is unique up to indistinguishability.
5 CONTINUOUS-TIME PORTFOLIO OPTIMIZATION 68
Since X
(t)/M(t)
t[0,T]
is a Q-supermartingale, this process is a Q-martingale
if, for instants,
E
Q
(T)
M(T)
= X
(0).
Obviously it is sucient to check if X
j
(t)/M(t)
t[0,T]
is a Q-martingale for
every j = 1, . . . , m.
5 Continuous-time Portfolio Optimization
5.1 Motivation
In the theory of option pricing the following problem is tackled:
Given: Payo C(T)
Wanted: Todays price C(0) and hedging strategy
In the theory of portfolio optimization the converse problem is considered:
Given: Initial wealth X(0)
Wanted: Optimal nal wealth X
(T) and optimal strategy
Question: What is optimal?

1st Idea. Maximze expected nal wealth, i.e. max
E[X
(T)].
Problem: Petersburg Paradox. Consider a game where a coin is ipped. If in
the n-th round, the rst time head occurs, then the player gets 2
n1
Euros, i.e. the
expected value is
1
2
1 +
1
4
2 +
1
8
4 +. . . +
1
2
n
2
n1
+. . . =
n=0
1
2
= .
Hence, investors who maximize expected nal wealth should be willing to pay
Euros to play this game! Therefore, maximization of expected nal wealth is usually
not a reasonable criterion.
Question: What is the reason for this result?
Answer: People are risk averse, i.e. if they can choose between
50 Euros or
100 Euros with probability 0.5 and 0 Euros with probability 0.5,
they choose the 50 Euros.
Bernoullis idea. Measure payos in terms of a utility function U(x) = ln(x), i.e.
1
2
ln(1) +
1
4
ln(2) +
1
8
ln(4) +. . . +
1
2
n
ln(2
n1
) +. . .
=
n=1
1
2
n
ln(2
n1
) = ln(2)
n=1
n 1
2
n
= ln(2).
Hence, an investor following this criterion would be willing to pay at most 2 Euros
to play this game.
This idea can be generalized: For a given initial wealth x
0
> 0, the investor
maximizes the following utility functional
sup
A
E[U(X
(T))], (27)
where / := : admissible and E[U(X
(T))
] < and
dX
(t) = X
(t)
(r(t) +(t)
((t) r(t)1))dt +(t)
dW(t)
, X
(0) = x
0
,
and the utility function U has the following properties:
Denition 5.1 (Utility Function)
A strictly concave, strictly increasing, and continuously dierentiable function U :
(0, ) IR is said to be a utility function if
U
(0) := lim
x0
U
(x) = , U
() := lim
x
U
(x) = 0.
Remarks.
The investors greed is reected by U
> 0, i.e. loosely speaking: More is better

than less.
The investors risk aversion is reected by the concavity of U which implies
that for any non-constant random variable Z, by Jensens inequality,
E[U(Z)] < U(E[Z]).
i.e. loosely speaking: A bird in the hand is worth two in the bush.
There exist two approaches to solve the dynamic optimization problem (27):
1. Martingale approach: Theorem 4.5 ensures that in a complete market any
claim can be replicated.
= Determine the optimal wealth via a static optimization problem and interpret
this wealth as a contingent claim
2. Optimal control approac: We know that expectations can be represented as
solutions to PDEs (Feynman-Kac theorem)
= Represent the functional (27) as PDE and solve the PDE
Standing Assumption. We consider a Brownian market (6) with bounded coef-
cients.
5.2 Martingale Approach
Assumption in the Martingale Approach: The market is complete.
The martingale approach consists of a three-step procedure:
1st step. Compute the optimal wealth X
(T) by solving a static optimization

problem.
2nd step. Prove that the static optimization problem is equivalent to the dynamic
optimization problem (27)
3rd step. Determine the replication strategy for X
(T) which, in the context of

portfolio optimization, is said to be the optimal portfolio strategy
.
1st step. Consider the constrained static optimization problem
sup
XC
E[U(A)],
subject to the so-called budget constraint (BC)
E[H(T)A] x
0
and ( := C(T) 0 : C(T) is T
T
-measurable and E[U(C(T))
] < .
The BC states that the present value of terminal wealth cannot exceed the initial
wealth x
0
(see Theorem 4.5 (i)).
Heuristic solution. Pointwise Lagrangian
L(A, ) := U(A) +(x
0
H(T)A).
First-order condition (FOC)
U
(A) H(T) = 0,
i.e. A
= V (H(T)), where V := (U
)
1
: (0, ) (0, ) is strictly decreasing.
Since the investor is greedy, the BC is satised as equality and we can choose
such that
() := E[H(T)V (H(T))] = x
0
(28)
leading to
= (x
0
) :=
1
(x
0
).
2nd step. Verify that the heuristic solution is indeed the optimal solution.
Lemma 5.1 (Properties of )
Assume () < for all > 0. Then is
(i) strictly decreasing,
(ii) continuous on (0, ),
(iii) (0) := lim
0
() = , and () := lim
() = 0.
Proof. (i) V is strictly decreasing on (0, ). Since H(t) > 0 for every t [0, T],
the function is also strictly decreasing.
(ii) follows from the continuity of H and V as well as the dominated convergence
theorem.
29
29
To check continuity at
0
, choose some
1
<
0
. Then due to (i) we can use (
1
) < as a
majorant.
(iii) Since
lim
y0
V (y) = , lim
y
V (y) = 0,
and V is strictly decreasing, the rst result follows from the monotonous convergence
theorem and the second from the dominated convergence theorem. 2
Remark. Hence, (x) =
1
(x) exists on (0, ) with 0 and
(0) := lim
x0
(x) = , () := lim
x
(x) = 0.
Therefore, one can always nd a such that (28) is satised as equality.
Lemma 5.2
For a utility function U with U
1
= V we get
U(V (y)) U(x) +y(V (y) x) for 0 < x, y < .
Proof. By concavity of U, Taylor approximation leads to U(x
2
) U(x
1
) +
U
(x
1
)(x
2
x
1
) and thus U(x
1
) U(x
2
) +U
(x
1
)(x
1
x
2
) implying
U(V (y)) U(x) +U
(V (y))(V (y) x) = U(x) +y(V (y) x).

2
Theorem 5.3 (Optimal Final Wealth)
Assume x
0
> 0 and () < for all > 0. Then the optimal nal wealth reads
A
= V ((x
0
) H(T))
and there exists an admissible optimal strategy
/ such that X
(T) = x
0
and
X
(T) = A
, P-a.s.
Proof. By denition of (x
0
), we get E[H(T)A
] = x
0
. Besides, A
> 0 because
V > 0. Hence, A
satises all constraints and, by Theorem 4.5 (ii), there exists a

strategy
replicating A
.
We now verify E[U(A
] < , i.e.
/. Due to Lemma 5.2,

U(A
) U(1) + (x
0
)H(T)(A
1).
Therefore,
30
E[U(A
] [U(1)[ +(x
0
) E[H(T)(A
+1)] [U(1)[ +(x

0
) (x
0
+E[H(T)]) < .
Finally, we show that
and A
are indeed optimal. Consider some /. Due

to Lemma 5.2,
U(A
) U(X
(T)) + (x
0
)H(T)(A
(T))
30
a b implies a
|b|.
and thus
E[U(A
)] E[U(X
(T))] + (x
0
)E[H(T)(A
(T))]
= E[U(X
(T))] + (x
0
) (x
0
E[H(T)(X
(T))])
. .. .
BC
0
which gives the desired result. 2
3rd step. To determine the optimal strategy
, one can apply a similar method

as in Proposition 4.22 (ii). This is demonstrated in the following example.
Example Power Utility with Constant Coecients.
We make the following assumptions:
Power utility function U(x) =
1
I = m = 1 (one stock, one Brownian motion)

The coecients r, , are constant.
> 0, i.e. the market is complete.
Since V (y) = (U
)
1
(y) = y
1
1
, the optimal nal wealth reads:
A
= V ((x
0
)H(T)) = ((x
0
)H(T))
1
1
,
where
H(T)
1
1
= exp
1
1
rT + 0.5
1
1
2
T +
1
1
W(T)
(29)
and = ( r)/. On the other hand, the nal wealth for a strategy is given by
X
(T) = x
0
exp
rT +
T
0
( r)
s
0.5
2
2
s
ds +
t
0
s
dW
s
. (30)
By comparing the diusion parts of (29) and (30) we guess the optimal strategy:
(t) =
1
1
r
2
implying
X
(T) = exp
rT +
1
1
2
T 0.5
1
(1)
2
2
T +
1
1
W(T)
.
Since () = E[H(T)V (H(T))] =
1
1
E[H(T)

1
] = x
0
, we get
(x
0
)
1
1
=
1
(x
0
) = x
0
exp

1
rT 0.5

1
2
T 0.5
2
T
.
Therefore,
A
= ((x
0
)H(T))
1
1
= x
0
exp
rT + 0.5
2
T 0.5
2
T +
1
1
W(T)
.
It is easy to check that
1
1
0.5
1
(1)
2
= 0.5 0.5
2
.
Hence, A
= X
(T) and due to completeness
is the unique optimal portfolio

strategy.
5.3 Stochastic Optimal Control Approach
Assumption. The coecients r, , are constant and I = m = 1.
Dene the value function
G(t, x) := sup
A
E
t,x
[U(X
(T))],
where E
t,x
[U(X
(T))] := E[U(X
(T))[X(t) = x]. Note that G(T, x) = U(x).

5.3.1 Heuristic Derivation of the HJB
We start with a heuristic derivation of the so-called Hamiltion-Jacobi-Bellman
equation (HJB) which is a PDE for G.
Assume that there exists an optimal portfolio strategy
. We now carry out the

following procedure:
(i) For given (t, x) [0, T] (0, ), we consider the following strategies over
[t, T]:
Strategy I: Use the optimal strategy
.
Strategy II: For an arbitrary use the strategy
=
on [t, t + ]
on (t + , T]
(ii) Compute E
t,x
[U(X
(T))] and E
t,x
[U(X

(T))].
(iii) Using the fact that
is at least as good as and taking the limit 0

yields the HJB.
We set X
:= X
and

X := X

.
ad (ii).
Strategy I. By assumption, E
t,x
[U(X
(T))] = G(t, x).

Strategy II.
E
t,x
[U(

X(T))] = E
t,x
[E
t+,

X(t+)
[U(

X(T))]] = E
t,x
[E
t+,

X(t+)
[U(X
(T))]]
= E
t,x
[G(t + ,

X(t + ))] = E
t,x
[G(t + , X
(t + ))],
for any admissible .
ad (iii). By denition of the strategies,
E
t,x
[G(t + , X
(t + ))] G(t, x), (31)

where we have equality if =
. Given that G C
1,2
, Itos formula yields
G(t + , X
(t + )) = G(t, x) +
t+
t
G
t
(s, X
s
)ds +
t+
t
G
x
(s, X
s
)dX
s
+0.5
t+
t
G
xx
(s, X
s
)d < X
>
s
= G(t, x) +
t+
t
G
t
(s, X
s
) +G
x
(s, X
s
)X
s
(r +
s
( r))
+0.5G
xx
(s, X
s
)(X
s
)
2
2
s
ds
+
t+
t
G
x
(s, X
s
)X
s

s
dW
s
Given E
t,x
[
t+
t
G
x
(s, X
s
)X
s

s
dW
s
] = 0, we can rewrite (31):
E
t,x

t+
t
G
t
(s, X
s
) +G
x
(s, X
s
)X
s
(r +
s
( r))
+0.5G
xx
(s, X
s
)(X
s
)
2
2
s
ds
0.
Dividing by and taking the limit 0 we get (Note: X
(t) = x)
G
t
(t, x) +G
x
(t, x)x(r +
t
( r)) + 0.5G
xx
(t, x)x
2
2
t
2
0
for any admissible . As above, we have equality for =
. Therefore, we get the

HJB
sup
G
t
(t, x) +G
x
(t, x)x(r +( r)) + 0.5G
xx
(t, x)x
2
= 0.
with nal condition G(T, x) = U(x).
Remarks.
Since (t, x) was arbitrarily chosen, the HJB holds for all (t, x) [0, T] (0, ).
Note that we take the supremum over a number and not over a process.
5.3.2 Solving the HJB
We consider the HJB for an investor with power utility U(x) =
1
:
sup
G
t
(t, x) +G
x
(t, x)x(r +( r)) + 0.5G
xx
(t, x)x
2
= 0.
with nal condition G(T, x) =
1
. Pointwise maximization over yields the

following FOC
G
x
(t, x)x( r) +G
xx
(t, x)x
2
(t)
2
= 0 (32)
implying
(t) =
r
2
G
x
(t, x)
xG
xx
(t, x)
.
Since the derivative of (32) w.r.t. is G
x
x(t, x)x
2
2
, the candidate
is the global
maximum if G
xx
< 0.
Substituting
into the HJB yields a PDE for G:

G
t
(t, x) +rxG
x
(t, x) 0.5
2
G
2
x
(t, x)
G
xx
(t, x)
= 0
with G(T, x) =
1
. We try the separation G(t, x) =

1
f(t)
1
for some function
f with f(T) = 1 and obtain
f
=

1
(r + 0.5
1
1
2
)
. .. .
=:K
f.
This is an ODE for f with the solution f(t) = exp(K(T t)). Hence, we get the
following candidates for the value function and the optimal strategy:
G(t, x) =
1
exp
(1 )K(T t)
, (33)
(t) =
1
1
r
2
.
Remark. Our candidate G is indeed a C
1,2
-function. Since < 1, we also get
G
xx
(t, x) = ( 1)x
1
f(t) < 0.
5.3.3 Verication
In this section, we show that our heuristic derivation of the HJB has lead to the
correct result.
Proposition 5.4 (Verication Result)
Assume that the coecients are constant and that the investor has a power utility
function. Then candidates (33) are the value function and the optimal strategy of
the portfolio problem (27), where for < 0 only bounded admissible strategies are
considered.
Proof. Let /. Since G C
1,2
G(T, X
(T)) = G(t, x) +
T
t
G
t
(s, X
s
)ds +
T
t
G
x
(s, X
s
)dX
s
+0.5
T
t
G
xx
(s, X
s
)d < X
>
s
= G(t, x) +
T
t
G
t
(s, X
s
) +G
x
(s, X
s
)X
s
(r +
s
( r))
+0.5G
xx
(s, X
s
)(X
s
)
2
2
s
ds +
T
t
G
x
(s, X
s
)X
s

s
dW
s
()
G(t, x) +
T
t
G
x
(s, X
s
)X
s

s
dW
s
,
where () holds because G satises the HJB.
The right-hand side is a local martingale in T. For > 0 we have G > 0 and thus
the right-hand side is even a supermartingale implying
E
t,x
[
1
(X
(T))
] G(t, x), (34)

where we use G(T, x) =
1
. Besides, since E
t,x
[
T
t
G
x
(s, X
s
)X
s
dW
s
] = 0
(exercise) and for
G
t
(s, X
s
) +G
x
(s, X
s
)X
s
(r +
s
( r)) + 0.5G
xx
(s, X
s
)(X
s
)
2
(
s
)
2
2
= 0,
we obtain
E
t,x
[
1
(X
(T))
] = G(t, x),
i.e. G is the value function and
is the optimal solution.

Since for a bounded strategy we get E[
T
t
G
x
(s, X
s
)X
s

s
dW
s
] = 0, the relation
(34) also holds for < 0 and bounded admissible strategies. 2
Remarks.
For < 0 the boundedness assumption of can be weakened.
Note that we have not explicitly used the fact that the market is complete.
5.4 Intermediate Consumption
So far the investor has maximized expected utility from terminal wealth only. We
now consider a situation where the investor maximizes
expected utility from terminal wealth as well as
expected utility from intermediate consumption.
Assumption. At time t, the investor consumes at a rate c(t) 0, i.e. over the
interval [0, T] he consumes

T
0
c(t) dt.
Remark. Hence, in our model consumption can be interpreted as a continuous
stream of dividends (See Exercise 40).
To formalize the problem we need to generalize the concept of admissibility:
Denition 5.2 (Consumption Process, Self-nancing, Admissible)
(i) A progressively measurable, real-valued process c
t
t[0,T]
with c 0 and
T
0
c(t) dt < P a.s. is said to be a consumption process.
(ii) Let be a trading strategy (see Denition 4.1) and c be a consumption process.
If
dX(t) = (t)
dS(t) c(t)dt
holds, then the pair (, c) is said to be self-nancing.
(iii) A self-nancing (, c) is said to be admissible if X(t) 0 for all t 0.
Hence, the investors goal function reads
sup
(,c)A
E
T
0
U
1
(t, c(t)) dt +U
2
(X
,c
(T))
, (35)
where U
1
is a utility function for xed t [0, T],
/ :=
(, c) : (, c) admissible and E
T
0
U
1
(t, c(t))
dt +U
2
(X
,c
(T))
<

and
dX
,c
(t) = X
,c
(t)
(r(t) +(t)
((t) r(t)1))dt +(t)
dW(t)
c(t)dt,
X
,c
(0) = x
0
.
We set V
1
:= (U
1
)
1
and V
2
:= (U
2
)
1
and dene
() := E
T
0
H(t)V
1
(t, H(t)) dt +H(T)V
2
(H(T))
It is straightforward to generalize the result of Section 5.2.

Theorem 5.5 (Optimal Consumption and Optimal Final Wealth)
Assume x
0
> 0 and () < for all > 0. Then the optimal consumption and
the optimal nal wealth reads
c
(t) = V
1
(t, (x
0
) H(t))
A
= V
2
((x
0
) H(T))
and there exists a portfolio strategy
such that (
, c
) / and X
,c
(T) = x
0
and X
,c
(T) = A
, P-a.s.
We will present the solution by using the stochastic control approach. The
value function now reads
G(t, x) := sup
(,c)A
E
t,x
T
t
U
1
(s, c(s)) ds +U
2
(X
,c
(T))
.
As in Section 5.3 we make the following
Assumption. The coecients r, , are constant and I = m = 1.
Assume that there exists an optimal strategy (
, c
). We now carry out the pro-

cedure of Section 5.3:
(i) For given (t, x) [0, T] (0, ), we consider the following strategies over
[t, T]:
Strategy I: Use the optimal strategy (
, c
).
Strategy II: For an arbitrary (, c) use the strategy
( , c) =
(, c) on [t, t + ]
(
, c
) on (t + , T]
(ii) Compute E
t,x
[
T
t
U
1
(s, c
(s)) ds + U
2
(X
,c
(T))] and E
t,x
[
T
t
U
1
(s, c(s)) ds +
U
2
(X
, c
(T))].
(iii) Using the fact that (
, c
) is at least as good as ( , c) and taking the limit

0 yields the HJB.
We set X
:= X
,c
and

X := X
, c
.
ad (ii).
Strategy I. By assumption, E
t,x
[
T
t
U
1
(s, c
(s)) ds +U
2
(X
(T))] = G(t, x).

Strategy II.
E
t,x
T
t
U
1
(s, c(s)) ds +U
2
(

X(T))
= E
t,x
E
t+,

X(t+)
T
t
U
1
(s, c(s)) ds +U
2
(

X(T))
= E
t,x
t+
t
U
1
(s, c(s)) ds + E
t+,

X(t+)
T
t+
U
1
(s, c(s)) ds +U
2
(

X(T))
= E
t,x
t+
t
U
1
(s, c(s)) ds + E
t+,

X(t+)
T
t+
U
1
(s, c
(s)) ds +U
2
(X
(T))
= E
t,x
t+
t
U
1
(s, c(s)) ds +G(t + , X
,c
(t + ))
for any admissible (, c).

ad (iii). By denition of the strategies,
E
t,x
t+
t
U
1
(s, c(s)) ds +G(t + , X
,c
(t + ))
G(t, x), (36)

where we have equality if (, c) = (
, c
). To shorten notations, we set X := X

,c
.
Given that G C
1,2
G(t + , X
t+
) = G(t, x) +
t+
t
G
t
(s, X
s
)ds +
t+
t
G
x
(s, X
s
)dX
s
+0.5
t+
t
G
xx
(s, X
s
)d < X >
s
= G(t, x) +
t+
t
G
t
(s, X
s
) +G
x
(s, X
s
)X
s
(r +
s
( r)) G
x
(s, X
s
)c
s
+0.5G
xx
(s, X
s
)(X
s
)
2
2
s
ds
+
t+
t
G
x
(s, X
s
)X
s
s
dW
s
Given E
t,x
[
t+
t
G
x
(s, X
s
)X
s
s
dW
s
] = 0, we can rewrite (36):
E
t,x

t+
t
U
1
(s, c
s
) ds +
t+
t
G
t
(s, X
s
) +G
x
(s, X
s
)X
s
(r +
s
( r))
G
x
(s, X
s
)c
s
+ 0.5G
xx
(s, X
s
)(X
s
)
2
2
s
ds
0.
Dividing by and taking the limit 0 we get (Note: X(t) = x)
U
1
(t, c
t
) +G
t
(t, x) +G
x
(t, x)x(r +
t
( r)) G
x
(t, x)c
t
+0.5G
xx
(t, x)x
2
2
t
2
0
for any admissible (, c). As above, we have equality for (, c) = (
, c
). Therefore,
we get the HJB
sup
,c
U
1
(t, c) +G
t
(t, x) +G
x
(t, x)x(r +( r)) G
x
(t, x)c (37)
+0.5G
xx
(t, x)x
2
= 0
with nal condition G(T, x) = U
2
(x).
Remarks.
Since (t, x) was arbitrarily chosen, the HJB holds for all (t, x) [0, T] (0, ).
Note that we take the supremum over a number (, c) and not over a process.
Example Power Utility. We consider the following power utility functions
U
1
(t, c) = e
t
1
,
U
2
(x) = e
T
1
,
where 0 resembles the investors time preferences. The HJB reads
sup
,c
e
t 1
+G
t
(t, x) +G
x
(t, x)x(r +( r)) G
x
(t, x)c
+0.5G
xx
(t, x)x
2
= 0.
with nal condition G(T, x) = e
T 1
. The FOCs read

G
x
(t, x)x( r) +G
xx
(t, x)x
2
(t)
2
= 0
e
t
c
1
G
x
(t, x) = 0
implying
(t) =
r
2
G
x
(t, x)
xG
xx
(t, x)
c
(t) = e

1
t
G
1
1
x
(t, x)
Substituting (
, c
) into the HJB yields

1
e

1
t
G
1
x
(t, x) +G
t
(t, x) +rxG
x
(t, x) 0.5
2
G
2
x
(t, x)
G
xx
(t, x)
= 0.
We try the separation G(t, x) =
1
(t) for some function f with f(T) = 1 and

some constant . Simplifying yields
1

1
t
f
+1
1
+

r + 0.5
1
1
f = 0
Choosing = 1 leads to
1

1
t
+
1
r + 0.5
1
1
f = 0,
which can be rewritten as
f
= Kf +g,
with K :=

1
(r +0.5
1
1
2
) and g(t) := e

1
t
. This is an inhomgeneous ODE
for f. Variations of Constants yields
f(t) = f
H
(t)
f(T)
T
t
g(s)
f
H
(s)
ds
,
where f
H
(t) = e
K(Tt)
is the solution of the corresponding homogeneous ODE
(see Section 5.3). Hence,
f(t) = e
K(Tt)
1 +
T
t
e

1
s+K(Ts)
ds
. .. .
0
= e
K(Tt)
1
e
KT
(1)
+K(1)
e
(

1
+K)T
e
(

1
+K)t
.
Especially, f > 0. Hence, we get the following candidates for the value function
and the optimal strategy:
G(t, x) =
1
f
1
(t), (38)
(t) =
1
1
r
2
,
c
(t) =
X
(t)
f(t)
e

1
t
.
Remark.
Our candidate G is indeed a C
1,2
-function. Since < 1, we also get G
xx
(t, x) =
( 1)x
1
f(t) < 0.
Clearly, it needs to be shown that these candidates are indeed the value function
and the optimal strategy! This can be done analogously to Proposition 5.4. For
the convenience of the reader, the proof of this verication result is also presented.
function. Then the candidates (38) are the value function and the optimal strategy
of the portfolio problem (35), where for < 0 only bounded portfolio strategies
are considered.
Proof. Let (, c) /. Since G C
1,2
G(T, X
,c
(T)) +
T
t
e
s 1
s
ds
= G(t, x) +
T
t
G
t
(s, X
,c
s
)ds +
T
t
G
x
(s, X
,c
s
)dX
,c
s
+0.5
T
t
G
xx
(s, X
,c
s
)d < X
,c
>
s
+
T
t
1
e
s
c
s
ds
= G(t, x) +
T
t
G
t
(s, X
,c
s
) +G
x
(s, X
,c
s
)X
,c
s
(r +
s
( r)) G
x
(s, X
,c
s
)c
s
+0.5G
xx
(s, X
,c
s
)(X
,c
s
)
2
2
s
2
+
1
e
s
c
ds
+
T
t
G
x
(s, X
,c
s
)X
,c
s

s
dW
s
(39)
It is straightforward to show that E
t,x
[
T
t
G
x
(s, X
s
)X
s
dW
s
] = 0 (exercise).
Besides, G satises the HJB, i.e. for (, c) = (
, c
)
G
t
(s, X
s
) +G
x
(s, X
s
)X
s
(r +
s
( r)) G
x
(s, X
s
)c
s
+0.5G
xx
(s, X
s
)(X
s
)
2
(
s
)
2
2
+
1
e
s
(c
s
)
= 0,
we obtain
E
t,x
[G(T, X
(T))] + E
t,x
T
t
e
s 1
(c
s
)
ds
= G(t, x). (40)

implying
E
t,x
[e
T 1
(T)
] + E
t,x
T
t
e
s 1
(c
s
)
ds
= G(t, x)
because G(T, x) = e
T 1
.
Finally, we need to show that for all (, c) /
E
t,x
[e
T 1
X
,c
(T)
] + E
t,x
T
t
e
s 1
s
ds
G(t, x).
Since G satises the HJB, for all (, c) / we get
G
t
(s, X
,c
s
) +G
x
(s, X
,c
s
)X
,c
s
(r +
s
( r)) G
x
(s, X
,c
s
)c
s
+0.5G
xx
(s, X
,c
s
)(X
,c
s
)
2
(
s
)
2
2
+
1
e
s
(c
s
)
0.
Therefore, (39) yields
G(T, X
,c
(T)) +
T
t
e
s 1
s
ds G(t, x) +
T
t
G
x
(s, X
,c
s
)X
,c
s

s
dW
s
(41)
and now two cases are distinguished.
1st case: > 0. The right-hand side of (41) is a local martingale in T. Since
for > 0 the left-hand side is greater than zero, this is even a supermartingale
implying
E
t,x
[G(T, X
,c
(T))] + E
t,x
T
t
e
s 1
s
ds
G(t, x) (42)
and thus
E
t,x
[e
T 1
X
,c
(T)
] + E
t,x
T
t
e
s 1
s
ds
G(t, x).
2nd case: < 0. Since for a bounded strategy we obtain
E[
T
t
G
x
(s, X
s
)X
s

s
dW
s
] = 0,
the result of the 1st case also holds for < 0 and bounded admissible strategies. 2
Remark. If we set c 0, the previous proof is identical to the proof of Proposition
5.4.
5.5 Innite Horizon
If the optimal portfolio decision over the life-time of an investor is to be deter-
mined, then the investment horizon lies very far into the future. For this reason, it
reasonable to approximate this situation by a portfolio problem with T = .
Hence, the investors goal function reads
sup
(,c)A
E

0
U
1
(t, c(t)) dt
, (43)
where
/ :=
(, c) : (, c) admissible and E

0
U
1
(t, c(t))
dt
<
and
dX
,c
(t) = X
,c
(t)
(r(t) +(t)
((t) r(t)1))dt +(t)
dW(t)
c(t)dt,
X
,c
(0) = x
0
.
As in the previous section we make the following
Assumption 1. The coecients r, , are constant and I = m = 1.
Additionally, we make the
Assumption 2. U
1
(t, c(t)) = e
t
U(c(t)), where 0 and U is a utility function.
The value function is given by
G(t, x) := sup
(,c)A
E
t,x

t
e
s
U(c(s)) ds
.
Lemma 5.7
The value function has the representation
G(t, x) = e
t
H(x),
where H is a function depending on wealth only.
Proof. We obtain
G(t, x)e
t
= sup
(,c)A
E
t,x

t
e
(st)
U(c(s)) ds
= sup
(,c)A
E
0,x

0
e
s
U(c(s)) ds
=: H(x)
2
Hence, the HJB (37) can be rewritten as follows:
sup
,c
e
t
U(c) +G
t
(t, x) +G
x
(t, x)x(r +( r)) G
x
(t, x)c
+0.5G
xx
(t, x)x
2
=
sup
,c
U(c) H(x) +H
x
(x)x(r +( r)) H
x
(x)c + 0.5H
xx
(x)x
2
= 0
Example Power Utility. Let U(c) =
1
. Then the FOCs read

H
x
(x)x( r) +H
xx
(x)x
2
(t)
2
= 0
c
1
H
x
(x) = 0
implying
(t) =
r
2
H
x
(x)
xH
xx
(x)
c
(t) = H
1
1
x
(x).
Substituting (
, c
) into the HJB leads to

1
1
x
(x) H(x) +rxH
x
(x) 0.5
2
H
2
x
(x)
H
xx
(x)
= 0.
We try the ansatz H(x) =
1
for some constants k and . Simplifying yields

1
+r + 0.5
1
1
2
= 0
Choosing = 1 leads to
k =

1
r 0.5
1
1
.
Our ansatz is only reasonable if k 0. For < 0 this is valid in any case. For
> 0 we need to assume that
(r + 0.5
1
1
2
). (44)
Hence, we get the following candidates for the value function and the optimal strat-
egy:
G(t, x) =
1
e
t
x
k
1
, (45)
(t) =
1
1
r
2
,
c
(t) = k X
(t).
Remark. In contrast to the previous section, the investor now consumes a time-
independent constant fraction of wealth.
function where for < 0 only bounded portfolio strategies are considered. If (44)
holds as strict inequality, i.e.
> (r + 0.5
1
1
2
), (46)
then the candidates (45) are the value function and the optimal strategy of the
portfolio problem (43).
Proof. Let (, c) /. Analogously to the proof of Proposition 5.6, we get the
relations (42) and (40) for the candidate G for the value function of this subsection
given by (45)
E
t,x
[G(T, X
,c
(T))] + E
t,x
T
t
e
s 1
s
ds
G(t, x)
E
t,x
[G(T, X
(T))] + E
t,x
T
t
e
s 1
(c
s
)
ds
= G(t, x).
Firstly note that, by the monotonous convergence theorem,
31
for any consumption
process c
lim
T
E
t,x
T
t
e
s
c
s
ds
= E
t,x

t
e
s
c
s
ds
. (47)
Secondly, since G(T, X
(T)) =
1
e
T
X
(T)
k
1
, we get
E
t,x
[G(T, X
(T))] =
1
k
1
e
T
E
t,x
[X
(T)
]
=
1
k
1
e
T
x
E
t,x
e
{r+
1
1
2
0.5
1
(1)
2
2
k}T+

1
W(Tt)
=
1
k
1
x
e
T+{r+
1
1
2
0.5
1
(1)
2
2
k+0.5

2
(1)
2
2
}T
=
1
k
1
x
e
kT
0 as T , (48)
since (46) ensures that k > 0. Therefore,
E
t,x

t
e
s 1
(c
s
)
ds
= G(t, x).
To prove
E
t,x

t
e
s 1
s
ds
G(t, x) (49)
two cases are distinguished.
1st case: > 0. Since in this case G(T, X
,c
(T)) 0, the relation (49) follows
immediately from (42).
2nd case: < 0. Since in this case G < 0, the proof is more involved. From (42)
it follows that
sup
,c
E
t,x
[G(T, X
,c
(T))]
+ E
t,x
T
t
e
s 1
s
ds
G(t, x)
and we obtain the following estimate
0 sup
,c
E
t,x
[G(T, X
,c
(T))]
= sup
,c
E
t,x
e
T
X
,c
(T)
k
1
()
sup
,c
E
t,x
e
T
X
,c
(T)
T
t
e
s 1
s
ds
k
1
sup
E
t,x
e
T
X
,0
(T)
k
1
= sup
E
t,x
X
,0
(T)
e
T
k
1
(+)
=
1
e
(r+0.5
1
1
2
)(Tt)
e
T
k
1
(46)
0 as T ,
where () is valid because

T
t
e
s 1
s
ds 0 and (+) follows from Proposition 5.4.
This establishes (49). 2
Remarks.
The relation
lim
T
E
t,x
[G(T, X
,c
(T))] = 0
is the so-called transversality condition.
31
We emphasize that 1/ is deliberately omitted because without this factor the argument of
the expected value is positive.
For > 0 the property (48) can be proved by using a dierent argument:
0 E
t,x
[G(T, X
(T))] = E
t,x
[
1
(T)
] e
T
k
1
E
t,x
[
1
,0
(T)
] e
T
k
1
sup
E
t,x
[
1
X
,0
(T)
] e
T
k
1
(+)
=
1
e
(r+0.5
1
1
2
)(Tt)
e
T
k
1
(46)
0 as T
where (+) follows from Proposition 5.4. Applying this argument one can avoid
some tedious calculations.
As in Proposition 5.4, the boundedness assumption for < 0 can be weakened.
References
Bingham, N. H. ; R. Kiesel (2004): Risk-neutral valuation, 2nd ed., Springer, Hei-
delberg.
Bjork, T. (2004): Arbitrage theory in continuous time, 2nd ed., Oxford University
Press, Oxford.
Dax, A. (1997): An elementary proof of Farkas lemma, SIAM Review 39, 503-507.
Due, D. (2001): Dynamic asset pricing theory, 3rd ed., Princeton University Press,
Princeton.
Karatzas, I.; S. E. Shreve (1991): Brownian motion and stochastic calculus, 2nd ed.,
Springer, New York.
Karatzas, I.; S. E. Shreve (1999): Methods of mathematical nance, Springer, New
York.
Korn, R. (1997): Optimal portfolios, World Scientic, Singapore.
Korn, R.; E. Korn (1999): Optionsbewertung und Portfolio-Optimierung, Vieweg,
Braunschweig/Wiesbaden.
Korn, R.; E. Korn (2001): Option pricing and portfolio optimization - Modern meth-
ods of nancial mathematics, AMS, Providence, Rhode Island.
Mangasarian, O. L. (1969): Nonlinear programming, New York.
Merton, R. C. (1990): Continuous-time nance, Basil Blackwell, Cambridge MA.
Protter, P. (2005): Stochastic integration and dierential equations, 2nd ed., Springer,
New York.
86

Blatt 11

Uploaded by

Document Information

Original Title

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

Blatt 11

Uploaded by

Copyright:

Available Formats

Lecture Notes

, where the i-th entry is

= S has a unique solution.

Theorem 2.9 (Completeness)

(Pricing with the equivalent martingale measure).

(0) < 0 und V

(t) = r B(t) with B(0) = 1 is

Remark. Let X be stochastic process.

() is dened for all , then X

can also be dened on , by setting

, dened on the set <

-measurable, and the stopped process X

is the random variable in (c), then E[Y [T

possesses a unique cadlag modication.

, P). Dene the collection of P-null sets by

for all t [0, T].

3 INTRODUCTION TO STOCHASTIC CALCULUS 37

Prop. 3.30 (ii)

Theorem and Denition 3.32 (Ito Integral for X L

(X) are indistinguishable.

be a linear mapping with properties (i) and (ii).

(X) are continuous processes, by Proposition

= 1, i.e. there exists an event

is compact implying that these functions are bounded as

(T) > 0) > 0.

(T) = C(T) P a.s.

((t) r(t)1))dt +(t)

with constants and . The interest rate r is constant as well.

(t)r (Otherwise there is an arbitrage opportunity with the money

< , the stochastic integral has zero expectation

(0) for all t [0, T]. (11)

(t) 0. By denition, H(t) 0 and thus

> 0 such that (t) K

for all t [0, T],

. Since M > 0, H > 0, and

dW(t), Z(0) = 1, (16)

(0) for all t [0, T].

((t) r(t)1)dt +(t)

(t) = 0 and (t)

) should be orthogonal to the vector (t) r(t)1. But the

) is Range((t)). Hence, (t) r(t)1 should be

(T) and optimal strategy

Question: What is optimal?

((t) r(t)1))dt +(t)

> 0, i.e. loosely speaking: More is better

(T) by solving a static optimization

(T) which, in the context of

(V (y))(V (y) x) = U(x) +y(V (y) x).

satises all constraints and, by Theorem 4.5 (ii), there exists a

/. Due to Lemma 5.2,

+1)] [U(1)[ +(x

are indeed optimal. Consider some /. Due

, one can apply a similar method

I = m = 1 (one stock, one Brownian motion)

(T) and due to completeness

is the unique optimal portfolio

(T))[X(t) = x]. Note that G(T, x) = U(x).

. We now carry out the

is at least as good as and taking the limit 0

(T))] = G(t, x).

(t + ))] G(t, x), (31)

. Therefore, we get the

. Pointwise maximization over yields the