SDENotes 2

STOCHASTIC DIFFERENTIAL EQUATIONS
BENJAMIN FEHRMAN
Abstract. These notes provide an essentially self-contained introduction to the theory of stochas-
tic differential equations, beginning with the theory of martingales in continuous time. These notes
were prepared for lecture courses at the University of Oxford during the Michaelmas Terms of Fall
2019/20. Each course consisted of sixteen fifty-minute lectures.
Contents
1. Preliminaries 2
1.1. Probability spaces, random variables, and independence 3
1.2. Conditional expectation 5
2. Martingales 8
2.1. Filtrations, martingales, and stopping times 10
2.2. The optional stopping theorem (discrete) 14
2.3. Uniform integrability 18
2.4. The optional stopping theorem (continuous) 20
2.5. Doob’s martingale inequality 23
2.6. Doob’s martingale convergence theorem 25
3. Itô’s Formula 33
3.1. Quadratic variation 38
3.2. Stochastic integration 41
3.3. The integration-by-parts formula 49
3.4. Proof of Itô’s formula 50
4. Applications of stochastic calculus 52
4.1. The stochastic exponential 53
4.2. Levy’s characterization of Brownian motion 56
4.3. The Dambis-Dubins-Schwarz theorem 58
4.4. The Burkholder-Davis-Gundy inequality 61
4.5. The martingale representation theorem 67
5. Local Times 71
5.1. Convex functions 72
5.2. The Meyer-Tanaka formula 74
5.3. Regularity of local times 80
6. The Girsanov theorem 83
7. Stochastic Differential Equations 87
7.1. The space of continuous paths 90
7.2. Existence of weak solutions by change of measure and time 92
7.3. Existence and uniqueness: Lipschitz continuous coefficients 96
7.4. Uniqueness by local times: Hölder continuous coefficients in one-dimension 102
8. Selected problems and solutions 105
1
2 BENJAMIN FEHRMAN
1. Preliminaries
In this course, we will assume basic concepts from measure theory including the notions of a
sigma algebra, a measure, and a measure space. Furthermore, we will assume the foundational
theorems of Lebesgue integration, including the monotone convergence theorem, Fatou’s lemma,
and the dominated convergence theorem. We will assume a familiarity with martingales in discrete
in continuous time, but all necessary concepts will be reviewed in the notes.
STOCHASTIC DIFFERENTIAL EQUATIONS 3
1.1. Probability spaces, random variables, and independence. A probability space is the
basic mathematical object used to describe random phenomena.
Definition 1.1. A probability space is a measure space (Ω, F, P) with total measure one—that is, it
is a set Ω, a sigma algebra F on Ω, and a nonnegative measure P on (Ω, F) that satisfies P(Ω) = 1.
Given a probability space (Ω, F, P), measurable subsets A ∈ F are called events. The simplest
nontrivial example is the two-point space Ω = {H, T } describing a single flip of a fair coin. The
set Ω is equipped with its power set F = P(Ω) for the sigma algebra, and it is equipped with the
probability measure P defined uniquely by
1
P[{H}] = P[{T }] = .
2
The event A = {H} represents an outcome of “heads,” and the event A = {T } represents the
outcome “tails.” The probability of both events is defined by their measure under P, and so both
events have probability 1/2.
A more expressive probability space, and one that is suitable for many purposes, is the unit
interval Ω = [0, 1] equipped with its Borel sigma algebra and Lebesgue measure. As an exercise,
show that on this probability space there exists a sequence of independent, identically distributed
random variables {Xn }n∈N satisfying
1
P[Xn = 1] = P[Xn = −1] = .
2
The functions Xn : Ω → R are random variables that represent the outcome of a coin-flip, where
we say that the nth flip is heads if Xn = 1 and is tails if Xn = −1. So, what we observe is that the
probability space [0, 1] is rich enough to model a countable infinity of independent coin-flips.
Definition 1.2. A real-valued random variable on a probability space (Ω, F, P) is a measurable
function X : (Ω, F) → (R, B(R)). That is,
X −1 (B) = {ω ∈ Ω : X(ω) ∈ B} ∈ F for every B ∈ B(R),
where B(R) is the Borel sigma-algebra of R.
Consider now a probability space that represents flipping a fair coin twice. The set is
Ω = {(H, H), (H, T ), (T, H), (T, T )},
the sigma algebra F = P(Ω) is the power set of Ω, and the measure P satisfies P[{ω}] = 1/4 for
every ω = (ω1 , ω2 ) ∈ Ω. Then consider the two random variables
1 if w1 = H, 1 if w2 = H,

X1 (ω1 , ω2 ) = and X2 (ω1 , ω2 ) =
− 1 if w1 = T, − 1 if w2 = T.
The sigma algebra generated by the random variables X1 is the smallest sub-sigma algebra σ(X1 ) ⊆
F with respect to which X1 is measurable. Precisely,
σ(X1 ) = {∅, Ω, {(H, H), (H, T )}, {(T, H), (T, T )}}.
Similarly, the sigma algebra σ(X2 ) is
σ(X2 ) = {∅, Ω, {(H, H), (T, H)}, {(H, T ), (T, T )}}.
The important thing to notice here is that these sigma algebras encode the dependencies of the
random variables X1 and X2 . The value of X1 only depends on the first flip, and so too do the
the sets in its sigma algebra. Similarly, the values of X2 depend only on the second flip, and the
same is true for the sets in its sigma algebra. In this way, we view the sigma algebra of a random
variables as carrying the “information” or the “dependencies of” the random variable.
4 BENJAMIN FEHRMAN
Definition 1.3. Let X : Ω → R be a random variable defined on a probability space (Ω, F, P).
The sigma algebra generated by X is the sigma algebra σ(X) ⊆ F defined by
σ(X) = {X −1 (B) : B ∈ B(R)}.
That is, σ(X) ⊆ F is the smallest sigma algebra on Ω with respect to which X is measurable.
Intuitively the random variables X1 and X2 defined above are independent. Information con-
cerning the first coin-flip tells us nothing about the outcome of the second. Mathematically we
make this notion precise through the notion of independence. Two events are independent if the
probability of both of them happening is the product of their probabilities:
1 1 1
P[{both flips heads}] = = · = P[{first flip heads}] · P[{second flip heads}].
4 2 2
In terms of events, if A = {first flip heads} = {(H, H), (H, T )} and B = {second flip heads} =
{(H, H), (T, H)} then A ∩ B = {both flips heads} = {(H, H)} and the above equality becomes
P[A ∩ B] = P[A]P[B].
This defines what it means for two events to be independent. We say that two random variables
are independent if their sigma algebras are independent.
Definition 1.4. Let (Ω, F, P) be a probability space. Two events A, B ∈ Ω are independent if
P[A ∩ B] = P[A]P[B].
Two random variables X1 , X2 : Ω → R are independent if
P[A ∩ B] = P[A]P[B] for every A ∈ σ(X1 ) and B ∈ σ(X2 ).
That is, two random variables are independent if their sigma algebras are independent.
Show that if X1 , X2 : Ω → R are bounded, independent random variables then
E[X1 X2 ] = E[X1 ]E[X2 ],
R
where the expectation E of a random variable Y is defined as the Lebesgue integral Ω Y dP. How-
ever, show that there exist bounded random variables X1 , X2 : Ω → R defined on some probability
space (Ω, F, P) which satisfy
E[X1 X2 ] = E[X1 ]E[X2 ],
but which are not independent. The point is that the expectation is rather course information
that tells you only about the averages of these random variables. Independence is a much finer
condition, which states that all of the information or dependencies of X1 are independent of those
of X2 .
1.2. Conditional expectation. Returning to the probability space describing two flips of a fair
coin, the conditional probability answers the following types of question. Given that the first flip is
heads what is the probability that both flips are heads? That is, what is the probability that both
flips are heads conditioned on the event that the first flip is heads? The answer should be 1/2, since
if we know that the first flip is heads the only randomness lies in the second flip. Mathematically,
P[{both flips heads and first flip heads}]
P[{both flips heads}|{given first flip heads}] = ,
P[{first flip heads}]
or,
P[{(H, H)} ∩ {(H, H), (H, T )}]
P[{(H, H)}|{given (H, H), (H, T )}] =
P[{(H, H), (H, T )}]
P[{(H, H)}]
=
P[{(H, H), (H, T )}]
1/4 1
= = .
1/2 2
In general, for an event B ∈ F satisfying P[B] > 0, we define the conditional probability of an
event A ∈ F with respect to B as
P[A ∩ B]
P[A|B] = .
P[B]
One could in fact define a new probability space (ΩB , FB , PB ) where ΩB = B, FB = F ∩ B, and
for every Ã ∈ FB satisfying Ã = A ∩ B for some A ∈ F, the new probability measure PB is defined
by
P[A ∩ B]
PB [Ã] = .
P[B]
In this new probability space, we are living in a world where the event B is guaranteed to occur,
and the corresponding probabilities are adjusted accordingly. Observe in particular that, if A, B
are independent events, then P[A|B] = P[A]. That is, the knowledge that the event B will occur
yields no information about the likelihood of A, as we should expect.
Conditional expectation generalizes the notion of conditional probability to random variables.
The sigma algebra σ(X) of a random variable X encodes the information or dependencies of the
random variable. The conditional expectation of a random variable Y with respect to X is our
best guess for Y using only the information or dependencies of X.
Definition 1.5. Let X : Ω → R be a random variable and let Y : Ω → R be an integrable random
variable on a probability space (Ω, F, P). The conditional expectation of Y given X is the unique
σ(X)-measurable random variable E[Y |X] or E[Y |σ(X)] that satisfies, for every A ∈ σ(X),
Z Z
E[Y : A] = Y dP = E[Y |X] dP = E[E[Y |X] : A].
A A
We can also take conditional expectations with respect to a sigma-algebra.

Definition 1.6. Let Y : Ω → R be an integrable random variable on a probability space (Ω, F, P)
and let G ⊆ F be a sub-sigma algebra. The conditional expectation of Y given G is the unique
G-measurable random variable E[Y |G] that satisfies, for every A ∈ G,
Z Z
E[Y : A] = Y dP = E[Y |G] dP = E[E[Y |G] : A].
A A
In this way, the conditional expectation of Y with respect to a random variable X is the conditional
expectation of Y with respect to the sub-sigma algebra σ(X).
6 BENJAMIN FEHRMAN
The existence and uniqueness of the conditional expectation follows from the integrability of Y
and the Radon-Nikodym theorem. Precisely, given a sub-sigma algebra G R⊆ F, the conditional
expectation is the Radon-Nikodym derivative of the finite measure µY (A) = A Y dP defined on G.
Uniqueness is a consequence of the fact that the conditional expectation must be G-measurable,
and that every two candidates E[Y |G] and Ẽ[Y |G] satisfy
Z
E[Y |G] − Ẽ[Y |G] dP = 0 for every A ∈ G.
A
The following properties are a consequence of the definition of the conditional expectation.
Proposition 1.7. Let Y1 , Y2 : Ω → R be integrable random variables on a probability space (Ω, F, P)
and let G1 , G2 ⊆ F be sub-sigma-algebras.
(a.) Linearity: for every α, β ∈ R,
E[αY1 + βY2 |G1 ] = αE[Y1 |G1 ] + βE[Y2 |G1 ].
(b.) Order-preserving: if Y1 ≤ Y2 almost surely then, almost surely,
E[Y1 |G1 ] ≤ E[Y2 |G1 ].
(c.) Tower property: if G1 ⊆ G2 then
E[E[Y1 |G2 ]|G1 ] = E[Y1 |G1 ].
Show that a random variable is measurable with respect to the trivial sigma algebra {∅, Ω} if and
only if that random variable is constant. Then, using the definition of the conditional expectation,
show that for every integrable random variable Y we have that E[Y |{∅, Ω}] = E[Y ] is constant.
Taking this one step further, if Y is independent of a sub-sigma algebra G in the sense that σ(Y ) and
G are independent sigma algebras, then the conditional expectation E[Y |G] should yield essentially
no information. The following proposition shows that this is the case: if Y is independent of G then
E[Y |G] = E[Y ] is constant and offers no information beyond what we deduce from the trivial sigma
algebra. Conversely, if Y is G-measurable, then G contains all of the dependencies or information
of Y and the conditional expectation E[Y |G] = Y describes Y exacty.
Proposition 1.8. Let Y, Z : Ω → R be integrable random variables on a probability space (Ω, F, P),
let G ⊆ F, and assume that the product Y Z : Ω → R is integrable.
(a.) Independence: if Y is independent of G then
E[Y |G] = E[Y ].
(b.) Factoring: If Y is G-measurable then
E[Y Z|G] = Y E[Z|G].
In particular, taking Z = 1 we have E[Y |G] = Y .
The following proposition is a version of Jensen’s inequality for the conditional expectation.
Observe that both sides of this inequality are random variables, and that the inequality holds for
almost every ω ∈ Ω.
Proposition 1.9. Let Y : Ω → R be an integrable random variable on a probability space (Ω, F, P),
let G ⊆ F be a sub-sigma algebra, and let f : R → R be a convex function. Then, almost surely,
f (E[Y |G]) ≤ E[f (Y )|G].
Finally, the conditional expectation is continuous with respect to L1 -convergence and satisfies
versions of the monotone convergence theorem, Fatou’s lemma, and the dominated convergence
theorem. Full details can be found in the Math B8.1 course notes, which are available on the
course website.
Proposition 1.10. Let {Yn : Ω → R}n∈N and Y : Ω → R be integrable random variables on a

probability space (Ω, F, P) and let G ⊆ F be a sub-sigma algebra. If Yn → Y in L1 (Ω) as n → ∞
then, as n → ∞,
E[Yn |G] → E[Y |G] in L1 (Ω).
8 BENJAMIN FEHRMAN
2. Martingales
A martingale is a mathematical model for a fair game, or a stochastic process where knowledge
of the past does not allow the player to predict the future. Think for example of flipping a coin.
The outcome of the first n flips does not reveal any information about the outcome of flip (n + 1).
More precisely, on a probability space (Ω, F, P), let {Xk }k∈N be independent, identically distributed
(i.i.d.) random variables which satisfy
1
P(X1 = −1) = P(X1 = 1) = .
2
We view 1 as representing a “heads” and −1 as representing “tails.” For every n ∈ N define the
sum
(2.1) Sn = X1 + . . . + Xn = (# “heads”) − (# “tails”).
For every n ∈ N let Fn ⊆ F be the sigma algebra generated by the random variables X1 , . . . , Xn ,
(2.2) Fn = σ(X1 , . . . , Xn ).
The sequence {Sn }n∈N forms a martingale with respect to the filtration {Fn }n∈N in the sense that
our best guess for the random variable Sn+1 given the information Fn is Sn . Or, in terms of the
conditional expectation,
E[Sn+1 |Fn ] = E[Xn+1 |Fn ] + E[Sn |Fn ] = E[Xn+1 ] + Sn = Sn .
A random process is called a submartingale if Sn provides a lower bound for the conditional ex-
pectation E[Sn+1 |Fn ], and a supermartingale if Sn provides an upper bound for the conditional
expectation. That is, roughly speaking, submartingales are “increasing” whereas supermartingales
are “decreasing.”
The process {Sn }n∈N is known as the simple random walk, and it provides our first example of a
martingale in discrete time. However, in this course, we will primarily be interested in martingales
in continuous time. Our first and most important example is a Brownian motion.
Definition 2.1. A Brownian motion on a probability space (Ω, F, P) is a stochastic process B : Ω×
[0, ∞) → R that satisfies the following four properties.
(a.) Beginning at zero: almost surely,
B0 = 0.
(b.) Normal distribution: for every s < t ∈ [0, ∞),
Bt − Bs is a normally distributed variable with mean zero and variance t − s.
(c.) Independent incrementes: for every N ∈ N and t0 ≤ t1 ≤ t2 ≤ . . . ≤ tn ∈ [0, ∞), the random
variables
{Bti − Bti−1 }N
i=1 are mutually independent.
(d.) Continuous sample paths: almost surely the map
t ∈ [0, ∞) 7→ Bt (ω) is continuous.
Not every probability space (Ω, F, P) is rich enough to support a Brownian motion, and the
details of its construction and an overview of its properties can be found in the Math B8.2 notes
provided on the course website. What we observe for now is that, if for each t ∈ [0, ∞) we define
Ft = σ(Bs : s ∈ [0, t]) to be the sigma algebra generated by the Brownian motion up to time t, then
we have for every s ≤ t ∈ [0, ∞),
E[Bt |Fs ] = E[Bt − Bs |Fs ] + E[Bs |Fs ].
Since (Bt − Bs ) is independent of Fs , and since Bs is Fs -measurable,
E[Bt |Fs ] = E[Bt − Bs ] + Bs = Bs .
This is the continuous martingale property, which states that our best guess for a Brownian motion
at time t ≥ s using only the dependencies or information of the Brownian motion up to time s is
Bs . Knowledge of the past does not allow you to predict the future.
10 BENJAMIN FEHRMAN
2.1. Filtrations, martingales, and stopping times. Let (Ω, F) be a measurable space, which
is to say that Ω is a set equipped with a sigma algebra F of subsets. We will view sigma algebras
as carrying information, where in the above the sigma algebra Fn defined in (2.2) carries the
information of the random variables X1 , . . . , Xn . This is to say that, given all of the information
of Fn , we can predict the random variables X1 , . . . , Xn exactly. Or, in terms of the conditional
expectation, for every m ∈ {1, . . . , n},
E[Xm |Fn ] = Xm .
A filtration is an increasing family of sub-sigma algebras, where the fact that the sub-sigma algebras
are increasing implies that the amount of information carried by the sub-sigma algebras is increas-
ing. We will be primarily interested in filtrations indexed by a continuous parameter t ∈ [0, ∞) or
a discrete parameter n ∈ N.
Definition 2.2. Let (Ω, F) be a measurable space and let I ⊆ R. A filtration on (Ω, F) indexed
by I is an increasing family of sub-sigma algebras {Ft }t∈I . That is, for every s ≤ t ∈ I we have
Fs ⊆ Ft ⊆ F.
Filtrations are often generated by a stochastic process, such as in (2.2) where the sub-sigma
algebra Fn was generated by the first n coin-flips. We will be primarily interested in continuous
processes indexed by t ∈ [0, ∞) or discrete processes indexed by n ∈ N.
Definition 2.3. Let (Ω, F, P) be a probability space, let (Θ, G) be a measurable space, and let
I ⊆ R. A stochastic process indexed by I taking values in Θ is a family of bimeasurable maps
{Xt : Ω → Θ}t∈I .
Every stochastic process {Xt }t∈[0,∞) generates a natural filtration
{Ft = σ(Xs : s ∈ [0, t])}t∈[0,∞) ,
since it follows by the measurability of the {Xt }t∈[0,∞) and by definition that Fs ⊆ Ft ⊆ F whenever
s ≤ t ∈ [0, ∞). We will often view the parameter t as a “time”-variable, and the sigma algebra Ft
as carrying the information of {Xs }s∈[0,∞) up to time t. This was the case in the discrete example
(2.2), where the time parameter was discrete, and was the case for Brownian motion.
A martingale is a stochastic process defined on a probability space with respect to a filtration. It
is important to understand that a martingale is only ever a martingale with respect to a filtration.
If the filtration changes, the martingale need not remain a martingale, and if there is no filtration
there is absolutely no martingale. We will be primarily interested in martingales defined over
t ∈ [0, ∞) or n ∈ N, and taking values some Euclidean space.
Definition 2.4. Let (Ω, F, P) be a probability space, let (Θ, G) be a measurable space, let I ⊆ R,
and let {Ft }t∈I be a filtration on (Ω, F, P).
(a.) A Θ-valued stochastic process (Mt )t∈I indexed by I is a martingale with respect to {Ft }t∈I if
(i) For every t ∈ I,
E[|Mt |] < ∞.
(ii) For every t ∈ I, Mt is Ft -measurable.
(iii) For every s ≤ t ∈ I,
E[Mt |Fs ] = Ms .
(b.) A Θ-valued stochastic process (Mt )t∈I indexed by I is a submartingale with respect to {Ft }t∈I
if
E[Mt+ ] = E[max(Mt , 0)] < ∞.

E[Mt |Fs ] ≥ Ms .
(c.) A Θ-valued stochastic process (Mt )t∈I indexed by I is a supermartingale with respect to {Ft }t∈I
if the stochastic process (−Mt )t∈I is a submartingale. That is, if
E[Mt− ] = E[min(Mt , 0)] > −∞.
E[Mt |Fs ] ≤ Ms .
Observe that a process (Mt )t∈[0,∞) is a martingale with respect to a filtration {Ft }t∈[0,∞) if
and only if (Mt )t∈[0,∞) is both a submartingale and a supermartingale with respect to {Ft }t∈[0,∞) .
Intuitively, the process (Mt )t∈[0,∞) is a martingale if our best guess for Mt given the information
Fs is Ms . A process is a submartingale if Ms provides a lower bound for our best guess of Mt given
Fs , and is a supermartingale if Ms provides an upper bound for our best guess of Mt given Fs .
Loosely speaking, we can therefore view submartingales as “increasing” and supermartingales as
“decreasing.”
An important concept in the study of martingales, and stochastic differential equations, is the
notion of a stopping time. Consider the example above, for i.i.d. random variables {Xk }k∈N which
satsify
1
P[X1 = 1] = P[X1 = −1] = ,
2
for the sigma algebras {Fn }n∈N defined by
Fn = σ(Xm : m ∈ {1, . . . , n}),
and for random variables {Sn }n∈N defined by S0 = 0 and
Sn = X1 + . . . + Xn .
The random variables {Sn }n∈N define a random walk on Z, and for a fixed integer z ∈ Z we will
be interested in the first time that the walk hits z. This is defined by the random variable
(2.3) Tz = inf{n ∈ N0 : Sn = z},
which is an example of a hitting time. The essential property of the variable Tz is that to know if
Tz ≤ n we only need to know information about the random variables X1 , . . . , Xn . This is to say
that
{Tz ≤ n} ∈ Fn ,
which is the defining property of a stopping time.
Definition 2.5. Let (Ω, F, P), let I = N0 or I = [0, ∞), and let {Ft }t∈I be a filtration on (Ω, F).
A stopping time is a measurable map T : Ω → I such that, for every t ∈ I,
{T ≤ t} ∈ Ft .
If we think of a stopping time T as an alarm clock, the condition {T ≤ t} ∈ Ft implies we don’t
need to look into the future to determine whether or not the alarm has rung. Hitting times like
(2.3) above are the most common and most important examples of stopping times that we will
encounter. Stopping times come with an associated sigma algebra, which consists of those events
that are Ft -measurable conditioned on the event {T ≤ t}.
Definition 2.6. Let (Ω, F, P) be a probability space, let I = N0 or I = [0, ∞), let {Ft }t∈I be a
filtration on (Ω, F), and let T : Ω → I be a stopping time. We define the sub-sigma algebra FT ⊆ F
by
FT = {A ∈ F : (A ∩ {T ≤ t}) ∈ Ft for every t ∈ I}.
12 BENJAMIN FEHRMAN
The following proposition proves that a stopped (sub/super) martingale remains a (sub/super)
martingale.
Proposition 2.7. Let (Ω, F, P) be a probability space, let I = N0 or I = [0, ∞), let {Ft }t∈I be a
filtration on (Ω, F), let (Mt )t∈I be an Ft -(sub/super) martingale, and let T : Ω → I be a stopping
time. Then the process
(MtT )t∈I = (Mt∧T )t∈I ,
is a (sub/super) martingale.
Proof. Let I = N0 and suppose that (Mn )n∈N0 is a martingale. For every n ∈ N0 ,
n
X
(2.4) MnT = MT ∧n = Mn 1{T >n} + Mk 1{T =k} .
k=1
Equation (2.4) proves that
n
X
T
E[ Mn ] ≤
E[|Mk |] < ∞,
k=1
and that MnT is Fn -measurable because every function on the righthand side of (2.4) is Fn -
measurable. Finally, by (2.4), for every n ∈ N0 ,
n+1
X
T
E[Mn+1 |Fn ] = E[Mn+1 1{T >n+1} |Fn ] + E[Mk 1{T =k} |Fn ]
k=1
n
X
= E[Mn+1 1{T >n} |Fn ] + E[Mk 1{T =k} |Fn ].
k=1
Since Mk 1{T =k} and 1{T >n+1} are Fn -measurable, the
n
X
T
E[Mn+1 |Fn ] = E[Mn+1 |Fn ]1{T >n} + Mk 1{T =n}
k=1
n
X
= Mn 1T >n + Mk 1{T =n}
k=1
= Mn ,
which completes the proof.
Stopping terms are sometimes referred to local times. We will encounter processes that are not
martingales, but which are locally a martingale when appropriately stopped. For example, let
{Bt }t∈[0,∞) be a standard Brownian motion, and let T1 denote the stopping time T1 = inf{t ∈
[0, ∞) : Bs = 1}. Since T1 is almost surely finite, the process
(
BT1 ∧( t ) if t ∈ [0, 1),
1−t
Wt =
1 if b ∈ [1, ∞),
is continuous, but it is not a martingale because
(
0 if t ∈ [0, 1),
E[Wt ] =
1 if t ∈ [1, ∞).
However, Proposition 2.7 proves that, for every k ∈ N, for the stopping times
τ−k = inf{t ∈ [0, ∞) : Wt = −k},
the stopped processes {W τk }k∈N are martingales. The process W is called a local martingale.
Definition 2.8. Let (Ω, F, P) be a probability space, let I = N0 or I = [0, ∞), a let {Ft }t∈I be a
filtration on (Ω, F). An Ft -adapted process M is called a local martingale if there exist a stopping
times {τk }k∈N such that
(i) For every k ∈ N,
τk ≤ τk+1 almost surely.
(ii) Almost surely, as k → ∞,
τk → ∞.
(iii) For every k ∈ N, the stopped process M τk 1{τk >0} is an Ft -martingale.
14 BENJAMIN FEHRMAN
2.2. The optional stopping theorem (discrete). In this section, we will first prove the optional
stopping theorem for discrete martingales, which we will use to prove the optional stopping theorem
for continuous martingales in Proposition 2.35 below. Let (Sn )n∈N be a simple random walk, as
defined in (2.1). This can be viewed as the outcome of a betting strategy, where for each coin flip
we bet £1 that the outcome will be “heads.” We thereby gain £1 for every “heads” and lose £1
for every “tails,” and the random variables {Sn }n∈N quantify our net profit/loss after n flips.
If we have an infinite bankroll, we can almost surely achieve an arbitrarily large profit. For every
m ∈ N, we simply play until the random walk {Sn }n∈N reaches m, which is an event of probability
one. In terms of stopping times, if
Tm = inf{n ∈ N0 : Sn = m},
then Tm is almost surely finite and we have
m = E[STm ] > E[S0 ] = 0.
However, if we instead have a finite bankroll £N it is no longer possible to play interminably, since
we’ll go bankrupt at the first time n ∈ N that Sn = −N . In this case, the stopping times {Tm }m∈N
must be replaced by
Tm,N = inf{n ∈ N0 : Sn = m or Sn = −N }.
A simple case of the optional stopping theorem below will prove that, for every m, N ∈ N,
E[STm,N ] = mP[STm,N = m] − N P[STm,N = −N ] = E[S0 ] = 0,
which since
P[STm,N = m] + P[STm,N = −N ] = P[Tm,N < ∞] = 1,
implies that
m
(2.5) P[STm,n = −N ] = .
N +m
This is a version of the gambler’s ruin estimate, which states that a gambler who repeatedly stakes
their bankroll to earn any profit, no matter how modest, will eventually lose everything. Indeed, it
follows from (2.5) that a gambler with an initial amount £N who repeatedly stakes the entirety of
their bankroll, including the profit from their previous bets, to earn a fixed profit £m goes bankrupt
with probability
∞
X N m
· = 1.
N + km N + (k + 1)m
k=0
The optional stopping theorem states that, if the gambler is playing a fair game, there is no strategy
(i.e. no stopping time) that gives them an advantage. The proof will use the concept of a predictable
process, which is a process for which the past determines the future.
Definition 2.9. Let (Ω, F) be a measurable space and let {Fn }n∈N0 be a filtration on (Ω, F). An
Fn -stochastic process (Hn )n∈N is called predictable if, for every n ∈ N,
Hn is Fn−1 -measurable.
In the following proposition, we define the discrete integral of a predictable process with respect
to a discrete martingale. The essential point is that the integral of a predictable process with
respect to a martingale is again a martingale. A fact we will see again in the context of integrals
with respect to continuous martingales.
Proposition 2.10. Let (Ω, F, P) be a probability space, let (Fn )n∈N0 be a filtration on (Ω, F), let
(Mn )n∈N0 be an Fn -martingale, and let (Hn )n∈N be a bounded Fn -predictable process. Then, defined
inductively, (
(H · M )0 = 0
(H · M )n = (H · M )n−1 + Hn (Mn − Mn−1 ) if n ∈ N,
is a Fn -martingale. If in addition (Hn )n∈N is nonnegative, then ((H · M )n )n∈N0 is a (sub/super)

martingale if (Mn )n∈N0 is a (sub/super)-martingale.
Proof. We will carry out the details in the case that (Mn )n∈N0 is an Fn -martingale. For every
n ∈ N,
X n
E[|(H · M )n |] ≤ max kHk kL∞ (Ω) E[|Mk |] < ∞,
k∈{1,...,n}
k=1
and since by definition
N
X
(H · M )n = Hk (Mn − Mn−1 ),
k=1
it follows that (H · M )n is Fn -measurable. Finally, the predictability of (Hk )k∈N and the martingale
property prove that, for every n ∈ N0 ,
n+1
X
E[(H · M )n+1 |Fn ] = E[Hk (Mk − Mk−1 )|Fn ]
k=1
n
X
= Hn E[(Mn+1 − Mn )|Fn ] + Hk (Mk − Mk−1 )
k=1
= (H · M )n ,
We now prove the optional stopping theorem in discrete time for bounded stopping times. We
will extend this theorem to continuous time martingales in the sections to follow. The result will
rely on the lemma below, which proves that is S ≤ T are stopping times then FS ⊆ FT .
Lemma 2.11. Let (Ω, F, P) be a probability space, let I = N0 or I = [0, ∞), let (Ft )t∈I be a
filtration on (Ω, F), and let S ≤ T be Ft -stopping times. Then FS ⊆ FT .
Proof. Let A ∈ FS . It is necessary to prove that, for every t ∈ I,
A ∩ {T ≤ t} ∈ Ft .
Since S ≤ T , it follows that
(A ∩ {T ≤ t}) = (A ∩ {S ≤ t}) ∩ ((A ∩ {S ≤ t}) \ {T > t}) ,
where the first term on the righthand side is in Ft by definition of FS , the second term is in Ft
because Ft is a sigma algebra and (A ∩ {S ≤ t}) and {T > t} = {T ≤ t}c are in Ft , and the union
is in Ft because Ft is a sigma algebra. This completes the proof.
Theorem 2.12. Let (Ω, F, P) be a probability space, let {Fn }n∈N0 be a filtration on (Ω, F), let
(Mn )n∈N0 be an Fn - (sub/super) martingale, and let σ ≤ τ be two bounded stopping times. Then,
(2.6) E[Mτ |Fσ ] = Mσ . (≥ / ≤)
In particular,
(2.7) E[Mτ ] = E[Mσ ]. (≥ / ≤)
Proof. We will carry out the details in the case that (Mn )n∈N0 is a martingale. Let σ ≤ τ be two
bounded stopping times, and let m ∈ N be such that, for almost every ω ∈ Ω,
(2.8) σ(ω) ≤ τ (ω) ≤ m.
We will first prove (2.7). Let (Hn )n∈N0 denote the process H0 = 0 and
Hn = 1{σ<n≤τ } = 1{n≤τ } − 1{n≤σ} .
16 BENJAMIN FEHRMAN
Observe that, for every n ∈ N,

{n ≤ τ } = (Ω \ {τ ≤ n − 1}) ∈ Fn−1 ,
and similarly {n ≤ τ } ∈ Fn−1 . Therefore, for every n ∈ N0 , the variable Hn is Fn−1 -measurable,
and (Hn )n∈N0 is a Fn -predictable process. Therefore, by Proposition 2.10, the process defined by
(H · M )0 = 0 and,for n ∈ N by
Xn
(2.9) (H · M )n = Hk (Mk − Mk−1 ) = Mτ ∧n − Mσ∧n ,
k=1
is a martingale. In combination (2.8) and (2.9) prove that, for every n ≥ m,
(H · M )n = Mτ − Mσ .
Therefore, since (H · M ) is a martingale,
0 = E[(H · M )0 ] = E[(H · M )m ] = E[Mτ ] − E[Mσ ],
which proves (2.7). It remains to prove (2.6).
Let B ∈ Fσ . It suffices to prove that
E[Mτ : B] = E[Mσ : B].
For m ∈ N defined in (2.8), define the stopping times τB , σB : Ω → N0 by
( (
τ (ω) if ω ∈ B, σ(ω) if ω ∈ B,
τB (ω) = c and σB (ω) =
m if ω ∈ B , m if ω ∈ B c .
Indeed, by definition of the sigma algebra Fσ and Lemma 2.11, for every n ∈ {0, 1, 2, . . . , m − 1},
{τB ≤ n} = ({τ ≤ n} ∩ B) ∈ Fn ,
and for every n ≥ m,
{τB ≤ n} = Ω ∈ Fn ,
and the identical argument proves that σB is a stopping time. Therefore, since τB and σB are
bounded stopping times, it follows from (2.7) that
(2.10) E[MτB ] = E[MτB : B] + E[MτB : B c ] = E[MσB : B] + E[MσB : B c ] = E[MσB ].
Since it follows by definition that
(2.11) E[MτB : B] = E[Mτ : B] and E[MσB : B] = E[Mσ : B],
and that
(2.12) E[MτB : B c ] = E[Mm : B c ] = E[MσB : B c ],
it follows from (2.10), (2.11), and (2.12) that
E[Mτ : B] = E[Mσ : B],
which completes the proof of (2.6), and therefore the proof.
The following corollary extends Theorem 2.12 to stopping times that are not necessarily bounded.
We will extend these results to continuous martingales in the sections to follow.
Corollary 2.13. Let (Ω, F, P) be a probability space, let {Fn }n∈N0 be a filtration on (Ω, F), let
(Mn )N0 be an Fn - (sub/super) martingale, and let σ ≤ τ be two almost surely finite stopping times.
Assume that |Mτ | and |Mσ | are integrable, and assume that
(2.13) lim E[|Mn | : τ > n] = 0.
n→∞
Then,
(2.14) E[Mτ |Fσ ] = Mσ . (≥ / ≤)
In particular,
(2.15) E[Mτ ] = E[Mσ ]. (≥ / ≤)
Proof. Let B ∈ Fσ . Theorem 2.12 proves that, for every n ∈ N, since the stopping times τ ∧ n and
σ ∧ n are bounded with σ ∧ n ≤ τ ∧ n, and since B ∩ {σ ≤ n} ∈ Fσ∧n ,
E[Mτ ∧n : B, σ ≤ n] = E[Mτ : B, τ ≤ n, σ ≤ n] + E[Mn : B, τ > n, σ ≤ n]
= E[Mσ : B, σ ≤ n]
= E[Mσ∧n : B, σ ≤ n].
Since
|E[Mn : B, τ > n, σ ≤ n]| ≤ E[|Mn | : τ > n],
it follows from (2.13) that
lim |E[Mn : B, τ > n, σ ≤ n]| = 0.
n→∞
The dominated convergence theorem, the fact that σ and τ are almost surely finite, and the
integrability of |Mτ | and |Mσ | prove that, after passing to the limit n → ∞,
E[Mτ : B] = E[Mσ : B].
This completes the proof of (2.14), which implies (2.15). This completes the proof.
18 BENJAMIN FEHRMAN
2.3. Uniform integrability. In this section, will now explain conditions which can be used to
upgrade the almost sure convergence or convergence in probability to strong convergence in L1 (Ω).
This will require the notion of uniform integrability, which when combined with almost sure con-
vergence (in fact, convergence in probability) implies convergence in L1 (Ω). This is a version of
Vitalli’s convergence theorem below.
Definition 2.14. Let (Ω, F, P) be a probability space. Let A be a set, and let {Xα }α∈A be a family
of random variables. The family {Xα }α∈A is uniformly integrable if the following two conditions
are satisfied.
(i) L1 -boundedness:
sup E[|Xα |] < ∞.
α∈A
(ii) No concentration:
lim sup E[|Xα | : {|Xα | ≥ K}] = 0.
K→∞ α∈A
Remark 2.15. The second property of Definition 2.14 guarantees that the family of random
variables {Xα }α∈A do not concentrate their mass in the following sense. Let ρ ∈ C∞
c (R) be a
nonnegative, smooth function which satisfies that
Z
ρ(x) dx = 1,
R
and for every ε ∈ (0, 1) define ρε (x)
= −1
ε ρ(x/ε).
The functions form a Dirac sequence in the sense
that, for every f ∈ C∞ (R),
Z Z
ε
lim f (x)ρ (x) dx = f (x)δ0 (x) dx = f (0),
ε→0 R R
where δ0 is the Dirac delta distribution at zero. That is, as ε → 0, as distributions,
(2.16) ρε * δ0 .
Since the distribution δ0 is not an L1 -function, we conclude that the family {ρε }ε∈(0,1) is not
precompact in L1 (R) despite the fact that, for every ε ∈ (0, 1),
kρε kL1 (R) = 1,
and despite the fact that ρε → 0 almost everywhere on R, as ε → 0. This lack of compactness is
due to the fact that the functions {ρε }ε∈(0,1) concentrate the entirety of their mass at the origin,
as ε → 0. In particular, they do not satisfy condition (ii) of Definition 2.14.
Remark 2.16. Definition 2.14 is the probabilistic version of uniform integrability, which is equiv-
alent to standard definition of uniform integrability. The reason we use this definition is that, as
we will see below, it is very convenient to verify when dealing with families of conditional expec-
tations. Definition 2.14 also relies on the fact that a probability space Ω is a finite measure space
with P(Ω) < ∞. On an infinite measure space, like Rd , it is also necessary to impose a tightness
condition that precludes mass from escaping to infinity. For example, on R and for the standard
convolution kernel ρ above the family of translates {ρ(· − n)}n∈Z converges pointwise to zero and
satisfies the conditions of Definition 2.14, but this family does not converge in L1 .
The following theorem is Vitalli’s convergence theorem specialized to the case of a probability
space.
Theorem 2.17. Let (Ω, F, P) be a probability space. Let I = N0 or let I = [0, ∞). Let {ft }t∈I be
family of random variables which satisfy the following two properties.
(i) Uniform integrability: the family {ft }t∈I is uniformly integrable.
(ii) Convergence in probability: there exists a random variable f such that, for every ε ∈ (0, 1),
lim P[|ft − f | > ε] = 0.
t→∞
Then f ∈ L1 (Ω) and, as t → ∞,
lim E[|ft − f |] = 0.
t→∞
That is, as t → ∞, the {ft }t∈I converge to f in L1 (Ω).
20 BENJAMIN FEHRMAN
2.4. The optional stopping theorem (continuous). In this section, we will prove a version of
the optional stopping theorem for uniformly integrable martingales in continuous time. As in the
case of martingales in discrete time, it is necessary to impose some conditions on the stopping time.
For instance, if (Bt )t∈[0,∞) is a standard Brownian motion then
T = inf{s ∈ [0, ∞) : Bs ≥ 1},
is an almost surely finite stopping time. However, in this case the conclusion of the optional
stopping theorem fails, since
E[BT ] = 1 6= 0 = E[B0 ].
We will therefore first prove the optional stopping theorem for bounded stopping times. We first
state a useful lemma which states that a family consisting of a collection of conditional expectations
of a fixed random variable is uniformly integrable.
Lemma 2.18. Let (Ω, F, P) be a probability space. Let A be a set, let {Gα }α∈A , and let X ∈ L1 (Ω).
Then the family
{E[X|Gα ]}α∈A ,
is uniformly integrable.
Proof. We will first prove property (i) of Definition 2.14. Since Jensen’s inequality proves that, for
every α ∈ A,
(2.17) P |E[X|Gα ]| ≤ E[|X| |Gα ],
it follows from X ∈ L1 (Ω) that, for every α ∈ A,
E[|E[X|Gα ]|] ≤ E[E[|X| |Gα ]] = E[|X|] < ∞.
Therefore,
(2.18) sup E[|E[X|Gα ]|] < ∞.
α∈A
It remains to prove property (ii). Since it follows from (2.17) that, for every α ∈ A and K ∈ (0, ∞),
{|E[X|Gα ]| ≥ K} ⊆ {E[|X| |Gα ] ≥ K},
it follows by properties of the conditional expectation and (2.17) that, for every α ∈ A and K ∈
(0, ∞),
E[|E[X|Gα ]| : {|E[X|Gα ]| ≥ K}] ≤ E[E[|X| |Gα ] : {E[|X| |Gα ] ≥ K}]
(2.19)
= E[|X| : {E[|X| |Gα ] ≥ K}].
By Chebyshev’s inequality and (2.18), for every α ∈ A and K ∈ (0, ∞),
1 1
(2.20) P[E[|X| |Gα ] ≥ K] ≤ E[E[|X| |Gα ]] ≤ E[|X|].
K K
We therefore conclude by continuity of the Lebesgue integral, X ∈ L1 (Ω), (2.19), and (2.20) that

lim sup E[|E[X|Gα ]| : {|E[X|Gα ]| ≥ K}] = 0,
K→∞ α∈A
Theorem 2.19. Let (Ω, F, P) be a probability space, let {Ft }t∈[0,∞) be a filtration on (Ω, F), let
(Mt )t∈[0,∞) be an Ft -martingale (sub/super), and let σ ≤ τ be two bounded bounded Ft -stopping
times. Then,
E[Mτ |Fσ ] = Mσ (≥ / ≤).
In particular,
E[Mτ ] = E[Mσ ] (≥ / ≤).
Proof. We will present the proof in the case that (Mt )t∈[0,∞) is a martingale. Let K ∈ N almost
surely satisfy τ ≤ M and σ ≤ M . For each k ∈ N let τk , σk : Ω → [0, ∞) be defined by
∞ ∞
X n+1 X n+1
τk = 1 n <τ ≤ n+1 and σk =
n o 1n n <σ≤ n+1 o .
2k 2 k 2k 2 k
2k 2k
n=0 n=0
The {τk }k∈N and {σk }k∈N are decreasing sequences of stopping times which satisfy τ ≤ τk and
σ ≤ σk for every k ∈ N and, almost surely as k → ∞,
(2.21) τk → τ and σk → σ.
The boundedness of τ and σ prove that τk and σk take only finitely many values and almost surely
satisfy τk ≤ K + 1 and σk ≤ K + 1. A repetition of the proof of Theorem 2.12 applied to the
stopping time τk and the constant stopping time K + 1, and the stopping time σk and constant
stopping time K + 1, proves that, for every k ∈ N,
(2.22) E[MK+1 |Fτk ] = Sτk and E[MK+1 |Fσk ] = Sσk .
Lemma 2.18 and (2.22) prove that that the families of random variables {Mτk }k∈N and {Mσk }k∈N
are uniformly integrable, and (2.21) proves that, almost surely as k → ∞,
Mτk → Mτ and Mσk → Mσ .
Vitalli’s convergence theorem therefore proves that, as k → ∞,
(2.23) Mτk → Mτ and Mσk → Mσ in L1 (Ω).
Since a repetition of the proof of Theorem 2.12 applied to the stopping time τk and σk proves that,
for each k ∈ N,
E[Mτk ] = E[Mσk ],
the convergence (2.23) proves that, for any two bounded stopping times σ ≤ τ ,
(2.24) E[Mτ ] = E[Mσ ].
Now let B ∈ Fσ and define the stopping times τB = τ 1B + K1B c and σB = τ 1B + K1B c . It
follows from Lemma 2.11, which proves that Fσ ⊆ Fτ , that τB and τσ are bounded stopping times.
Therefore, the definitions and (2.24) prove that
E[MτB ] = E[Mτ : B] + E[MK : B c ] = E[Mσ : B] + E[MK : B c ] = E[MσB ].
We therefore conclude that, for every B ∈ Fσ ,
E[Mτ : B] = E[Mσ : B].
Since Mσ is Fσ -measurable, this proves that E[Mτ |Fσ ] = Mσ and completes the proof.
In this course, we will most often be dealing with stopping times that are not bounded. The
following corollary will therefore be very useful in our applications of the optional stopping theorem.
Corollary 2.20. Let (Ω, F, P) be a probability space, let {Ft }t∈[0,∞) be a filtration on (Ω, F), let
(Mt )t∈[0,∞) be an Ft -martingale (sub/super), and let σ ≤ τ be two finite Ft -stopping times. Assume
that Mτ and Mσ are integrable and that
(2.25) lim E[|Mn | : τ > n] = 0.
n→∞
Then,
E[Mτ |Fσ ] = Mσ (≥ / ≤).
In particular,
E[Mτ ] = E[Mσ ] (≥ / ≤).
22 BENJAMIN FEHRMAN
Proof. Let B ∈ Fσ . For every n ∈ N, since σ ∧ n ≤ τ ∧ n are stopping times and since B ∩ {σ ≤
n} ∈ Fσ∧n , Theorem 2.19 proves that
E[Mτ ∧n : B, σ ≤ n] = E[Mτ : B, τ ≤ n, σ ≤ n] + E[MN : B, τ > n, σ ≤ n]
= E[Mσ : B, σ ≤ n]
= E[Mσ∧n : B, σ ≤ n].
The integrability of Mτ and Mσ , the finiteness of the stopping times, and the dominated convergence
theorem prove that
lim E[Mτ : B, τ ≤ n, σ ≤ n] = E[Mτ : B] and lim E[Mσ : B, σ ≤ n] = E[Mσ : B].
N →∞ N →∞
Assumption (2.25) proves that
lim |E[Mn : B, σ ≤ n, τ > n]| ≤ lim E[|Mn | : τ > n] = 0.
N →∞ N →∞
Therefore, for every B ∈ Fσ ,
E[Mτ : B] = E[Mσ : B],
Finally, we will see in Proposition 2.35 below that there is also a version of the optional stopping
theorem for uniformly integrable martingales that completely avoids imposing integrability condi-
tions on the stopping time itself. However, in this course, the martingales we encounter will not
usually be uniformly integrable. For this reason, and because the stopping times we encounter will
not usually be bounded, it is good to keep Corollary (2.20) in mind.
2.5. Doob’s martingale inequality. In this section, we will establish fundamental inequalities
for submartingales. In a sense that will be made more precise by the Doob-Meyer decomposition
to follow in the next section, recall that a submartingale (Xt )t∈[0,∞) is “increasing” in the sense
that, for every s ≤ t ∈ [0, ∞),
Xs ≤ E[Xt |Fs ].
This implies in particular that, for every s ≤ t ∈ [0, ∞),
E[Xs ] ≤ E[Xt ].
Doob’s martingale inequality proves that the running maximum XT∗ of a nonnegative submartingale,
definedy for every T ∈ [0, ∞) by
(2.26) XT∗ = max |Xt | ,
0≤t≤T
can be estimated in Lp -norms by XT . We will first present the results for discrete martingales. We
will extend these results to continuous martingales in the next section.
Definition 2.21. Let (Ω, F, P) be a probability space, and let X : Ω → R be a random variable.
For every p ∈ [1, ∞),
Z 1
p
p
kXkp = |X| dP ,
Ω
and
kXk∞ = ess supω∈Ω |X(ω)| .
Definition 2.22. Let (Ω, F, P) be a probability space, and let X : Ω → R be a random variable.
For every B ∈ F,
Z
E[X : B] = X dP.
B
Proposition 2.23. Let (Ω, F, P) be a probability space, let {Ft }t∈[0,∞) be a filtration on (Ω, F),
and let (Xt )t∈[0,∞) be a nonnegative Ft -submartingale. Then, for every T ∈ [0, ∞) and λ ∈ [0, ∞),
λP[XT∗ ≥ λ] ≤ E[XT : {XT∗ ≥ λ}].
Proof. Let T ∈ [0, ∞) and λ ∈ [0, ∞). Define the stopping time τ by
τ = (inf{s ∈ [0, ∞) : Xs ≥ λ}) ∧ T.
Since the constant function T is also a stopping time, and since by definition τ ≤ T , the optional
stopping theorem, the fact that (Xt )t∈[0,∞) is a submartingale, and the definition of τ prove that
E[XT ] ≥ E[Xτ ]
= E[XT : {XT∗ < λ}] + E[Xτ : {XT∗ ≥ λ}]
≥ E[XT : {XT∗ < λ}] + λP[XT∗ ≥ λ].
Therefore, by linearity of the expectation,
E[XT : {XT∗ ≥ λ}] ≥ λP[XT∗ ≥ λ],
In the following proposition, we prove for every p ∈ (1, ∞) that the Lp -norm of the running
maximum XT∗ is controlled from above by the Lp -norm of XT . The case p = 2 is the Doob-
Kolmogorov inequality.
24 BENJAMIN FEHRMAN
Proposition 2.24. Let (Ω, F, P) be a probability space, let {Ft }t∈[0,∞) be a filtration on (Ω, F),
and let (Xt )t∈[0,∞) be a nonnegative Ft -submartingale. Then, for every T ∈ [0, ∞) and λ ∈ [0, ∞),
(2.27) λP[XT∗ ≥ λ] ≤ E[XT : {XT∗ ≥ λ}] ≤ E[XT ],
and, for every p ∈ (1, ∞),

p
(2.28) kXT∗ kp ≤ kXT kp .
1−p
Proof. Estimate (2.29) is an immediate consequence of Proposition 2.23. It remains to prove (2.30).
Let p ∈ (1, ∞) and T ∈ [0, ∞). It follows from (2.29) that
Z ∞
∗ p
E[|XT | ] = p λp−1 P[XT∗ ≥ λ] dλ
0
Z ∞
≤p λp−2 E[XT : {XT∗ ≥ λ}] dλ
Z0 Z ∞
=p λp−2 XT 1{XT∗ ≥λ} dλ dP
Ω 0 Z
p
= (XT∗ )p−1 XT dP.
p−1 Ω
p
Hölder’s inequality with exponents p and p−1 then proves that

p p−1 1
E[|XT∗ |p ] ≤ E[|XT∗ |p ] p E[|XT |p ] p .
p−1
p
In terms of L -norms, this implies that

p
kXT∗ kpp ≤ kXT∗ kp−1 kXT kp ,
p−1
which proves that, after dividing by kXT∗ kp−1 ,

∗ p
kXT kp ≤ kXT kp .
p−1
This completes the proof.
The following corollary is an immediate consequence of Jensen’s inequality and Propositions 2.23
and 2.24.
Corollary 2.25. Let (Ω, F, P) be a probability space, let {Ft }t∈[0,∞) be a filtration on (Ω, F), and
let (Xt )t∈[0,∞) be a Ft -martingale. Then, for every N ∈ N0 and λ ∈ [0, ∞),
(2.29) λP[XT∗ ≥ λ] ≤ E[|XT | : {XT∗ ≥ λ}] ≤ E[|XT |],
and, for every p ∈ (1, ∞),

p
(2.30) kXT∗ kp ≤ kXT kp .
1−p
Proof. Jensen’s inequality proves that (|Xt |)t∈[0,∞) is a submartingale, since the absolute value
function is convex. The proof then follows from Propositions 2.23 and 2.24.
2.6. Doob’s martingale convergence theorem. In this section, we will prove the martingale
convergence theorem. The convergence is obtained by proving that, in the long run, an integrable
martingale cannot oscillate to infinity. This is most easily seen on the level of L2 -bounded mar-
tingales. Precisely, suppose that (Mn )n∈N0 is an L2 -bounded martingale on a probability space
(Ω, F, P) in the sense that
(2.31) sup E[Mn2 ] < ∞.

n∈N0
Since martingale increments in some ways behave like independent random variables, in the sense
that, for every n ∈ N0 ,
 !2 
n
X
E[Mn2 ] = E  (Mk − Mk−1 ) 
k=1
 
n n
" #
X X
=E (Mk − Mk−1 )2 + 2E  (Mk − Mk−1 )(Mj − Mj−1 )
k=0 j<k=1
 
n n
" #
X X
=E (Mk − Mk−1 )2 + 2E  E[(Mk − Mk−1 )(Mj − Mj−1 )|Fk−1 ]
k=0 j<k=1
" n  
n
#
X X
=E (Mk − Mk−1 )2 + 2E  (E[Mk |Fk−1 ] − Mk−1 )(Mj − Mj−1 )
k=0 j<k=1
" n #
X
=E (Mk − Mk−1 )2 .
k=0
Therefore, after passing to the limit n → ∞, it follows from (2.31) that
∞
" #
X
2
(2.32) E (Mk − Mk−1 ) < ∞.
k=0
Since the same computation proves that, for every n ≤ m ∈ N0 ,
n
" #
X
2 2
E[(Mn − Mm ) ] = E (Mk − Mk−1 ) ,
k=m+1
it follows from (2.32) that {Mn }n∈N0 is a Cauchy sequence in L2 (Ω). Therefore, there exists
M∞ ∈ L2 (Ω) such that, as n → ∞,
Mn → M∞ strongly in L2 (Ω).
The estimate (2.32) implies almost surely that the sequence {Mn }n∈N0 does not oscillate to infinity
in the sense that the differences |Mk − Mk−1 |2 decay in an summable fashion.
We will make the notion of oscillation precise by defining an upcrossing. Let a < b ∈ R, and
suppose that (Mt )t∈[0,∞) is a continuous (sub/super)martingale. We then define the following
26 BENJAMIN FEHRMAN
infinite sequence of stopping times {τk }k∈N by

τ0 = 0
τ1 = inf{t ∈ [0, ∞) : Mt ≤ a}
τ2 = inf{t ∈ [0, ∞) : t ≥ τ1 and Mt ≥ b}
τ3 = inf{t ∈ [0, ∞) : t ≥ τ2 and Mt ≤ a}
τ4 = inf{t ∈ [0, ∞) : t ≥ τ3 and Mt ≥ b}
..
.
τ2k−1 = inf{t ∈ [0, ∞) : t ≥ τ2k−2 and Mt ≤ a}
τ2k = inf{t ∈ [0, ∞) : t ≥ τ2k−1 and Mt ≥ b}
..
.
We are interested in the number of instances over the interval [0, T ] that the martingale moves from
below the value a to above the value b. Such an event is called an upcrossing from a to b. For a
(sub/super) martingale (Mt )t∈[0,∞) and for T ∈ (0, ∞), we define U (a, b; M, T ) to be the number
of upcrossings on [0, T ],
(2.33) U (a, b; M, T ) = sup{k ∈ N : τ2k ≤ T }.
The following proposition, which is Doob’s upcrossing inequality, estimates the expectation of
U (a, b; M, T ) for a super martingale (Mt )t∈[0,∞) .
Proposition 2.26. Let (Ω, F, P) be a probability space, let (Ft )t∈[0,∞) be a filtration on (Ω, F),
and let (Mt )t∈[0,∞) be a continuous super martingale. Then, for every T ∈ (0, ∞) and a < b ∈ R,
1
E[U (a, b; M, T )] ≤ E[(XT − a)− ],
b−a
where (MT − a)− = − min((MT − a), 0).
Proof. Let T ∈ (0, ∞) and let K ∈ N. Then,
K
X K
X

Mτ2K ∧T = Mτ2k ∧T − Mτ2k−1 ∧T + Mτ2k−1 ∧T − Mτ2k−2 ∧T + M0 .
k=1 k=1
We first observe that, by continuity of the process (Mt )t∈[0,∞) ,

K
X
Mτ2k ∧T − Mτ2k−1 ∧T
k=1
K∧U (a,b;M,T ) K
X X
= Mτ2k ∧T − Mτ2k−1 ∧T + Mτ2k ∧T − Mτ2k−1 ∧T
k=1 k=U (a,b;M,T )+1
= (b − a)(U (a, b; M, T ) ∧ K) + (MT − Mτ2U (a,b;M,T )+1 )1{τ2U (a,b;M,T )+1 ≤T }
= (b − a)(U (a, b; M, T ) ∧ K) + (MT − a)1{τ2U (a,b;M,T )+1 ≤T }
Therefore, after rearranging the sums and observing the cancellation of the terms involving M0 ,
(b − a)(U (a, b; M, T ) ∧ K) + (MT − Mτ2U (a,b;M,T )+1 )1{τ2U (a,b;M,T )+1 ≤T }
K
X
= Mτ2k ∧T − Mτ2k−1 ∧T .
k=1
Since the super martingale property and the optional stopping theorem applied to the bounded
stopping times τi ∧ T prove that
"K # K
X X
E Mτ2k ∧T − Mτ2k−1 ∧T = E[Mτ2k ∧T ] − E[Mτ2k−1 ∧T ] ≤ 0,
k=1 k=2
it follows that
(2.34) (b − a)E[(U (a, b; M, T ) ∧ K)] + E[(MT − a)1{τ2U (a,b;M,T )+1 ≤T } ] ≤ 0.
We therefore conclude from (2.34) that, for every K ∈ N,

1 1
E[U (a, b; M, T ) ∧ K] ≤ E[−(MT − a)1{τ2U (a,b;M,T )+1 ≤T } ] ≤ E[(MT − a)− ].
(b − a) (b − a)
Since by continuity of (Mt )t∈[0,∞) , which implies the uniform continuity of (Mt )t∈[0,∞) on [0, T ], it
follows that, for almost every ω ∈ Ω,
U (a, b; M, T )(ω) < ∞,
the monotone convergence theorem proves that, after passing to the limit K → ∞,
1
E[U (a, b; M, T )] ≤ E[(MT − a)− ],
b−a
Upcrossings are convenient for supermartingales, and downcrossings are convenient for sub-
martingales. That is, given a (sub/super) martingale (Mt )t∈[0,∞) define the stopping times
τ0 = 0
τ1 = inf{t ∈ [0, ∞) : Mt ≥ b}
τ2 = inf{t ∈ [0, ∞) : t ≥ τ1 and Mt ≤ a}
τ3 = inf{t ∈ [0, ∞) : t ≥ τ2 and Mt ≥ b}
τ4 = inf{t ∈ [0, ∞) : t ≥ τ3 and Mt ≤ a}
..
.
τ2k−1 = inf{t ∈ [0, ∞) : t ≥ τ2k−2 and Mt ≥ b}
τ2k = inf{t ∈ [0, ∞) : t ≥ τ2k−1 and Mt ≤ a}
..
.
and let D(b, a; M, T ) = inf{k ∈ N0 : τ2k ≤ T }. This is the total number of downcrossings of
(Mt )t∈[0,∞) from b to a. The following is Doob’s downcrossing inequality.
and let (Mt )t∈[0,∞) be a continuous submartingale. Then, for every T ∈ (0, ∞) and a < b ∈ R,
1
E[D(b, a; M, T )] ≤ E[(MT − b)+ ],
b−a
where (MT − a)+ − = max((MT − a), 0).
Proof. Since (Mt )t∈[0,∞) is a submartingales, its negative (−Mt )t∈[0,∞) is a supermartingale. Fur-
thermore, a down crossing of (Mt )t∈[0,∞) from b to a is an upcrossing of (−Mt )t∈[0,∞) from −b to
28 BENJAMIN FEHRMAN
−a. That is, D(b, a; M, T ) = U (−b, −a; −M, T ). Therefore, by Proposition 2.26,
E[D(b, a; M, T )] = E[U (−b, −a; M, T )]
1
≤ E[(−MT − (−b))− ]
((−a)) − (−b))
1
= E[(Mt − b)+ ],
(b − a)
Given a (sub/super) martingale (Mt )t∈[0,∞) and a < b ∈ R, we define the total number of
upcrossings from a to b to be
U (a, b; M ) = lim U (a, b; M, T ),
T →∞
and the total number of downcrossings to be
D(b, a; M ) = lim D(b, a; M, T ).
T →∞
This limit always exists, as the limit of an increasing sequence, but it may be infinite. We will now
prove the first version of the martingale convergence theorem, which states that an L1 -bounded
submartingale converges almost surely as t → ∞.
and let (Mt )t∈[0,∞) be a continuous submartingale. Assume that M0 ∈ L1 (Ω) and that
(2.35) sup E[Mt+ ] < ∞.
t∈[0,∞)
Then, there exists M∞ ∈ L1 (Ω) such that, for almost every ω ∈ Ω, as t → ∞,

(2.36) Mt (ω) → M∞ (ω).
Remark 2.29. Note carefully that (2.40) implies only almost sure convergence, it does not im-
ply convergence in L1 (Ω). Convergence in L1 (Ω) requires the additional assumption of uniform
integrability, as explained below.
Proof. It follows from Proposition 2.27 and (2.39) that there exists c ∈ (0, ∞) such that, for every
T ∈ (0, ∞) and a < b ∈ R,
1 1 c + |a|
E[D(b, a; M, T )] ≤ E[(XT − b)+ ] ≤ |b| + E[XT+ ] ≤ .
(b − a) (b − a) b−a
Therefore, for every a < b ∈ R, after passing to the limit T → ∞ using the monotone convergence
theorem,
c + |a|
E[D(b, a; M )] ≤ .
b−a
We therefore conclude that, for every a < b ∈ R, there exists a subset Ωa,b ⊆ Ω of full probability
such that
D(b, a; M )(ω) < ∞ for every ω ∈ Ω.
0
We then define the subset Ω ⊆ Ω of full probability
Ω0 = ∩a<b∈Q Ωa,b ,
and claim that, for every ω ∈ Ω0 ,
(2.37) lim Mt (ω) exists.
t→∞
Since the limit (2.37) exists if and only if
lim sup Mt (ω) = lim inf Mt (ω),
t→∞ t→∞
suppose by contradiction that there exist ω ∈ Ω0 and a, b ∈ [−∞, ∞] such that

(2.38) a = lim inf Mt (ω) < lim sup Mt (ω) = b.
t→∞ t→∞
By density of the rationals, there exist a < b ∈ Q such that a < a < b < b, and by definition of the
lim inf and lim sup it follows from (2.38) that
D(b, a; M )(ω) = ∞,
which contradicts the assumption that ω ∈ Ω0 ⊆ Ωa,b .
We therefore define the limit M∞ for every ω ∈ Ω0 by
M∞ (ω) = lim Mt (ω),
t→∞
and conclude using Fatou’s lemma, (2.39), and the submartingale property that
E[|M∞ |] = E[( lim |Mt |)] ≤ lim inf (E[|Mt |])
t→∞ t→∞
= lim inf 2E[Mt+ ] − E[Mt ]

t→∞
≤ lim inf 2E[Mt+ ] − E[M0 ]

t→∞
≤ lim inf 2E[Mt+ ] + E[|M0 |] < ∞,

t→∞
which proves that M∞ ∈ L1 (Ω). This completes the proof.

and let (Mt )t∈[0,∞) be a continuous super martingale. Assume that M0 ∈ L1 (Ω) and that
(2.39) sup E[Mt− ] < ∞.
t∈[0,∞)
Then, there exists M∞ ∈ L1 (Ω) such that, for almost every ω ∈ Ω, as t → ∞,

(2.40) Mt (ω) → M∞ (ω).
Proof. The proof is identical to the proof of Proposition 2.28, with the exception that Proposi-
tion 2.26 is used in place of Proposition 2.27. Also, in this case, to prove the integrability of M∞
we observe using the supermartingale property that
E[|M∞ |] = E[( lim |Mt |)] ≤ lim inf (E[|Mt |])
t→∞ t→∞
= lim inf 2E[Mt− ] + E[Mt ]

t→∞
≤ lim inf 2E[Mt− ] + E[M0 ]

t→∞
≤ lim inf 2E[Mt− ] + E[|M0 |] < ∞,

t→∞
Remark 2.31. In particular, it follows from Proposition 2.30 that nonnegative super martingales
with an integrable initial condition always converge in almost sure sense to an integrable function.
Again, I emphasize that this is only almost sure convergence. In Theorem 2.32 below, we will
upgrade this convergence to strong convergence in L1 (Ω) for uniformly integrable martingales.
The following theorem is the Doob’s martingale convergence theorem for a uniformly integrable
martingale. It follows from either Proposition 2.28 or Proposition 2.30 that if (Mt )t∈[0,∞) is uni-
formly integrable then, in an almost sure sense, M∞ (ω) = limt→∞ Mt (ω) exists and M∞ ∈ L1 (Ω).
The following theorem proves that not only is this true, but if (Mt )t∈[0,∞) is uniformly integrable
then limt→∞ E[|Mt − M∞ |] = 0 and Mt = E[M∞ |Ft ] for every t ∈ [0, ∞).
30 BENJAMIN FEHRMAN
Theorem 2.32. Let (Ω, F, P) be a probability space, let (Ft )t∈[0,∞) be a filtration on (Ω, F), and
let (Mt )t∈[0,∞) be a continuous Ft -martingale. Then, the following three statements are equivalent.
(i) (Mt )t∈[0,∞) is closed: there exists M∞ ∈ L1 (Ω) such that, for every t ∈ [0, ∞),
Mt = E[M∞ |Ft ].
(ii) The family (Mt )t∈[0,∞) is uniformly integrable.
(iii) The family (Mt )t∈[0,∞) converges, as t → ∞, almost surely and in L1 (Ω).
Proof. We will first prove that (i) implies (ii). Since M∞ ∈ L1 (Ω), and since Mt = E[M∞ |Ft ] for
every t ∈ [0, ∞), it follows from Lemma 2.18 that (Mt )t∈[0,∞) is uniformly integrable. We will now
prove that (ii) implies (iii). Since the uniform integrability implies that (Mt )t∈[0,∞) is bounded in
L1 (Ω), Proposition 2.28 proves that there exists M∞ ∈ L1 (Ω) such that, as t → ∞,
Mt → M∞ almost surely.
Since almost sure convergence implies convergence in probability, the uniform integrability and
Theorem 2.17 prove that
lim E[|Mt − M∞ |] = 0.
t→∞
Finally, we will prove that (iii) implies (i). By assumption, there exists M∞ ∈ L1 (Ω) such that
(Mt )t∈[0,∞) converges, as t → ∞, to M∞ almost surely and strongly in L1 (Ω). Since the martingale
property implies that, for every s ≤ t ∈ [0, ∞),
(2.41) Ms = E[Mt |Fs ],
and since the conditional expectation is stable with respect to convergence in L1 (Ω), after passing
to the limit t → ∞ in (2.41) we have that, for every s ∈ [0, ∞),
Ms = E[M∞ |Fs ].
The following corollary applies Theorem 2.32 to martingales bounded in Lp , for p ∈ (1, ∞). It
is a consequence of the following lemma, which proves that a family of random variables that is
bounded in Lp (Ω), for some p ∈ (1, ∞), is necessarily uniformly integrable.
Lemma 2.33. Let (Ω, F, P) be a probability space, let A be a set, and let {Xα }α∈A be a family of
random variables. Assume that, for p ∈ (1, ∞),
(2.42) sup kXα kp < ∞.
α∈A
Then the family {Xα }α∈A is uniformly integrable.
Proof. We will first prove condition (i) of Definition 2.14. It follows from Hölder’s inequality and
(2.42) that
sup kXα kL1 (Ω) ≤ sup kXα kp < ∞.
α∈A α∈A
We will no prove condition (ii). For every α ∈ A and K ∈ (0, ∞), Hölder’s inequality, Chebyshev’
inequality, and p ∈ (1, ∞) prove that
p−1
p−1 1
E[|Xα | : {|Xα | ≥ K}] ≤ kXkp P[|Xα | ≥ K] p ≤ kXkp kXkp = K −(p−1) kXkpp .
K
Therefore, it follows from (2.42) and p ∈ (1, ∞) that

lim sup E[|Xα | : {|Xα | ≥ K}] ≤ lim K −(p−1) sup kXα kpp = 0.
K→∞ α∈A K→∞ α∈A
let (Mt )t∈[0,∞) be a continuous Ft -martingale. Assume that there exists p ∈ (1, ∞) such that
(2.43) sup kMt kp < ∞.
t∈[0,∞)
Then there exists M∞ ∈ Lp (Ω) such that

lim kMt − M∞ kp = 0.
t→∞
Proof. It follows from Lemma 2.33 that (Mt )t∈[0,∞) is uniformly integrable. Therefore, by Theo-
rem 2.32, there exists M∞ ∈ L1 (Ω) such that, as t → ∞,
(2.44) Mt → M∞ almost surely and strongly in L1 (Ω).
It remains to prove that M∞ ∈ Lp (Ω) and that the convergence takes place in Lp (Ω). For the first
point, Fatou’s lemma proves that
E[|M∞ |p ] = E[( lim |Mt |p )] ≤ lim inf E[|Mt |p ] = lim inf kMt kpp < ∞.
t→∞ t→∞ t→∞
Then, since it follows from Doob’s martingale inequality that

p p
sup |Mt | ≤ lim sup kMt kp < ∞,

t∈[0,∞) p − 1 t→∞
p
and since we have for every t ∈ [0, ∞) that

!
|M∞ − Mt | ≤ |M∞ | + sup |Mt | ∈ Lp (Ω),
t∈[0,∞)
it follows from (2.44) and the dominated convergence theorem that

lim kMt − M∞ kp = 0.
t→∞
In the final proposition of this section, we extend the optional stopping theorem to uniformly
integrable martingales. Suppose that (Mt )t∈[0,∞) is a continuous, uniformly integrable martingale.
Then, by Theorem 2.32, there exists M∞ ∈ L1 (Ω) such that, as t → ∞,
Mt → M∞ almost surely and strongly in L1 (Ω).
Therefore, given a possibly infinite stopping time T : Ω → [0, ∞], we define
MT = MT 1{T <∞} + M∞ 1{T =∞} .
The following proposition proves that, when dealing with a uniformly integrable martingale, the
optional stopping theorem applies even to stopping times that are infinite with positive probability.
This stands in stark contrast to the example of the simple random walk above, which is not
uniformly integrable, and for which the optional stopping theorem fails for a stopping time that is
almost surely finite.
and let (Mt )t∈[0,∞) be a continuous Ft -martingale. Assume that (Mt )t∈[0,∞) is uniformly integrable.
Then for every pair of Ft -stopping times S ≤ T : Ω → [0, ∞],
E[MT |FS ] = MS .
In particular,
E[MT ] = E[MS ].
32 BENJAMIN FEHRMAN
Proof. For every bounded stopping time T , Theorem 2.19 proves that
E[MN |FT ] = MT .
The martingale convergence theorem then proves that, after passing to the limit N → ∞,
E[M∞ |FT ] = MT .
Lemma 2.18 proves that the family of random variables MT , for T a bounded stopping time, is
uniformly integrable. A repetition of the proof of Theorem 2.19 then proves that, for every stopping
time T : Ω → [0, ∞],
E[M∞ |FT ] = MT ,
which proves using Lemma 2.18 that the family of random variables MT , for T a stopping time,
is uniformly integrable. Then, for every pair of stopping times S ≤ T , the tower property of
conditional expectation and FS ⊆ FT prove that
E[MT : FS ] = E[E[M∞ |FT ]|FS ] = E[M∞ |FS ] = MS ,
3. Itô’s Formula
In this section, we will develop the theory of stochastic integration on the way to proving Itô’s
formula, which is the fundamental theorem of the calculus of stochastic processes. We first observe
that the integral with respect to a process of bounded variation is almost surely well-defined.
Observe in particular that a nondecreasing or nonincreasing process is necessarily of finite variation.
Definition 3.1. Let (Ω, F, P) be a probability space and let a < b ∈ R. A stochastic process
(At )t∈[0,∞) has bounded variation on [a, b] if
n(∆)
X
V (a, b; A) = sup At − At < ∞ almost surely,
k k−1
∆⊆[a,b] k=1
where the sup is over all partitions ∆ = {a = t0 < t1 < . . . < tn(δ)−1 < tn(∆) = b} of [a, b].
Given a process of finite variation (At )t∈[0,∞) and a continuous stochastic (Xt )t∈[0,∞) , we can
define the integral of (Xt )t∈[0,∞) with respect to (At )t∈[0,∞) as the limit
Z b n(∆)
X
Xt dAt = lim Xtk−1 (Atk − Atk−1 ),
a |∆|→0
k=1
where the facts that (At )t∈[0,∞) has finite variation and that (Xt )t∈[0,∞) is continuous prove that
the above limit is almost surely well-defined. We aim to extend the above definition to martingales,
and later to semimartingales.
Definition 3.2. Let (Ω, F, P) be a probability space and let (Ft )t∈[0,∞) be a filtration on (Ω, F).
A stochastic process (Xt )t∈[0,∞) is a continuous semimartingale if there exists a continuous process
of finite variation (At )t∈[0,∞) and a continuous local martingale (Mt )t∈[0,∞) such that
Xt = Mt + At .
Given a martingale (Mt )t∈[0,∞) and a continuous process (Xt )t∈[0,∞) , we will ultimately define
the stochastic integral of (Xt )t∈[0,∞) with respect to (Mt )t∈[0,∞) to be the limit
Z b n(∆)
X
Xt dMt = lim Xtk−1 (Mtk − Mtk−1 ).
a |∆|→0,∆⊆[a,b]
k=1
However, unlike the finite variation case, the above limit is not well-defined in a pointwise sense. The
above limit is well-defined because of certain stochastic cancellations that rely on the martingale
property. These will imply in particular that the stochastic integral is itself a martingale beginning
from zero, and therefore that, for every a ≤ b ∈ [0, ∞),
Z b
E Xt dMt = 0.
a
We have seen that a martingale is of finite variation if and only if it is constant. Specifically, for
a Brownian motion (Bt )t∈[0,∞) , the limit along a sequence of partitions {∆k ⊆ [0, t]}k∈N satisfying
|∆k | → 0 as k → ∞,
n(∆k )
X
lim (Btk − Btk−1 )2 = b − a = hBib − hBia ,
k→∞
k=1
exists almost surely. The continuous increasing process (hBit = t)t∈[0,∞) is the quadratic variation
of Brownian motion. We will define the quadratic variation of a general local martingale (Mt )t∈[0,∞)
in the section to follow.
34 BENJAMIN FEHRMAN
We will use the quadratic variation and a stochastic integration to prove Itô’s formula, which is
the fundamental theorem of the calculus of semimartingales. Let
(Mt = (Mt1 , . . . , Mtd ))t∈[0,∞) ,
be a d-dimensional continuous martingale beginning from zero, and let f ∈ C2 (Rd ) be a twice-
differentiable function. We are essentially interested in establishing a fundamental theorem of
calculus for the process f ((Mt ))t∈[0,∞) . A first guess, based on the classical fundamental theorem
of calculus, would be that
Z t
f (Mt ) = f (M0 ) + f 0 (Ms ) dMs .
0
However, we can see immediately that such a formula cannot hold true unless M = 0 is constant.
Indeed, if we choose f (x) = |x|2 , since stochastic integral is a martingale, for every t ∈ [0, ∞),
Z t
2
E[|Mt | ] = E[|M0 |] + E f (Ms ) dMs = E[|M0 |2 ].
0
0
This implies that
E[(M0 + Mt ) · (Mt − M0 )] = 0,
from which it follows by the continuity of (Mt )t∈[0,∞) that M is constant. We therefore seek a
higher order approximation of the process.
A second-order Taylor expansion of f about M0 proves for every t ∈ [0, ∞) that
1
f (Mt ) = f (M0 ) + ∇f (M0 ) · (Mt − M0 ) + h∇2 f (M0 )(Mt − M0 ), (Mt − M0 )i + O(|Mt − M0 |3 ),
2
where
d
X ∂f
∇F (M0 ) · (Mt − M0 ) = (M0 )(Mti − M0i ),
∂xi
i=1
and where
d
∂2f
(M0 )(Mti − M0i )(Mtj − M0j ).
X
2
h∇ f (M0 )(Mt − M0 ), (Mt − M0 )i =
∂xi ∂xj
i,j=1
Therefore, on the interval [0, 1], for a partition

∆ = {0 = t0 < t1 < . . . < tn(∆)−1 < tn(∆) = 1},
we have
(3.1)
n(∆)
X
f (M1 ) − f (M0 ) = (f (Mk ) − f (Mk−1 ))
k=1
n(∆)
X
= ∇f (Mtk−1 ) · (Mtk − Mtk−1 )
k=1
n(∆) n(∆)
1 X 2 X 3
+ h∇ f (Mtk−1 )(Mtk − Mtk−1 ), (Mtk − Mtk−1 )i + O(Mtk − Mtk−1 ).
2
k=1 k=1
As the mesh |∆| → 0, we expect by stochastic integration that

n(∆) d Z 1
X X ∂f
∇f (Mtk−1 ) · (Mtk − Mtk−1 ) → (Ms ) dMsi .
0 ∂xi
k=1 i=1
Motivated by the quadratic variation, we expect that, as |∆| → 0,

n(∆) d Z
1 X 2 1 X 1 ∂2f
h∇ f (Mtk−1 )(Mtk − Mtk−1 ), (Mtk − Mtk−1 )i → (Ms ) dhM i , M j is ,
2 2 0 ∂xi ∂xj
k=1 i,j=1
where will define (hM i , M j it )t∈[0,∞) to be the quadratic covariation of the one-dimensional martin-
gales (Mti )t∈[0,∞) and (Mtj )t∈[0,∞) , for each i, j ∈ {1, . . . , d}. Finally, since the quadratic variation
is bounded, we expect by continuity that, as |∆| → 0,
n(∆)
X 3
O(Mtk − Mtk−1 ) → 0.
k=1
In combination this will prove Itô’s formula, which states that, for every t ∈ [0, ∞),
d Z t d Z
X ∂f i 1 X t ∂2f
f (Mt ) − f (M0 ) = (Ms ) dMs + (Ms ) dhM i , M j is .
0 ∂xi 2 0 ∂xi ∂xj
i=1 i,j=1
In particular, if (Bt )t∈[0,∞) is a one-dimensional Brownian motion,

Z t
1 t 00
Z
f (Bt ) = f (0) + f (Bx ) dBs + f (Bs ) ds,
0 2 0
which implies in particular that
Z t
Bs dBs = Bt2 − t.
0
Notice as well that Itô’s formula implies that the process (f (Bt )t )t∈[0,∞) is a semimartingale, because
the stochastic integral is martingale and the deterministic integral is a of bounded variation. This
fact is true in general, we will prove below that if (Zt )t∈[0,∞) is a semimartingale then (f (Zt ))t∈[0,∞)
is also a semimartingale.
Indeed, the integration theory will be developed within the class of semimartingales. The theory
will therefore apply to the case of Brownian motion, but it will not in general cover the case of
integration with respect to fractional Brownian motion. We define a real-valued centered Gaussian
process to be a real valued process (Xt )t∈[0,∞) on a probability space (Ω, F, P) with finite dimen-
sional distributions that are normally distributed and mean zero. A centered Gaussian process
(Xt )t∈[0,∞) is a fractional Brownian motion with Hurst parameter h ∈ (0, 1) if P[X0 = 0] = 1 and
if, for every s, t ∈ [0, ∞),
1 2h
E[Xt Xs ] = t + s2h − |t − s|2h .
2
A fractional Brownian motion satisfies the following four properties.
(a) For every s ≤ t ∈ [0, ∞), the increment Xt − Xs has mean zero and variance |t − s|2h .
(b) For every p ∈ (0, ∞) there exists cp ∈ (0, ∞) such that, for every s ≤ t ∈ [0, ∞),
E[|Xt − Xs |p ] = cp |t − s|hp .
(c) For every γ ∈ (0, h), Fractional Brownian motion has a γ-Hölder continuous modification.
(d) Factional Brownian motion is not a semimartingale if h 6= 12 .
Proof. We will first prove (a). Let h ∈ (0, 1). It follows by assumption that, since the finite
dimensional distributions are normally distributed with mean zero, for every s ≤ t ∈ [0, ∞),
E[Xt − Xs ] = 0.
36 BENJAMIN FEHRMAN
For the variance, for every s ≤ t ∈ [0, ∞),

E[|Xt − Xs |2 ] = E[Xt2 ] + E[Xs2 ] − 2E[Xs Xt ]
= t2h + s2h − (t2h + s2h − |t − s|2h )
= |t − s|2h .
We will now prove (b). Let h ∈ (0, 1). Let s ≤ t ∈ [0, ∞). Since (Xt − Xs ) is normally distributed
with mean zero and variance |t − s|2h , it follows for every p ∈ (0, ∞) that
!
2
|x|
Z
1
E[|Xt − Xs |p ] = |x|p (2π |t − s|2h )− 2 exp − dx
R 2 |t − s|2h
!
2
|x|
Z
1
= |t − s|hp |x|p (2π)− 2 exp − dx,
R 2
where the final equality follows from the change of variables x̃ = x/|t−s|h . Therefore, for every
p ∈ (0, ∞), after defining
!
2
|x|
Z
1
cp = |x|p (2π)− 2 exp − dx,
R 2
it follows for every p ∈ (0, ∞) that
E[|Xt − Xs |p ] = cp |t − s|hp .
We will now prove (c). Let h ∈ (0, 1). The Kolmogorov continuity criterion states that, since for
every p ∈ (0, ∞),
E[|Xt − Xs |p ] = cp |t − s|hp ,
fractional Brownian motion has a Hölder continuous modification with Hölder exponent γ ∈ (0, h −
1
p ) for every p ∈ (0, ∞). That is, fractional Brownian motion is almost surely Hölder continuous for
any Hölder exponent γ ∈ (0, h). We will now prove (d). We will first identify the scaling properties
of fractional Brownian motion. Let h ∈ (0, 1) and λ ∈ (0, ∞). We aim to identify α ∈ R, such
that the process (λXλα s )s∈[0,∞) is a fractional Brownian motion. Since this process is a centered
Gaussian process with mean zero, it is sufficient to identify α ∈ R such that the covariance satisfies
E[λXλα t λXλα s ] = t2h + s2h − |t − s|2h .
For this, notice that

E[λXλα t λXλα s ] = λ2 E[Xλα t Xλα s ] = λ2−2αh t2h + s2h − |t − s|2h .
We therefore require that λ2−2αh = 1, which implies that α = −1/h. We conclude that, for every
α ∈ (0, ∞), if (Xt )t∈[0,∞) is a fractional Brownian motion with Hurst parameter h ∈ (0, ∞) then
(λXλ−1/h t )t∈[0,∞) is also a fractional Brownian motion with the same Hurst parameter.
We will now identify p ∈ [1, ∞) depending on h ∈ (0, 1) such that fractional Brownian motion
with Hurst parameter h ∈ (0, 1) has finite p-variation. By this we mean that, if {∆n = {0 = t0 <
t1 < . . . < tkn−1 < tkn = 1}}n∈N is a sequence of partitions of [0, 1] such that |∆n | → 0 as n → ∞,
then
Bt − Bt p exists and is finite almost surely.
X
lim k k−1
k→∞
∆n
For simplicity, we consider the sequence of partitions
{∆n = 0 < 1/2n < 2/2n < . . . < 2n −1/2n < 1}n∈N ,
where, for every n ∈ N, the scaling properties of fractional Brownian motion prove that in law, for
every p ∈ [1, ∞),
X2n p X2n p 2n
X
−nph nh −nph
(3.2) B kn − B k−1 = 2 2 B kn − B k−1 =2 |Bk − Bk−1 | .

2 n 2 n 2 2
k=1 k=1 k=1
The sequence {Bk − Bk−1 }k∈N is a stationary and ergodic sequence. This means that the random
variables {Bk − Bk−1 }k∈N are identically distributed, so that in particular we have, for every n ∈ N,
since B0 = 0,
" 2n #
X p
(3.3) B kn − B k−1 = 2−nph · 2n E[|B1 |] = 2n(1−ph) E[|B1 |].

E n
2 2
k=1
(For the problem sheet, observing that the above equality implies that, if the p-variation is finite
and nonzero, then the above inequality implies that p = 1/h is sufficient.) The ergodicity states
that, for the measure W h induced the space of continuous paths beginning from zero C0 ([0, ∞))
by fractional Brownian motion, the only measurable subsets A ⊆ C0 ([0, ∞)) left invariant by the
shift operators {τk }k∈N0 defined by
τk (σ)(t) = σ(k + t) − σ(k) for every t ∈ [0, ∞),
have measure zero or measure one. The ergodic theorem and the estimate (8.6) prove that, almost
surely and in L1 (Ω), as n → ∞,
2n
X
−n
2 |Bk − Bk−1 | → E[|X1 |].
k=1
Therefore, returning to (8.5), it follows that in probability
0 if ph > 1,

n
2 p 
1
X 
lim B kn − B k−1 = E[|X1 |] if p = ,

n→∞
k=1
2 2n 
 h
∞ if ph < 1.

In particular, because semi-martingales have finite quadratic variation, it follows that fractional
Brownian motion is a semi-martingale if and only if h = 12 . That is, if and only if the fractional
Brownian motion is a Brownian motion.
38 BENJAMIN FEHRMAN
3.1. Quadratic variation. In this section, we will define the quadratic variation (hM it )t∈[0,∞) of
a continuous local martinagle (Mt )t∈[0,∞) as a continuous increasing process that is defined by the
limit, along a sequence of partitions locally finite partitions
∆n = {0 = t0n < t1n < . . . < tkn < t(k+1)n < . . .},
of [0, ∞) satisfying |∆n | → 0 as n → ∞ by
∞
X
hM it = lim (Mtkn ∧t − Mt(k−1)n ∧t )2 .
n→∞
k=1
The construction is based on approximating this limit by the corresponding discrete sums. Let
∆ ⊆ [0, ∞) be a locally finite partition
∆ = {0 = t0 < t1 < t2 < . . . < tk < tk+1 < . . .},
and define the process (Tt∆ (M ))t∈[0,∞) by
∞
X
Tt∆ (M ) = (Mtk ∧t − Mtk−1 ∧t )2 .
k=1
That is, if tn ≤ t < tn+1 then
n
X
Tt∆ (M ) = (Mt2k − Mt2k−1 )2 + (Mt − Mtn )2 .
k=1
The following proposition proves that, for every partition ∆, the process (Mt2 − Tt∆ (M ))t∈[0,∞) is
a martingale.
Proposition 3.3. Let (Ω, F, P) be a probability space, let (Ft )t∈[0,∞) be a filtration on (Ω, F), and
let (Mt )t∈[0,∞) be a bounded Ft -martingale. Then, for every locally finite partition ∆ ⊆ [0, ∞),
(Mt2 − Tt∆ (M ))t∈[0,∞) is a martingale.
Proof. Let ∆ ⊆ [0, ∞) be a locally finite partition
∆ = {0 = t0 < t1 < t2 < . . . < tk < tk+1 < . . .}.
Since (Mt )t∈[0,∞) is bounded, it follows by definition that, for every t ∈ [0, ∞),
E[Mt2 − Tt∆ (M )] < ∞.

Let s ≤ t ∈ [0, ∞). Fix n ∈ N such that tn−1 ≤ s < tn . Since properties of the conditional
expectation and the martingale property prove that, for every k > n,
E[(Mtk − Mtk−1 )2 |Fs ] = E[Mt2k − Mt2k−1 |Fs ],
and since
E[(Mtn − Mtn−1 )2 |Fs ] = E[Mt2n |Fs ] − 2Ms Mtn−1 + Mt2n−1
= E[Mt2n |Fs ] + (Ms − Mtn−1 )2 − Ms2
it follows that
∞
X n−1
X
E[Tt∆ (M )|Fs ] = E[(Mtk ∧t − Mtk−1 )2 ∧t |Fs ] + E[(Mtn − Mtn−1 ) |Fs ] + 2
(Mtk − Mtk−1 )2
k=n+1 k=1
= E[Mt2 |Fs ] − E[Mt2n |Fs ] + E[Mt2n |Fs ] − Ms2 + Ts∆ (M ).

Therefore, we conclude that
E[Mt2 − Ts∆ (M )|Fs ] = Ms2 − Ts∆ (M ),
Observe that the processes {Tt∆ (M )}∆⊆[0,∞) are not necessarily increasing. However, whenever
k ≤ j and tk , tj ∈ ∆, it follows by definition that Tt∆
k
(M ) ≤ Tt∆
j
(M ). This observation and the
boundedness of the martingale allow us to pass strongly to the limit |∆| → 0 in the following
proposition, and thereby construct a continuous, increasing process (hM i)t∈[0,∞) such that the
process (Mt2 − hM it )t∈[0,∞) is a martingale.
let (Mt )t∈[0,∞) be a bounded continuous Ft -martingale. Then there exists a unique nondecreasing
process (hM it )t∈[0,∞) which vanishes at zero such that
(Mt2 − hM it )t∈[0,∞) is a martingale.
Proof. To prove uniqueness, observe that if (At )t∈[0,∞) and (Bt )t∈[0,∞) are two increasing processes
vanishing zero such that
(Mt2 − At )t∈[0,∞) and (Mt2 − Bt )t∈[0,∞) are martingales,
then the difference (At − Bt )t∈[0,∞) is a martingale of finite variation. Hence, the difference is
constant and identically equal to zero. To prove existence, choose a nested sequence of partitions
∆1 ⊆ ∆2 ⊆ ∆3 ⊆ . . . satisfying |∆k | → 0 and k → ∞. Prove that the sequence {Tt∆k (M )}k∈N
converges strongly in L2 (Ω), and that the limit is satisfies the desired properties. (See Math B8.2
notes for full details, which will shortly be added here.)
This theorem has the following important extension to continuous local martingales.
Theorem 3.5. Let (Ω, F, P) be a probability space, let (Ft )t∈[0,∞) be a filtration on (Ω, F), and let
(Mt )t∈[0,∞) be a continuous local Ft -martingale. Then there exists a unique continuous increasing
process (hM it )t∈[0,∞) such that (Mt2 − hM it )t∈[0,∞) is a continuous local martingale. Furthermore,
for every t ∈ [0, ∞), for any sequence of partitions {∆k }k∈N of [0, t] satisfying |∆k | → 0 as k → ∞,
!
∆
lim sup Ts k (M ) − hM is = 0 in probability.
k→∞ s∈[0,T ]
We can now define the quadratic covariation or bracket process of two local martingales (Mt )t∈[0,∞)
and (Nt )t∈[0,∞) . Based upon the computations above, for t ∈ [0, ∞) and for a sequence of partitions
{∆k }k∈N of [0, T ] satisfying |∆k | → 0 as k → ∞, we expect to define
n(∆k )
X
hM, N it = lim (Mtk − Mtk−1 )(Ntk − Ntk−1 ).
k→∞
k=1
where by convention we define hM, M it = hM it . Indeed, by polarization, the work is already done.
(Mt )t∈[0,∞) and (Nt )t∈[0,∞) be a continuous local Ft -martingales. Then there exists a unique finite
variation process (hM it )t∈[0,∞) such that
(Mt Nt − hM, N it )t∈[0,∞) is a continuous local martingale.
Furthermore, for every t ∈ [0, ∞), for any sequence of partitions {∆n }n∈N of [0, t] satisfying |∆n | →
0 as n → ∞, !
∆
lim sup Ts n (M, N ) − hM is = 0 in probability,
n→∞ s∈[0,T ]
where
∞
X
Ts∆n (M, N ) = (Mtk ∧s − Mtk−1 ∧s )(Ntk ∧s − Ntk−1 ∧s ).
k=1
40 BENJAMIN FEHRMAN
Proof. Since (Mt + Nt )t∈[0,∞) and (Mt − Nt )t∈[0,∞) are continuous local martingales, we define by
polarization the bracket
1
hM, N it = (hM + N, M + N it − hM − N, M − N it ) .
4
The formula
1
(M + N )2t − (M − N )2t ,

Mt Nt =
4
proves that (hM, N it )t∈[0,∞) satisfies the desired properties.
In the final proposition of this section, we prove an important estimate for the integral of measur-
able processes with respect to the quadratic covariation. This is the Kunita-Watanabe inequality.
let (Mt )t∈[0,∞) and (Nt )t∈[0,∞) be a continuous local Ft -martingales. Let (Hs )s∈[0,∞) and (Ks )s∈[0,∞)
be measurable stochastic processes. Then, for every t ∈ [0, ∞),
Z t Z t 21 Z t 12
2 2
|Hs | |Ks | d |hM, N i|s ≤ Hs dhM is Ks dhN is almost surely.
0 0 0
Furthermore, for every p, q ∈ [1, ∞) satisfying 1/p + 1/q = 1,
Z t Z
t 12

Z
t 12

2 2

|Hs | |Ks | d |hM, N i| ≤
Hs dhM is

K dhN is .

s 0 s

0 1 0
p q
Proof. Let t ∈ [0, ∞). By density, it suffices to prove that statement for functions of the form
X n X n
K= Kk 1[tk−1 ,tk ) and H = Hk 1[tk−1 ,tk ) ,
k=1 k=1
where {0 = t0 < t1 < . . . < tn−1 < tn = t} is a partition of [0, T ], and where {Kk }k∈{1,...,n} and
{Hk }k∈{1,...,n} are bounded random variables. It then follows by definition that
Z t n
X
|Hs | |Ks | d |hM, N i|s = |Hk | |Kk | |hM, N i|ttkk−1 ,
0 k=1
where |hM, N i|ttkk−1 = |hM, N i|tk − |hM, N i|tk−1 . Since it follows for every r ∈ Q that, for each
k ∈ {1, . . . , n}, almost surely,
hM ittkk−1 − 2rhM, N ittkk−1 + r2 hN ittkk−1 = hM + rN, M + rN ittkk−1 ≥ 0,
and therefore by continuity this inequality holds for every r ∈ R. Minimizing the lefthand side of
r ∈ R proves that, for every k ∈ {1, . . . , n}, almost surely,
1 1
hM, N ittkk−1 ≤ hM ittkk−1 hN ittkk−1 .
2 2
It then follows by Hölder’s inequality that

N
! 12 n
!1
Z t X X 2
|Hs | |Ks | d |hM, N i|s ≤ Hk2 hM ittkk−1 Kk2 hN ittkk−1

0 k=1 k=1
Z t 21 Z t 12
= Hs2 dhM is Ks2 dhN is .
0 0
This proves the first statement. The final statement is then a consequence of Hölder’s inequality.
3.2. Stochastic integration. In this section, we will develop the stochastic integral with respect to
a continuous martingale. We henceforth fix a probability space (Ω, F, P) and a filtration (Ft )t∈[0,∞)
on (Ω, F). The construction is based on duality. Indeed, we will consider the class of integrators
( )
H= (Mt )t∈[0,∞) : (Mt )t∈[0,∞) is a martingale with sup E[Mt2 ] < ∞ .
t∈[0,∞)
For every (Mt )t∈[0,∞) ∈ H, it follows from Theorem 2.34 that there exists M∞ ∈ L2 (Ω) such that,
as t → ∞,
Mt → M∞ almost surely and in L2 (Ω).
Furthermore, it follows from Doob’s inequality that
" #
E sup Mt2 ≤ 4E[M∞
2
].
t∈[0,∞)
We therefore define an inner product and norm on H for M, N ∈ H by

hhM, N ii = E[M∞ N∞ ] and kM k2H = E[M∞
2
],
where, since (Mt − hM it )t∈[0,∞) is a martingale,
2
E[M∞ ] = E[hM i∞ ].
The following proposition characterizes the space H as a complete Hilbert space isomorphic to an
L2 -space with respect to the sigma algebra F∞ = ∪t∈[0,∞) Ft .
Proposition 3.8. The space (H, hh·, ·ii) is a complete Hilbert space isomorphic to L2 (Ω, F∞ , P).
Proof. Let (M n )n∈N be a Cauchy sequence in H. It follows by definition of k·kH that (M∞n)
n∈N
2 2
is a Cauchy sequence in L (Ω, F∞ , P). Therefore, there exists M∞ ∈ L (Ω, F∞ , P) such that, as
n → ∞,
n
M∞ → M∞ strongly in L2 (Ω).
We define the L2 -bounded martingale, for t ∈ [0, ∞),
Mt = E[M∞ |Ft ],
and conclude by definition of k·kH that (Mt )t∈[0,∞) ∈ H and that, as n → ∞,
(Mtn )t∈[0,∞) → (Mt )t∈[0,∞) strongly in H.
These arguments prove that the maps
(Mt )t∈[0,∞) ∈ H 7→ M∞ ∈ L2 (Ω, F∞ , P),
M ∈ L2 (Ω, F∞ , P) 7→ (E[M |Ft ])t∈[0,∞) ∈ H,
define a Hilbert space isomorphism between H and L2 (Ω, F∞ , P). This completes the proof.
We will integrate with respect to continuous L2 -bounded martingales, and therefore define the
subspace H ⊆ H of continuous L2 -bounded martingales
H = {(Mt )t∈[0,∞) ∈ H : t ∈ [0, ∞) 7→ Mt is almost surely continuous.}.
The following proposition proves that (H, hh·, ·ii) remains is a complete Hilbert subspace of H.
Proposition 3.9. The space (H, hh·, ·ii) is a complete Hilbert space.
42 BENJAMIN FEHRMAN
Proof. It is necessary to prove that H is complete. Suppose that (M n )n∈N is a Cauchy sequence in
H. Then, there exists M∞ ∈ L2 (Ω, F∞ , P) such that, as n → ∞,
n
M∞ → M∞ strongly in L2 (Ω, F∞ , P).
Define the L2 -bounded martingale, for t ∈ [0, ∞),
Mt = E[M |Ft ].
Since Jensen’s inequality proves for every n ∈ N that |M n − M | is a submartingale, Doob’s in-
equality proves for every n ∈ N that
" #
E sup |Mtn − Mt |2 ≤ 4E[|M∞
n
− M∞ |2 ].
t∈[0,∞)
Therefore, there exists a subsequence {nk }k∈N such that, as k → ∞,

sup |Mtnk − Mt |2 → 0 almost surely.
t∈[0,∞)
This proves that t ∈ [0, ∞) 7→ Mt is almost surely the uniform limit of continuous functions, which
proves that (Mt )t∈[0,∞) is almost surely continuous. This implies that (Mt )t∈[0,∞) ∈ H, which
completes the proof that H is complete.
The following two propositions provide a useful characterizations of H.
Proposition 3.10. Let (Mt )t∈[0,∞) be a continuous local martingale. Then, (Mt )t∈[0,∞) ∈ H if and
only if the following two conditions are satisfied.
(i) We have that
M0 ∈ L2 (Ω).
(ii) We have that
E[hM i∞ ] < ∞.
Furthermore, in this case, (Mt2 − hM it )t∈[0,∞) is uniformly integrable.
Proof. If (Mt )t∈[0,∞) ∈ H then by L2 -boundedness we have that M0 ∈ L2 (Ω) and that (Mt2 −
hM it )t∈[0,∞) is a martingale. And, by the L2 -boundedness and martingale converge theorem,
2
E[hM i∞ ] = E[M∞ ] < sup E[Mt2 ] < ∞.
t∈[0,∞)
For the converse, because (Mt )t∈[0,∞) is a continuous local martingale, there exist an increasing
sequence of stopping times {τn }n∈N satisfying almost surely that τn → ∞ as n → ∞. For every
n ∈ N, since by assumption M τn = (Mt∧τn 1{τn >0} )t∈[0,∞) is a martingale, and since
hM τn i = hM iτn 1{τn >0} ,
it follows from M0 ∈ L2 (Ω) and E[hM i∞ ] < ∞ that
((M τn )2t − hM τn it )t∈[0,∞) is a martingale.
The martingale property and the fact that the quadratic variation is nondecreasing prove that, for
every n ∈ N and t ∈ [0, ∞),
E[Mτ2n ∧t ] = E[M02 ] + E[hM τn it ] ≤ E[M02 ] + E[hM it ] ≤ E[M02 ] + E[hM i∞ ].
We then pass to the limit using Fatou’s lemma. Indeed, since the stopping times diverge to infinity
almost surely, we have that, as n → ∞, for every t ∈ [0, ∞),
Mτn ∧t → Mt almost surely,
and therefore, for every t ∈ [0, ∞),

E[Mt2 ] = E[M02 ] + E[ lim Mτ2n ∧t ] ≤ E[M02 ] + lim inf E[Mτ2n ∧t ] < E[M02 ] + E[hM i∞ ],
n→∞ n→∞
which proves that (Mt )t∈[0,∞) is a uniformly bounded L2 (Ω) and therefore uniformly integrable,
from which it follows that (Mt )t∈[0,∞) is a continuous martingale. This completes the proof that
(Mt )t∈[0,∞) ∈ H.
We will now prove that if (Mt )t∈[0,∞) ∈ H then (Mt2 − hM it )t∈[0,∞) is uniformly integrable. The
martingale convergence theorem proves that there exists M∞ ∈ L2 (Ω) such that, as t → ∞,
This implies by Hölder’s inequality that
Mt2 → M∞ 2
almost surely and strongly in L1 (Ω),

using the equality Mt2 − M∞ 2 = |(M − M )| |(M + M )|. Therefore, since the bracket process
t ∞ t ∞
is an increasing function, the dominated convergence theorem with dominating function hM∞ i∞
proves that, as t → ∞,
hM it → hM i∞ almost surely and strongly in L1 (Ω).
In combination, this implies that, as t → ∞,
Mt2 − hM it → M∞
2
− hM∞ i∞ almost surely and strongly in L1 (Ω).
The equivalent properties of the martingale convergence theorem then prove that (Mt2 −hM it )t∈[0,∞)
is uniformly integrable.
Proposition 3.11. Let (Mt )t∈[0,∞) be a continuous local martingale. Then, (Mt )t∈[0,∞) converges
on the set {hM i∞ < ∞}.
Proof. For every n ∈ N define the stopping time
σn = inf{t ∈ [0, ∞) : hM it ≥ n}.
For every n ∈ N, consider the stopped local martingale (Mσn ∧t )t∈[0,∞) . It follows by definition of
the {σn }n∈N that, for every n ∈ N,
E[hM σn i∞ ] ≤ n.
The above proposition proves that there exists M∞,n ∈ L2 (Ω) such that, as t → ∞,
Mσn ∧t → M∞,n almost surely and strongly in L2 (Ω).
Observe that for every n ≤ m ∈ N, it follows that
(3.4) M∞,n = M∞,m on the set {hM i∞ < n}.
We therefore define
(
M∞,n (ω) if ω ∈ {hM i∞ < n} for some n ∈ N,
M∞ (ω) =
0 if ω ∈ {hM i∞ = ∞}.
It then follows from (8.7) that, for almost every ω ∈ {hM i∞ < ∞}, as t → ∞,
Mt (ω) → M∞ (ω),
44 BENJAMIN FEHRMAN
Let (Mt )t∈[0,∞) ∈ H, and suppose that
∆ = {0 = t0 < t1 < t2 < t3 < . . .},
is a locally finite partition of [0, ∞). Here locally finite means that for every compact set K ⊆ [0, ∞)
the intersection K ∩ ∆ is finite. For uniformly bounded random variables {Ki }i∈N satisfying that
Ki is Fti -measurable for every i ∈ N, consider the simple process, for t ∈ [0, ∞),
∞
X
Kt = Ki 1(ti ,ti+1 ] (t).
i=0
The stochastic integral of the process (Kt )t∈[0,∞) with respect to (Mt )t∈[0,∞) should mimic the
discrete stochastic integral, and were therefore expect that
Z ∞ X∞
Ks dMs = Ki (Mti+1 − Mti ).
0 i=0
Since the random variables {Ki }i∈N are uniformly bounded and since Ki is Fti -measurable for every
i ∈ N, it follows from Proposition 2.10 that the process
∞
Z t !
X
Ks dMs = Ki (Mti+1 ∧t − Mti ∧t ) ,
0 i=0 t∈[0,∞)
is a martingale. Furthermore, by the martingale property and the predictability and uniform
boundedness of (Kt )t∈[0,∞) ,
"Z 2 # "∞ #
∞ X
2 2 2
2
Ks dMs =E Ki (Mti+1 − Mti ) ≤ sup Kt E M∞ .

E
0 i=1
t∈[0,∞)
∞
We therefore conclude that the stochastic integral of a uniformly bounded, predictable simple
process with respect to an element of H is again an element of H. In the following proposition, we
observe that the stochastic integral of a discrete process defines an isometry by which we come to
extend the integral to more general processes.
Proposition 3.12. Let (Mt )t∈[0,∞) and (Nt )t∈[0,∞) be elements of H, let
∆ = {0 = t0 < t1 < t2 < t3 < . . .},
be a locally finite partition of [0, ∞), and for uniformly bounded random variables {Ki }i∈N satisfying
that Ki is Fti -measurable for every i ∈ N let (Kt )t∈[0,∞) be defined by
∞
X
Kt = Ki 1(ti ,ti+1 ] .
i=0
Then,
Z · Z ∞
hh Ks dMs , N· ii = E Ks dhM, N is .
0 0
Proof. Since
Z t Z ∞
lim Ks dMs = Ks dMs strongly in L2 (Ω),
t→∞ 0 0
it follows by definition and the martingale property that

Z · Z ∞
hh Ks Ms , N· ii = E Ks dMs N∞
0 0
 ! ∞ 
X∞ X
= E Ki (Mti+1 − ti )  (Ntj+1 − Ntj )
i=0 j=1
∞
" #
X
=E Ki (Mti+1 − Mti )(Nti+1 − Nti )
i=0
∞
" #
X
=E Ki (Mti+1 Nti+1 − Mti Nti ) .
i=0
We will now analyze the final term on the righthand side of this equality. The martingale property
and the predictability of (Kt )t∈[0,∞) prove that, for every i ∈ N0 ,
E[Ki (Mti+1 Nti+1 − Mti Nti )] = E[Ki (E[Mti+1 Nti+1 |Fti ] − Mti Nti )].
Since the process (Mt Nt − hM, N it )t∈[0,∞) is a martingale,
E[Mti+1 Nti+1 |Fti ] = E[hM, N iti+1 |Fti ] + Mti Nti − hM, N iti .
Therefore, by the martingale property and predictability of (Kt )t∈[0,∞) , for every i ∈ N0 ,
E[Ki (Mti+1 Nti+1 − Mti Nti )] = E[Ki (E[hM, N iti+1 |Fti ] − hM, N iti )]
= E[Ki (hM, N iti+1 − hM, N iti )].
Therefore, by definition of the deterministic integral with respect to the quadratic covariation,
Z · "∞ #
X
hh Ks Ms , N· ii = E Ki (hM, N iti+1 − hM, N iti )
0 i=0
Z ∞
=E Ks dhM, N is .
0
Observe in particular that, by choosing
Z t
Nt = Ks dMs ∈ H,
0 t∈[0,∞)
the above computation proves that
"Z 2 #
∞ Z ∞
E Ks dMs =E Ks2 dhM is .
0 0
This equality is the Itô isometry, and it suggests the class of processes for which we can define the
stochastic integral with respect to an element (Mt )t∈[0,∞) ∈ H.
Definition 3.13. For every (Mt )t∈[0,∞) ∈ H, we define the space
Z ∞
2
L (M ) = (Kt )t∈[0,∞) : (Kt )t∈[0,∞) is predictable and satisfies E Ks2 dhM is <∞ .
0
We define, for every (Kt )t∈[0,∞) ,
Z ∞
kKkL2 (M ) = E Ks2 dhM is .
0
46 BENJAMIN FEHRMAN
In the following theorem, for every (Mt )t∈[0,∞) ∈ H and (Kt )t∈[0,∞) ∈ L2 (M ), we construct the
stochastic integral
Z t
Ks dMs ,
0 t∈[0,∞)
as an element of H. The proof is a consequence of the Riesz representation theorem, and it is for
this reason that we proved that H is a complete Hilbert space.
Proposition 3.14. For every (Mt )t∈[0,∞) ∈ H and (Kt )t∈[0,∞) ∈ L2 (M ), there exists a unique
element ((K • M )t )t∈[0,∞) ∈ H which satisfies for every (Nt )t∈[0,∞) that
Z ∞
hhK • M, N ii = E Ks dhM, N is .
0
We will henceforth write, for every t ∈ [0, ∞),
Z t
(K • M )t = Ks dMs .
0
Proof. The Kunita-Watanabe inequality proves that, for every (Nt )t∈[0,∞) ∈ H,
Z ∞ Z ∞

E
Ks dhM, N is ≤ E
|Ks | d |hM, N i|s
0 0
"Z
∞ 1 # "Z ∞ 1 #
2 2
2
≤E Ks dhM is E dhN is
0 0
1
= kKkL2 (M ) E[hN i∞ ] . 2
= kKkL2 (M ) kN kH .
We therefore conclude that the map
Z ∞
(Nt )t∈[0,∞) ∈ H 7→ E Ks dhM, N is ∈ R
0
is a continuous linear functional on H. The Riesz representation theorem therefore proves that
there exists a unique elements ((K • M )t )t∈[0,∞) ∈ H which satisfies that, for every (Nt )t∈[0,∞) ∈ H,
Z ∞
hhK • M, N ii = E Ks dhM, N is .
0
Observe in particular that, for (Mt )t∈[0,∞) ∈ H and K ∈ L2 (M ), by choosing (Nt = (K •
M )t )t∈[0,∞) we have
"Z 2 #
∞ Z ∞
2
hhK • M, N ii = E Ks dMs =E Ks dhM is ,
0 0
which is the general Itô isometry. This isometry implies that if (K n )n∈N ⊆ L2 (M ) and K ∈ L2 (M )
satisfy that Z ∞
n n 2
lim kK − K kL2 (M ) = lim E |Ks − Ks | dhM is = 0,
n→∞ n→∞ 0
then it follows that "Z
∞ Z ∞ 2 #
lim E Ksn dhM is − Ks dhM is = 0.
n→∞ 0 0
This fact can be used to justify the more standard construction of the stochastic integral based
upon Riemann sum approximations. Precisely, for a sequence of locally finite partitions
{∆n = {0 = tn0 < tn1 < tn2 < . . .}}n∈N
which satisfy that |∆n | → 0 as n → ∞, define suitable simple approximations
∞
X
Ktn = Kin 1(tni ,tni+1 ] (t),
i=0
which satisfy
lim kK n − KkL2 (M ) = 0.
n→∞
Since for simple processes we know that, for every t ∈ [0, ∞),
Z t X∞
n
Ks dMs = Kin (Mtni+1 ∧t − Mtni ∧t )
0 i=0
it follows that, as n → ∞,
∞
!
X Z t
Kin (Mtni+1 ∧t − Mtni ∧t ) → Ks dMs strongly in H.
i=0 0 t∈[0,∞)
t∈[0,∞)
We will now extend the definition of the stochastic integral to the class of continuous local
martingales. If (Mt )t∈[0,∞) is a continuous local martingale, then there exists a nondecreasing
sequence of stopping times {taun }n∈N such that τn → ∞ almost surely as n → ∞, and such that
for every n ∈ N we have
(Mtτn )t∈[0,∞) = (Mτn ∧t 1{τn >0} )t∈[0,∞) ∈ H.
Since it holds that
hM τn it = hM 1{τn >0} iτn ∧t ,
we define the space L2loc (M ) to be the space of predictable processes (Kt )t∈[0,∞) for which there
exists a nondecreasing sequence of stopping times {σn }n∈N satisfying almost surely that σn → ∞
as n → ∞ such that Z σn Z ∞
σn
E Ks dhM is = E Ks dhM is < ∞.
0 0
The following proposition is then a consequence of localization.
Proposition 3.15. Let (Mt )t∈[0,∞) be a continuous local martingale and let (Kt )t∈[0,∞) ∈ L2loc (M ).
Then, there exists a unique continuous local martingale ((K • M )t )t∈[0,∞) such that, for every
continuous local martingale (Nt )t∈[0,∞) there exists an nondecreasing sequence of stopping times
{τn }n∈N satisfying almost surely τn → ∞ as n → ∞ such that, for every n ∈ N,
(Mtτn )t∈[0,∞) , (Ntτn )t∈[0,∞) ∈ H and (Kt )t∈[0,∞) ∈ L2 (M τn ),
such that, for every n ∈ N,
Z t
Ks dhM is τn
= ((K • M )τt n )t∈[0,∞) ∈ H,
0 t∈[0,∞)
and such that, for every n ∈ N,

Z ∞
τn τn τn
hh(K • M ) , N ii = E Ks dhM , N is .
0
In the final definition of this section, we define the stochastic integral with respect to a continuous
semimartingale. The integral with
48 BENJAMIN FEHRMAN
Definition 3.16. Let (Xt = At + Mt )t∈[0,∞) be a continuous semimartingale, where (At )t∈[0,∞) is
a continuous process of bounded variation, and where (Mt )t∈[0,∞) is a continuous local martingale.
For every bounded process (Kt )t∈[0,∞) ∈ L2loc (M ), we define the continuous semimartingale
Z t Z t Z t
Ks dXs = Ks dAs + Ks dMs .
0 t∈[0,∞) 0 0 t∈[0,∞)
3.3. The integration-by-parts formula. In this section, we will prove an integration-by-parts

formula that plays the role of the Leibniz rule in stochastic calculus. We will use this formula in
the proof of Itô’s formula below. The proof is a straightforward consequence of the definition of
the quadratic variation and the existence of the stochastic integral.
Proposition 3.17. Let (Mt )t∈[0,∞) and (Nt )t∈[0,∞) be continuous semimartingales. Then the pro-
cess (Mt Nt )t∈[0,∞) is a continuous semimartingale which satisfies, for every t ∈ [0, ∞),
Z t Z t
Mt Nt = M0 N0 + Ns dMs + Ms dNs + hM, N it .
0 0
Or, in differential notation,
d(Mt Nt ) = Nt dMt + Mt dNt + dhM, N it .
Proof. Let ∆ ⊆ [0, ∞) be a locally finite partition
∆ = {0 = t0 < t1 < t2 < . . .}.
For every i ∈ N0 , observe that
Mti+1 Nti+1 − Mti Nti = Mti (Nti+1 − Nti ) + Nti (Mti+1 − Mti ) + (Mti+1 − Mti )(Nti+1 − Nti ).
Therefore, for every t ∈ [0, ∞),
X∞
Mt Nt = M0 N0 + (Mti+1 ∧t Nti+1 ∧t − Nti ∧t Mti ∧t )
i=0
∞
X
= M0 N0 + Mti (Nti+1 − Nti ) + Nti (Mti+1 − Mti ) + (Mti+1 − Mti )(Nti+1 − Nti ) .
i=0
Since the fact that (Mt )t∈[0,∞) and (Nt )t∈[0,∞) are continuous implies that they are respectively in
L2loc (N ) and L2loc (M ), it follows after passing to the limit |∆| → 0 that, for every t ∈ [0, ∞),
Z t Z t
Mt Nt = M0 N0 + Ms dNs + Ns dMs + hM, N it ,
0 0
50 BENJAMIN FEHRMAN
3.4. Proof of Itô’s formula. We are now prepared to prove Itô’s formula, which is essentially the
fundamental theorem of the calculus of semimartingales. There are many approaches to prove this
statement, much like for the construction of the stochastic integral, including a direct argument
based on Taylor expansions and discrete approximations along a sequence of partitions whose mesh
approaches zero. The approach we take is based on the integration-by-parts formula and the Stone-
Weierstrass theorem, which states that polynomials are dense in the space of continuous functions
on a compact set of Euclidean space.
Theorem 3.18. Let (Xt )t∈[0,∞) be a continuous, d-dimensional semimartingale and let f ∈ C2 (Rd ).
Then (f (Xt ))t∈[0,∞) is a semimartingale which satisifes
d Z t d Z
X ∂f 1 X t ∂2f
(3.5) f (Xt ) = f (X0 ) + (Xs ) dXsi + (Xs ) dhX i , X j is .
0 ∂xi 2 0 ∂xi ∂xj
i=1 i,j=1
Proof. We may assume without loss of generality that the semimartingale (Xt )t∈[0,∞) takes values
in a compact subset of Rd . For, if not, we can introduce the stopping times {TR }R∈(0,∞) defined
by
TR = inf{t ∈ [0, ∞) : Xt ∈
/ BR },
prove the theorem for the stopped process (XtTR )t∈[0,∞) , and pass to the limit R → ∞ using the
continuity of (Xt )t∈[0,∞) .
We observe that for every i ∈ {1, . . . , d} equation (3.5) is satisfied by the function fi (x) = xi ,
since
Z t
Xti = X0i + dXsi .
0
Similarly, formula (3.5) is satisfied by the constant function f = 1 and if functions f, g ∈ C2 (Rd )
and c ∈ R then by linearity if f and g satisfy (3.5) so too do cf and f + g. It remains to prove that
the product f g satisfy (3.5).
Assume that f, g ∈ C2 (Rd ) satisfy (3.5). Then, the processes (f (Xt ))t∈[0,∞) and (g(Xt ))t∈[0,∞)
are semimartingales, and the integration-by-parts formula proves that
(3.6) d(f (Xt )g(Xt )) = f (Xt ) dg(Xt ) + g(Xt ) df (Xt ) + dhf (Xt ), g(Xt )it ,
where
d d
X ∂g 1 X ∂2g
(3.7) dg(Xt ) = (Xt ) dXti + (Xt ) dhX i , X j it ,
∂xi 2 ∂xi ∂xj
i=1 i,j=1
where
d d
X ∂f i 1 X ∂2f
(3.8) df (Xt ) = (Xt ) dXt + (Xt ) dhX i , X j it ,
∂xi 2 ∂xi ∂xj
i=1 i,j=1
and where
d
X ∂g ∂f
(3.9) dhf (X), g(X)it = (Xt ) (Xt ) dhX i , X j it
∂xi ∂xj
i,j=1
Returning to (3.6), it follows from the identities

∂(f g) ∂f ∂g ∂(f g) ∂2f ∂f ∂g ∂f ∂g ∂2g
= g+f and = g+ + +f ,
∂xi ∂xi ∂xi ∂xi ∂xj ∂xi ∂xj ∂xi ∂xj ∂xj ∂xi ∂xi ∂xj
from the symmetry of the quadratic covariation, and from (3.7), (3.8), and (3.9) that
d d
X ∂(f g) 1 X ∂ 2 (f g)
d(f (Xt )g(Xt )) = (Xt ) dXsi + (Xt ) dhX i , X j it ,
∂xi 2 ∂xi ∂xj
i=1 i,j=1
which completes the proof that the product (f g) satisfies (3.5). We therefore conclude that every
polynomial function satisfies (3.5).
Let f ∈ C2 (Rd ) be arbitrary and fix R ∈ (0, ∞) such that Xt ∈ BR for every t ∈ [0, ∞). Then,
by the Stone-Weierstrass theorem, there exists a sequence of polynomials {pn }n∈N such that, in the
C2 -topology,
!
2 2

lim sup |f (x) − pn (x)| + |∇f (x) − ∇pn (x)| + ∇ f (x) − ∇ pn (x) = 0.
n→∞ x∈BR
Since each of the polynomials {pn }n∈N satisfies (3.5), after passing to the limit n → ∞ the uniform
C2 -convergence of the {pn }n∈N to f and the Itô isometry prove that f satisfies (3.5). This completes
the proof.
In the case of a standard d-dimensional Brownian motion (Bt )t∈[0,∞) , it follows by choosing
f (x) = |x|2 in Itô’s formula that
d Z t
X d Z t
X
|Bt |2 = 2 Bsi dBsi + dhB i , B j it
i=1 0 i,j=1 0
d Z
X t d Z
X t
=2 Bsi dBsi + dt .
i=1 0 i=1 0
d Z
X t
=2 Bsi dBsi + dt
i=1 0
Hence, the square of the Euclidean norm of a d-dimensional Brownian motion satisfies
d Z t
X 1
Bsi dBsi = |Bt |2 − dt .
0 2
i=1
In particular, in one-dimension, Z t
1
Bs dBs = (Bt2 − t),
0 2
2
which reproves the statement that (Bt − t)t∈[0,∞) is a martingale.
52 BENJAMIN FEHRMAN
4. Applications of stochastic calculus

In this section, we will explain some important applications of stochastic calculus. The first
is the stochastic analogue of the exponential function, which we will subsequently use to prove
Levy’s characterization of Brownian motion and the martingale representation theorem. We will
use Levy’s theorem to prove the Dambis-Dubins-Schwarz theorem. We will use Itô’s formula to
provide a proof of the Burkholder-Davis-Gundy inequality.
4.1. The stochastic exponential. The exponential function Yt = exp(t) is the unique solution
to the ordinary differential equation
dYt = Yt dt in (0, ∞),

Y0 = 1,
and in this way plays an essential role in the theory of ordinary and partial differential equations.
For instance, we can reduce the nonlinear partial differential equation
∂t u = ∆u + |∇u|2 in Rd × (0, ∞),
(
u = u0 on Rd × {0},
by making the exponential transformation
v(x, t) = exp(u(x, t)),
whereby the chain rule implies that v solves the linear heat equation
in Rd × (0, ∞),
(
∂t v = ∆v
v = exp(u0 ) on Rd × {0},
which can be solved explicitly in terms of the heat kernel. We will see below, for the simpler
example of a geometric Brownian motion, that the stochastic exponential defined below can be
similarly used to simplify terms appearing stochastic differential equations.
Proposition 4.1. Let (Xt )t∈[0,∞) be a continuous semimartingale that vanishes at zero. Then the
continuous semimartingale (E(X)t )t∈[0,∞) defined by

hXit
E(X)t = exp Xt − ,
2
is the unique continuous semimartingale solution of the stochastic differential equation
dZt = Zt dXt in (0, ∞),

(4.1)
Z0 = 1.
The process (E(X)t )t∈[0,∞) is the stochastic exponential of (Xt )t∈[0,∞) .
Proof. We first observe formally that if (Zt )t∈[0,∞) solves (4.1) then Itô’s formula would prove that
(log(Zt ))t∈[0,∞) solves
1 1
d(log(Zt )) = dZt − dhZit
Zt 2Zt2
1
= dXt − dhXit .
2
This would imply that
hXit
log(Zt ) = Xt − ,
2
from which it follows that

hXit
(4.2) Zt = exp Xt − .
2
This argument, however, is not justified because we do not know a priori that a semimartingale
solution to (4.1) exists, and because the logarithm is not a C2 -function on the whole of R. The
latter of these points is not so serious, however, because we could instead analyze approximating
54 BENJAMIN FEHRMAN
processes that are stopped before hitting a neighborhood of the origin. Nevertheless, equation (4.2)
provides a candidate for the solution, and we define the stochastic exponential (E(X)t )t∈[0,∞) by

hXit
E(X)t = exp Xt − .
2
It then follows by definition that E(X)0 = 0 and by using Itô’s formula that
1 1
dE(X)t = E(X)t dXt − E(X)t dhXix + E(X)t dhXit
2 2
= E(X)t dXt ,
which complets the proof that (E(X)t )t∈[0,∞) solves (4.1).
In order to prove uniqueness, observe using Itô’s formula that the inverse process (E(X)−1t )t∈[0,∞)
defined by
−1 hXit
E(X)t = exp −Xt + ,
2
solves the equation
dE(X)−1 −1 −1
t = −E(X)t dXt + E(X)t dhXit .
If (Zt )t∈[0,∞) is a continuous semimartingale solution of (4.1), then the integration-by-parts formula
proves that
d(Zt E(X)−1 −1 −1 −1
t ) = Zt dE(X)t + E(X)t dZt + dhE(X) , Zt it
= −Zt E(X)−1 −1 −1 −1
t dXt + Zt E(X)t dXt + Zt E(X)t dhXit − Zt E(X)t dhXit
= 0.
Zt E(X)−1 −1
t = Z0 E(X)t = 1 and Zt = E(X)t ,
A geometric Brownian motion is a stochastic process (St )t∈[0,∞) that satisfies the stochastic
differential equation, for a standard one dimensional Brownian motion (Bt )t∈[0,∞) , for some µ, σ ∈
R,
dSt = µSt dt + σSt Bt in (0, ∞),

(4.3)
S0 = 1.
Versions of such processes play an important role in mathematical finance, where (St )t∈[0,∞) rep-
resents the value of a stock or bond, where the parameter µ is the interest rate, and where the
parameter σ quantifies the volatility of a financial market. In the following proposition, we use the
stochastic exponential to solve equation (4.3). Observe, in particular, that the solution is positive.
Proposition 4.2. Let (Bt )t∈[0,∞) be a standard one-dimensional Brownian motion and let µ, σ ∈ R.
Then the stochastic process (St )t∈[0,∞) defined by
σ2

(4.4) St = exp µ− t + σBt ,
2
is the unique continuous semimartingale solution of (4.3).
Proof. The fact that (St )t∈[0,∞) is a solution of (4.3) is an immediate consequence of Itô’s formula.
Now, suppose that (St )t∈[0,∞) is an arbitrary continuous semimartingale solution of (4.3), and let
(E(σB)t )t∈[0,∞) be the unique solution of the equation
(
dE(σB)t = σE(σB)t dBt in (0, ∞),
E(σB)0 = 1,
defined in Proposition 4.1 by

σ2t

E(σB)t = exp σBt − ,
2
since hBit = t. It was shown in Proposition 4.1 that the inverse process (E(σB)−1
t )t∈[0,∞) satisfies
E(σB)−1 −1 2
t = −σE(σB)t dBt + σ E(σB)t dt.
Therefore, by the integration-by-parts formula,
d(St E(σB)−1 −1 −1 −1
t ) = St dE(σB)t + E(σB)t dSt + dhS, E(σB) it
= −σSt E(σB)−1 −1 −1
t dBt + σSt E(σB)t σ dBt + µSt E(σB)t dt
+ σ 2 St E(σB)−1 2 −1
t dt − σ St E(σB)t dt
= µSt E(σB)−1
t dt.
We therefore conclude that
St E(σB)−1
t = exp(µt),
and therefore that
σ2

St = exp µ− t + σBt .
2
56 BENJAMIN FEHRMAN
4.2. Levy’s characterization of Brownian motion. In this section, we will use stochastic calcu-
lus to prove that Brownian motion is uniquely characterized in the class of continuous local martin-
gales by its quadratic variation. Observe in particular that for general martingales (Mt )t∈[0,∞) the
quadratic variation process (hM it )t∈[0,∞) is a random process, in the sense that for every t ∈ (0, ∞)
the random variable hM it is not deterministic, which is not the case for a standard Brownian
motion.
Theorem 4.3. Let (Mt )t∈[0,∞) be a continuous local martingale vanishing at zero. Then (Mt )t∈[0,∞)
is a standard Brownian motion if and only if, for every t ∈ [0, ∞),
hM it = t.
Proof. If (Mt )t∈[0,∞) = (Bt )t∈[0,∞) is a Brownian motion, then for every T ∈ (0, ∞) and N ∈ N,
X∞ 2
B k+1 ∧T − B k ∧T ,
N N
k=0
is a sum of independent random variables with mean 1/N . Hence, after rewriting this sum,
∞ ∞
!
X 2 1 X 2
B k+1 ∧T − B k ∧T = N B k+1 ∧T − B k ∧T ,
N N N N N
k=0 k=0
the law of large numbers proves that, as N → ∞,
∞
!
1 X 2
N B k+1 ∧T − B k ∧T → T almost surely,
N N N
k=0
from which we conclude that hBiT = T , for every T ∈ (0, ∞).
Now suppose that (Mt )t∈[0,∞) is a continuous local martingale satisfying hM it = t for every
t ∈ [0, ∞). It then follows that, for every T ∈ (0, ∞), the stopped martingale (MtT )t∈[0,∞) =
(Mt∧T )t∈[0,∞) satisfies hM T i∞ = hM iT = T . We therefore conclude that, for every T ∈ (0, ∞),
the martingale (MtT )t∈[0,∞) is L2 -bounded and hence uniformly integrable. This implies that
(MtT )t∈[0,∞) is a martingale, for every T ∈ (0, ∞), and hence that (Mt )t∈[0,∞) is a martingale.
We will not prove that (Mt )t∈[0,∞) is a Brownian motion.
Since it follows by definition that M0 = 0 and that (Mt )t∈[0,∞) is almost surely continuous, it
remains only to prove that Mt − Ms is normally distributed with mean zero and variance t − s, for
every s ≤ t, and that Mt − Ms is independent of Fs , for every s ≤ t. Let s ≤ t. For the proof, we
will essentially study the characteristic function, for each α ∈ R,
E[exp(iα(Mt − Ms ))],
which is related to the complex exponential
α2 α2 r

exp iαMr + hM ir = exp iαMr + .
2 t∈[0,∞) 2 r∈[0,∞)
Observe that this is the stochastic exponential of the complex martingale (iαMr )r∈[0,∞) .
We say that a complex-valued process is a complex martingale if both its real and complex parts
are martingales. In the case of the complex exponential, for every r ∈ [0, ∞),
α2 r
2
α r
exp iαMr + = exp (cos(αMr ) + i sin(αMr )) .
2 2
Since the argument for both the real and complex parts are the same, we will prove that
2
α r
exp sin(αMr )
2 r∈[0,∞)
is a martingale. Define the process of bounded variation

(Ar )r∈[0,∞) = (sin(αMr ))r∈[0,∞) ,
and define the process (Zr )r∈[0,∞) by
Zr = sin(αMr ).
. The integration-by-parts formula proves that
2
α r
d exp sin(αMr ) = Ar dZr + Zr dAr + dhA, Zir ,
2
where Itô’s formula and hM ir = r prove that
α2
dZr = α cos(αMr ) dMr − sin(αMr ) dr,
2
where ordinary calculus proves that
α2 α2 r

dAr = exp ,
2 2
and where dhA, Zir = 0 since (Ar )r∈[0,∞) has bounded variation. Hence,
2 2
α r α r
d exp sin(αMr ) = α exp cos(αMr ) dMr ,
2 2
which implies using the boundedness of the cos function that the lefthand side is a martingale.
And, therefore, we conclude that the complex stochastic exponential is an exponential martingale.
The martingale property proves that
α2 t α2 s

E exp iαMt + |Fs = exp iαMs + .
2 2
Since the deterministic functions are Fs -measurable, and since Ms is Fs -measurable, this implies
that
α2

E [exp(iα(Mt − Ms ))|Fs ] = exp − (t − s) .
2
This implies that (Mt − Ms ) is normally distributed and independent of Fs , which completes the
proof. (The full proof will be added after the next problem session.)
58 BENJAMIN FEHRMAN
4.3. The Dambis-Dubins-Schwarz theorem. In the case of a standard Brownian motion, the
oscillations of the process effectively act as a clock in the sense that the quadratic variation hBit = t
for every t ∈ [0, ∞). In this section, we will prove that every continuous martingale (Mt )t∈[0,∞) that
“oscillates enough” in the sense that hM i∞ = ∞ almost surely is merely a time-changed Brownian
motion. That is, on the event {hM it > t} the martingale is behaving like a Brownian motion
that has been “sped-up”, and on the event {hM it < t} the martingale is behaving like a Brownian
motion that has been “slowed-down.” We make the notion of a time change precise in the next
definition.
Definition 4.4. Let (Ω, F, P) be a probability space, and let (Ft )t∈[0,∞) be a filtration on (Ω, F).
A time change is a nondecreasing sequence of right-continuous Ft -stopping times (σt )t∈[0,∞) which
satisfy σs → ∞ as s → ∞ almost surely.
Observe in particular that, given a uniformly integrable martingale (Mt )t∈[0,∞) and a time change
(σt )t∈[0,∞) , we can define the time-changed process (Mσt )t∈[0,∞) . The optional stopping theorem
proves that, for every s ≤ t ∈ [0, ∞),
E[Mσt |Fσs ] = Mσs .
That is, the time-changed process (Mσt )t∈[0,∞) is a martingale with respect to (Fσt )t∈[0,∞) , which
is the filtration generated by the time change (σt )t∈[0,∞) .
We will now show that every continuous local martingale (Mt )t∈[0,∞) that satisfies hM i∞ = ∞
almost surely is a time-changed Brownian motion. The time change is determined by the quadratic
variation process (hM it )t∈[0,∞) . We will “slow down” the martingale if its quadratic variation is
larger than t, and we will “speed up” the martingale if the quadratic variation is smaller than t.
The theorem follows immediately after the next two lemmas.
Lemma 4.5. Let (Mt )t∈[0,∞) be a continuous local martingale vanishing at zero. Then Mt = 0 for
every t ∈ [0, ∞) if and only if hM it = 0 for every t ∈ [0, ∞).
Proof. If Mt = 0 for every t ∈ [0, ∞) then it follows by definition that hM it = 0 for every t ∈ [0, ∞).
To prove that, suppose that hM it = 0 for every t ∈ [0, ∞). Then, since (Mt2 −hM it )t∈[0,∞) is a local
martingale, it follows that there exist a sequence of stopping times {τn }n∈N that satisfy τn → ∞
almost surely as n → ∞ such that
2
E[Mt∧τ n
] = 0.
Therefore, by continuity, it follows almost surely that Mt2 = 0 for every t ∈ [0, ∞). This completes
the proof.
Lemma 4.6. Let (Mt )t∈[0,∞) be a continuous local martingale vanishing at zero. Then, the inter-
vals of constancy of (Mt )t∈[0,∞) and (hM it )t∈[0,∞) coincide almost surely. That is, almost surely,
Mt (ω) = Mb (ω) for every t ∈ [a, b] if and only if hM it (ω) = hM ib (ω) for every t ∈ [a, b].
Proof. It follows by definition that (hM it )t∈[0,∞) is constant on an interval [a, b] if (Mt )t∈[0,∞) is
constant on [a, b]. Let a ∈ Q ∩ [0, ∞) and let Ta denote the stopping time
Ta = inf{s ∈ (a, ∞) : hM is > hM ia }.
A repetition of the above argument proves almost surely that Mt = Ma for every t ∈ [0, Ta ]. It
then follows by density of the rationals and continuity that if hM it is constant on an interval [a, b]
then almost surely Mt is constant on [a, b]. This completes the proof.
Theorem 4.7. Let (Mt )t∈[0,∞) be a continuous local martingale vanishing at zero that satisfies
hM i∞ = ∞ almost surely. Define the stopping times {σt }t∈[0,∞) by
σt = inf{s ∈ [0, ∞) : hM is > t} = sup{s ∈ [0, ∞) : hM is = t}.
Then, (Mσt )t∈[0,∞) is a standard Brownian motion (Bt )t∈[0,∞) and, for every t ∈ [0, ∞),
Mt = BhM it .
Proof. First, we observe that the stopping times (σt )t∈[0,∞) are indeed stopping times. This relies
on the right continuity of the filtration (Ft )t∈[0,∞) , which means that, for every t ∈ [0, ∞),
Ft = ∩s>t Fs .
Let Mσ denote the time-changed process (Mσt )t∈[0,∞) . Then, Mσ is a Fσt -martingale. Indeed, for
every t ∈ [0, ∞), the stopped process (Mσt ∧s )s∈[0,∞) is L2 -bounded and hence uniformly integrable
since hM σt i∞ = hM iσt = t almost surely. Therefore, the optional stopping theorem applies and,
for every s ≤ t ∈ [0, ∞),
E[Mσt |Fσs ] = Mσs .
Since t ∈ [0, ∞) was arbitrary, this completes the proof that (Mσt )t∈[0,∞) is a Fσt -martingale.
Let M σ denote the Fσt -martingale (Mσt )t∈[0,∞) . We aim to apply Levy’s characterization of
Brownian motion, and for this it is necessary to prove continuity and to prove that the quadratic
variation satisfies hM σ it = t. For the quadratic variation, since the assumption hM i∞ = ∞ implies
that σt < ∞ almost surely, we have that by definition of the time change, for every t ∈ [0, ∞),
hM σ it = hM σt i∞ = hM iσt = t.
To prove continuity, for every t ∈ [0, ∞), the definition of the time change proves that
lim Mσs = Mt ,
s→t−
where by continuity of (hM it )t∈[0,∞) ,
t = sup{s ∈ [0, ∞) : hM is < t} = min{s ∈ [0, ∞) : hM is = t}.
Similarly,
lim Mσs = Mt ,
s→t−
where
t = inf{s ∈ [0, ∞) : hM is > t} = max{s ∈ [0, ∞) : hM is = t} = σt .
Since the intervals of constancy of the quadratic variation and the continuous local martingale
almost surely coincide, and since the quadratic variation is constant on the interval [t, t] = [t, σt ],
we conclude that
M t = M t = Mσ t ,
and therefore that
lim Mσs = Mσt .
s→t
Levy’s characterization of Brownian motion implies that (Mσt )t∈[0,∞) is a standard Fσt -Brownian
motion (Bt )t∈[0,∞) where by definition Bt = Mσt .
It remains to prove that, for every t ∈ [0, ∞),
Mt = BhM it .
By definition, for every t ∈ [0, ∞),
BhM it = BσhM it ,
where by definition, for every t ∈ [0, ∞),
σhM it = inf{s ∈ [0, ∞) : hM is > hM it } = sup{s ∈ [0, ∞) : hM is = hM it }.
Therefore, for every t ∈ [0, ∞), it follows that σhM it ≥ t and that the quadratic variation is
constant on the interval [t, σhM it ]. Since the intervals of constancy of the quadratic variation and
the martingale almost surely coincide, we conclude by definition of (Bt )t∈[0,∞) that
Mt = MσhM it = BhM it ,
60 BENJAMIN FEHRMAN

4.4. The Burkholder-Davis-Gundy inequality. We have already observed that if (Mt )t∈[0,∞)
is an L2 -bounded continuous martingale, then the quadratic variation process (hM it )t∈[0,∞) exists
and we have, for every t ∈ [0, ∞),
E[Mt2 ] = E[hM it ].
If we define the running maximum
Mt∗ = sup |Ms | ,
s∈[0,t]
it is then immediate that
E[hM it ] = E[Mt2 ] ≤ E[(Mt∗ )2 ].
Alternately, Doob’s inequality proves for every t ∈ [0, ∞) that, since (Mt2 )t∈[0,∞) is a submartingale,
E[(Mt∗ )2 ] ≤ 4E[Mt2 ] = 4E[hM it ].
In combination, therefore, we have for every t ∈ [0, ∞) that
1
E[(Mt∗ )2 ] ≤ E[hM it ] ≤ E[(Mt∗ )2 ],
4
which implies that the L2 -norm of the running maximum and the L1 -norm of the quadratic variation
are equivalent norms. The Burkholder-Davis-Gundy inequalities prove that this is the case for every
Lp -norm, for every p ∈ (0, ∞), in the sense that the L2p -norm of the running maximum and the
Lp -norm of the quadratic variation are equivalent.
Theorem 4.8. For every p ∈ (0, ∞) there exist constants cp , Cp ∈ (0, ∞) such that, for every
continuous martingale (Mt )t∈[0,∞) vanishing at zero, for every t ∈ [0, ∞),
(4.5) cp E[(Mt∗ )2p ] ≤ E[hM ipt ] ≤ Cp E[(Mt∗ )2p ].
Proof. Let (Mt )t∈[0,∞) be a continuous martingale vanishing at zero. For every n ∈ N let Tn be the
stopping time defined by
Tn = inf{t ∈ [0, ∞) : |Mt | ≥ n or hM it ≥ n}.
Observe that if we prove that the theorem holds for the stopped martingales (MtTn )t∈[0,∞) , for
constants that are independent of n ∈ N, then we can apply the monotone convergence theorem to
deduce that the theorem holds for (Mt )t∈[0,∞) . The monotone convergence theorem applies because,
for every t ∈ [0, ∞), the functions (M Tn )∗t and hM Tn it are nondecreasing functions of n ∈ N. We can
therefore assume without loss of generality that (Mt )t∈[0,∞) is a bounded martingale with bounded
quadratic variation. This assumption guarantees that, in the applications of Itô’s formula to follow,
all of the integrals appearing are true martingales, as opposed to only local martingales.
Henceforth, let (Mt )t∈[0,∞) be a bounded continuous martingale vanishing at zero with bounded
quadratic variation. We have already proven (4.5) in the case p = 1. Suppose now that p ∈ (1, ∞).
Then, since the function f (x) = |x|2p is twice continuously differentiable, it follows from Itô’s
formula and the fact that (Mt )t∈[0∞) vanishes at zero that
Z t Z t
2p
|Mt | = 2p Ms2p−1
sgn(Ms ) dMs + p(2p − 1) |Ms |2(p−1) dhM is .
0 0
Hence, after taking the expectation, using the fact that the first term on the righthand side is a
martingale,
h i Z t h i
2p 2(p−1)
E |Mt | = p(2p − 1)E |Ms | dhM is ≤ p(2p − 1)E (Mt∗ )2(p−1) hM it .
0
Hence, by Hölder’s inequality with exponents p and p/p−1,
h i p−1 1
E |Mt |2p ≤ p(2p − 1)E (Mt∗ )2p p E [hM ipt ] p .

62 BENJAMIN FEHRMAN
Since Jensen’s inequality implies that |Mt |2p is a submartingale, it follows by Doob’s inequality
that 2p
2p
∗ 2p
E[(Mt ) ] ≤ E[|Mt |2p ],
2p − 1
and therefore we have that
2p
2p p−1 1
∗ 2p
E[(Mt ) ] ≤ p(2p − 1)E[(Mt∗ )2p ] p E[hM ipt ] p .
2p − 1
p−1
After dividing through by E[(Mt∗ )2p ] p ,
2p
1 2p 1
E (Mt∗ )2p p ≤ p(2p − 1)E [hM ipt ] p ,

2p − 1
from which we conclude that
h i 2p 3p
E |Mt |2p
≤ (p(2p − 1))p E [hM ipt ] .
2p − 1
This completes the proof of the leftmost inequality in (4.5) with
−3p
2p
cp = (p(2p − 1))−p .
2p − 1
For the rightmost inequality in (4.5), define the process
Z t p−1
Nt = hM is 2 dMs .
0
The reason for considering this process is that its quadratic variation is defined by
Z t
1
hN it = hM ip−1
s dhM is = hM ipt .
0 p
Therefore, owing to the fact that (Nt − hN it )t∈[0,∞) is a martingale vanishing at zero,
1
E[hM ipt ] = E[hN it ] = E[Nt2 ].
p
p−1
However, we observe using the integration-by-parts formula that, since (hM it 2 )t∈[0,∞) is a process
of bounded variation owing to the facts that p ∈ (1, ∞) and that (hM it )t∈[0,∞) is a bounded process
of bounded variation,
p−1
Z t p−1
Z t p−1

Mt hM it =
2
hM is dMs +
2
Ms d hM is 2
0 0
Z t p−1

= Nt + Ms d hM is 2
.
0
We therefore conclude that
p−1
Z t p−1
p−1
|Nt | ≤ |Mt | hM it2
+ |Ms | d hM is 2
≤ 2Mt∗ hM it 2 .
0
Hence, again applying Hölder’s inequality with exponents p and p/p−1,
1 1 p−1
E[hM ipt ] = E[Nt2 ] ≤ 4E[(Mt∗ )2 hM ip−1
t ] ≤ 4E[(Mt∗ )2p ] p E[hM ipt ] p .
p
We conclude that
1 1
E[hM ipt ] p ≤ 4pE[(Mt∗ )2p ] p ,
and therefore that

E[hM ipt ] ≤ (4p)p E[(Mt∗ )2p ].
This completes the proof of the rightmost inequality of (4.5) with
Cp = (4p)p ,
and therefore the proof of (4.5) in the case p ∈ (1, ∞). We have by now completed the proof for
every p ∈ [1, ∞). It remains to consider the case p ∈ (0, 1).
Let p ∈ (0, 1). For every α ∈ (0, 1) let (Ntα )t∈[0,∞) be defined by
Z t
p−1
α
Nt = (hM is + α) 2 dMs .
0
The quadratic variation process (hN α i t )t∈[0,∞) satisfies
Z t
1
(4.6) hN α it = (hM is + α)p−1 dhM is = (hM it + α)p ,
0 p
where the final equality uses the fact that d (hM it + α) = dhM it . Furthermore, since it follows by
definition that
p−1
dNtα = (hM it + α) 2 dMt ,
it follows that
1−p
dMt = (hM it + α) 2 dNtα .
The integration-by-parts formula proves that, since the process
1−p

(hM it + α) 2
t∈[0,∞)
is a process of bounded variation,

1−p
Z t t 1−p 1−p
Z

Ntα (hM it + α) 2 = (hM is + α) + 2 dNsα
Nsα d (hM is + α) 2
0 0
Z t
1−p

= Mt + Nsα d (hM is + α) 2 .
0
Therefore,
1−p
Z t 1−p
1−p
(4.7) |Mt | ≤ |Ntα | (hM it + α) 2 + |Ntα | d (hM is + α) 2 ≤ 2(N α )∗t (hM it + α) 2 .
0
Let t ∈ [0, ∞). It follows from (4.7) that, for every s ∈ [0, t],
1−p 1−p
|Ms | ≤ 2(N α )∗s (hM is + α) 2 ≤ 2(N α )∗t (hM it + α) 2 ,
from which it follows that
1−p
|Mt∗ | ≤ 2(N α )∗t (hM it + α) 2 .
It then follows from Hölder’s inequality with exponents 1/p and 1/1−p that
h i p
E[(Mt∗ )2p ] ≤ 22p E (Ntα,∗ )2p (hM it + α)p(1−p) ≤ 22p E (Ntα,∗ )2 E [(hM it + α)p ]1−p .

(4.8)
It then follows from Doob’s inequality and (4.6) that
4
E (Ntα,∗ )2 ≤ 4E[(Ntα )2 ] = 4E[hN α it ] = E [(hM it + α)p ] .

p
Returning to (4.8), we have that
22(p+1)
E[(Mt∗ )2p ] ≤ E [(hM it + α)p ] .
p
64 BENJAMIN FEHRMAN
The monotone convergence theorem then proves that, after passing to the limit α → 0,
22(p+1)
E[(Mt∗ )2p ] ≤ E [hM ipt ] ,
p
which completes the leftmost inequality of (4.5) in the case p ∈ (0, 1) with
cp = p2−2(p+1) .
It remains to prove the rightmost inequality of (4.5) in the case p ∈ (0, 1). Let α ∈ (0, ∞). Since
we have that p
hM ipt = hM it (Mt∗ + α)2(p−1) (Mt∗ + α)2p(1−p) ,
it follows from Hölder’s inequality with exponents 1/p and 1/1−p that
h ip h i1−p
E[hM ipt ] ≤ E hM it (Mt∗ + α)2(p−1) E (Mt∗ + α)2p .
Let (Ntα )t∈[0,∞) be the process
Z t
Ntα = (Ms∗ + α)p−1 dMs ,
0
for which
Z t
(4.9) hN α it = (Ms∗ + α)2(p−1) dhM is ≥ (Mt∗ + α)2(p−1) hM it .
0
The integration-by-parts formula proves that, since as a nondecreasing process (Mt∗ )t∈[0,∞) is a
bounded process of bounded variation,
Z t Z t
∗ ∗
p−1 p−1
Ms d (Ms∗ + α)p−1

Mt (Mt + α) = (Ms + α) dMs +
0 0
Z t
= Ntα + (p − 1) Ms (Ms∗ + α)p−2 dMs∗ .
0
We therefore conclude that, since p ∈ (0, 1),
Z t
|Ntα | ≤ |Mt | (Mt∗ p−1
+ (1 − p)
+ α) |Ms | (Ms∗ + α)p−2 dMs∗
0
Z t
∗ ∗
(4.10) ≤ Mt (Mt + α) p−1
+ (1 − p) (Ms∗ + α)p−1 dMs∗ .
0
1
≤ Mt∗ (Mt∗ + α)p−1 + − 1 (Mt∗ + α)p .
p
Since it follows from (4.9) and Hölder’s inequality with exponents 1/p and 1/1−p that
E[hM ipt ] ≤ E[hN α ipt (Mt∗ + α)2p(1−p) ]
≤ E[hN α it ]p E[(Mt∗ + α)2p ]1−p
= E[(Ntα )2 ]p E[(Mt∗ + α)2p ]1−p ,
and since it follows from (4.10) that
" 2 #
1
E (Nt ) ≤ E (Mt∗ )2 (Mt∗ + α)2(p−1) + − 1 (Mt∗ + α)2p ,
α 2
p
we have that
" 2 #p
1
E[hM ipt ] ≤ E (Mt∗ )2 (Mt∗ + α)2(p−1) + − 1 (Mt∗ + α)2p E[(Mt∗ + α)p ]1−p .
p
By the monotone convergence theorem, after passing to the limit α → 0, we conclude that
E[hM ipt ] ≤ p−2p E (Mt∗ )2p ,

which completes the proof of the rightmost inequality of (4.5) in the case p ∈ (0, 1) with
Cp = p−2p .
In the following three corollaries, we will show that the Burkholder-Davis-Gundy inequality
applies equally to stopped continuous martingales, continuous local martingales, and to integrals
of bounded predictable processes.
Corollary 4.9. For every p ∈ (0, ∞) there exist constants cp , Cp ∈ (0, ∞) such that, for every
continuous martingale (Mt )t∈[0,∞) vanishing at zero, for every almost surely finite stopping time T ,
cp E[(MT∗ )2p ] ≤ E[hM ipT ] ≤ Cp E[(MT∗ )2p ].
Proof. Let p ∈ (0, ∞), let (Mt )t∈[0,∞) be a continuous martingale vanishing at zero, and let T be an
almost surely finite stopping time. Then, since the stopped process (MtT )t∈[0,∞) is a martingale, it
follows from Theorem 4.8 that there exist universal constants cp , Cp ∈ (0, ∞) such that, for every
t ∈ [0, ∞),
∗
cp E[(Mt∧T )2p ] ≤ E[hM ipt∧T ] ≤ Cp E[(Mt∧T
∗
)2p ].
Since the stopping time is almost surely finite, the monotone convergence theorem proves that,
after passing to the limit t → ∞,
cp E[(MT∗ )2p ] ≤ E[hM ipT ] ≤ Cp E[(MT∗ )2p ],
continuous local martingale (Mt )t∈[0,∞) vanishing at zero, for every almost surely finite stopping
time T ,
cp E[(MT∗ )2p ] ≤ E[hM ipT ] ≤ Cp E[(MT∗ )2p ].
Proof. Let p ∈ (0, ∞), let (Mt )t∈[0,∞) be a continuous local martingale vanishing at zero, and let
T be an almost surely finite stopping time. Then there exist stopping times {τn }n→∞ such that
τn → ∞ almost surely as n → ∞, and such that
(Mtτn )t∈[0,∞) = (Mt∧τn 1{τn >0} )t∈[0,∞) is a martingale.
Therefore, by the previous Corollary, there exist universal constants cp , Cp ∈ (0, ∞) such that
cp E[(Mτ∗n ∧T )2p 1{τn >0} ] ≤ E[hM ipτn ∧T 1{τn >0} ] ≤ Cp E[(Mτ∗n ∧T )2p 1{τn >0} ].
Since the stopping time is almost surely finite, the monotone convergence theorem proves that,
after passing to the limit n → ∞,
cp E[(MT∗ )2p ] ≤ E[hM ipT ] ≤ Cp E[(MT∗ )2p ],
continuous local martingale (Mt )t∈[0,∞) vanishing at zero, for every bounded predictable process
(Ht )t∈[0,∞) , for every almost surely finite stopping time T ,
" Z t 2 !p # "Z
T p # " Z t 2 !p #
cp E sup Hs dMs ≤E Hs2 dhM is ≤ Cp E sup Hs dMs .
t∈[0,T ] 0 0 t∈[0,T ] 0
66 BENJAMIN FEHRMAN
Proof. The statement is an immediate consequence of the previous corollary, since the boundedness
of the process (Ht )t∈[0,∞) guarantees that
Z t
Hs dMs is a local martingale
0 t∈[0,∞)
with quadratic variation process
Z · Z t
h Hs dMs it = Hs2 dhM is .
0 t∈[0,∞) 0 t∈[0,∞)

4.5. The martingale representation theorem. Let (Ω, F, P) be a probabilty space, and let
X : Ω → R be a random variable. Then, it follows that for every random variable Y : Ω → R that
is measurable with respect to σ(X), there exists a measurable function fY : R → R such that
(4.11) Y = fY (X).
This fact emphasizes the view that σ(X) carries all of the information of X, in the sense that every
random variable measurable with respect to σ(X) is actually a function of X.
The martingale representation theorem is the analogue of this fact for Brownian motion. Suppose
that (Bt )t∈[0,∞) is a standard Brownian motion on a probability space (Ω, F, P) with respect to a
filtration (Ft )t∈[0,∞) . We define the sigma algebra
F∞ = σ(∪s∈[0,∞) Fs ),
which carries the information of the entire Brownian path. We therefore expect that every random
variable M ∈ L2 (Ω, F∞ , P), that is every L2 -random variable that is measurable with respect to
F∞ , is a “function of Brownian motion.” We will see that M = I(Bt ) is a functional of the Brownian
paths, and that the functional is a stochastic integral. That is, for every M ∈ L2 (Ω, F∞ , P) there
exists a unique process (Hs )s∈[0,∞) ∈ L2 (B) such that
Z ∞
(4.12) M = E[M ] + Ms dBs .
0
In particular, if we define the L2 -bounded martingale (Mt )t∈[0,∞) by

Z t
Mt = E[M |Ft ] = E[M ] + Hs dBs ,
0
it follows that (Mt )t∈[0,∞) is a solution to the stochastic differential equation
(
dMt = Ht dBt in (0, ∞),
M0 = E[M ]
Much like the proof of (4.11), the proof of (4.12) is done by proving the density of certain simple
functions, which in this case are provided by the stochastic exponential.
Remark 4.12. Some care if necessary when applying the above intuition. It is not true, in general,
that Brownian motion can be replaced by an arbitrary martingale (Xt )t∈[0,∞) . In particular, the
proof presented below uses the fact that Brownian motion has a deterministic quadratic variation
process hBit = t.
Define the set of simple functions
n−1
( )
X
S= f= λi 1(ti ,ti+1 ] : λi ∈ R and 0 = t0 < t1 < . . . < tn < ∞ .
i=0
Then, for every f ∈ S, let (Mtf )t∈[0,∞) denote the martingale

Z t n−1
Mtf
X
= fs dBs = λi (Bti+1 ∧t − Bti ∧t ),
0 i=0
and the stochastic exponential (E(M f )t )t∈[0,∞) defined by

n−1 n−1
!
X X λ2i ((ti+1 ∧ t) − (ti ∧ t))
E(M f )t = exp λi (Bti+1 ∧t − Bti ∧t ) − .
2
i=0 i=0
68 BENJAMIN FEHRMAN
solves
dE(M f )t = E(M f )t dMtf = E(M f )t ft dBt in (0, ∞),
(
E(M f )0 = 1.
We will write E(M f ) = E(M f )∞ , and observe that this is to say
Z ∞
E(M f ) = 1 + E(M f )t ft dBt ,
0
which takes the form (4.12) for Ht = E(M f )t ft .

In order to prove (4.12), it will therefore suffice to
prove that the linear span of {E(M f )}f ∈S is dense in L2 (Ω, F, P).
Proposition 4.13. The linear span of {E(M f )∞ }f ∈S is dense in L2 (Ω, F∞ , P).
Proof. Let Y ∈ L2 (Ω, F∞ , P). It suffices to prove that if
E[Y E(M f )] = 0 for every f ∈ S if and only if Y = 0.
It is clear that the expectation vanishes if Y = 0. To prove the converse, suppose that Y ∈
L2 (Ω, F∞ , P) satisfies that, for every f ∈ S,
E[Y E(M f )] = 0.
This implies that for every λ0 , . . . , λn−1 ∈ R and 0 = t0 < t1 < . . . < tn < ∞,
n−1 n−1
" ! #
X X λ2 (ti+1 − ti )
i
E exp λi (Bti+1 − Bti ) − Y = 0,
2
i=0 i=0
and therefore that

n−1
" ! #
X
(4.13) E exp λi (Bti+1 − Bti ) Y = 0.
i=0
Fix a sequence 0 = t0 < t1 < . . . < tn < ∞. Since Y ∈ L2 (Ω, F∞ , P), it follows that the function
n−1
" ! #
X
(4.14) (z0 , . . . , zn−1 ) ∈ Cn 7→ E exp zi (Bti+1 − Bti ) Y is holomorphic.
i=0
Since the holomorphic function (4.14) vanishes on the connected open set Rn ⊆ Cn by (4.13), the
function (4.14) is identically zero. In particular, for every λ0 , . . . , λn−1 ∈ R,
n−1
" ! #
X
(4.15) E exp i λi (Bti+1 − Bti ) Y = 0.
i=0
Define the measure dQ = Y dP and define the measurable map F : Ω → Rn by

F (ω) = Bt1 (ω) − Bt0 (ω), Bt2 (ω) − Bt1 (ω), . . . , Btn (ω) − Btn−1 (ω)

= Bt1 (ω), Bt2 (ω) − Bt1 (ω), . . . , Btn (ω) − Btn−1 (ω) .
Let F∗ Q be the pushforward measure on Rn defined for every Borel measurable set A ⊆ Rn by
F∗ Q(A) = Q(F −1 (A)) = Q Bt1 , Bt2 − Bt1 , . . . , Btn − Btn−1 ∈ A .

(4.16)
It follows by definition and (4.15) that, for every λ = (λ0 , . . . , λn−1 ) ∈ Rn ,
n−1
" ! # Z
X
E exp i λi (Bti+1 − Bti ) Y = exp(ihλ, xi)F∗ Q( dx) = 0.
i=0 Rn
We conclude that the Fourier transform of F∗ Q vanishes, and therefore that F∗ Q is the zero measure
on Rn . Returning to (4.16), it follows from the definition of Q that, for every set B ∈ σ(Bt1 , Bt2 −
Bt1 , . . . , Btn − Btn−1 ) = σ(Bt1 , Bt2 , . . . , Btn ),
E[Y 1B ] = 0.
Since the collection
{B ∈ F∞ : B ∈ σ(Bt1 , Bt2 , . . . , Btn ) for some 0 = t0 < t1 < . . . < tn < ∞}
is a π-system on which the measure dQ = Y dP vanishes, and since
F∞ = σ ({B ∈ F∞ : B ∈ σ(Bt1 , Bt2 , . . . , Btn ) for some 0 = t0 < t1 < . . . < tn < ∞}) ,
we conclude that dQ = Y dP is the zero measure on F∞ . This implies that Y = 0, since Y is
F∞ -measurable, which completes the proof.
Theorem 4.14. Let M ∈ L2 (Ω, F∞ , P). Then there exists a unique process (Hs )s∈[0,∞) ∈ L2 (B)
such that Z ∞
M = E[M ] + Ht dBt .
0
Proof. We observe that if (Hs )s∈[0,∞) ∈ L2 (B) is a predictable process, and if X is the random
variable defined for some c ∈ R by
Z ∞
X =c+ Ht dBt ,
0
then it follows that Z ∞
c = E[X] since E Ht dBt = 0.
0
We therefore consider the linear subspace
Z ∞
I = X = E[X] + Ht dBt : (Hs )s∈[0,∞) ∈ L (B) ⊆ L2 (Ω, F∞ , P),
2
0
where the rightmost inclusion follows from the fact that (Hs )s∈[0,∞) ∈ L2 (B) and from the definition
of F∞ . The fact that I is a linear subspace follows from the linearity of the expectation, the linearity
of the stochastic integral, and the fact that L2 (B) is a vector space.
We aim to prove that I = L2 (Ω, F∞ , P). For this, it is sufficient to prove (i) that I is closed
and (ii) that I contains a dense subset. For (i), suppose that {Xn }n∈N is a Cauchy sequence in I
corresponding to a sequence {(Hsn )s∈[0,∞) }n∈N in L2 (B). It follows from the Itô isometry, Young’s
inequality, and Hölder’s inequality that, for every n, m ∈ N,
Z ∞ "Z 2 #
∞
E |Hsn − Hsm |2 ds = E (Hsn − Hsm ) dBs
0 0
h i
≤ 2 E |Xn − Xm |2 + |E[Xn ] − E[Xm ]|2 .
h i
≤ 4E |Xn − Xm |2 .
We therefore conclude that {(Hsn )}s∈[0,∞) is a Cauchy sequence in L2 (B).

It follows that there exists X∞ ∈ L2 (Ω, F∞ , P) and (Hs∞ )s∈[0,∞) such that, as n → ∞,
Xn → X∞ strongly in L2 (Ω, F∞ , P),
such that, as n → ∞,
E[Xn ] → E[X∞ ],
70 BENJAMIN FEHRMAN
and such that, as n → ∞,

Z ∞ Z ∞
Hsn dBs → Hs∞ dBs strongly in L2 (Ω, F∞ , P),
0 0
where this final convergence is again a consequence of the Itô isometry. After passing to the limit
n → ∞ in the equation Z ∞
Xn = E[Xn ] + Hsn dBs ,
0
it follows that Z ∞
X∞ = E[X∞ ] + Hs∞ dBs in L2 (Ω, F∞ , P).
0
Therefore, we have X∞ ∈ I and we conclude that I is closed.
To prove (ii), since for every f ∈ S properties of the stochastic exponential prove that
Z ∞ Z ∞
f f f
E(M ) = 1 + E(M )s dBs = E[E(M )] + E(M f )s dBs ,
0 0
we conclude that {E(M f )}f ∈S ⊆ I. Proposition 4.13 then proves that I contains a dense subset.
This completes the proof of (ii), and therefore the proof that I = L2 (Ω, F∞ , P). That is, for every
M ∈ L2 (Ω, F∞ , P), there exists (Hs )s∈[0,∞) ∈ L2 (B) such that
Z ∞
M = E[M ] + Hs dBs .
0
This completes the proof. (The proof of uniqueness will be added after the third problem session.)

5. Local Times
In this section, we will develop the notion of the local time associated to a one-dimensional
semimartingale (Xt )t∈[0,∞) . Roughly speaking, for a ∈ R, the local time (Lat )t∈[0,∞) or sometimes
written (Lat (X))t∈[0,∞) measures the time (Xt )t∈[0,∞) spends at the point a up to time t. However,
time in this case is measured with respect to the quadratic variation process (hXit )t∈[0,∞) . As
motivated by the Dambis-Dubins-Schwarz theorem, the quadratic variation process acts as an
internal clock for the semimartingale (Xt )t∈[0,∞) .
Let A ⊆ Rd . The random time tA the process (Xt )t∈[0,∞) spends occupying A up to time t is
defined by the integral Z t
tA = 1A (Xs ) dhXis .
0
In particular, tR = hXit is simply the total time elapsed as measured by the quadratic variation,
since the process resides always in the whole space R. However, we expect that this occupation
time can be similarly expressed in terms of local times. That is,
Z
tA = Lat da,
A
or, more generally, for any nonnegative Borel measurable function Φ : R → R,
Z t Z
Φ(Xs ) dhXis = Φ(a)Lat da.
0 R
This is the occupation formula, which we prove below. The construction of the local time will rely
on properties of convex functions, and an important extension of Itô’s formula from C2 -functions
to functions that can be written as the differenceof two convex functions.
72 BENJAMIN FEHRMAN
5.1. Convex functions. In this section, we will prove that a convex function f : R → R is always
twice-differnetiable, at least in a distributional sense.
Definition 5.1. A function f : R → R is convex if, for every a, b ∈ R and every λ ∈ [0, 1],
f (λa + (1 − λ)b) ≤ λf (a) + (1 − λ)f (b).
We observe that the following proposition implies, in particular, that finite convex functions are
always locally bounded and continuous.
Proposition 5.2. Let f : R → R be convex. Then the left derivative f−0 : R → R defined by
f (x) − f (x − h)
f−0 (x) = lim exists and is nondecreasing and left continuous.
h→0 h
Similarly, the right derivative f+0 (x) : R → R defined by
f (x + h) − f (x)
f+0 (x) = lim exists and is nondecreasing and right continuous.
h→0 h
Furthermore, for all φ ∈ C∞c (R),
Z Z Z
f (x)φ (x) dx = − f− (x)φ(x) dx = − f+0 (x)φ(x) dx.
0 0
R R R
Proof. We will prove the statement for the left derivative, since the case of the right derivative is
virtually identical. Let h1 < h2 ∈ (0, ∞). Then, by convexity,

h1 h1
f (x − h1 ) ≤ f (x − h2 ) + 1 − f (x).
h2 h2
Hence, after reordering this inequality,
f (x) − f (x − h2 ) f (x) − f (x − h1 )
≤ ,
h2 h1
which implies that, as h → 0+ ,
f (x) − f (x − h)
is nondecreasing.
h
Furthermore, for every h ∈ (0, ∞), it follows by convexity that
1 h
f (x) ≤ f (x − h) + f (x + 1).
1+h 1+h
Hence, after reordering this inequality,
f (x) − f (x − h)
≤ f (x + 1) − f (x),
h
which implies that
f (x) − f (x − h)
is bounded from above.
h h∈(0,∞)
We therefore conclude that, as an increasing sequence that is bounded from above,
f (x) − f (x − h)
f−0 (x) = lim exists.
h→0 h
A repetition of the above arguments prove that, for every x < y ∈ R and h ∈ (0, |x − y|),
f (x) − f (x − h) f (y − h) − f (x) f (y) − f (y − h)
≤ ≤ ,
h y−x−h h
which proves after passing to the limit h → 0 that f−0 (x) ≤ f (y)−f (x)/y−x ≤ f−0 (y). This completes
the proof that f−0 is nondecreasing. Finally, since the sequence is decreasing in both h, h0 ∈ (0, ∞),
f (x − h) − f (x − h − h0 )

lim f−0 (x − h) = lim lim

h→0 h→0 h0 →0 h0
f (x − h) − f (x − h − h0 )

= lim lim
h0 →0 h→0 h0
f (x) − f (x − h0 )
= lim
h0 →0 h0
0
= f− (x),
which completes the proof of left continuity.
Finally, suppose that φ ∈ C∞ 0
c (R). Then, since f and f− are locally bounded, the dominated
convergence theorem and a change of variables prove that

φ(x + h) − φ(x)
Z Z
0
f (x)φ (x) dx = lim f (x) dx
R h→0 R h
Z
f (x) − f (x − h)
= − lim φ(x) dx
h→0 R h
Z
= − f−0 (x)φ(x) dx,
R
In the final proposition of this section, we prove that the second derivative of a convex function
exists in a distributional sense. Precisely, the second derivative is the Riemann-Stietjes measure
associated to f−0 . We emphasize that the choice to work with the left derivative as opposed to the
right derivative is not essential, but it is necessary to choose one or the other.
Proposition 5.3. Let f : R → R be a convex function. Let µf denote the nonnegative Riemann-
Stieljes measure associated to the nondecreasing left derivative f−0 defined for every a ≤ b ∈ R
by
µf [(a, b)] = f−0 (b) − f−0 (a+) = f−0 (b) − lim f−0 (a + h) .
h→0
Then, for all φ ∈ C∞
c (R),
Z Z Z
f (x)φ00 (x) dx = − f−0 (x)φ0 (x) dx = φ00 (x)µf ( dx).
R R R
Proof. The first equality was proven above. It remains only to prove that
Z Z
0 0
− f− (x)φ (x) dx = φ00 (x)µf ( dx),
R R
which, owing to the fact that φ ∈ C∞
c (R), is the integration-by-parts formula for Riemann-Stieltjes
integrals.
74 BENJAMIN FEHRMAN
5.2. The Meyer-Tanaka formula. We will now define the local time of a continuous semimartin-
gale (Xt )t∈[0,∞) , which will lead to an important generalization of Itô’s formula. In particular, the
next proposition proves that a convex function of a semimartingale is again a semimartingale, much
like Itô’s formula proved that a C2 -function of a semimartingale is again a semimartingale.
Proposition 5.4. Let (Xt )t∈[0,∞) be a continuous semimartingale, and let f : R → R be convex.
Then, there exists a nondecreasing process (Aft )t∈[0,∞) such that
Z t
f (Xt ) = f (X0 ) + f−0 (Xs ) dXs + Aft .
0
Proof. By localization using stopping times, we can assume without loss of generality that (Xt =
Mt + At )t∈[0,∞) , for a local martingale (Mt )t∈[0,∞) and a process of bounded variation (At )t∈[0,∞) ,
is bounded in the sense that there exists N ∈ N such that, almost surely,
Z t
sup |Xt | + | dAt | + hM it < N.
t∈[0,∞) 0
Let ρ ∈ C∞
c (R) be a smooth function supported on the set (−∞, 0] ⊆ R. For every ε ∈ (0, 1] define
ρε : R → R by
x
ρε (x) = ε−1 ρ .
ε
For every ε ∈ (0, 1] define f ε : R → R to be the convolution
Z Z
ε ε ε
f (x) = (f ∗ ρ )(x) = f (y)ρ (y − x) dy = f (y + x)ρε (y) dy.
R R
Since the convexity of f implies that f is locally bounded and continuous, the functions {f ε }ε∈(0,1]
are smooth. Therefore, Itô’s formula proves that, for every t ∈ [0, ∞),
Z t
1 t ε 00
Z
ε ε ε 0
f (Xt ) = f (X0 ) + (f ) (Xs ) dXs + (f ) (Xs ) dhXis .
0 2 0
For the left derivative f−0 of f and for the Riemann-Stieljes measure µf of f−0 , we have that
Z
ε 0
(f ) (x) = f−0 (x + y)ρε (y) dy,
R
and Z
(f ε )00 (x) = ρε (y − x)µf ( dx).
R
In particular, we observe that the convolution preserves the convexity in the sense that the noneg-
ativity of the measure µf implies that, for every x ∈ R,
Z
ε 00
(f ) (x) = ρε (y − x)µf ( dx) ≥ 0,
R
fε
and therefore that the process (At )t∈[0,∞) defined by
1 t ε 00
Z
fε
At = (f ) (Xs ) dhXis is nondecreasing.
2 0
Since f is continuous and hence uniformly continuous on compact sets, it follows using properties
of the convolution that !
lim sup |f ε (x) − f (x)| = 0.
ε→0 {|x|≤N }
Similarly, the triangle inequality, the Itô isometry prove, the construction of the convolution kernel,
the left continuity of f−0 , the local boundedness of f−0 , the dominated convergence theorem, and
Doob’s inequality prove that, for every t ∈ [0, ∞),
2 # 12
 " 
Z s
lim E sup (f ε )0 (Xr ) − f−0 (Xr ) dXr 
ε→0 s∈[0,t] 0
2 # 21
 
Z t 12 "Z
t
2 2
(f ε )0 (Xs ) − f−0 (Xs ) (f ε )0 (Xs ) − f−0 (Xs ) | dAs |

= lim 4E dhXis +E 
ε→0 0 0
= 0.
After passing to a subsequence {εk }k∈N , we conclude almost surely that, for every t ∈ [0, ∞),
Z t Z t !
sup (f εk )0 (Xs ) dXs − f−0 (Xs ) dXs

lim = 0.
k→∞ t∈[0,∞) 0 0
Since by definition, for every k ∈ N ,

εk
Z t
Aft = ftεk (Xt ) − f εk (X0 ) − (f εk )0 (Xs ) dXs ,
0
we conclude that, as a limit of nondecreasing processes the process (Aft )t∈[0,∞) defined by
εk
Aft = lim Aft exists almost surely,
k→∞
and almost surely satisfies, for every t ∈ [0, ∞),

Z t
f (Xt ) = f (X0 ) + f−0 (Xs ) dXs + Aft ,
0
Roughly speaking, for a ∈ R we expect the local time (Lat )t∈[0,∞) of a continuous semimartingale
(Xt )t∈[0,∞) to be defined by
Z t
a
Lt = δ0 (Xs − a) dhXis ,
0
where δ0 denotes the Dirac distribution at zero. Since for a convex f the process (Aft )t∈[0,∞)
constructed above is in spirit given by
1 t 00
Z
Aft = f (Xs ) dhXis ,
2 0
we expect that the local time will be defined by the process (Aft )t∈[0,∞) for f (x) = |x − a|, since in
this case
f−0 (x) = 1{x>a} − 1{x≤a} and f 00 (x) = δ0 (x − a) as distributions.
This is the content of the next proposition. We write sgn for the left continuous version of the sign
function
sgn(x) = 1{x>0} − 1{x≤0} .
Again, this choice is made because we work with the left derivative as opposed to the right derivative.
The following is the Tanaka formula.
76 BENJAMIN FEHRMAN
Proposition 5.5. Let (Xt )t∈[0,∞) be a continuous semimartingale. For every a ∈ R there exists an
increasing process (Lat )t∈[0,∞) called the local time of (Xt )t∈[0,∞) in a such that, for every t ∈ [0, ∞),
Z t
1
(Xt − a)+ = (X0 − a)+ + 1{Xs >a} dXs + Lat ,
0 2
Z t
1
(Xt − a)− = (X0 − a)− + 1{Xs ≤a} dXs + Lat ,
0 2
Z t
|Xt − a| = |X0 − a| + sgn(Xs − a) dXs + Lat .
0
Proof. By convexity of the functions f (x)+ = (x − a)+ and f (x)− = (x − a)− , there exist increasing
−
processes (A+
t )t∈[0,∞) and (At )t∈[0,∞) such that
Z t
(Xt − a)+ = (X0 − a)+ + 1{Xs >a} dXs + A+
t ,
0
Z t
(Xt − a)− = (X0 − a)− + 1{Xs ≤a} dXs + A−
t .
0
Subtracting these equalities proves that
Z t
−
Xt = X0 + dXs + (A+
t − At ),
0
and therefore by continuity that, almost surely,
−
A+
t = At for every t ∈ [0, ∞).
We therefore define (Lat )t∈[0,∞) = (2A+
t )t∈[0,∞) , and conclude by subtracting the previous equalities
that Z t
|Xt − a| = |X0 − a| + sgn(Xs ) dXs + Lat ,
0
Proposition 5.6. Let (Xt )t∈[0,∞) be a continuous semimartingale and let (Lat )a∈R,t∈[0,∞) denote
the local times of (Xt )t∈[0,∞) . There exists a B(R × [0, ∞)) ⊗ F-measurable process
(a, t, ω) → L̃(a, t, ω)
such that, for every a ∈ R, the processes (t, ω) 7→ L̃(a, t, ω) and (t, ω) 7→ Lat (ω) are indistinguishable.
Proof. The proof is an immediate consequence of Fubini’s theorem for stochsatic integrals. The
proof will be added after the fourth problem session.
Remark 5.7. Given a continuous semimartingale (Xt )t∈[0,∞) we by indistinguishability we can
and will henceforth assume without loss of generality that the process (a, t, ω) 7→ Lat (ω) is B(R ×
[0, ∞)) ⊗ F-measurable.
Again, the intuition is that (Lat )t∈[0,∞) measures the time (Xt )t∈[0,∞) spends at a. The fol-
lowing proposition proves that the random Riemann-Stieltjes measure associated to (Lat )t∈[0,∞) is
supported on the random set {Xt = a}. However, it is clear that, in general, the support of the
measure will be smaller than the set {Xt = a}. This can be seen, for instance, by considering the
constant process Xt = 0. In this case, for a = 0, we have that {Xt = 0} = [0, ∞). But, the local
time L0t = 0 vanishes identically as well, due to the fact “no time elapses” because the quadratic
variation hXit = 0 vanishes identically.
Proposition 5.8. Let (Xt )t∈[0,∞) be a continuous local martingale, let a ∈ R, and let (Lat )t∈[0,∞)
be the local time of (Xt )t∈[0,∞) in a. Then the random Riemann-Stieltjes measure dLat on [0, ∞) is
almost surely supported on the random set {Xt = a}.
Proof. The proof is based on the equality (Xt − a)2 = |Xt − a| · |Xt − a|. On the lefthand side, Itô’s
formula proves that
Z 2 Z t
2 2
(Xt − a) = (X0 − a) + 2 (Xs − a) dXs + dhXis
0 0
Z t
= (X0 − a) + 2 (Xs − a)2 dXs + hXit .
2
0
On the righthand side, the integration-by-parts formula and Tanaka’s formula prove that
Z t
(Xt − a)2 = (X0 − a)2 + 2 |Xs − a| d (|Xs − a|) + h|Xt − a|it ,
0
Z t Z t Z t
2 a
= (X0 − a) + 2 sgn(Xs − a) |Xs − a| dXs + |Xs − a| dLs + dhXis
0 0 0
Z t Z t
2
= (X0 − a) + 2 (Xs − a) dXs + |Xs − a| dLas + hXit .
0 0
Therefore, we conclude almost surely that, for every t ∈ [0, ∞),
Z t
|Xs − a| dLas = 0,
0
We are now prepared to present the Meyer-Tanaka formula, which generalizes Itô’s formula to
functions f that are the difference of two convex functions. We emphasize that the above reasoning
can equally be applied to a concave function, which is nothing more than the negative of a convex
function.
Theorem 5.9. Let (Xt )t∈[0,∞) be a continuous semimartingale, let (Lat )a∈R,t∈[0,∞) be the local
times of (Xt )t∈[0,∞) , let f : R → R be the difference of two convex functions, and let µf denote the
Riemann-Stieljes integral of f−0 . Then, for every t ∈ [0, ∞),
Z t
1 t a
Z
f (Xt ) = f (X0 ) + f−0 (Xs ) dXs + L µf ( da).
0 2 0 t
Proof. By linearity, it suffices to prove the theorem in the case that f is convex. By a stopping
time argument, we can assume without loss of generality that, almost surely, for some N ∈ N,
sup (|Xt | + hXit ) < N.
t∈[0,∞)
Define the convex function f N : R → R by

 0
 f (N ) + f− (−N )(x + N ) if x ∈ (−∞, −N ],

f N (x) = f (x) if x ∈ [−N, N ],
f (N ) + f−0 (N )(x − N )

if x ∈ [N, ∞).

The convexity of f proves that f N is convex, and by definition µf N has compact support and
satisfies
f N = f, (f N )0− = fN
0
, and µf N = µf on [−N, N ].
Since µf N as compact support, define the function
Z
g(x) = |x − a| µf N ( da),
R
78 BENJAMIN FEHRMAN
and observe by a direct computation using the compact support of µf N that

Z Z
0

g− (x) = 1{x>a} − 1{x≤a} µf N ( da) = sgn(x − a)µf N ( da).
R R
Then, by Fubini’s theorem, for every φ ∈ C∞c (R),
Z Z Z Z Z
0
g− (x)φ0 (x) dx = sgn(x − a)φ0 (x)µf N ( da) dx = sgn(x − a)φ0 (x) dxµf N ( da).
R R R R R
Since as distributions sgn0 (x) = 2δ0 (x), it follows that
Z Z Z
0
g− (x)φ0 (x) dx = −2 φ(a)µf N ( da) = −2 φ(x)µf N ( dx).
R R R
That is, as a distribution,
1 00

N
f − g = 0 on R.
2
This implies that there exists a, b ∈ R such that
Z
N 1 1
f (x) = a + bx + g(x) = a + bx + |x − a| µf N ( da).
2 2 R
Therefore, by Tanaka’s formula,
Z
N 1
f (Xt ) = a + bXt + |Xt − a| µf N ( da)
2 R
Z Z t
1 a
= a + bXt + |X0 − a| + sgn(Xs − a) dXs + Lt µf N ( da).
2 R 0
Since Z
1
(f N )0− (x) = b + g 0 (x) = b + sgn(x − a)µf N ( da),
2 R
and since Z
N 1
f (0) = a + bX0 + |X0 − a| µf N ( da),
2 R
it follows by Fubini’s theorem that
Z Z t Z
1 1
f N (Xt ) = f (X0 ) + b(Xt − X0 ) + sgn(Xs − a) dXs µf N ( da) + La µ N ( da)
2 R 0 2 R t f
Z t Z Z
N 1 1
= f (X0 ) + b(Xt − X0 ) + sgn(Xs − a)µf N ( da) dXs + Lat µf N ( da)
0 2 R 2 R
Z t Z Z
1 1
= f N (X0 ) + b+ sgn(Xs − a)µf N ( da) dXs + La µ N ( da)
0 2 R 2 R t f
Z t Z
N N 0 1
= f (X0 ) + (f ) (Xs )Xs + Lat µf N ( da).
0 2 R
Therefore, since almost surely we have
sup (|Xt | + hXit ) < N,
t∈[0,∞)
and since the fact that the support of dLat is almost surely contained in the set {Xt = a} implies
that Lat = 0 for every |a| ≥ N , it follows by the construction of f N that
Z t Z
0 1
f (Xt ) = f (X0 ) + f− (Xs ) dXs + Lat µf ( da),
0 2 R
The equivalence of Itô’s formula and the Meyer-Tanaka formula for twice differentiable convex
functions now yields the occupation formula that motivated our study of local times.
Proposition 5.10. Let (Xt )t∈[0,∞) be a continuous semimartingale and let (Lat )a∈R,t∈[0,∞) be the
local times of (Xt )t∈[0,∞) . Then, for every nonnegative Borel measurable function Φ : R → R,
Z Z
Φ(Xs ) dhXis = Φ(a)Lat da.
R R
Proof. Let f : R → R be a twice continuously differentiable convex function. Then, it follows from
Itô’s formula and the Tanaka-Meyer formula that, on the one hand,
Z t
1 t 00
Z
0
f (Xt ) = f (X0 ) + f (Xs ) dXs + f (Xs ) dhXis ,
0 2 0
and on the other hand that
Z t Z
1
f (Xt ) = f (X0 ) + f 0 (Xs ) dXs + f 00 (a)Lat da,
0 2 R
where we have used the fact that f being twice continuously differentiable implies that f 0 = f−0
and that µf ( da) = f 00 (a) da. This implies that
Z t Z
00
f (Xs ) dhXis = f 00 (a)Lat da.
0 R
The proof now follows by smooth approximation.
80 BENJAMIN FEHRMAN
5.3. Regularity of local times. In this section, we will prove that the local times (Lat )a∈R,t∈[0,∞)
admit a modification that is continuous in time and Càdlàg in a. A process defined on R is called
Càdlàg if it is right continuous with left limits.
Definition 5.11. Let (Xa )a∈R be a stochastic process. We say that (Xa )a∈R is Càdlàg if almost
surely the map a → Xa is right continuous and satisfies
Xa− = lim Xa−h exists.
h→0+
That is, the left limit of (Xa )a∈R exists for every a ∈ R.
Proposition 5.12. Let (Xt )t∈R be a continuous semimartingale. Let (Lat )a∈R,t∈[0,∞) be the local
times of (Xt )t∈[0,∞) Then, the local times (Lat )a∈R,t∈[0,∞) admit a modification (L̂at )a∈R,t∈[0,∞) that
is continuous in time and Càdlàg in a. Furthermore,
Z t Z t
a a−
L̂t − L̂t = 2 1{Xs =a} dVs = 2 1{Xs =a} dXs .
0 0
Proof. Let T ∈ (0, ∞). By construction, for every a ∈ R the local time (Lat )t∈[0,∞) is a continuous
function of time. We will therefore apply the Kolmogorov continuity criterion to the map
a ∈ R 7→ (Lat )t∈[0,T ] ∈ C([0, T ]; R),
where C([0, T ]; R) is the Banach space of continuous functions from [0, T ] to R equipped with the
norm
kf − gkC([0,T ];R) = sup |f (s) − g(s)| .
s∈[0,T ]
Let (Xt = Mt +At )t∈[0,∞) be the semimartingale decomposition of (Xt )t∈[0,∞) for a local martingale
(Mt )t∈[0,∞) and a process of bounded variation (At )t∈[0,∞) . By definition, for every a ∈ R,
Z t Z t
a
Lt = 2 (Xt − a)+ − (X0 − a)+ − 1{Xs >a} dMs − 1{Xs >a} dAs .
0 0
We first observe that
(Xt − a)+ − (X0 − a)+ ,
is a continuous function of t ∈ [0, ∞) and a ∈ R. For every a ∈ R, let (M̃ta )t∈[0,∞) denote the
continuous local martingale
Z t
a
M̃t = 1{Xs >a} dMs .
0
The Burkholder-Davis-Gundy inequality and the occupation formula prove that, for every k ∈
(0, ∞), for every a ≤ b ∈ R,
"
2k
# "Z
T k #
E sup M̃t − M̃tb
a
≤ Ck E 1{a<Xs ≤b} dhM it

t∈[0,T ] 0
"Z
T k #
= Ck E 1{a<Xs ≤b} dhXit
0
"Z
b k #
= Ck E Lx∞ dx
a
" Z b k #
1
= Ck (b − a)k E Lx∞ dx .
b−a a
Jensen’s inequality and Fubini’s theorem then prove that

" # " Z b k #

a
2k
b k 1 x
E sup M̃t − M̃t ≤ Ck (b − a) E L dx
t∈[0,T ] b−a a ∞
" #
2k
a b
≤ E sup M̃t − M̃t
(5.1) t∈[0,T ]
Z b
1
≤ Ck (b − a)k E (Lx∞ )k dx
b−a a
h i
≤ Ck (b − a)k sup E (L∞ x ) k
.
x∈R
If for some k ∈ (1, ∞) we have that

h i
sup E (L∞
x )
k
< ∞,
x∈R
it then follows from (5.1) and the Kolmogorov continuity criterion that the map
a ∈ R → (M̃ta )t∈[0,T ] ∈ C([0, T ]; R),
admits a continuous modification. We will now show that we can always reduce to this case using
a stopping time argument. Precisely, since for every x ∈ R and t ∈ [0, ∞),
Z t
x
Lt = 2 (Xt − a)+ − (X0 − a)+ − 1{Xs >a} dXs ,
0
and since
|(Xt − x)+ − (X0 − x)+ | ≤ |X − X0 | ,
we have for every k ∈ (1, ∞) using the Burkholder-Davis-Gundy inequality that there exists con-
stants c1,k , c2,k ∈ (0, ∞) such that
" Z ∞ k Z ∞ k #
x k k
E(L∞ ) ≤ c1,k E sup |Xt − X0 | + sup 1{Xs >x} dMs + | dAs |
t∈[0,∞) t∈[0,T ] 0 0
(5.2) " k #
k
Z ∞
≤ c2,k E sup |Xt − X0 |k + hM i∞
2
+ | dAs | .
t∈[0,∞) 0
We observe in particular that the righthand side of (5.2) is independent of x ∈ R. Therefore, if we

introduce the stopping times (Tn )n∈N defined by
Z t
Tn = inf{t ∈ [0, ∞) : sup |Xs − X0 | + | dAs | + hM it ≥ n},
s∈[0,t] 0
it follows from (5.1) and (5.2) that the stopped process (M̃tTn )t∈[0,∞) admits a continuous modifi-
cation for every n ∈ N. Since Tn → ∞ as n → ∞ we conclude that (M̃t )t∈[0,∞) admits a continuous
modification.
Let (M̂ta )a∈R,t∈[0,∞) denote the continuous modification of (M̃ta )a∈R,t∈[0,∞) . Let (L̂at )a∈R,t∈[0,∞)
be defined by
Z t
a a
L̂t = 2 (Xt − a)+ − (X0 − a)+ − M̂t − 1{Xs >a} dAs .
0
We observe that (L̂at )a∈R,t∈[0,∞)

is a modification of (Lat )a∈R,t∈[0,∞) .
It remains to prove that the
integral defined by the process of bounded variation is a Càdlàg process. For every a ∈ R and
82 BENJAMIN FEHRMAN
t ∈ [0, ∞), the dominated convergence theorem proves that

Z t Z t
lim 1{Xs >b} dAs = 1{Xs ≥a} dAs .
b→a− 0 0
Similarly, the dominated convergence proves that, for every a ∈ R and t ∈ [0, ∞),
Z t Z t
lim 1{Xs >b} dAs = 1{Xs >a} dAs .
b→a+ 0 0
It then follows from the continuity of (M̂ta )a∈R,t∈[0,∞)
and the continuity of (Xt − a)+ that
Z t
a− b a
(5.3) L̂t = lim L̂t = 2 (Xt − a)+ − (X0 − a)+ − M̂t − 1{Xs ≥a} dAs exists.
b→a− 0
Similarly, we have that
Z t
+
(5.4) L̂at = lim L̂bt a
= 2 (Xt − a)+ − (X0 − a)+ − M̂t − 1{Xs >a} dAs = Lat .
b→a+ 0
Equations (5.3) and (5.4) prove that (L̂at )a∈R,t∈[0,∞)

is a Càdlàg process. Furthermore, we have
that Z t Z t
−
L̂at = L̂at = 2 1{Xs =a} dVs = 2 1{Xs =a} dXs ,
0 0
where the final inequality follows from the fact that the occupation times formula proves that
Z t
2 1{Xs =a} dhM is = 0,
0
and therefore that Z t
2 1{Xs =a} dMs = 0.
0
The following corollary is an immediate consequence of Proposition 5.12.
Corollary 5.13. Let (Xt )t∈R be a continuous local martingale. Let (Lat )a∈R,t∈[0,∞) be the local times
of (Xt )t∈[0,∞) Then, the local times (Lat )a∈R,t∈[0,∞) admit a continuous modification (L̂at )a∈R,t∈[0,∞) .
6. The Girsanov theorem

We have seen in the previous sections that the class of semimartingales is stable with respect to
addition, multiplication, and composition with twice-differentiable and convex functions. In this
section, given a local martingale (Mt )t∈[0,∞) defined on a probability space (Ω, F, P) with respect
to a filtration (Ft )t∈[0,∞( , we will analyze the effect of making an absolutely continuous change of
measure.
Suppose that Q is a mutually absolutely continuous measure with respect to P. Then, for
each t ∈ [0, ∞), the restriction Q|Ft is mutually absolutely continuous with P|Ft . We will write
(Dt )t∈[0,∞) for the corresponding Radon-Nikodym derivatives. That is, for every t ∈ [0, ∞) and
A ∈ Ft , Z
Q(A) = Dt dP.
A
In general, if (Mt )t∈[0,∞) is a local Ft -martingale with respect to P then there is no reason to ex-
pect that (Mt )t∈[0,∞) will be a local Ft -martingale with respect to Q. However, we will see that
(Mt )t∈[0,∞) is a semimartingale with respect to Q, and its corresponding semimartingale decompo-
sition is described by the Girsanov theorem to follow.
Definition 6.1. Let (Ω, F, P) be a probability space and let (Ft )t∈[0,∞) be a filtration on (Ω, F).
A pair of probability measures (Q, P) on F∞ is called a Girsanov pair if Q and P are mutually ab-
solutely continuous and if the Radon-Nikodym derivative (Dt )t∈[0,∞) is a continuous P-martingale.
Remark 6.2. Observe that the Radon-Nikodym derivative (Dt )t∈[0,∞) is a P-martingale by defini-
tion. However, the martingale (Dt )t∈[0,∞) is not in general continuous. Therefore, in Definition 6.1,
the essential assumption is that the process (Dt )t∈[0,∞) is continuous. This assumption is not es-
sential for the results to follow, although it is for us because we have thus far focused entirely on
continuous martingales.
The following theorem is the Girsanov theorem, which describes the effect of making an absolutely
continuous change of measure on pathspace. In particular, it describes how local P-martingales are
transformed into local Q-martingales.
(P, Q) be a Girsanov pair. Then, every continuous P-semimartingale is a continuous Q-semimartingale.
Precisely, if (Mt )t∈[0,∞) is a continuous P-semimartingale then (M̃ )t∈[0,∞) defined by
Z t
M̃t = Mt − Ds−1 dhM, Dis ,
0
is a continuous Q-semimartingale.
Proof. We will first prove that the P-martingale (Dt )t∈[0,∞) is strictly positive Q-almost surely. Let
T denote the stopping time
T = inf{s ∈ [0, ∞) : Ds = 0}.
For every N ∈ N the optional stopping theorem applied to the bounded stopping times N and
T ∧ N proves that, for the expectation E with respect to P,
E[MN ] = E[MT ∧N ] = E[MN : T > N ].
Therefore, using the linearity of the expectation, the definition of Q, and {T ≤ N } ∈ FN ,
E[MN : T ≤ N ] = Q[{T ≤ N }] = 0.
Since {T < ∞} = ∪N ∈N {T ≤ N } we conclude that Q[{T < ∞}] = 0.
We first observe that (M̃t )t∈[0,∞) is a Q-local martingale if and only if (M̃t Dt )t∈[0,∞) is a P-local
martingale. We will therefore prove that (M̃t Dt )t∈[0,∞) is a P-local martingale. For the stopping
84 BENJAMIN FEHRMAN
times {Tn }n∈N defined above, we observe that (Dt−1 hM, Dit )Tn t∈[0,∞) is a semimartingale as the

produce of semimartingales. The integration-by-parts formula then proves that

Z Tn ∧t Z Tn ∧t
Tn
(M̃ D)t = M0 D0 + M̃s dDs + Ds dM̃s + hD, M it
0 0
Z Tn ∧t Z Tn ∧t
= M 0 D0 + M̃s dDs + Ds dMs .
0 0
This completes the proof that (M̃ D)Tt n )t∈[0,∞) is a P-martingale, and therefore the proof that
(M̃t Dt )t∈[0,∞) is a P-local martingale. Hence, by the above, (M̃t )t∈[0,∞) is a Q-local martingale.
Since process of bounded variation remain processes of bounded variation with respect to an abso-
lutely continuous chang eof measure, this completes the proof.
In the following proposition we characterize the Girsanov theorem in terms of the stochastic
exponential. Precisely, we will show that the Radon-Nikodym process can be uniquely expressed as
the stochastic exponential of a continuous local martingale (Lt )t∈[0,∞) , and we then use the process
(Lt )t∈[0,∞) to simplify the correction appearing due to the change of measure.
Proposition 6.4. Let (Ω, F, P) be a probability space, let (Ft )t∈[0,∞) be a filtration on (Ω, F ), and
let (Dt )t∈[0,∞) be a strictly positive Ft -martingale. Then, there exists a unique continuous local
martingale (Lt )t∈[0,∞) such that
Dt = E(L)t .
Proof. Since the martingale (Dt )t∈[0,∞) is almost surely strictly positive, Itô’s formula proves that
Z t
1 t −2
Z
log(Dt ) = log(D0 ) + Ds−1 dDs − D dhDis .
0 2 0 s
Let (Lt )t∈[0,∞) be defined by
Z t
Lt = log(D0 ) + Ds−1 dDs ,
0
for which it follows that Z t
hLit = Ds−2 dhDis .
0
We therefore conclude that, after exponentiating the first equation,
Dt = E(L)t ,
let (Lt )t∈[0,∞) be a continuous local P-martingale. If for every t ∈ [0, ∞) we have dQ = E(L)t dP
on Ft , then for every continuous P-martingale (Mt )t∈[0,∞) the process (M̃t )t∈[0,∞) defined by
M̃t = Mt − hM, Lit ,
is a continuous local Q-martingale. Furthermore, dP = E(−L̃)t dQ on Ft .
Proof. Let (Dt )t∈[0,∞) be defined by Dt = E(L)t . In view of the Girsanov theorem, for the first
statement it remains only to prove that
Z t
Ds−1 dhM, Dis = hM, Lit .
0
Since by definition, for every t ∈ [0, ∞),
Z t
Lt = log(D0 ) + Ds−1 dDs ,
0
we have
Z t
hM, Lit = Ds−1 dhM, Dis .
0
For the second statement, since by definition and the first statement
L̃ = L − hL̃, Lit ,
is a Q-local martingale with hL̃it = hLit , we have that

1
E(−L̃) = exp(−L − hLit ) = E(L)−1
t .
2
Therefore, dP = E(−L̃) dQ on Ft for every t ∈ [0, ∞). This completes the proof.
The transformation from P-local martingales from Q-local martingales will play an important
role in our construction of weak solutions to stochastic differential equations. It is particularly
important that for a Girsanov pair (P, Q) the Girsanov transformation maps the continuous local
P-martingales bijectively onto the continuous local Q-martingales.
Definition 6.6. Let (Ω, F, P) be a probability space, let (Ft )t∈[0,∞) be a filtration on (Ω, F), and
let (P, Q) be a Girsanov pair. The map from continuous P-local martingales to continuous Q-local
martingales defined by
(Mt )t∈[0,∞) 7→ (M̃t )t∈[0,∞) ,
is called the Girsanov transformation from P to Q.
let (P, Q) be a Girsanov pair. Then, the Girsanov transformation maps the space of continuous
local P-martingales bijectively onto the space of continuous local Q-martingales.
Proof. Let (Lt )t∈[0,∞) be the unique continuous local P-martingale satisfying dQ = E(L)t dP on Ft
for every t ∈ [0, ∞). Then, the Girsanov transformation from P to Q is defined by
M̃t = Mt − hM, Lit .
However, since dP = E(−L̃)t dQ on Ft for every t ∈ [0, ∞) the inverse transformation is defined by
Mt0 = Mt + hM, L̃it = Mt + hM, Lit .

It is then clear that every continuous local P-martingale satisfies
(M̃ )0t = Mt .
In the next proposition, we prove that the Girsanov transformation commutes with stochastic
integration. That is, the Girsanov transform of the stochastic integral is the same as the stochastic
integral with respect to the Girsanov transform.
and let (P, Q) be a Girsanov pair. Let (Mt )t∈[0,∞) be a continuous local P-martingale and let
H ∈ L2loc (M ). Then,
Z t ˜ Z t
Hs dMs = Hs dM̃s .
0 t∈[0,∞) 0 t∈[0,∞)
86 BENJAMIN FEHRMAN
Proof. By density and stopping, it suffices to prove the statement for a bounded (Ht )t∈[0,∞) and for
an L2 -bounded (Mt )t∈[0,∞) . In this case, if dQ = E(L)t dP on Ft for a continuous local martingale
(Lt )t∈[0,∞) , for each t ∈ [0, ∞),
Z t ˜ Z t Z t Z t
Hs dMs = Hs dMs − Hs dhM, Lis = Hs dM̃s .
0 0 0 0
The following theorem describes how Brownian motion is transformed by an absolutely contin-
uous change of measure.
let (P, Q) be a Girsanov pair. Let (Bt )t∈[0,∞) be a standard P-Brownian motion. Then the process
(B̃t )t∈[0,∞) defined by
B̃t = Bt − hL, Bit ,
is a standard Q-Brownian motion.
Proof. By the Girsanov theorem (B̃t )t∈[0,∞) is a continuous Q-local martingale. The statement is
now an immediate consequence of Levy’s characterization of Brownian motion, using the fact that
an absolutely continuous change of measure does not change the quadratic variation.
Finally, in order to construct weak solutions to stochastic differential equations, we will oftentimes
seek to change the measure ourselves. For this, it is useful to have a criterion that guarantees the
stochastic exponential of a continuous local martingale is again a continuous local martingale. In
general, the stochastic exponential is only a supermartingale. The following two conditions are
known as Kazamaki’s criterion and Novikov’s criterion respectively.
Proposition 6.10. Let (Lt )t∈[0,∞) be a continuous local martingale. If (exp( 12 Lt ))t∈[0,∞) is a
uniformly integrable submartingale then the stochastic exponential (E(L)t )t∈[0,∞) is a uniformly
integrable martingale.
Proof. The proof will be added after the fourth problem session.
Proposition 6.11. If (Lt )t∈[0,∞) is a continuous local martingale which satisfies

1
E exp hLi∞ < ∞,
2
then (E(L))t∈[0,∞) is a uniformly integrable martingale.
Proof. The proof will be added after the fourth problem session.
Corollary 6.12. If (Lt )t∈[0,∞) is a continuous local martingale which satisfies, for every t ∈ [0, ∞),

1
E exp hLit < ∞,
2
then (E(L))t∈[0,∞) is a martingale.
Proof. We apply Novikov’s criterion applied to the stopped local martingale (Ls∧t )s∈[0,∞) , for every
t ∈ [0, ∞). This completes the proof.
7. Stochastic Differential Equations

Stochastic differential equations model quantities subject to random noise. The noise can come
from collisions on the microscopic level, such as those experienced by a particle suspended in fluid,
the molecules of a perfume being diffused by air particles, or the value of a portfolio subject to
fluctuations in the stock market. These fluctuations are modeled, for instance, by a d-dimensional
Brownian motion (Bt )t∈[0,∞) and a diffusion coefficient σ taking values in the space of (d × d)-
matrices. You can think of the coefficient σ as shaping then noise, by increasing or decreasing
its variance in certain directions. The situation just described leads to a stochastic differential
equation of the type
dXt = σ(Xt ) dBt ,
where in general the noise (Bt )t∈[0,∞) and the solution (Xt )t∈[0,∞) can have different dimensions.
In this case, σ will not be a square matrix.
However, such quantities are also subject to deterministic effects, which describe the local mean
motion. That is, on average, particles suspended in a fluid move in the direction of the current. On
average, the perfume will move in the direction of the wind. And, perhaps, the average appreciation
of a stock will be governed by a deterministic interest rate. For a d-dimensional process, such motion
is described by a drift b taking values in Rd , which leads to the more general equation
(7.1) dXt = b(Xt ) dt + σ(Xt ) dBt .
This is the prototypical model of the kind of stochastic differential equations that we will study
in this course. Most generally, we will study equations of the type, for a d-dimensional process
(Xt )t∈[0,∞) , a m-dimensional semimartingale (Zt )t∈[0,∞) , and for f taking values in the space of
(d × m)-matrices,
(7.2) dXt = f (Xt ) dZt .
By choosing
t

Zt = ,
Bt
and by choosing
f (x) = (b(x), σ(x)),
it follows that (7.1) can be written in the form (7.2).
Let (Bt )t∈[0,∞) be a standard d-dimensional Brownian motion with respect to the filtration
(FtB )t∈[0,∞) defined by
FtB = σ(Bs : s ∈ [0, T ]).
That is, the filtration (FtB )t∈[0,∞) is the filtration generated by the Brownian motion, and carries
no additional information. We will first define in which senses we judge processes (Xt )t∈[0,∞) and
(Yt )t∈[0,∞) to be the same. Immediately following, we define in which senses a process (Xt )t∈[0,∞)
can be a solution of (7.1).
Definition 7.1. Let (Xt )t∈[0,∞) and (Yt )t∈[0,∞) be d-dimensional continuous stochastic processes.
(i) We say that (Xt )t∈[0,∞) and (Yt )t∈[0,∞) are indistinguishable if
P (Xt = Yt for every t ∈ [0, ∞)) = 1.
(ii) We say that (Xt )t∈[0,∞) is a modification of (Yt )t∈[0,∞) if
P (Xt = Yt ) = 1 for every t ∈ [0, ∞)
(iii) We say that (Xt )t∈[0,∞) and (Yt )t∈[0,∞) have the same finite-dimensional distributions if, for
every 0 ≤ t0 < t1 < . . . < tn < ∞) and for every measurable A ⊆ (Rd )n ,
P((Xt1 , . . . , Xtn ) ∈ A) = P((Yt1 , . . . , Ytn ) ∈ A).
88 BENJAMIN FEHRMAN
Definition 7.2. Let d, m ∈ N, let C([0, ∞); Rd ) denote the space of continuous functions from
[0, ∞) to Rd , and let

b : [0, ∞) × C([0, ∞), Rd ) → Rd and σ : [0, ∞) × C([0, ∞); Rd ) → Rd×m ,
be measurable functions. A solution to the stochastic differential equation, for a probability measure
µ on Rd ,
dXt = b(t, X· ) dt + σ(t, X· ) dBt in (0, ∞),

X0 = µ,
is a m-dimensional Brownian motion (Bt )t∈[0,∞) and a d-dimensional continuous semimartingale
(Xt )t∈[0,∞) which satisfy that, for every measurable A ⊆ Rd ,
P(X0 ∈ A) = µ(A),
and that, almost surely,
Z t Z t
Xt = X0 + b(s, X· ) ds + σ(s, X· ) dBs .
0 0
We say that the solutions (Xt , Bt )t∈[0,∞) is a strong solution if (Xt )t∈[0,∞) is FtB -adapted. If
(Xt )t∈[0,∞) is not FtB -adapted, then we say the the solution (Xt , Bt )t∈[0,∞) is a weak solution.
Remark 7.3. Observe in particular that the above definition applies to measurable maps

b̃ : [0, ∞) × Rd → Rd and σ̃ : [0, ∞) × Rd → Rd×m ,
by defining b(s, X· ) = b̃(s, Xs ) and σ(s, X· ) = σ̃(s, Xs ). This is certainly the most important case
that will be considered in this course, and yields the stochastic differential equation
dXt = b̃(t, Xt ) dt + σ̃(t, Xt ) dBt ,
which is a time-dependent version of the standard type (7.1) discussed above.
The following definition extends Definition 7.2 to the case of a general semimartingale (Zt )t∈[0,∞) .
The definition is identical, except that in this case a strong solution is adapted to the filtration
(FtZ )t∈[0,∞) generated by (Zt )t∈[0,∞) .
f : [0, ∞) × C([0, ∞); Rd ) → Rd×m ,
be a measurable function. Let (Zt )t∈[0,∞) be a continuous semimartingale. A solution to the
stochastic differential equation, for a probability measure µ on Rd ,
dXt = f (t, X· ) dZt in (0, ∞),

X0 = µ,
is a d-dimensional continuous semimartingale (Xt )t∈[0,∞) which satisfies that, for every measurable
A ⊆ Rd ,
P(X0 ∈ A) = µ(A),
and that, almost surely, Z t
Xt = X0 + f (s, X· ) dZs.
0
We say that the solutions (Xt , Zt )t∈[0,∞) is a strong solution if (Xt )t∈[0,∞) is FtZ -adapted. If
(Xt )t∈[0,∞) is not FtZ -adapted, then we say the the solution (Xt , Zt )t∈[0,∞) is a weak solution.
We will now define two notions of uniqueness for solutions in the sense of Definition 7.2.

b : [0, ∞) × C([0, ∞), Rd ) → Rd and σ : [0, ∞) × C([0, ∞); Rd ) → Rd×m ,
be measurable functions. For a probability measure µ on Rd , consider the stochastic differential

equation
dXt = b(t, X· ) dt + σ(t, X· ) dBt in (0, ∞),

(7.3)
X0 = µ.
(i) We say that there is pathwise uniqueness for (7.3) if whenever (Xt , Bt )t∈[0,∞) and (Xt0 , Bt0 )t∈[0,∞)
are two solutions defined on the same probability space which satisfy that X0 = X00 almost
surely and that (Bt )t∈[0,∞) and (Bt0 )t∈[0,∞) are indistinguishable m-dimensional Brownian mo-
tions, we have that (Xt )t∈[0,∞) and (Xt0 )t∈[0,∞) are indistinguishable.
(ii) We say that there is uniqueness in law for (7.3) if whenever (Xt , Bt )t∈[0,∞) and (Xt0 , Bt0 )t∈[0,∞)
are two solutions (defined for possibly different m-dimensional Brownian motions on possibly
different probability spaces) for which X0 = X00 in distribution, we have that (Xt )t∈[0,∞) and
(Xt0 )t∈[0,∞) have the same law.
90 BENJAMIN FEHRMAN
7.1. The space of continuous paths. Let C([0, ∞); Rd ) denote the space of continuous functions
X : [0, ∞) → Rd . This space comes equipped with the metric
∞
X
(7.4) d(X, Y ) = 2−k max |Xs − Ys | ∧ 1 ,
s∈[0,k]
k=1
which induces the topology of locally uniform convergence. This is to say that a sequence of
continuous paths {Xn }n∈N converges to a path X if and only if for every T ∈ [0, ∞),

lim max |Xsn − Xs | = 0.
n→∞ s∈[0,T ]
The space C([0, ∞); Rd ) comes equipped with the metric topology induced by (7.4) and the Borel
σ-algebra B(C([0, ∞); Rd )) generated by the metric topology. The Borel σ-algebra comes equipped
with a natural filtration {Ft }t∈[0,∞) . For every t ∈ [0, ∞) the σ-algebra Ft is the sigma algebra
generated by continuous paths up to time t, which can be defined using the projection map
πt : C([0, ∞); Rd ) → C([0, t]; Rd ),
and letting Ft be defined by

(7.5) Ft = πt−1 B(C([0, t]; Rd )) .
More explicitly, this definition if equivalent to defining

Ft = {A ∈ B(C([0, ∞); Rd )) : X· ∈ A if and only if X·∧t ∈ A},
where X·∧t denotes the path X stopped at time t (which is again an element of C([0, ∞); Rd ))).
We can therefore view C([0, ∞); Rd ) equipped with the σ-algebra B(C([0, ∞); Rd )) and filtration
{Ft }t∈[0,∞) as a filtered measurable space. Now, suppose that (Ω, G, P) is a probability space
with a filtration {Gt }t∈[0,∞) . Suppose that (Bt )t∈[0,∞) be a Gt -Brownian motion and suppose that
(Xt )t∈[0,∞) is a weak solution of the stochastic differential equation, defined for bounded measurable
functions σ and b,
(7.6) dXt = σ(Xt ) dBt + b(Xt ) dt in (0, ∞) with X0 = 0.
The solution map is a measurable function S : Ω → C([0, ∞); Rd ) that is almost surely defined by
S(ω) = X· (ω). Observe as well that, using the definition, we have S −1 (Ft ) ⊆ Gt for every t ∈ [0, ∞).
We can then define the pushforward measure P = S∗ P on C([0, ∞); Rd ) by
P (A) = P[Xt (ω) ∈ A],
and define the probability space (C([0, ∞)), B(C([0, ∞); Rd )), P ) with filtration {Ft }t∈[0,∞) . And
we can define a Ft -Brownian motion (B̃t )t∈[0,∞) on (C([0, ∞)), B(C([0, ∞); Rd )), P ) by defining
−1
(
0 if SB (X· ) = ∅,
B̃t (X· ) = −1
Bt (ω) if ω ∈ SB (X· ),
where SB : Ω → C([0, ∞); Rd ) denotes the measurable map SB (ω) = B· (ω). Then, with respect to
P , paths X· ∈ C([0, ∞); Rd ) almost surely satisfy, for every t ∈ [0, ∞),
Z t Z t
(7.7) Xt = b(Xs ) ds + σ(Xs ) dB̃s (X· ).
0 0
We can therefore view a weak solution of (7.6) as a measure P on (C([0, ∞)), B(C([0, ∞); Rd )) and
an Ft -Brownian motion (B̃t )t∈[0,∞) on (C([0, ∞)), B(C([0, ∞); Rd ), P ) such that P -almost every
path satisfies (7.7).
This formulation leads to the so-called martingale formulation of the equation (7.6). Restricting
to the case of one-dimension, if f is twice-continuously differentiable with bounded derivatives up
to order two it follows from Itô’s formula that
Z t Z t
0 1 2 00
f (Xt ) − f (X0 ) − b(Xs )f (Xs ) + σ (Xs )f (Xs ) ds = σ(Xs )f 0 (Xs ) dB̃s .
0 2 0
That is, the process
Z t
0 1 2 00
(7.8) f (Xt ) − f (X0 ) − b(Xs )f (Xs ) + σ (Xs )f (Xs ) ds
0 2 t∈[0,∞)
is a P -martingale. This is the so-called martingale problem associated to the coefficients (b, σ 2 ).
A solution to the martingale problem with initial data defined by a probability measure µ on R
is a measure P on (C([0, ∞)), B(C([0, ∞); Rd )) such that (i) for every measurable A ⊆ R we have
P [X0 ∈ A] = µ and (ii) the process (7.8) is a P -martingale for every twice-differentiable function f
with bounded derivatives up to order two. While we will not develop this approach in these notes,
do keep it in mind for the future. The martingale problem provides a very general and powerful
means for proving the existence of weak solutions to (7.6) and for proving that the solutions are
unique in law. A full account of this approach can be found in the book Multidimensional Diffusion
Processes by Stroock and Varadhan.
In the following section, we will use the following important proposition concerning measures on
the space (C([0, ∞)), B(C([0, ∞); Rd )) equipped with the filtration (Ft )t∈[0,∞) .
Proposition 7.6. Let (C([0, ∞)), B(C([0, ∞); Rd )) be the space of continuous paths equipped with
its Borel sigma algebra and let (Ft )t∈[0,∞) be the filtration (7.5). Assume that for ever t ∈ [0, ∞)
there exists a probability measure Qt on the measurable space (C([0, ∞); Rd ), Ft ), and assume that
the probability measures satisfy the compatibility condition, for every s ≤ t ∈ [0, ∞),
Qt |Fs = Qs (that is, Qt (A) = Qs (A) for every A ∈ Fs ).
Then there exists a probability measure Q on (C([0, ∞)), B(C([0, ∞); Rd )) such that, for every t ∈
[0, ∞),
Q|Ft = Qt .
Proof. The statement can be found in Appendix 6 of Revuz and Yor, and the proof appears as
special case of Theorem 1.3.5 in Stroock and Varadhan.
92 BENJAMIN FEHRMAN
7.2. Existence of weak solutions by change of measure and time. In this section, we will
construct weak solutions to stochastic differential equations of the form
(7.9) dXt = b(Xt ) dt + σ(Xt ) dBt ,
using the Girsanov theorem and change of measure and using random time change. Precisely, given
a one-dimensional Brownian motion (Bt )t∈[0,∞) and a positive bounded function σ : R → R we will
show that there exists a time change (τt )t∈[0,∞) and a new Brownian motion (B̃t )t∈[0,∞) such that
the process (Xt )t∈[0,∞) defined by
Xt = Bτt ,
is a solution to the equation
(7.10) dXt = σ(Xt ) dB̃t .
Then, given a bounded measurable function b : R → R, we will prove using the Girsanov theorem
that there exists a new Brownian motion (Bt0 )t∈[0,∞) defined with respect to a new measure such
that, with respect to this measure, the process (Xt )t∈[0,∞) is a solution to the equation
dXt = b(Xt ) dt + σ(Xt ) dBt0 .
That is, beginning from a Brownian motion (Bt )t∈[0,∞) , we can construct weak solutions to (7.9)
for a general class of coefficients. In the sections below, we will establish conditions that imply
pathwise uniqueness holds for (7.9) and therefore using the theorem below that the weak solutions
are in fact strong solutions.
Theorem 7.7. Let (Bt )t∈[0,∞) be a standard n-dimensional Brownian motion. Let σ : C([0, ∞); Rd )×
[0, ∞) → Rd×n and b : C([0, ∞), Rd ) × [0, ∞) → Rd be measurable predictable functions. If pathwise
uniqueness holds for the equation
dXt = b(X· , t) dt + σ(X· , t) dBt in (0, ∞),

(7.11)
X0 = x,
then uniqueness in law holds for (7.11) and every weak solution is a strong solution.
We will now prove that above outline. In the first statement, we essentially use a random time
change to transform a Brownian motion into a solution to a more general diffusion equation of the
type (7.10).
Proposition 7.8. Let (Bt )t∈[0,∞) be a standard Brownian motion. Let σ : R → R be a bounded
measurable function which satisfies, for some ε ∈ (0, ∞),
0 < ε ≤ σ.
Then, for each x ∈ R, the equation
1
(
dXt = σ 2 (Xt ) dBt in (0, ∞),
X0 = x
has a weak solution and satisfies uniqueness in law.
Proof. We assume without loss of generality that x = 0. Since σ is bounded and strictly positive,
define the family of time changes (τt )t∈[0,∞) by
Z s
−1
τt = inf s ∈ [0, ∞) : σ (Br ) dr ≥ t. .
0
The boundedness and positivity of σ proves that the process
Z τt
1
t ∈ [0, ∞) 7→ σ − 2 (Bs ) dBs ,
0
is a local martingale. Furthermore, it follows by definition of (τt )t∈[0,∞) that the quadratic variation
process satisfies Z τ· Z τt
− 12
h σ (Bs ) dBs it = σ −1 (Bs ) ds = t.
0 0
Therefore, by Levy’s characterization of Brownian motion, there exists a Brownian motion (B̃t )t∈[0,∞)
such that Z τt
1
B̃t = σ − 2 (Bs ) dBs .
0
Or, equivalently, that
1
dB̃t = σ − 2 (Bτt ) dBτt ,
which implies that
1
dBτt = σ 2 (Bτt ) dB̃t .
Hence, after defining the process (Xt )t∈[0,∞) by Xt = Bτt we conclude that (Xt , B̃t )t∈[0,∞) is a
solution to the equation
1
(
dXt = σ 2 (Xt ) dB̃t in (0, ∞),
(7.12)
X0 = 0.
For the converse, the positivity of σ proves that, as t → ∞,
Z t
hXit = σ(Xs ) ds → ∞.
0
Therefore, by the Dambis-Dubins-Schwarz theorem, every solution to (7.12) is a time-changed
Brownian motion. This completes the proof.
In the next two propositions, we will use the Girsanov theorem to solve a general diffusion
equation with drift. This will be achieved by changing the underlying measure. We assume without
loss of generality that (Ω, F, P) is the probability space (C([0, ∞)), B(C([0, ∞)), P ) constructed in
Section 7.1.
Proposition 7.9. Let (Bt )t∈[0,∞) be a standard Ft -Brownian motion with respect to some filtration
(Ft )t∈[0,∞) . Let σ : R → R and b : R → R be bounded measurable functions. Then for each x ∈ R
there exist weak solutions to the equation
dXt = σ(Xt ) dBt in (0, ∞),

(7.13)
X0 = x
if and only if there exist weak solutions to the equation
dXt = σ(Xt )b(Xt ) dt + σ(Xt ) dBt in (0, ∞),

(7.14)
X0 = x
Furthermore, equation (7.13) satisfies uniqueness in law if and only if (7.14) satisfies uniqueness
in law.
Proof. Without loss of generality we assume that x = 0. Suppose that (Xt )t∈[0,∞) is a weak solution
of (7.13). Since b is bounded, define the martingale (Lt )t∈[0,∞) by
Z t
Lt = b(Xs ) dBs
0
and using Proposition 7.6 we define the measure Q by dQ|Ft = E(L)t dP|Ft . The Girsanov theorem
proves that the process (B̃t )t∈[0,∞) defined by
Z t
B̃t = Bt − hL, Bit = Bt − b(Xs ) ds,
0
94 BENJAMIN FEHRMAN
is a Q-Brownian motion. And furthermore, since (Xt )t∈[0,∞) solves (7.13), we have that
dXt = σ(Xt ) dBt
= σ(Xt ) dBt − σ(Xt )b(Xt ) dt + σ(Xt )b(Xt ) dt
= σ(Xt ) dB̃t + σ(Xt )b(Xt ) dt.
That is, with respect to the measure Q we have that (Xt , B̃t )t∈[0,∞) is a solution of (7.14). For
the reverse direction, if a process (Xt )t∈[0,∞) is a solution of (7.14) with respect to the Brownian
motion (Bt )t∈[0,∞) we invert the previous transformation by replacing the process (Lt )t∈[0,∞) by
the process (−Lt )t∈[0,∞) . This completes the proof.
Remark 7.10. We emphasize that the measure Q constructed in the proof is in general singular
with respect to P, despite the fact that Q|Ft is mutually absolutely continuous with respect to P|Ft
for every t ∈ [0, ∞). Indeed, the Radon-Nikodym derivative is E(L)t . To see that Q can be singular,
consider the two equations
dXt1 = dBt and dXt2 = dBt + dt with X01 = X02 = 0,
for which we have Xt1 = Bt and Xt2 = Bt + t.
By the Girsanov theorem, the laws of the processes X 1 and X 2 on C([0, T ]; R) are mutually
absolutely continuous for every T ∈ [0, ∞). However, if Q1 denotes the pushforward measure of X 1
on C([0, ∞); R) and Q2 the pushforward measure of X 2 , then
0 = Q1 [lim sup Xt/t = 1] 6= Q2 [lim sup Xt/t = 1] = 1.
t→∞ t→∞
That is,Q1 Q2
and are singular measures on C([0, ∞); R). We should expect this, since the solution
Xt2 is almost surely running to infinity as t → ∞ but the solution Xt1 is a Brownian motion, and
hence returns to zero infinity often as t → ∞. The interesting statement is that on every finite
interval [0, T ], the laws of the two solutions are mutually absolutely continuous.
Proposition 7.11. Let (Bt )t∈[0,∞) be a standard Brownian motion. Let σ : R → R and b : R → R
be bounded measurable functions and assume that σ satisfies, for some ε ∈ (0, ∞),
0 < ε ≤ σ.
Then for each x ∈ R there exist weak solutions to the equation

(7.15)
X0 = x
if and only if there exist weak solutions to the equation
dXt = b(Xt ) dt + σ(Xt ) dBt in (0, ∞),

(7.16)
X0 = x
Furthermore, equation (7.15) satisfies uniqueness in law if and only if (7.16) satisfies uniqueness
in law.
Proof. The proof is an immediate consequence of Proposition 7.9 applied to the drift
b̃ = σ −1 b,
where the positivity of σ is used to ensure that this function is bounded. This completes the
proof.
We are now prepared to present the main theorem of this section, which proves that for strictly
positive diffusion coefficients we can always find weak solutions to equations of the form (7.9). And
furthermore, such equations satisfy uniqueness in law.
Theorem 7.12. Let (Bt )t∈[0,∞) be a standard Brownian motion. Let σ : R → R and b : R → R be
bounded measurable functions and assume that σ satisfies, for some ε ∈ (0, ∞),
0 < ε ≤ σ.
Then for each x ∈ R there exist a weak solution to the equation
dXt = b(Xt ) dt + σ(Xt ) dBt in (0, ∞),

(7.17)
X0 = x
Furthermore, equation (7.17) satisfies uniqueness in law.
Proof. The proof is an immediate consequence of Proposition 7.8 and Proposition 7.11. That is,
assuming without loss of generality that x = 0, given a Brownian motion (Bt )t∈[0,∞) we first define
the time change (τt )t∈[0,∞) by
Z t
−2
τt = inf s ∈ [0, ∞) : σ (Bs ) ds = 1 .
0
Then by Proposition 7.8 there exists a Brownian motion (B̃t )t∈[0,∞) such that the process (Xt )t∈[0,∞)
defined by Xt = Bτt satisfies X0 = 0 and is a weak solution of
dXt = σ(Xt ) dB̃t .
Then we define the martingale (Lt )t∈[0,∞) by
Z t
Lt = σ −1 (Xs )b(Xs ) dB̃s ,
0
and the corresponding measure Q by dQFt = E(L)t dPFt . It then follows from the Girsanov theorem
and Proposition 7.11 that the process (Bt0 )t∈[0,∞) defined by
Z t
0
Bt = B̃t − σ −1 (Xs )b(Xs ) ds,
0
is a Q-Brownian motion and that 0
(Xt , Bt )t∈[0,∞) satisfies X0 = 0 and is a weak solution to
dXt = σ(Xt ) dBt0 + b(Xt ) dt.
Since Brownian motion satisfies uniqueness in law, we conclude that (7.17) satisfies uniqueness in
law.
Remark 7.13. We emphasize that the solutions constructed in Theorem 7.12 are in general weak
solutions, but not strong solutions, despite the fact that we started with a strong solution. That is,
despite the fact that we started with a Brownian motion. The reason for this is that the time-change
introduces a new filtration, and the application of the Girsanov theorem changes the underlying
measure and thereby introduces a new Brownian motion.
Remark 7.14. The conclusion of Theorem 7.12 remains true in higher dimensions, however in
this case the assumption that σ is positive is replaced by the assumption that the matrix σσ t is
uniformly elliptic. This is to say that, for some constants λ ≤ Λ ∈ (0, ∞), as symmetric matrices,
λI ≤ σσ t ≤ ΛI.
Or, in terms of eigenvalues, this condition states that the eigenvalues of σσ t are bounded below by
λ and above by Λ. The proof in the higher dimensional case is significantly more difficult.
96 BENJAMIN FEHRMAN
7.3. Existence and uniqueness: Lipschitz continuous coefficients. In this section, we will
prove pathwise uniqueness for solutions of the stochastic differential equation
dXt = f (t, X· ) dZt ,
for a continuous semimartingale (Zt )t∈[0,∞) , and a for a Lipschitz continuous function f on the
space of continuous paths. We recall that Lipschitz continuity is also the condition required to
prove existence and uniqueness for ordinary differential equations of the type
dXt = f (Xt ) dt.
Indeed, the equation p
dXt = Xt dt with X0 = 0,
has infinitely many solutions. In the sections to follow, we will prove that, at least for one-
dimensional equations, the Lipschitz continuity can be relaxed for the case of Brownian motion
due to the regularizing effect of its quadratic variation. We will present the proof of existence and
uniqueness immediately following Grönwall’s inequality.
Proposition 7.15. Let φ : [0, ∞) → [0, ∞) be a nonnegative, locally bounded function which satis-
fies, for some a, b ∈ [0, ∞), for every t ∈ [0, ∞),
Z t
φ(t) ≤ a + b φ(s) ds.
0
Then, for every t ∈ [0, ∞),
φ(t) ≤ a exp(bt).
Proof. Let t ∈ [0, ∞). By assumption,
Z t Z t Z s
φ(t) ≤ a + b φ(s) ds ≤ a + b a+ φ(r) dr .
0 0 0
Therefore, Z tZ s
φ(t) ≤ a + abt + φ(r) dr ds.
0 0
Proceeding inductively, it follows for every n ∈ N that
n Z t Z t1 Z tn
X a(bt)k n+1
φ(t) ≤ +b ... φ(tn+1 ) dtn+1 dtn . . . dt1 ,
k! 0 0 0
k=0
where the local boundedness of φ proves that, for some c ∈ (0, ∞) depending on t ∈ [0, ∞),
Z t Z t1 Z tn
cbn+1 tn+1
bn+1 ... φ(tn+1 ) dtn+1 dtn . . . dt1 ≤ .
0 0 0 (n + 1)!
We therefore conclude that, after passing to the limit n → ∞,
∞
X a(bt)k
φ(t) ≤ = a exp(bt),
k!
k=0
Theorem 7.16. Let m, d ∈ N. Let (Zt )t∈[0,∞) be a m-dimensional continuous Ft -semimartingale.
Let
f : [0, ∞) × C([0, ∞), Rd ) → Rd×m ,
be a locally bounded measurable function that is Lipschitz continuous in the sense that, for some
K ∈ (0, ∞),
|f (t, X· ) − f (t, Y· )| ≤ K sup |Xs − Ys | ,
s∈[0,t]
for every t ∈ [0, ∞) and X, Y ∈ C([0, ∞); Rd ). For every x ∈ Rd , here exists a unique up to being
indistinguishable FtZ -adapted process (Xt )t∈[0,∞) such that
Z t
Xt = x + f (s, X· ) dZs .
0
Proof. We will consider the case m = d = 1. The additional difficulties in the general setting are
only notational. The proof follows by Picard iteration. Let (Xt0 )t∈[0,∞) denote the process Xt0 = x
for every t ∈ [0, ∞). Then, for every n ∈ N, define inductively the process (Xtn )t∈[0,∞) by
Z t
Xtn =x+ f (s, X·n−1 ) dZs .
0
Let (Zt = Mt +At )t∈[0,∞) denote the semimartingale decomposition of (Zt )t∈[0,∞) , where (At )t∈[0,∞)
is a process of bounded variation and (Mt )t∈[0,∞) is a continuous local martingale. We will first
consider the case that the measures dhM it and | dAt | are dominated by the Lebesgue measure.
The general case will then follow using a time-change argument.
Fix t ∈ [0, ∞). We first observe that, if n = 1, for each s ∈ [0, t],
Z s 2 Z s 2 !
1
Xs − Xs0 ≤ 2

f (r, x) dMr + f (r, x) dAr ,
0 0
and, if n ∈ {2, 3, . . .}, for each s ∈ [0, t],

Z t 2
n n−1
2 n−1 n−2

Xs − Xs ≤ 2 f (r, X· ) − f (r, X· ) dMr

0
Z t 2
n−1 n−2

+ 2
f (r, X· ) − f (r, X· ) dAr .
0
It therefore follows that, for each n ∈ {2, 3, . . .},

" # " Z r 2 #
n n−1
2 n−1 n−2

E sup Xt − Xt ≤ 2E sup f (r, X· ) − f (r, X· ) dMr
s∈[0,t] s∈[0,t] 0
" Z s 2 #
n−1 n−2

+ 2E sup f (r, X· ) − f (r, X· ) dAr .
s∈[0,t] 0
The Burkholder-David-Gundy inquality with p = 2 (or, really, Doob’s inequality in this case) and
Hölder’s inequality prove that
" # Z t
n n−1
2 n−1 n−2
2
E sup Xs − Xs ≤ 8E f (s, X· ) − f (s, X· ) dhM is
s∈[0,t] 0
Z t 2

+ 2tE f (s, X·n−1 ) − f (s, X·n−2 ) | dAs | .
0
It then follows from the Lipschitz condition and the assumption that dhM it ≤ dt and | dAt | ≤ dt
that
" # Z t " #
2 2
E sup Xs − Xsn−1 ≤ 2K 2 (4 + t)
n
E sup Xsn−1 − Xsn−2 ds.

s∈[0,t] 0 r∈[0,s]
98 BENJAMIN FEHRMAN
Applying this inequality inductively, using the fact that t ∈ [0, ∞) was arbitrary, we conclude using
the definition of (Xt0 )t∈[0,∞) that, for every n ∈ N,
" # Z t Z t1 Z tn " #
2 n 2
E sup Xs − Xsn−1 ≤ 2K 2 (4 + t)
n
E sup Xr1 − Xr0 ds

...
s∈[0,t] 0 0 0 r∈[0,s]
n " #
2K 2 (4 + t) tn 1 2
≤ E sup Xs − x .
n! s∈[0,t]
Since for every t ∈ [0, ∞) the local boundedness of f , the definition of (Xt1 )t∈[0,∞) , and the fact
that dhM it < ∞ imply that
" #
1 2
E sup Xs − x < ∞,
s∈[0,t]
we conclude that
∞
" #
X 2
sup Xs − Xsn−1 < ∞.
n
E
n=1 s∈[0,t]
Therefore, almost surely, we have that

∞
X
sup |Xsn − Xtn | < ∞,
n=1 s∈[0,t]
which implies that there exists a FtZ -measurable continuous process (Xt )t∈[0,∞) , which is continuous
since it is the uniform limit of continuous processes, and which is FtZ -measurable since each of the
(Xtn )t∈[0,∞) are FtZ -measurable, such that, almost surely, for every t ∈ [0, ∞), as n → ∞,
(7.18) X·n → X· uniformly on [0, t].
Since a repetition of the above estimates prove that
Z t Z t
n
(7.19) lim f (s, X· ) dZs = f (s, X· ) dZs ,
n→∞ 0 0
it follows from (7.18) and (7.19) that (Xt )t∈[0,∞) is a solution of

dXt = f (t, X· ) dZt in (0, ∞),

(7.20)
X0 = x
To prove uniqueness, suppose that (Xt )t∈[0,∞) and (Yt )t∈[0,∞) are two solutions of (7.20). For
every n ∈ N let Tn denote the FtZ -stopping time
Tn = inf{t ∈ [0, ∞) : |Xt | ≥ n or |Yt | ≥ n}.
It then follows that the stopped processes (XtTn )t∈[0,∞) and (YtTn )t∈[0,∞) are solutions of (7.20) for
the semimartingale (ZtTn )t∈[0,∞) . It then follows by a repetition of the above estimates that
" # Z t " #
2 2
E sup Xs n − YsTn ≤ 2K 2 (4 + t)
T
E sup XrTn − YrTn dr.

s∈[0,t] 0 s∈[0,r]
Grönwall’s inequality therefore proves that, for every t ∈ [0, ∞) and n ∈ N,

" #
2
E sup Xs n − YsTn = 0.
T
s∈[0,t]
By letting t → ∞ and n → ∞, the monotone convergence theorem proves that

" #
E sup |Xs − Ys |2 = 0,
s∈[0,∞)
which proves that (Xt )t∈[0,∞) and (Yt )t∈[0,∞) are indistinguishable.
It remains to treat the general case. For this, define the time change {Ct }t∈[0,∞) by
Z s
Ct = inf s ∈ [0, ∞) : t + | dAr | + hM it ≥ t.
0
Consider the time changed process (Z̃t )t∈[0,∞) by

Z̃t = ZCt ,
and define the function
f˜: ([0, ∞) × C([0, ∞); R)) → R
by f˜(s, X· ) = f (Cs , X· ). Since it follows by definition that Ct ≤ t for every t ∈ [0, ∞), we have by
definition that

˜
f (t, X· ) − f˜(t, Y· ) = |f (Ct , X· ) − f (Ct , Y· )| ≤ K sup |Xs − Ys | ≤ K sup |Xs − Ys | ,

s∈[0,Ct ] s∈[0,t]
and that f˜ is locally bounded. Therefore, since (Z̃t = Ãt + M̃t )t∈[0,∞) for (Ãt = ACt )t∈[0,∞) and
(M̃t = MCt )t∈[0,∞) is a semimartingale satisfying by definition of the time change (Ct )t∈[0,∞) that

dhM̃ it ≤ dt and dÃt ≤ dt,

there exists a unique solution (X̃t )t∈[0,∞) to the stochastic differential equation
dX̃t = f˜(t, X̃· ) dZ̃t in (0, ∞),
(
X̃0 = x.
Define the inverse process (At )t∈[0,∞) by
At = inf{s ∈ [0, ∞) : Cs = t}.
It then follows by the change of variables formula that the process (Xt )t∈[0,∞) defined by
Xt = X̃At ,
is a solution of the stochastic differential equation
dXt = f (t, X· ) dZt in (0, ∞),

X0 = x,
The following theorem proves that the solution can be constructed to depend continuously on the
initial condition as well as time under the additional assumption that the coefficient f is globally
bounded.
Theorem 7.17. Let m, d ∈ N. Let (Zt )t∈[0,∞) be a m-dimensional continuous Ft -semimartingale.
Let
f : [0, ∞) × C([0, ∞), Rd ) → Rd×m ,
be a bounded measurable function that is Lipschitz continuous in the sense that, for some K ∈
(0, ∞),
|f (t, X· ) − f (t, Y· )| ≤ K sup |Xs − Ys | ,
s∈[0,t]
100 BENJAMIN FEHRMAN
for every t ∈ [0, ∞) and X, Y ∈ C([0, ∞); Rd ). Then there exists a unique FtZ -adapted process
{Xtx }t∈[0,∞),x∈Rd that is continuous is both variables (x, t) ∈ Rd × [0, ∞) such that almost surely,
for every x ∈ Rd and t ∈ [0, ∞),
Z t
x
(7.21) Xt = x + f (s, X·x ) dZs .
0
Proof. The issue is proving continuity in space. For every x ∈ R there exists a continuous in time
solution (Xtx )t∈[0,∞) of (8.12). Let t ∈ (0, ∞). We will consider the solution as a map from R to
the Banach space fo continuous functions C([0, T ]; R) equipped with the norm
kf kC([0,t];R) = sup |f (s)| .
s∈[0,t]
We will then apply the Kolmogorov continuity criterion to the map

x ∈ R 7→ (Xsx )s∈[0,t] ∈ C([0, t]; R).
Let p ∈ [2, ∞). The inequality |a + b + c|p ≤ 3p−1 (|a|p + |b|p + |c|p ) and equation (8.12) prove that,
for each x, y ∈ Rd ,
sup |Xsx − Xsy |p
s∈[0,t]
≤ 3p−1 |x − y|p
p p !
s s
Z Z
p−1
µ(r, Xrx ) µ(r, Xry ) dr σ(r, Xrx ) y

+3 sup − + sup − σ(r, Xr ) dBr .
s∈[0,t] 0 s∈[0,t] 0
The Burkholder-Davis-Gundy inequality, Hölder’s inequality, and p ∈ [2, ∞) prove that there exists
Cp ∈ (0, ∞) such that
" Z s p # "Z
t p2 #
σ(r, Xrx ) − σ(r, Xry ) dBr ≤ Cp E ((σ(s, Xsx ) − σ(s, Xsy ))2 ds

E sup
s∈[0,t] 0 0
"Z
t p2 #
≤ Cp Kp2 E |Xsx − Xsy |2 ds
0
" #
p−2
Z t
≤ Cp K2p t p E sup |Xsx − Xsy |p dr.
0 s∈[0,r]
Similarly, it follows by Jensen’s inequality and p ∈ [2, ∞) that

Z s p Z t
µ(r, Xr ) − µ(r, Xr ) dr ≤ tp−1
x y
|µ(r, Xrx ) − µ(r, Xry )|p dr

sup
s∈[0,t] 0 0
Z t
p−1 p
≤ t K2 sup |Xsx − Xsy |p dr.
0 s∈[0,r]
The previous three inequalities prove that, for every p ∈ [2, ∞) and t ∈ [0, ∞),
" #
E sup |Xsx − Xsy |p
s∈[0,t]
" # !
p−2 Z t
≤ 3p−1 |x − y|p + Cp K2p t p + K2p tp−1 E sup |Xsx − Xsy |p dr .
0 s∈[0,r]
Grönwall’s inequality proves that, for every p ∈ [2, ∞) and t ∈ [0, ∞) there exists a constant
c(t, p) ∈ (0, ∞) that is locally bounded in t and p such that
" #
E sup |Xsx − Xsy |p ≤ c(t, p) |x − y|p .
s∈[0,t]
The Kolmoogorov continuity criterion that there exists a continuous modification (X̂sx )s∈[0,t] of the
process x ∈ R 7→ (Xsx )s∈[0,t] ∈ C([0, T ]; R). It follows that the modification solves (8.12), because
the zero set is independent of time. That is, for every x ∈ R we have
P[X̂tx = Xtx for every t ∈ [0, ∞)] = 1.
7.4. Uniqueness by local times: Hölder continuous coefficients in one-dimension. In

the previous section, for instance, we proved the pathwise well-posedness of stochastic differential
equations of the form
dXt = b(t, Xt ) dt + σ(t, Xt ) dBt ,
for coefficients b and σ that were locally bounded and globally Lipschitz continuous. In general,
these assumptions cannot be relaxed for the drift b, since the equations
p
ẋ(t) = x(t) and ẋ(t) = x(t)2 ,
are either have non-unique solutions or blow-up in finite time. However, in this section, we will
prove using local times that the these assumptions can be relaxed for the diffusion coefficient σ. In
particular, we will prove that the equation
( p
dXt = Xt dBt in (0, ∞),
(7.22)
X0 = 0
is well-posed and that zero is its unique solution.
Uniqueness for equations like (7.22) will be obtained the local time of the solution at zero. The
idea is that on the one hand a solution (Xt )t∈[0,∞) of (7.22) begins at zero and therefore if (Xt )t∈[0,∞)
is non-constant we expect that L0t (X) > 0 for every t ∈ (0, ∞). However, on the other hand, the
quadratic variation process
dhXit = Xt dt,
vanishes whenever Xt = 0. Therefore, time isn’t running when Xt = 0, and we expect that
L0t (X) = 0 for all t ∈ [0, ∞). That is, from the point of the view of the local time, the process
(Xt )t∈[0,∞) spends no time at zero. The only way to reconcile these two perspectives is for (Xt )t∈[0,∞)
to be constantly equal to zero.
Proposition 7.18. Let (Bt )t∈[0,∞) be a one-dimensional Brownian motion and let b, σ : [0, ∞) ×
R → R be predictable processes. Let (Xt1 )t∈[0,∞) and (Xt2 )t∈[0,∞) satisfy that X01 = X02 almost
everywhere and that, for each i ∈ {1, 2},
(7.23) dXti = b(Xti ) dt + σ(Xti ) dBt .
Then X 1 ∨ X 2 is a solution of (7.23) if and only if L0t (X 2 − X 1 ) vanishes identically.
Proof. We observe using Tanaka’s formula applied to the semimartingale (Xt1 − Xt2 )t∈[0,∞) with
a = 0 that
Z t
1 2 1 2 1 1
Xt ∨ Xt = Xt + (Xt − Xt )+ = Xt + 1{Xt2 >Xt1 } d(Xt2 − Xt1 ) + L0t (X 2 − X 1 ).
0
Since, for each i ∈ {1, 2} we have that
dXti = b(Xti , t) dt + σ(Xti , t) dBt ,
it follows that, since X01 = X02 almost everywhere,
Z t
1 2 1
Xt ∨ Xt = X0 + 1{Xs1 ≥Xs2 } b(Xs1 , s) + 1{Xs2 >Xs1 } b(Xs2 , s) ds
0
Z t
+ 1{Xs1 ≥Xs2 } σ(Xs1 , s) + 1{Xs2 >Xs1 } σ(Xs2 , s) dBs + L0t (X 2 − X 1 )
0
Z t Z t
1 2 1 2
= (X0 ∨ X0 ) + b(Xs ∨ Xs , s) ds + σ(Xx1 ∨ Xs2 , s) dBs + L0t (X 1 − X 2 ).
0 0
We therefore conclude that 1
(Xt ∨ Xt2 )t∈[0,∞)
is a solution of (7.23) if and only if L0t (X 2 − X 1 )
vanishes identically. This completes the proof.
We will now prove that uniqueness in law and the vanishing of the local times implies pathwise
uniqueness.
Proposition 7.19. Let (Bt )t∈[0,∞) be a one-dimensional Brownian motion and let b, σ : [0, ∞) ×
R → R be predictable processes. Suppose that we have uniqueness in law for the equation
(7.24) dXti = b(Xti ) dt + σ(Xti ) dBt ,
and that any two solutions (X 1 )t∈[0,∞) and (Xt2 )t∈[0,∞) that satisfy X01 = X02 almost everywhere
satisfy L0t (X 2 − X 1 ) = 0. Then, pathwise uniqueness holds for (7.24).
Proof. Suppose that (X 1 )t∈[0,∞) and (Xt2 )t∈[0,∞) are two solutions of (7.24) defined with respect
to the same Brownian motion that satisfy X01 = X02 almost everywhere. Then, since L0t (X 2 − X 1 )
vanishes identically, we have by Proposition 7.18 that X 1 ∨ X 2 is also a solution. Hence, by
uniqueness in law, it follows that the law of (Xt1 )t∈[0,∞) and the law of (Xt1 ∨ Xt2 )t∈[0,∞) are the
same, which occurs if and only if (Xt1 )t∈[0,∞) and (Xt2 )t∈[0,∞) are indistinguishable.
In the following to statements we are essentially imposing a Hölder continuity condition on σ.
That is, for ρ(x) = |x|α we are requiring that
|σ(x) − σ(y)| ≤ |x − y|α ,
or, equivalently, that
|σ(x) − σ(y)|2 ≤ |x − y|2α .
The integrability condition√requires that 2α ≥ 1 and therefore that α ≥ 1/2. So, in particular, the
theorem applies to σ(x) = x.
Proposition 7.20. Let ρ : [0, ∞) → [0, ∞) be a nondecreasing function that satisfies, for every
ε ∈ (0, 1),
Z ε
1
(7.25) ds = ∞.
0 ρ(s)
Let (Xt )t∈[0,∞) be a continuous semimartingale such that for some ε ∈ (0, 1), for every θ ∈ (0, ∞),
Z t
1
(7.26) 1{0<Xs ≤ε} dhXis < ∞.
0 ρ(X s)
Then L0t (X) = 0 almost surely.

Proof. By the occupation formula,
Z t Z ε
1 1 a
1{0<Xs ≤ε} dhXis = L da.
0 ρ(Xs ) 0 ρ(s) t
By the right continuity of Lat in a, it follows from assumption (7.25) that the righthand side of this
equality is infinite on the set {L0t > 0}. We therefore conclude by (7.26) that L0t = 0 almost surely,
for every t ∈ (0, ∞). Since L0t is continuous and increasing in t, this implies that L0t almost surely
vanishes identically.
Proposition 7.21. Let σ, b1 , b2 : [0, ∞)×R → R be a predictable processes. Assume that there exists
a nondecreasing ρ : [0, ∞) → [0, ∞) which satisfies (7.25) such that, for all (x, t), (y, t) ∈ R × [0, ∞),
|σ(x, t) − σ(y, t)|2 ≤ ρ(|x − y|).
Then, for a one-dimensional Brownian motion (Bt )t∈[0,∞) , if for each i ∈ {1, 2} the process
(Xti )t∈[0,∞) is a solution of
we have that L0t (X 1 − X 2 ) = 0 for every t ∈ [0, ∞).
Proof. Since for each i ∈ {1, 2} we have that

it follows that 2
dhX 1 − X 2 it = σ(Xt1 , t) − σ(Xt2 , t) dt ≤ ρ(Xt1 − Xt2 ) dt.

Therefore, for every t ∈ (0, ∞),

Z t Z t
1 1 2
1{X 1 >X 2 } dhX − X is ≤ 1{X 1 >X 2 } dt ≤ t.
0 ρ(Xs1 − Xx2 ) 0

We may now state the main theorem of this section.
Theorem 7.22. Let ρ : [0, ∞) → [0, ∞) be a nondecreasing function which satisfies (7.25), let
(Bt )t∈[0,∞) be a one-dimensional Brownian motion, and let σ, b : [0, ∞) × R → R be predictable
functions. Then, pathwise uniqueness for the equation
dXt = b(Xt , t) dt + σ(Xt , t) dBt ,
under the following two conditions.
(i) Both σ(t, x) = σ(x) and b(t, x) = b(x) are time-independent, |σ(x) − σ(y)|2 ≤ ρ(|x − y|)there
exists ε > 0 such that 0 < ε ≤ σ, and σ and b are bounded.
(ii) |σ(x, t) − σ(y, t)|2 ≤ ρ(|x − y|) and b is locally Lipschitz continuous.
Proof. In case (i), these conditions prove that the solution is unique in law and that the local time
at zero of the difference of any two solutions must vanish. Therefore, the solution is pathwise
unique. In case (ii), we may assume without loss of generality by a stopping time argument that
σ is bounded and that b is globally Lipschitz continuous. It then follows by Proposition 7.21
that, for any two solutions (Xt1 )t∈[0,∞) and (Xt2 )t∈[0,∞) defined with respect to the same Brownian
motion that satisfy X01 = X02 almost surely, we have almost surely that L0t (X 1 − X 2 ) = 0 for every
t ∈ [0, ∞). Therefore, by the Tanaka formula applied to Xt1 − Xt2 at a = 0,
Z t
1
Xt − Xt2 = sgn(Xt1 − Xt2 ) d(Xt1 − Xt2 ).

0
The boundedness of σ and the equation then prove that
Z t
1
Xt − Xt2 − (b(Xs1 , t) − b(Xs2 , s)) ds is a martingale.

0
Therefore, for every t ∈ [0, ∞), there exists c ∈ (0, ∞) depending on the Lipschitz constant of b
such that Z t
E[Xt1 − Xt2 ] ≤ c E[Xs1 − Xs2 ] ds.

0
It then follows by Grönwall’s inequality that E[Xt1 − Xt2 = 0 for every t ∈ [0, ∞). Therefore, by
continuity, we conclude that the processes (Xt1 )t∈[0,∞) and (Xt2 )t∈[0,∞) are indistinguishable. This
completes the proof.
8. Selected problems and solutions

In this section, I’ve included some selected problems and solutions.
(i) Let (Bt )t∈[0,∞) be a standard Brownian motion on a probability space (Ω, F, P). Prove using
the properties of Brownian motion that the following three processes are martingales.
(a) The process
(Bt )t∈[0,∞) .
(b) The process
Bt2 − t t∈[0,∞) .

(c) And for every α ∈ R, the process

α2 t

exp αBt − .
2 t∈[0,∞)
Proof. We first prove (a). To prove that integrability, we observe that B0 = 0 and, for every
t ∈ (0, ∞), !
|x| |x|2
Z
E[|Bt |] = 1 exp − dx < ∞.
R (2πt) 2 2t
To prove the martingale property, the independence of Brownian increments and properties
of the conditional expectation prove that, for every s ≤ t ∈ [0, ∞) that
E[Bt |Fs ] = E[Bt − Bs |Fs ] + E[Bs |Fs ] = E[Bt − Bs ] + Bs = Bs .
We will now prove (b). To prove the integrability, we have B0 = 0 and the integration by
parts formula proves that, for every t ∈ (0, ∞),
!
2 2
|x| |x|
Z
E[|Bt |2 ] = 1 exp − dx
R (2πt) 2 2t
!!
|x|2
Z
x d
= −t 1 exp − dx
(8.1) R (2πt) 2 dx 2t
!
|x|2
Z
− 21
= t (2πt) (exp − dx
R 2t
= t.
To prove that martingale property, it follows from (8.8) and properties of the conditional
expectation that that for every s ≤ t ∈ [0, ∞)¡
(8.2) E[Bt2 − t|Fs ] = E[Bt2 − Bs2 − t|Fs ] + E[Bs2 |Fs ] = E[(Bt − Bs )2 |Fs ] − t + Bs2 .
Since (Bt − Bs ) is a mean zero variance (t − s) normal random variable that is independent
of Fs , we have that
E[(Bt − Bs )2 |Fs ] = E[(Bt − Bs )2 ] = t − s.
Therefore, returning to (8.2),
E[Bt2 − t|Fs ] = Bs2 − t + (t − s) = Bs2 − s.
We will now prove (c). Let α ∈ R. It follows that exp(αB0 ) = 1 and that, for every t ∈ (0, ∞),
!
2
α2 t 2t
Z
1 α |x|
E exp αBt − = (2πt)− 2 exp αx − exp − dx < ∞.
2 R 2 2t
To prove the martingale property, properties of the conditional expectation and the indepen-
dence of Brownian increments prove that, for every s ≤ t ∈ [0, ∞),
α2 t α2 (t − s) α2 s

E exp αBt − |Fs = E exp α(Bt − Bs ) − |Fs exp αBs −
2 2 2
(8.3) 2
2

α (t − s) α s
= E exp α(Bt − Bs ) − exp αBs − .
2 2
Since (Bt − Bs ) is a normal random variable with mean zero and variance (t − s),
!
2
α2 (t − s) 2 (t − s)
Z
1 α |x|
E exp α(Bt − Bs ) − = (2π(t − s))− 2 exp αx − − dx
2 R 2 2(t − s)
!
|x − (t − s)α|2
Z
− 12
= (2π(t − s)) exp − dx
R 2(t − s)
!
|x|2
Z
− 12
= (2π(t − s)) exp − dx
R 2(t − s)
= 1.
Returning to (8.3), we conclude that, for every s ≤ t ∈ [0, ∞),
α2 t α2 s

E exp αBt − |Fs = exp αBs − ,
2 2
(ii) Let (Bt )t∈[0,∞) be a standard Brownian motion on R2 and for R ∈ (0, ∞) let BR denote the
ball of radius R centered at the origin. For every t ∈ (0, ∞), compute
P[Bt ∈ BR ],
and thereby prove that
/ B√2λt ) = e−λ ,
(i) P(Bt ∈
(ii) and that, for the Lebesgue measure |BR | of BR ,
P[Bt ∈ BR ] 1
lim = .
R→0 |BR | 2πt
What happens in dimension three?
Proof. We will first prove (a). We observe that using polar coordinates that, for every R ∈
(0, ∞) and t ∈ (0, ∞),
!
2
|x|
Z
P[Bt ∈ BR ] = (2πt)−1 exp − dx
BR 2t
Z R 2
r r
= exp − dr
(8.4) 0 t 2t
2 r=t
r
= − exp −
2t
r=0
R2

= 1 − exp − .
2t
So, for every R ∈ (0, ∞) and t ∈ (0, ∞),

R2

P[Bt ∈
/ BR ] = 1 − P[Bt ∈ BR ] = exp − .
2t
In particular, for every λ ∈ (0, ∞) and t ∈ (0, ∞),
/ B√2λt ] = exp (−λ) .
P[Bt ∈
We will no prove (b). For every t, R ∈ (0, ∞), it follows from (8.4) and Taylor expansion that
∞ 2 k
R2
X
k+1 1 R
P[Bt ∈ BR ] = 1 − exp − = (−1) .
2t k! 2t
k=1
Therefore,
∞
P[Bt ∈ BR ] 1X 1 R2k−2
= (−1)k+1 .
|BR | π k! (2t)k
k=1
Thus, as R → 0 the sum converges to its first term, which implies that
P[Bt ∈ BR ] 1
lim = .
R→0 |BR | 2πt
In three dimensions, polar coordinates and the integration by parts formula prove that, for
every t, R ∈ (0, ∞),
!
|x|2
Z
− 32
P[Bt ∈ BR ] = (2πt) exp − dx
BR 2t
Z R 2
− 32 2 r
= 4π (2πt) r exp − dx
0 2t
Z R 2
− 32 d r
= −4πt (2πt) r exp − dr
0 dr 2t
Z R
R2
2
4πt − 23 r
=− 3 R exp − + 4πt (2πt) exp − dr
(2πt) 2 2t 0 2t
Z R 2
R2

4πt r
= 3 exp − dr − R exp − .
(2πt) 2 0 2t 2t
This probability cannot be computed explicitly, so the best we can say is that for every
λ, t ∈ (0, ∞),
P[Bt ∈ / B√2λt ] = 1 − P[Bt ∈ B√2λt ].
More interesting is the limit. Since we have, for every R ∈ (0, ∞),
4
|BR | = πR3 ,
3
we Taylor expand to third order in R to see that
Z R 2 Z R
r2 R3

r
exp − dr = 1− + . . . dr = R − + ...,
0 2t 0 2t 6t
and that
R2 R3

R exp − =R− + ....
2t 2t
Hence, to third order in R,
R3 R3

4πt 1 4 3
P[Bt ∈ BR ] = 3 R− −R+ + ... = 3 · πR + . . . .
(2πt) 2 6t 2t (2πt) 2 3
Therefore,
P[Bt ∈ BR ] 1
lim = 3 .
R→0 |BR | (2πt) 2

(iii) Let (Bt )t∈[0,∞) be a standard Brownian motion on a probability space (Ω, F, P).
(i) Prove that Wt = tB 1 is a standard Brownian motion on (Ω, F, P).
t
(ii) For every a ∈ [0, ∞), let Ta denote the stopping time
Ta = inf{t ∈ [0, ∞) : Bt = a}.
Show using the reflection principle that
P(Ta < t) = 2P(Bt > a),
and thereby deduce that the probability density function fa of Ta on [0, ∞) is
2
a a
fa (t) = √ exp − .
2πt3 2t
(iii) For every a ∈ [0, ∞), define the random time
Sa = sup{t ∈ [0, ∞) : Bt = at}.
Is Sa a stopping time? Show that Sa = 1/Ta in distribution and find E[Sa ] and E[BBSa ].
Proof. We will first prove (i). We simply define W0 = 0, and we observe by continuity
of Brownian motion that, almost surely, Wt is continuous on (0, ∞). It remains to check
continuity at zero. For this, use the fact that the law of the iterated logarithm states that,
almost surely,
√
tB tB

|Bt | 1
t
1
t
lim sup p = lim sup p = lim sup q √ = 1.
t→∞ 2t log log(t) t→0 2 log log(1/t) t→0 2 t log log(1/t)
Since properties of the logarithm prove that the denominator of the penultimate term vanishes
in the limit t → 0, it follows almost surely that

0 ≤ lim sup tB 1 = 0.

t→0 t
Hence, almost surely,

lim Wt = lim tB 1 = 0 = W0 ,
t→0+ t→0+ t
which proves that (Wt )t∈[0,∞) is almost surely continuous. For the remainder of the proof,
we will use that fact that Brownian motion is the unique Gaussian process with covariance
function C(s, t) = s ∧ t. That is, we must prove that, for every s ≤ t ∈ [0, ∞),
E[Ws Wt ] = s ∧ t.
For this, since s ≤ t implies that 1/t ≤ 1/s,

1 1 st
E[Ws Wt ] = stE[B 1 B 1 ] = st ∧ = = s = (s ∧ t).
s t s t t
We will now prove (ii). Let a ∈ (0, ∞). Since the events
{Ta < t} = { sup Bs < t},
s∈[0,t)
the reflection principle proves that

P[Ta < t] = P[ sup Bs < t] = 2P[Bt > a].
s∈[0,t)
Therefore, if a probability density function fa of Ta exists, then for every s ≤ t ∈ [0, ∞) we

would have that
Z t
fa (r) dr = P[s ≤ Ta < t]
s
= P[Ta < t] − P[Ta < s]
= 2P[Bt > a] − 2P[Bs > a]
! !!
|x|2 |x|2
Z ∞
− 12 − 12
= (2πt) exp − − (2πs) exp − dx.
a 2t 2s
And, by Lebesgue’s differentiation theorem, for almost every t ∈ R,

Z t
1
fa (t) = lim fa (r) dr
s→t− t − s s
! !!
Z ∞ 2 2
1 1 |x| 1 |x|
= lim (2πt)− 2 exp − − (2πs)− 2 exp − dx
s→t− t − s a 2t 2s
!!
|x|2
Z ∞
d − 12
= (2πt) exp − dx.
a dt 2t
This implies that

! !!
∞
|x|2 x2 |x|2
Z
− 12 1 3
fa (t) = (2π) − t− 2 exp − + 5 exp − dx
a 2 2t 2t 2 2
!!
∞
|x|2
Z
3 − 21 d
= −(2πt ) x exp − dx.
a dx 2t
2

a a
=√ exp − .
2πt3 2t
Then, working backward, we conclude that fa is the probability density function of Ta .
We will now prove (iii). First observe by the law of the iterated logarithm that Sa is almost
surely finite. The random time Sa is not a stopping time with respect to the filtration defining
the Brownian motion (Bt )t∈[0,∞) , since it is necessary to look arbitrarily far into the future in
order to determine the event {Sa ≤ t}, for some t ∈ (0, ∞). However, the random time Sa will
be a stopping time with respect to the filtration defining the Brownian motion (Wt )t∈[0,∞) ,
since, for every t ∈ (0, ∞),
Sa = sup{t ∈ [0, ∞) : Bt = at} = inf{t ∈ [0, ∞) : B1/t = a/t} = inf{t ∈ [0, ∞) : tB1/t = a}.
That is,
Sa = inf{t ∈ [0, ∞) : Wt = a}.
To prove that Sa has the same distribution as 1/Ta , observe that since (Bt )t∈[0,∞) and (Wt )[t∈[0,∞)
have the same distributions, for every t ∈ [0, ∞),

P[Sa ≥ t] = P sup(Bs − as) ≥ 0 = P sup(Ws − as) ≥ 0 = P sup(sB1/s − as) ≥ 0 .
s≥t s≥t s≥t
Therefore,

P[Sa ≥ t] = P sup(B1/s − a) ≥ 0 = P inf (Bs − a) ≥ 0 = P[Ta ≤ 1/t] = P[1/Ta ≥ t],
s≥t s≤1/t
which proves that Sa and 1/Ta have the same distribution. Therefore,
E[Sa ] = E[1/Ta ]
Z ∞
1 1
= fa ( /t) dt
0 t
Z ∞ 2
a − 52 a
=√ t exp − dt
2π 0 2t
Z ∞
2 1
= 1 t 2 exp(−t) dt,
2
π2a 0
2
where the final equality follows from the change of variables t̃ = a2t . The final integral is the
Γ-function (look up this value on Wikipedia, for instance), for which we have
1
2 2 π2 1
E[Sa ] = 1 Γ(3/2) = 1 · = 2.
π 2 a2 π 2 a2 2 a
Then, by definition,
1
E[BSa ] = aE[Sa ] = ,
a
(iv) A real-valued centered Gaussian process is a real valued process (Xt )t∈[0,∞) on a probability
space (Ω, F, P) with finite dimensional distributions that are normally distributed and mean
zero. A centered Gaussian process (Xt )t∈[0,∞) is a fractional Brownian motion with Hurst
parameter h ∈ (0, 1) if P[X0 = 0] = 1 and if, for every s, t ∈ [0, ∞),
1 2h
E[Xt Xs ] = t + s2h − |t − s|2h .
2
(a) Show that, for every s ≤ t ∈ [0, ∞), the increment Xt − Xs has mean zero and variance
|t − s|2h .
(b) Show that for every p ∈ (0, ∞) there exists cp ∈ (0, ∞) such that, for every s ≤ t ∈ [0, ∞),
E[|Xt − Xs |p ] = cp |t − s|hp .
(c) Show that fractional Brownian motion has a continuous modifcation, and determine its
Hölder exponent. (Hint: Use Kolmogorov’s continuity criterion.)
(d) For each h ∈ (0, 1), determine the value p ∈ (0, ∞) so that fractional Brownian motion
with Hurst parameter h has finite and nonzero p-variation. Is fractional Brownian motion
a semimartingale for any h ∈ (0, 1)?
Proof. We will first prove (a). Let h ∈ (0, 1). It follows by assumption that, since the finite
dimensional distributions are normally distributed with mean zero, for every s ≤ t ∈ [0, ∞),
E[Xt − Xs ] = 0.
For the variance, for every s ≤ t ∈ [0, ∞),

E[|Xt − Xs |2 ] = E[Xt2 ] + E[Xs2 ] − 2E[Xs Xt ]
= t2h + s2h − (t2h + s2h − |t − s|2h )
= |t − s|2h .
We will now prove (b). Let h ∈ (0, 1). Let s ≤ t ∈ [0, ∞). Since (Xt − Xs ) is normally
distributed with mean zero and variance |t − s|2h , it follows for every p ∈ (0, ∞) that
!
|x|2
Z
p p 2h − 12
E[|Xt − Xs | ] = |x| (2π |t − s| ) exp − dx
R 2 |t − s|2h
!
2
|x|
Z
1
= |t − s|hp |x|p (2π)− 2 exp − dx,
R 2
where the final equality follows from the change of variables x̃ = x/|t−s|h . Therefore, for every
p ∈ (0, ∞), after defining
!
2
|x|
Z
1
cp = |x|p (2π)− 2 exp − dx,
R 2
it follows for every p ∈ (0, ∞) that

E[|Xt − Xs |p ] = cp |t − s|hp .
We will now prove (c). Let h ∈ (0, 1). The Kolmogorov continuity criterion states that, since
for every p ∈ (0, ∞),
E[|Xt − Xs |p ] = cp |t − s|hp ,
fractional Brownian motion has a Hölder continuous modification with Hölder exponent γ ∈
(0, h − p1 ) for every p ∈ (0, ∞). That is, fractional Brownian motion is almost surely Hölder
continuous for any Hölder exponent γ ∈ (0, h). We will now prove (d). We will first identify
the scaling properties of fractional Brownian motion. Let h ∈ (0, 1) and λ ∈ (0, ∞). We aim
to identify α ∈ R, such that the process (λXλα s )s∈[0,∞) is a fractional Brownian motion. Since
this process is a centered Gaussian process with mean zero, it is sufficient to identify α ∈ R
such that the covariance satisfies
E[λXλα t λXλα s ] = t2h + s2h − |t − s|2h .
For this, notice that

E[λXλα t λXλα s ] = λ2 E[Xλα t Xλα s ] = λ2−2αh t2h + s2h − |t − s|2h .
We therefore require that λ2−2αh = 1, which implies that α = −1/h. We conclude that,
for every α ∈ (0, ∞), if (Xt )t∈[0,∞) is a fractional Brownian motion with Hurst parameter
h ∈ (0, ∞) then (λXλ−1/h t )t∈[0,∞) is also a fractional Brownian motion with the same Hurst
parameter.
We will now identify p ∈ [1, ∞) depending on h ∈ (0, 1) such that fractional Brownian
motion with Hurst parameter h ∈ (0, 1) has finite p-variation. By this we mean that, if
{∆n = {0 = t0 < t1 < . . . < tkn−1 < tkn = 1}}n∈N is a sequence of partitions of [0, 1] such
that |∆n | → 0 as n → ∞, then
Bt − Bt p exists and is finite almost surely.
X
lim k k−1
k→∞
∆n
For simplicity, we consider the sequence of partitions

{∆n = 0 < 1/2n < 2/2n < . . . < 2n −1/2n < 1}n∈N ,
where, for every n ∈ N, the scaling properties of fractional Brownian motion prove that in
law, for every p ∈ [1, ∞),
n
2 2n 2n
X p X p X
−nph nh −nph
(8.5) B − B k−1 = 2 2 B − B k−1 =2 |Bk − Bk−1 | .

k k
2n n2 2n 2n
k=1 k=1 k=1
The sequence {Bk − Bk−1 }k∈N is a stationary and ergodic sequence. This means that the
random variables {Bk − Bk−1 }k∈N are identically distributed, so that in particular we have,
for every n ∈ N, since B0 = 0,
" 2n #
X p
(8.6) B kn − B k−1 = 2−nph · 2n E[|B1 |] = 2n(1−ph) E[|B1 |].

E n
2 2
k=1
(For the problem sheet, observing that the above equality implies that, if the p-variation is
finite and nonzero, then the above inequality implies that p = 1/h is sufficient.) The ergodicity
states that, for the measure W h induced the space of continuous paths beginning from zero
C0 ([0, ∞)) by fractional Brownian motion, the only measurable subsets A ⊆ C0 ([0, ∞)) left
invariant by the shift operators {τk }k∈N0 defined by
τk (σ)(t) = σ(k + t) − σ(k) for every t ∈ [0, ∞),
have measure zero or measure one. The ergodic theorem and the estimate (8.6) prove that,
almost surely and in L1 (Ω), as n → ∞,
2n
X
−n
2 |Bk − Bk−1 | → E[|X1 |].
k=1
Therefore, returning to (8.5), it follows that in probability

0 if ph > 1,

n
2 p 
1
X 
lim B kn − B k−1 = E[|X1 |] if p = ,

n→∞
k=1
2 2n 
 h
∞ if ph < 1.

In particular, because semi-martingales have finite quadratic variation, it follows that frac-
tional Brownian motion is a semi-martingale if and only if h = 21 . That is, if and only if the
fractional Brownian motion is a Brownian motion.
(v) Let (Mt )t∈[0,∞) be a continuous local martingale with M0 = 0.

(a) Prove that if E[hM i∞ ] < ∞ then M is L2 -bounded in the sense that
sup E[Mt2 ] < ∞.
t∈[0,∞)
In this case, conclude that the process Mt2 − hM it t∈[0,∞) is uniformly integrable.

(b) Prove that M converges almost surely as t → ∞ on the set {hM i∞ < ∞}.
Proof. We will first prove (a). Because (Mt )t∈[0,∞) is a continuous local martingale, there exist
an increasing sequence of stopping times {τn }n∈N satisfying almost surely that τn → ∞ as
n → ∞. For every n ∈ N, since by assumption M τn = (Mt∧τn 1{τn >0} )t∈[0,∞) is a martingale,

and since
hM τn i = hM iτn 1{τn >0} ,
it follows that ((M τn )2t − hM τn it )t∈[0,∞) , is a martingale. The martingale property, M0 = 0,
and the fact that the bracket is an increasing function therefore imply that, for every n ∈ N
and t ∈ [0, ∞),
E[Mτ2n ∧t ] = E[hM τn it ] ≤ E[hM it ] ≤ E[hM i∞ ].
We then pass to the limit using Fatou’s lemma. Indeed, since the stopping times diverge to
infinity almost surely, we have that, as n → ∞, for every t ∈ [0, ∞),
Mτn ∧t → Mt almost surely,
and therefore, for every t ∈ [0, ∞),
E[Mt2 ] = E[ lim Mτ2n ∧t ] ≤ lim inf E[Mτ2n ∧t ] < E[hM i∞ ].
n→∞ n→∞
Since (Mt )t∈[0,∞) is a uniformly bounded L2 (Ω)

martingale, we conclude that (Mt2 −hM it )t∈[0,∞)
is a martingale. The martingale convergence theorem proves that there exists M∞ ∈ L2 (Ω)
such that, as t → ∞,
This implies by Hölder’s inequality that
Mt2 → M∞ 2
almost surely and strongly in L1 (Ω),
2
2 = |(M − M )| |(M + M )|. Therefore, since the bracket pro-
using the equality Mt − M∞ t ∞ t ∞
cess is an increasing function, the dominated convergence theorem with dominating function
hM∞ i∞ proves that, as t → ∞,
hM it → hM i∞ almost surely and strongly in L1 (Ω).
In combination, this implies that, as t → ∞,
Mt2 − hM it → M∞
2
− hM∞ i∞ almost surely and strongly in L1 (Ω).
The equivalent properties of the martingale convergence theorem then prove that (Mt2 −
hM it )t∈[0,∞) is uniformly integrable.
We will now prove (b). For every n ∈ N define the stopping time
σn = inf{t ∈ [0, ∞) : hM it ≥ n}.
For every n ∈ N, consider the stopped local martingale (Mσn ∧t )t∈[0,∞) . It follows by definition
of the {σn }n∈N that, for every n ∈ N,
E[hM σn i∞ ] ≤ n.
Therefore, part (a) proves that there exists M∞,n ∈ L2 (Ω) such that, as t → ∞,
Mσn ∧t → M∞,n almost surely and strongly in L2 (Ω).
Observe that for every n ≤ m ∈ N, it follows that
(8.7) M∞,n = M∞,m on the set {hM i∞ < n}.
We therefore define
(
M∞,n (ω) if ω ∈ {hM i∞ < n} for some n ∈ N,
M∞ (ω) =
0 if ω ∈ {hM i∞ = ∞}.
It then follows from (8.7) that, for almost every ω ∈ {hM i∞ < ∞}, as t → ∞,
Mt (ω) → M∞ (ω),
(vi) Suppose that (Mt )t∈[0,∞) is a bounded continuous martingale with finite variation. Prove that
Z t
2 2
Mt = M0 + 2 Ms dMs ,
0
where the final integral is almost surely well-defined. (Hint: Cite here known results about
one-dimensional functions of bounded variation.) Deduce that M is almost surely constant.
Proof. Since the process (Mt )t∈[0,∞) is bounded with finite variation, it follows that the process
(Mt2 )t∈[0,∞) is bounded with finite variation since, for every T ∈ (0, ∞), for every partition
0 = t0 < t1 < . . . < tN −1 < tN = T ,
N
X X N
2 2

Mtk − Mtk−1 = Mt − Mt Mt + Mt ≤ sup kMt k Var(M, T ).

k k−1 k k−1 ∞
k=1 k=1 t∈[0,T ]
Similarly, the process (M̃t2 = Mt2 − M02 )t∈[0,∞) is of finite variation. Therefore, since M̃0 = 0,
almost surely, for every t ∈ (0, ∞),
Z t Z t
d 2
M̃t2 = M̃r dr = 2 M̃r dMr ,
0 dr 0
where the fact that (Mt )t∈[0,∞) is bounded with finite variation implies that, almost surely,
for partitions ∆ = {0 = t0 < t1 < . . . < tN (∆)−1 < tN (∆) = t} of [0, T ], for the mesh
|∆| = supk∈{1,...,N (∆)} |tk − tk−1 |,
Z t N (∆)
X
M̃r dMr = lim M̃tk−1 Mtk − Mtk−1 .
0 |∆|→0
k=1
For every partition ∆ of [0, T ], the martingale property proves that
 
N (∆) N (∆) h h ii
X X
E  M̃tk−1 Mtk − Mtk−1 =  E E M̃tk−1 Mtk − Mtk−1 |Ftk−1 = 0.
k=1 k=1
We therefore conclude that, for every t ∈ [0, ∞),

E[M̃t2 ] = 0,
and therefore by continuity that (Mt )t∈[0,∞) is almost surely constant.
(vii) We say that a function u : R2 → R is C2 -bounded if u is twice-differentiable and satisfies

sup |u(x)| + |∇u(x)| + ∇2 u(x) < ∞.

x∈R2
Prove that every C2 -bounded function u satisfying

∂2u ∂2u
∆u = + = 0,
∂2x ∂2y
is constant. (Hint: Apply Itô’s formula to the composition (u(Bt ))t∈[0,∞) , for (Bt )t∈[0,∞)
a standard two-dimensional Brownian motion.) Conclude that every bounded holomorphic
function f : C → C is constant.
Proof. Let (Bt )t∈[0,∞) be a standard two-dimensional Brownian motion. Itô’s formula proves
that, for every t ∈ [0, ∞),
2 Z t 2 Z
X ∂u i 1 X t ∂2u
u(Bt ) = u(0) + (Bs ) dBs + (Bs ) dhB i , B j is
i=1 0 ∂xi 2 0 ∂xi ∂xj
i,j=1
2 Z t Z t
X ∂u 1
= u(0) + (Bs ) dBsi + ∆u(Bs ) ds.
0 ∂xi 2 0
i=1
Therefore, since u is harmonic,
2 Z 1
X ∂u
u(Bt ) = u(0) + (Bs ) dBsi .
0 ∂xi
i=1
Since u is C2 -bounded it follows that that (u(Bt ))t∈[0,∞) is a martingale, since the stochastic
integrals are martingales. Furthermore, because u is bounded, the martigale (u(Bt ))t∈[0,∞) is
bounded and hence L2 -bounded and hence uniformly integrable. The martingale convergence
theorem then proves that there exists M∞ ∈ L∞ (Ω, F∞ , P) such that, almost surely as t → ∞,
(8.8) u(Bt ) → M∞ .
To conclude, we use the fact that Brownian motion in two dimensions is recurrent. That is,
almost surely, for every non-empty open set A ⊆ R2 ,
sup{t ∈ [0, ∞) : Bt ∈ A} = ∞.
By contradiction, suppose that u is not constant. Then, there exist non-empty open subsets
A1 , A2 ⊆ R2 such that
sup (u(x)) < inf (u(x)).
x∈A1 x∈A2
Since Brownian motion is recurrent, this implies almost surely that

lim inf u(Bt ) ≤ sup (u(x)) < inf (u(x)) ≤ lim sup u(Bt ).
t→∞ x∈A1 x∈A2 t→∞
and hence almost surely that

lim u(Bt ) does not exist,
t→∞
which contradicts (8.8). We therefore conclude that u must be constant. For the final state-
ment, we recall that bounded holomorphic functions are C2 -bounded (in fact, they are Ck -
bounded for every k ∈ N) in the sense that both their real and imaginary parts are Ck -bounded
for every k ∈ N. Furthermore, the Cauchy-Schwarz equations prove that the real and imagi-
nary parts are holomorphic, and therefore by the above they are constant.
(viii) Let (Mt )t∈[0,∞) be a continuous local martingale vanishing at zero.

(a) Show that the intervals of constancy of the maps t 7→ Mt and t 7→ hM it coincide almost
surely.
(b) Show that if, for every ξ ∈ R, for every s ≤ t ∈ [0, ∞),
2
ξ (t − s)
E[exp(iξ(Mt − Ms ))|Fs ] = exp − ,
2
then (Mt )t∈[0,∞) is a Brownian motion.
Proof. We will first prove (a). First observe that a continuous martingale (Mt )t∈[0,∞) is
constant if and only if its quadratic variation is zero. If (Mt )t∈[0,∞) is constant, then its
quadratic variation is zero. Conversely, if the quadratic variation is zero, for every t ∈ [0, ∞),
E[Mt2 ] = E[M02 ] and therefore E[(Mt + M0 )(Mt − M0 )] = 0,
which implies by continuity that for almost every ω ∈ Ω we have Mt (ω) = ±M0 (ω) for every
t ∈ [0, ∞). By continuity this implies almost surely that Mt (ω) = M0 (ω).
Let t ∈ [0, ∞) be arbitrary and let
n o
∆n = {0 = tn0 < tn1 < . . . < tnN (n)−1 < tnN (n) = t} ,
n∈N
be a sequence of partitions of [0, T ] which satisfy |∆n | → 0 as n → ∞. For every n ∈ N define

the process (Ts∆n (M ))s∈[0,∞) by
N (n)−1
X
Ts∆n (M ) = (Mtnk+1 ∧s − Mtnk ∧s )2 .
k=0
Then we know that, as n → ∞,

sup Ts∆n (M ) − hM is → 0 in proability.

s∈[0,T ]
This implies that there exists subsequence {nk }k∈N satisfying nk → ∞ as k → ∞ such that,
as k → ∞,
∆n
sup Ts k (M ) − hM is → 0 almost surely.

s∈[0,T ]
This proves that, on a subset of full probability, if Mt (ω) = Mb (ω) for every t ∈ [a, b] then
hM it (ω) = hM ib (ω) for every t ∈ [a, b].
To prove the converse, let q ∈ Q ∩ [0, ∞) and define the translated Ft+q -martingale
(Nt )t∈[0,∞) by
Nt = Mt+q − Mq .
Then, define the stopping time
T = inf{s ∈ [0, ∞) : hN is > 0},
which is a stopping time because the filtration is right continuous. Since the quadratic
variation of (NtT )t∈[0,∞) is identically zero, by the above we conclude almost surely that
NT ∧t (ω) = MT ∧(t+q) (ω) − Mq (ω) = 0 for every t ∈ [0, ∞). Since we have
hN T it = hM iT ∧(t+q) − hM iq ,
we conclude that there exists a subset Ωq ⊂ Ω of full probability such that, if hM it (ω) =
hM iq (ω) for every t ∈ [q, b], for some b ∈ (q, ∞), then Mt (ω) = Mq (ω) for every t ∈ [q, b].
Define the subset of full probability
Ω0 = ∩q∈Q∩[0,∞) Ωq .
It then follows by density of the rational that, for every ω ∈ Ω0 , if hM it (ω) = hM ib (ω) for
every t ∈ [a, b], for some a ≤ b ∈ [0, ∞), then Mt (ω) = Mb (ω) for every t ∈ [a, b]. This
completes the proof of (a).
For (b), let s ≤ t ∈ [0, ∞). We first observe by taking the expectation that, for every ξ ∈ R,
2
ξ (t − s)
E[exp(iξ(Mt − Ms ))] = exp − .
2
This proves that (Mt − Ms ) is a normal random variable with mean zero and variance (t − s).
Since M0 = 0 and since (Mt )t∈[0,∞) is almost surely continuous by assumption, it remains
only to prove that (Mt − Ms ) is independent of Fs .
Let A ∈ Fs satisfy P(A) > 0. Then, since
2
ξ (t − s)
E[exp(iξ(Mt − Ms ))|Fs ] = exp − ,
2
it follows that 2
ξ (t − s)
E[exp(iξ(Mt − Ms ))1A ] = P[A] exp − .
2
In particular, if we define the measure PA (B) = P(A∩B)/P(A), we have that
2
E[exp(iξ(Mt − Ms ))1A ] ξ (t − s)
EPA [exp(iξ(Mt − Ms ))] = = exp − ,
P(A) 2
which implies that with respect to the measure PA the random variable (Mt − Ms ) is normally
distributed with mean zero and variance (t−s). Therefore, we conclude that for every bounded
measurable function g : R → R,
E[g(Mt − Ms )1A ]
EPA [g(Mt − Ms )] = = E[g(Mt − Ms )].
P(A)
That is, we have
E[g(Mt − Ms )1A ] = P(A)E[g(Mt − Ms )].
By choosing g = 1{x≤α} for α ∈ R, this implies that
P [{(Mt − Ms ) ≤ α} ∩ A] = P[A]P[(Mt − Ms ) ≤ α].
Since sets of the form {{(Mt − Ms ) ≤ α}}α∈R form a π-system, this equality proves that, for
every B ∈ σ(Mt − Ms ),
P[B ∩ A] = P(A)P(B).
Hence, since A ∈ Fs was arbitrary, we conclude that (Mt − Ms ) is independent of Fs . This
(ix) Let (Bt = (Bt1 , . . . , Btd ))t∈[0,∞) be a standard d-dimensional Brownian motion. Let (Ft =
(Ft1 , . . . , Ftd ))t∈[0,∞) be a continuous, adapted, d-dimensional stochastic process that satisfies,
for every i ∈ {1, . . . , d}, for every t ∈ (0, ∞),
Z t
i 2
Fs ds < ∞.
0
(a) Prove that, for every i, j ∈ {1, . . . , d}, for every t ∈ (0, ∞),
t if i = j,

i j
hB , B it = δij t =
0 if i 6= j.
(b) Prove that, for every i, j ∈ {1, . . . , d}, for every t ∈ (0, ∞),
Z · Z · Z t
i i j j
h Fs dBs , Fs dBs it = δij Fsi Fsj ds.
0 0 0
(c) Prove that the process (Xt )t∈[0,∞) defined by
d Z t
!2 d Z
X X t 2
i i
Xt = Fs dBs − Fsi ds,
i=1 0 i=1 0
is a martingale.
(d) Prove that, for every λ, t ∈ (0, ∞),

" d Z ! # d Z t
X s X
sup Fr dBr ≥ λ ≤ 4λ−2
i i
E[(Fti )2 ] ds.

P
s∈[0,t]
i=10 0 i=1
Proof. We first prove (a). We will use the defining property of the quadratic variation. On
the previous problem set, we proved that, for every i ∈ {1, . . . , d},
((Bti )2 − t)t∈[0,∞) is a martingale.
Therefore hB i it = t because it is the unique increasing process for which ((B i )2t − t)t∈[0,∞) is
a martingale. Similarly, for every i 6= j ∈ {1, . . . , d}, for every s ≤ t ∈ [0, ∞),
E[Bti Btj |Fs ] = E[(Bti − Bsi )(Btj − Bsj )|Fs ] + E[Bti |Fs ]Bsj + E[Btj |Fs ]Bsi − Bsi Bsj .
Since the increments of (Bt )t∈[0,∞) are independent, and since the individual components
{(Bti )t∈[0,∞) }i∈{1,...,d} are independent, the martingale property proves that
E[Bti Btj |Fs ] = E[(Bti − Bsi )(Btj − Bsj )] + Bsi Bsj + Bsj Bsi − Bsi Bsj
= E[(Bti − Bsi )]E[(Btj − Bsj )] + Bsi Bsj
= Bsi Bsj .
Therefore, if i 6= j then (Bti Btj )t∈[0,∞) is a martingale, which proves that hB i , B j it = 0. This
completes the proof of (a).
We now prove (b). This follows immediately from the defining property of the stochastic
integral. Namely that for any local martingales (Mt )t∈[0,∞) and (Nt )t∈[0,∞) , for any F ∈ L2 (M )
and G ∈ L2 (N ), and for any s ≤ t ∈ [0, ∞),
Z t Z t Z t
E Fr dMr Gr dNr = E Fr Gr dhM, N ir .
s 0 s
We now prove part (c). Let (X̃t )t∈[0,∞) be the process defined by
d Z
X t
X̃t = Fsi dBsi .
i=1 0
Then, (X̃t )t∈[0,∞) is a martingale, and the linearity of the quadratic variation and part (b)
prove that, for every t ∈ [0, ∞),
d
X Z · Z · d Z t
X
i i j j
hX̃it = h Fr dBr , Fr dBr it = (Fri )2 dr.
i,j=1 0 0 i=1 0
We therefore conclude using the integrability assumption on the {(Fsi )s∈[0,∞) }i∈{1,...,d} that

X̃ 2 − hX̃it
t∈[0,∞)
is a martingale, which completes the proof of (c).

We finally prove (d). Let λ, t ∈ (0, ∞). It follows from Chebyshev’s inequality that
!  !2 
d Z s d Z s
" #
X X
sup Fri dBri ≥ λ = P  sup Fri dBri ≥ λ2 

P
s∈[0,t] i=1 0 s∈[0,t] i=1 0
 d Z !2 
X s
≤ λ−2 E  sup Fri dBri 

s∈[0,t] i=1 0

d Z s
!2 
X
= λ−2 E  sup Fri dBri 
s∈[0,t] i=1 0
Jensen’s inequality, the Itô isometry, and the integrability assumptions on (Ft )t∈[0,∞) prove
that 
d Z t
!2 
X
 Fri dBri  is a submartingale.
i=1 0
t∈[0,∞)
Therefore, Doob’s inequality, the martingale property, and part (c) prove that
" d Z ! # 
d Z t
!2 
X s X
sup Fri dBri ≥ λ ≤ 4λ−2 E  Fri dBri 

P
s∈[0,t]
i=10 0 i=1
d Z
X t
= 4λ−2 E (Fri )2 dr,

i=1 0
which completes the proof of (d).
(x) Let (Bt )t∈[0,∞) be a standard Brownian motion on a filtered probability space
(Ω, G, (Gt )t∈[0,∞) , P).
Let X be a finite G0 -measurable positive random variable that is independent of the Brownian
motion. Let (Mt = BtX )t∈[0,∞) and define the filtration (Ft )t∈[0,∞) by
Ft = σ(BsX : s ∈ [0, t]).
(a) Show that M is a local martingale with respect to (Ft )t∈[0,∞) .
√
(b) Show that M is a martingale if and only if E[ X] < ∞.
(c) Calculate (hM it )t∈[0,∞) .
(d) Let (At )t∈[0,∞) be an increasing process vanishing at zero that is independent of (Bt )t∈[0,∞) .
Define the filtration (FtA )t∈[0,∞) by
FtA = σ(BAs : s ∈ [0, t]).
Show that (BAt )t∈[0,∞) is a local FtA -martingale, find conditions that guarantee that
(BAt )t∈[0,∞) is a FtA -martingale, and compute its quadratic variation process.
Pn
Proof. We will first prove (a). It suffices to prove the case that X = i=1 αi 1Ai is a bounded
simple function. Then, for every s ≤ t ∈ [0, ∞),
n
X
E[BtX |Fs ] = E[Btαi 1Ai |Fs ].
i=1
We observe that by the independence of X and (Bt )t∈[0,∞) the process (Bαi t 1Ai )t∈[0,∞) is a
Brownian motion scaled by αi in time with respect to the measure PAi (B) = P(Ai ∩B)/P(Ai ) and
with respect to the filtration Fti = {Ai ∩ B : B ∈ Ft }. Therefore,
n
X
E[BtX |Fs ] = E[Bsαi 1Ai |Fs ] = E[BsX : Fs ] = BsX .
i=1
This completes the proof that (BtX )t∈[0,∞) is a Ft -martingale. In general, (BtX )t∈[0,∞) is only
a local martingale due to the fact that X may not be bounded.
We will now prove (b). In this case, the issue is one of integrability. Indeed, for every
t ∈ (0, ∞) we have by independence, the nonnegativity of X, and Brownian scaling that
Z Z
1 √
E[|BtX |] = E[Btα ]µX ( dα) = E[|Bt |] α 2 µX ( dα) = E[|Bt |]E[ X].
R R
Hence, we see that the process (BtX )t∈[0,∞) is integrable and a martingale if and only if
√
E[ X] < ∞. This completes the proof of (b).
We will now prove (c). It suffices to prove the case that X = ni=1 αi 1Ai is a bounded
P
simple function. In this case, since the conditioned process is a Brownian motion,
Xn
2
E[BtX |Fs ] = E[(Btαi )2 1Ai |Fs ]
i=1
n
X
E[ (Bsαi )2 + αi (t − s) 1Ai |Fs ]

=
i=1
2
= E[BsX − (t − s)X|Fs ]
2
= BsX − sX + tX.
We therefore have that
2
(BtX − tX)t∈[0,∞) is a martingale.
This implies that hB·X it = tX, which completes the proof of (c).
We will now prove (d). The identical arguments apply in this case, where previously we
considered the case As = sX. Indeed, if for every n ∈ N we define
Tn = inf{t ∈ [0, ∞) : BAt ≥ n},
then by a repetition of the arguments from part (a), replacing everywhere sX with As , it
follows that (BAt ∧Tn )t∈[0,∞) is a martingale and hence that (BAt )t∈[0,∞) is a local martingale.
For every t ∈ [0, ∞), the independence proves that
h 1 i
/2
E[BAt ] = E At E[|B1 |],
1/2
and therefore that (BAt )t∈[0,∞) is a martingale if E[At ] < ∞ for every t ∈ [0, ∞). Finally, in
this case, the quadratic variation process is given by At . This completes the proof.
(xi) Let (Bt )t∈[0,∞) , (Wt )t∈[0,∞) be two independent standard Brownian motions. Find the sto-
chastic differential equations satisfied by the following processes (Xt )t∈[0,∞) , and determine
which are martingales.
(a) Xt = exp( 2t ) cos(Bt )
(b) Xt = tBt
(c) Xt = (Bt + t) exp(−Bt − 2t )
(d) Xt = (Bt )2 + (Wt )2
Proof. For (a), the integration-by-parts formula and Itô’s formula provee that, since the pro-
cess (exp(t/2))t∈[0,∞) is of bounded variation,

1 t 1 t
dXt = Xt dt − exp sin(Bt ) dBt − Xt dt = − exp sin(Bt ) dBt .
2 2 2 2
We conclude by the boundedness of the sin function that (Xt )t∈[0,∞) is a martingale.
For (b), the integration-by-parts formula and Itô’s formula prove that, since the process
(t)t∈[0,∞) is of bounded variation,
dXt = Bt dt + t dBt .
In this case, (Xt )t∈[0,∞) is a semimartingale but not a martingale.
For (c), we write the process (Xt )t∈[0,∞) in the form

t t
Xt = (Bt exp(−Bt )) exp − + exp(−Bt ) t exp − .
2 2
The second term in each product on the righthand side is a process of bounded variation.
Therefore, the integration-by-parts formula and Ito’s formula prove that

t t
dXt = exp −Bt − − Bt exp −Bt − dBt
2 2

1 t t
+ Bt exp −Bt − − 2 exp −Bt − dt
2 2 2

1 t
− Bt exp −Bt − dt
2 2

t t t
− t exp −Bt − dBt + exp −Bt − dt
2 2 2

t t t
+ exp −Bt − − exp −Bt − dt.
2 2 2
Therefore,
t
dXt = (1 − t − Bt ) exp −Bt − dBt ,
2
and (Xt )t∈[0,∞) is a martingale.
For (d), it follows from Itô’s formula that
dXt = Bs dBs + Ws dWs + 2 ds.
In this case (Xt )t∈[0,∞) is a submartingale but not a martingale.
(xii) Let (Bt )t∈[0,∞) be a standard d-dimensional Brownian motion. Let (Xt )t∈[0,∞) be the process
q
Xt = kBt k = (Bt1 )2 + . . . + (Btd )2 .
(a) Find the SDE satisfied by (Xt )t∈[0,∞) and show that
Z t
d−1
Xt = X0 + ds + Wt ,
0 2Xs
where (Wt )t∈[0,∞) is standard one-dimensional Brownian motion.
(b) Let βk (t) = E[|Xt |2k ] for every k ∈ N0 and t ∈ (0, ∞). Prove that
Z t
βk (t) = k(2(k − 1) + d) βk−1 (s) ds.
0
(c) Calculate E[kBt k4 ] = E[kBt k6 ].
Proof. We will first prove (a). Since X0 = 0, Itô’s formula, the fact that hB i , B j it = δi,j t, and
the definition of (Xt )t∈[0,∞) prove that
d Z t d Z t
X Bsi X 1 (B i )2
Xt = i
dBs + − s 3 ds
Xs 2Xs 2Xs
i=1 0 i=1 0
d Z t Z t
X Bsi d−1
= dBsi + ds.
0 X s 0 2Xs
i=1
Define the process (Wt )t∈[0,∞) by

d Z t
X Bsi
Wt = dBsi .
0 Xs
i=1
Since for every i ∈ {1, . . . , d}

i
B
0 ≤ s ≤ 1,
Xs
it follows that (Wt )t∈[0,∞) is a continuous local martingale. Furthermore, by Q1 fo this sheet,
for every t ∈ [0, ∞),
d Z t t
(Bsi )2
X Z
hW it = ds = ds = t.
0 Xs2 0
i=1
Levy’s characterization of Brownian motion therefore proves that (Wt )t∈[0,∞) is a standard
one-dimensional Brownian motion. Hence, we conclude that X0 = 0 and
d−1
(8.9) dXt = dt + dWt ,
2Xt
which completes the proof of (a).
We will now prove (b). Let k ∈ N. Since it follows from (8.14) that hXit = t, Itô’s formula
proves that, since (Xt )t∈[0,∞) is nonnegative, for every t ∈ [0, ∞),
Xt2k = 2kXt2k−1 dXt + k(2k − 1)X 2(k−1) dhXit
2(k−1)
= 2kXt2k−1 dWt + k(d − 1)Xt dt + k(2k − 1)X 2(k−1) dt
= 2kX 2k−1 dWt + k(2(k − 1) + d)X 2(k−1) dt.
We therefore conclude that, since the first term on the righthand side is L2 -bounded by the
Itô isometry and hence uniformly integrable and hence a martingale,
Z t h i
2k 2(k−1)
E[Xt ] = k(2(k − 1) + d) E Xt dt.
0
This completes the proof of (b).
For part (c), since E[Xt0 ] = 1,
Z t
E[Xt2 ] =d ds = td.
0
And therefore,
Z t
E[Xt4 ] = 2(2 + d)d s ds = t2 (2 + d)d.
0
Finally,
Z t
E[Xt6 ] = 3(4 + d)(2 + d)d s2 ds = t3 (4 + d)(2 + d)d.
0
1
We therefore conclude that E[Xt4 ] = E[Xt6 ] when t = 4+d . This completes the proof.
(xiii) Let (Bt )t∈[0,∞) be a standard one-dimensional Brownian motion. Prove that, for every x ∈ R,
Z t
x
Xt = sgn(Bs − x) dBs ,
0
is a Brownian motion where
1 if y ≥ 0,

sgn(y) =
− 1 if y < 0.
Proof. The proof is an immediate consequence of Levy’s characterization of Brownian motion.

For every x ∈ R, it follows that (Xtx )t∈[0,∞) is a continuous local martingale vanishing at zero
which satisfies, for every t ∈ [0, ∞),
Z t Z t
hXtx it = sgn(Bs − x)2 ds = ds = t.
0 0
Therefore, Levy’s theorem proves that (Xtx )t∈[0,∞) is a standard one-dimensional Brownian
motion, which completes the proof.
(xiv) Let (Bt1 , Bt2 )t∈[0,∞) be a standard two-dimensional Brownian motion. Prove that the process
((Xt1 , Xt2 ))t∈[0,∞) defined by
Z t Z t
1 1 1
dXt = cos(Bs ) dBs − sin(Bs1 ) dBs2 ,
0 0
Z t Z t
2 1 1
dXt = sin(Bs ) dBs + cos(Bs1 ) dBs2 ,
0 0
is a standard two-dimensional Brownian motion.
Proof. The proof is again a consequence of Levy’s characterization of Brownian motion. The
processes (Xt1 )t∈[0,∞) and (Xt2 )t∈[0,∞) are continuous local martingales vanishing at zero by
the boundedness of the sin and cos functions. A repetition of the arguments leading to Q2 of
this sheet prove that
Z t Z t
hX 1 it = cos2 (Bs1 ) + sin2 (Bs1 ) ds = ds = t.
0 0
Similarly,
hX 1 it and hX 1 , X 2 it = 0.
Levy’s theorem therefore proves that ((Xt1 , Xt2 ))t∈[0,∞) is a standard two-dimensional Brown-
ian motion, which completes the proof.
(xv) Let (Xt )t∈[0,∞) and (Yt )t∈[0,∞) be continuous semimartingales. Define the stochastic exponen-
tial (E(X)t )t∈[0,∞) to be the process

hXit
E(X)t = exp Xt − .
2
Prove that there exists a unique continuous semimartingale (Zt )t∈[0,∞) such that
Z t
(8.10) Z t = Yt + Zs dXs ,
0
and that
Z t Z t
−1 −1
(8.11) Zt = E(X)t Y0 + E(X)s dYs + E(X)s dhX, Y is .
0 0
Proof. Suppose that (Zt )t∈[0,∞) is defined by (8.11). Define the continuous semimartingale
(Nt )t∈[0,∞) by
Z t Z t
−1
Nt = Y0 + E(X)s dYs − E(X)−1
s dhX, Y is .
0 0
The integration-by-parts formula then proves that
dZt = E(X)t dNt + Nt dE(X)t + dhN, E(X)it .
Since the stochastic exponential satisfies
dE(X)t = E(X)t dXt ,
and since by definition
dNt = E(X)−1 −1
t dYt − E(X)t dhX, Y it ,
it follows that
dhN, E(X)it = dhX, Y it .
Therefore, we conclude that
dZt = E(X)t Nt dXt = Zt dXt .
Since Z0 = Y0 by definition, this prove that the process (Zt )t∈[0,∞) satisfies (8.10).
Now, suppose that (Zt )t∈[0,∞) is an arbitrary solution of (8.10) in the sense that Z0 = Y0
and
dZt = dYt + Zt dXt .
Since we computed in class that
dE(X)−1 −1 −1
t = −E(X)t dXt + E(X)t dhXit ,
the integration-by-parts formula proves that
d E(X)−1 −1 −1

t Zt ) = −Zt E(X)t dXt + Zt E(X)t dhXit
+ E(X)−1 −1 −1
t dYt + E(X)t Zt dXt + dhE(X) , Zit .
Since the quadratic covariation is defined by
dhE(X)−1 , Zit = −Zt E(X)−1 −1
t dhXit − E(X)t dhX, Y it ,
it follows that
d E(X)−1 −1 −1

t Zt ) = E(X)t dYt − E(X)t dhX, Y it .
Since at t = 0 we have
E(X)−1
0 Z0 = Y0 ,
it follows that, for every t ∈ [0, ∞),

Z t
E(X)−1t Z t = Y 0 + +E(X)−1 −1
t dYt − E(X)t dhX, Y it .
0
Z t
−1 −1
Zt = E(X)t Y0 + +E(X)t dYt − E(X)t dhX, Y it ,
0
which proves that the solution is unique.
(xvi) For every probability measure µ on Rd let µ̂ denote the Fourier transform defined for every
ξ ∈ Rd by Z
µ̂(ξ) = exp (ihx, ξi) µ( dx).
Rd
In particular, if X is an Rd -valued random variable with distribution µX , then the Fourier
transform is the characteristic function of X in the sense that
Z
E [exp(ihX, ξi)] = exp(ihx, ξi)µX ( dx) = µ̂X (ξ).
Rd
For probability measures µ and ν on Rd , prove that µ = ν if and only if µ̂ = ν̂. Hint: For a
Schwarz function φ, compute Z
φ(ξ)µ̂(ξ) dξ.
Rd
Show that if X is a normally distributed random variable with mean zero and variance t ∈
(0, ∞) then 2
ξ t
µ̂X (ξ) = exp − .
2
Proof. Let S denote the Schwarz space (that is, the space of all C∞ -functions all of whose
derivatives decay faster than polynomially at infinity). In particular, C∞ d
c (R ) ⊂ S. The
essential feature is that the Fourier transform is a bijection from S to S with inverse given by
the inverse Fourier transform.
Let φ ∈ S. It then follows from Fubini’s theorem that
Z Z Z √ Z
φ(ξ)µ̂(ξ) dξ = φ(ξ) exp(ihx, ξi) dξ µ( dx) = 2π φ̂(x)µ( dx).
Rd Rd Rd Rd
Since µ̂ = ν̂, we conclude that, for every φ ∈ S,
Z Z
φ̂(x)µ( dx) = φ̂(x)ν( dx).
Rd Rd
Then, since the Fourier transform maps S bijectively into S, for every φ ∈ S,
Z Z
φ(x)µ( dx) = φ(x)ν( dx),
Rd Rd
which proves that µ = ν. Clearly, if µ = ν we have µ̂ = ν̂.
For the final part, we have
! !
2 2 Z 2
|x| |x −
Z
d ξ t d itξ|
(2πt)− 2 exp ihx, ξi − dx = exp − (2πt)− 2 exp − dx.
Rd 2t 2 Rd 2t
We observe that the function
!
2
|x −
Z
d z|
z ∈ C 7→ (2πt)− 2 exp − dx is holomorphic,
Rd 2t
and satisfies for every z = iz̃ for z̃ ∈ R that

!
2
|x
Z
d + tz̃|
(2πt)− 2 exp − dx = 1.
Rd 2t
We therefore conclude that, for every z ∈ C,
!
|x − itξ|2
Z
− d2
(2πt) exp − dx = 1,
Rd 2t
and hence
!
|x|2
Z 2
− d2 ξ t
µ̂X (ξ) = (2πt) exp ihx, ξi − dx = exp − ,
Rd 2t 2
(xvii) Let (Bt )t∈[0,∞) be a standard Ft -Brownian motion. Let (Mt )t∈[0,∞) be an L2 -bounded Ft -
martingale in the sense that
sup E Mt2 < ∞.

t∈[0,∞)
Prove that there exists a unique predictable process (Ht )t∈[0,∞) ∈ L2 (B) such that, for every
t ∈ [0, ∞),
Z t
Mt = E[M0 ] + Hs dBs .
0
Proof. The martingale convergence theorem proves that there exists M∞ ∈ L2 (Ω, F∞ , P) such
that, almost surely, as t → ∞,
Mt → M∞ almost surely and in L2 (Ω, F∞ , P)..
The martingale representation theorem proves that there exists (Hs )s∈[0,∞) ∈ L2 (B) such that
Z ∞ Z ∞
M∞ = E[M∞ ] + Hs dBs = E[M0 ] + Hs dBs ,
0 0
where the final equality uses the martingale property. Therefore, using properties of the
stochastic integral, for every t ∈ [0, ∞),
Z t
Mt = E[M∞ |Ft ] = E[M0 ] + Hs dBs .
0
Finally suppose that (Hs )s∈[0,∞) and (H̃s )s∈[0,∞) are two such processes. The Itô-isometry
then proves that
"Z 2 # Z ∞
∞
E (Hs − H̃s ) dBs = E[(Hs − H̃)2 ] ds = 0.
0 0
This implies that (Hs )s∈[0,∞) = (H̃s )s∈[0,∞) in L2 (B), which completes the proof.
(xviii) Let (Bt )t∈[0,∞) be a standard one-dimensional Brownian motion. Let C([0, ∞); R) denote
the space of continuous paths from [0, ∞) into R. Let µ, σ : [0, ∞) × C([0, ∞), R) → R be
bounded functions in the sense that there exists K1 ∈ (0, ∞) such that, for every t ∈ [0, ∞)
and continuous path (Xt )t∈[0,∞) ,
(|µ(t, X· )| + |σ(t, X· )|) ≤ K1 ,
and which are Lipschitz continuous in the sense that there exists K2 ∈ (0, ∞) such that, for
every t ∈ [0, ∞), for every pair of continuous paths (Xt )t∈[0,∞) and (Yt )t∈[0,∞) ,
(|σ(t, X· ) − σ(t, Y· )| + |µ(t, X· ) − µ(t, Y· )|) ≤ K2 sup |Xs − Ys | .

s∈[0,t]
Prove that there exists a jointly continuous process (Xtx )x∈Rd ,t∈[0,∞) such that, for every
x ∈ Rd and t ∈ [0, ∞),
Z t Z t
(8.12) Xtx = x + µ(s, X·x ) ds + σ(s, X·x ) dBs almost surely.
0 0
Proof. The issue is proving continuity in space. For every x ∈ R there exists a continuous in
time solution (Xtx )t∈[0,∞) of (8.12). Let t ∈ (0, ∞). We will consider the solution as a map
from R to the Banach space fo continuous functions C([0, T ]; R) equipped with the norm
kf kC([0,t];R) = sup |f (s)| .

s∈[0,t]
We will then apply the Kolmogorov continuity criterion to the map
x ∈ R 7→ (Xsx )s∈[0,t] ∈ C([0, t]; R).
Let p ∈ [2, ∞). The inequality |a + b + c|p ≤ 3p−1 (|a|p + |b|p + |c|p ) and equation (8.12) prove
that, for each x, y ∈ Rd ,
sup |Xsx − Xsy |p

s∈[0,t]
≤ 3p−1 |x − y|p
Z s
p Z s
p !
p−1
µ(r, Xrx ) µ(r, Xry ) dr σ(r, Xrx ) σ(r, Xry ) dBr

+3 sup − + sup − .
s∈[0,t] 0 s∈[0,t] 0
The Burkholder-Davis-Gundy inequality, Hölder’s inequality, and p ∈ [2, ∞) prove that there
exists Cp ∈ (0, ∞) such that
" Z s p # "Z
t p2 #
σ(r, Xrx ) − σ(r, Xry ) dBr ≤ Cp E ((σ(s, Xsx ) − σ(s, Xsy ))2 ds

E sup
s∈[0,t] 0 0
"Z p # t 2
≤ Cp Kp2 E |Xsx − Xsy |2 ds
0
" #
p−2
Z t
≤ Cp K2p t p E sup |Xsx − Xsy |p dr.
0 s∈[0,r]
Similarly, it follows by Jensen’s inequality and p ∈ [2, ∞) that

Z s p Z t
x y p−1
|µ(r, Xrx ) − µ(r, Xry )|p dr

sup µ(r, Xr ) − µ(r, Xr ) dr ≤ t

s∈[0,t] 0 0
Z t
≤ tp−1 K2p sup |Xsx − Xsy |p dr.
0 s∈[0,r]
The previous three inequalities prove that, for every p ∈ [2, ∞) and t ∈ [0, ∞),
" #
E sup |Xsx − Xsy |p
s∈[0,t]
" # !
p−2 Z t
≤ 3p−1 |x − y|p + Cp K2p t p + K2p tp−1 E sup |Xsx − Xsy |p dr .
0 s∈[0,r]
Grönwall’s inequality proves that, for every p ∈ [2, ∞) and t ∈ [0, ∞) there exists a constant
c(t, p) ∈ (0, ∞) that is locally bounded in t and p such that
" #
E sup |Xsx − Xsy |p ≤ c(t, p) |x − y|p .
s∈[0,t]
The Kolmoogorov continuity criterion that there exists a continuous modification (X̂sx )s∈[0,t]
of the process x ∈ R 7→ (Xsx )s∈[0,t] ∈ C([0, T ]; R). It follows that the modification solves
(8.12), because the zero set is independent of time. That is, for every x ∈ R we have
P[X̂tx = Xtx for every t ∈ [0, ∞)] = 1.
(xix) Let (Wt )t∈[0,∞) = (Wt1 , Wt2 , Wt3 )t∈[0,∞) be a three-dimensional Brownian motion, and assume
that W0 takes values in Rd \ {0} and that W0 is independent of (Wt − W0 )t∈[0,∞) . Define the
Euclidean norm
1
|W | = (Wt1 )2 + (Wt2 )2 + (Wt3 )2 2 .
(a) Show that (|Wt |−1 )t∈[0,∞) is a local martingale. Hint: If d ≥ 3, the function |x|2−d is
harmonic on Rd \ {0}.
(b) Suppose that W0 = y ∈ Rd and for every t ∈ [0, ∞) define Mt = |W1+t − y|−1 . Prove by
direct calculation that E[Mt2 ] = 1/1+t. Deduce that (Mt )t∈[0,∞) is L2 -bounded and hence
uniformly integrable.
(c) Show that (Mt )t∈[0,∞) is a local martingale and a supermartingale.
(d) Use the martingale convergence theorem to prove that (Mt )t∈[0,∞) is not a martingale.
Proof. We will first prove (a). For every n ∈ N let Tn denote the stopping time

1
Tn = inf t ∈ [0, ∞) : |Wt | ≤ .
n
It then follows from Itô’s formula, the fact that W0 is supported on R3 \ {0}, and the fact
−1
that the function |x|−1 is harmonic on R3 that the stopped process (WtTn )t∈[0,∞) is a

martingale. Since Tn → ∞ as n → ∞, using the fact that three-dimensional Brownian motion
almost surely does not hit zero,
P[Wt 6= 0 for every t ∈ [0, ∞)] = 1,
−1
it follows that (|Wt | )t∈[0,∞) is a local martingale.
We will now prove (b). it follows by definition that
!
2
|x −
Z
3 y|
E[Mt2 ] = (2π(1 + t))− 2 |x − y|−2 exp − dx
R3 2(1 + t)
!
2
|x|
Z
3
= (2π(1 + t))− 2 |x|−2 exp − dx.
R3 2(1 + t)
√
The change of variables x̃ = x/ proves that
1+t
!
2
|x|
Z
1 3
E[Mt2 ] = (2π)− 2 |x|−2 exp − dx
1 + t R3 2
Z ∞ 12 2
1 2 r
= exp − dr
1+t 0 π 2
1
= .
1+t
We therefore conclude that (Mt )t∈[0,∞) is L2 -bounded and hence uniformly integrable.
We will no prove part (c). It follows from part (a) that (Mt )t∈[0,∞) is a local martingale.
The nonnegativity of (Mt )t∈[0,∞) implies that (Mt )t∈[0,∞) is a supermartingale. Precisely, let
{τn }n∈N be stopping times such that (Mtτn )t∈[0,∞) is a martingale and such that τn → ∞ as
n → ∞. Then, for every s ≤ t ∈ [0, ∞), for every A ∈ Fs ,
E[Mtτn 1A ] = E[Msτn 1A ].
We now use the nonnegativity of (Mt )t∈[0,∞) to apply Fatou’s lemma, which proves that, since
for every n ∈ N we have {τn > s} ∈ Fs ,
E[Mt 1A ] ≤ lim inf E[Mtτn 1{τn >s} 1A ] = lim inf E[Msτn 1{τn >s} 1A ] = lim inf E[Ms 1{τn >s} 1A ].
n→∞ n→∞ n→∞
Finally, because Ms ∈ L2 (Ω, P), the dominated convergence theorem proves that
E[Mt 1A ] ≤ lim E[Ms 1{τn >s} 1A ] = E[Ms 1A ].
n→∞
We will now prove part (d). Assume by contradiction that (Mt )t∈[0,∞) is a martingale.
Then, there would exist M∞ ∈ L2 (Ω, F∞ , P) such that, as t → ∞,
Mt → M∞ almost surely and strongly in L2 (Ω, F∞ , P).
This implies that
1
lim E[Mt2 ] = lim 2
= 0 = E[M∞ ].
t→∞ t→∞ 1 + t
Hence M∞ = 0 and we have, for every t ∈ [0, ∞),
Mt = E[M∞ |Ft ] = 0,
which is a contradiction. Therefore (Mt )t∈[0,∞) is not a martingale.
(xx) Let (Bt )t∈[0,∞) be a standard one-dimensional Brownian motion. Prove that
Z t
4 2
12(t − s)Bs + 4Bs3 dBs .

Bt = 3t +
0
Proof. We observe by Itô’s formula that
Z t Z t
Bt4 = 4 (Bs )3 dBs + 6 (Bs )2 ds.
0 0
We furthermore have Z t
2
(Bt ) = 2 Bs dBs + t.
0
Hence,
Z t Z t Z s Z t Z s
4 3 2 3
Bt = 4 (Bs ) dBs + 6 2 Br dBr + s ds = 3t + 4(Bs ) + 12 Br dBr ds.
0 0 0 0 0
It follows from Fubini’s thoerem that

Z tZ s Z t Z t−s Z t
Br dBr ds = dr Bs dBs = (t − s)Bs dBs .
0 0 0 0 0
(xxi) Let d1 , d2 ∈ N. Let (Bt )t∈[0,∞) be a standard d2 -dimensional Brownian motion. Let µ be a
constant (d1 × d1 )-matrix and let σ be a constant (d1 × d2 )-matrix.
(a) For every x ∈ Rd , find the unique strong solution (Xtx , Bt )t∈[0,∞) to the equation
dXtx = µXtx dt + σ dBt in (0, ∞),

(
(8.13)
X0x = x.
Hint: For a d1 × d1 -matrix A, use properties of the matrix exponential
∞ k k
X t A
exp(tA) = .
k!
k=0
The solution itself will be expressed in terms of a stochastic integral.

(b) Find the distribution of Xtx for every t ∈ [0, ∞).
(c) Let d1 = d2 = 1. Prove that, for every bounded measurable function f : Rd → R,
" s !#
x 0 σ2
E [f (Xt )] = E f x exp(µt) + N (exp(2tµ) − 1) ,
2µ
where E0 denotes the expectation on any probability space (Ω0 , F 0 , P0 ) carrying a normally
distributed random variable N with mean zero and variance one.
Proof. We will first prove (a). Let Yt = exp(−tµ) for every t ∈ [0, ∞). It follows by definition
that
dYt = −Yt µ = −µYt .
We observe that if (Xt )t∈[0,∞) solves (8.13) then
d(Yt Xt ) = Yt µXt dt − Yt µXt dt + Yt µσ dBt = Yt σ dBt .
We therefore conclude that, since Y0 is the identity matrix and since X0 = x,
Z t Z t
Yt Xt = x + Ys σ dBs = x + exp(−sµ)σ dBs .
0 0
Hence,
Z t
(8.14) Xt = exp(tµ)x + exp((t − s)µ)σ dBs .
0
Since a direct calculation proves that the righthand side of (8.14) is a solution of (8.13), this
We will now prove (b). You an assume without proof that the stochastic integral is a normal
random variable with mean zero. So, it remains only to calculate its covariance. For this, we
have using the Itô isometry that
"Z 2 # Z t
t
exp((t − s)µ)σσ t exp((t − s)µ) ds.

E exp((t − s)µ)σ dBs =
0 0
We therefore conclude that Xtx is normally distributed on Rd1 with mean exp(tµ)x and co-
variance matrix Z t
exp((t − s)µ)σσ t exp((t − s)µ) ds.

0
We will make this formula more precise in dimension one below.
We will now prove (c). If d1 = d2 = 1 then we can compute
Z t Z t
t
2 σ2
exp((t − s)µ)σσ exp((t − s)µ) ds = σ exp((t − s)2µ) ds = (exp(2tµ) − 1) .
0 0 2µ
Therefore, in one dimension Xtx is normally distributed with mean exp(tµ)x and variance
Z t
2 σ2
σ exp((t − s)2µ) ds = (exp(2tµ) − 1) .
0 2µ
That is, for any normal random variable N with mean zero and variance one we have that in
distribution Xtx is equal to
s
σ2
exp(tµ)x + N (exp(2tµ) − 1).
2µ
This proves the final claim, and completes the proof.
(xxii) Let λ, ν, σ ∈ (0, ∞) and let µ : R → R be defined by

µ(x) = λ(ν − x).
Let (Bt )t∈[0,∞) be a standard one-dimensional Brownian motion.
(a) For every x ∈ R, find the unique strong solution (Xtx , Bt )t∈[0,∞) to the equation
dXtx = µ(Xtx ) dt + σ dBt in (0, ∞),
(
X0x = x.
(b) Calculate the mean and variance of Xtx , for each x ∈ Rd and t ∈ [0, ∞).
(c) Let ν = 0. Show that (Ytx = (Xtx )2 , Bt )t∈[0,∞) is a strong solution to the equation
( p
dYtx = −2λYt + σ 2 dt + 2σ Yt dBt in (0, ∞),

Y0x = x2 .
Proof. We will first prove (a). By definition, we are solving the equation
dXtx = λ (ν − Xtx ) dt + σ dBt in (0, ∞),
(
(8.15)
X0x = x.
We therefore observe that, for Yt = exp(λt), if (Xt )t∈[0,∞) solves (8.15) then
d(Xt Yt ) = λνYt dt + σYt dBt .
Therefore, as before,
Z t Z t
Xt Yt = x + λν Ys ds + σ Ys dBs .
0 0
Therefore,
Z t
(8.16) Xt = exp(−λt)x + ν (1 − exp(−λt)) + σ exp((s − t)λ) dBs .
0
Since a direct computation proves that (Xt )t∈[0,∞) solves (8.15), this completes the proof.
We will now prove (b). You can assume without proof that the stochastic integral is a
normal random variable with mean zero. The variance is, by the Itoı̂sometry,
" Z 2 #
t Z t
2 σ2
E σ exp((s − t)λ) dBs =σ exp((s − t)2λ) ds = (1 − exp(−2λt)) .
0 0 2λ
We therefore conclude that Xtx is a normal random variable with mean
exp(−λt)x + ν (1 − exp(−λt)) ,
and variance
σ2
(1 − exp(−2λt)) .
2λ
We will now prove (c). It follows by definition that Y0x = x2 and by Itô’s formula that
Z t Z t
Ysx ds = (Xsx )2 ds
0 0
Z t
=2 Xsx dXsx + hX x it
0
Z tp Z tp
Ysx dBs + σ 2 t.
p
=2 x x
Ys µ( Ys ) ds + 2σ
0 0
Z t Z tp Z tp
= −2λ Ysx ds + λν Ysx ds + 2σ Ysx dBs + σ 2 t.
0 0 0
Since ν = 0, this completes the proof.
(xxiii) Let µ, σ ∈ R. Let (Bt )t∈[0,∞) be a standard one-dimensional Brownian motion. For every
x ∈ R, use the stochastic exponential to find a strong solution (Xtx , Bt )t∈[0,∞) to the equation
dXtx = µ dt + σXtx dBt in (0, ∞),
(
(8.17)
X0x = x.
Proof. Let (E(σB)t )t∈[0,∞) denote the stochastic exponential

σ2t

E(σB)t = exp σBt − .
2
Since we have that
d E(B)−1 = −σE(B)−1 2 −1

t t dBt + σ E(B)t dt,
it follows that if (Xt )t∈[0,∞) is a solution of (8.17) then the integration-by-parts formula proves
d XE(B)−1 = µE(B)−1 2 −1 −1

t t dt + σ E(B)t Xt dt + dhX, E(B) it
= µE(B)−1 2 −1 −1
t dt + σ E(B)t Xt dt − σXt µE(B)t dt
= µE(B)−1
t dt.
Therefore, we have that
Z t
Xt E(B)−1
t =x+µ E(B)−1
s ds.
0
Hence,
Z t
−1
(8.18) Xt = E(B)t x + µ E(B)s ds .
0
And indeed, if we define (Xt )t∈[0,∞) in this way, an integration-by-parts calculation proves
that (Xt )t∈[0,∞) is the unique solution. This completes the proof.
(xxiv) Let (Mt )t∈[0,∞) be a continuous local martingale vanishing at zero, and let (L0t )t∈[0,∞) denote
the local time of (Mt )t∈[0,∞) at zero.
(a) Prove that
inf{t ∈ [0, ∞) : L0t > 0} = inf{t ∈ [0, ∞) : hM it > 0} almost surely.
(b) Prove that if α ∈ (0, 1) and (Mt )t∈[0,∞) is not identically zero, then (|Mt |α )t∈[0,∞) is not
a semimartingale.
Proof. We will first prove (a). We define the random times t1 , t2 by

t1 = inf{t ∈ [0, ∞) : L0t > 0} and t2 = inf{t ∈ [0, ∞) : hM it > 0}.
Since we have almost surely that
Z t
1
L0t = lim 1{−ε<Ms <ε} dhM is ,
ε→0 2ε 0
it follows almost surely that t1 ≥ t2 since by assumption hM it2 = 0 and for each ε ∈ (0, 1) we
have that Z t2

1{−ε<Ms <ε} dhM is ≤ hM it2 = 0.
0
Conversely, since the local time vanishes on [0, t1 ] we have by Tanaka’s formula that
Z t∧t1
|Mt∧t1 | = sgn(Ms ) dMs .
0
This implies that (|Mt∧t1 |)t∈[0,∞) is a local martingale. That is there exist a sequence of
stopping times τn → ∞ as n → ∞ such that
E[Mτn ∧t1 ] = 0.
We therefore conclude almost surely that Mt = 0 for every t ∈ [0, t1 ]. This implies almost
surely that hM it1 = 0 and therefore that t2 ≥ t1 almost surely. This completes the proof.
We will now prove (b). Let α ∈ (0, 1) and assume that (|Mt |α )t∈[0,∞) is a continuous
semimartingale. The Tanaka-Meyer formula proves that
Z t
α(α − 1) t α−2 a
Z
α α−1
|Mt | = α |Mt | d |Mt | + a Lt (|Mt |) da.
0 2 0
Since α ∈ (0, 1) the right continuity of the local time proves that, on the event {L0t (|Mt |) 6= 0},
Z t
aα−2 Lat (|Mt |) da = ∞.
0
We therefore have that L0t (|Mt |) = 0 almost surely, from which it follows using part (a) that
Mt = 0 vanishes identically and is constant.
(xxv) Let (Mt )t∈[0,∞) be a continuous local martingale that vanishes at zero. Let (E(M )t )t∈[0,∞)
denote the stochastic exponential.
(a) Show that (E(M )t )t∈[0,∞) is a nonnegative continuous local martingale.
Proof. It follows from Itô’s formula that almost surely, for each t ∈ [0, ∞),
Z t
E(M )t = 1 + E(M )s dMs .
0
Therefore, the claim follows due to the fact that the stochastic integral is a continuous
local martingale.
(b) Show that (E(M )t )t∈[0,∞) is a supermartingale with E[E(M )t )] ≤ 1 for every t ∈ [0, ∞).
Proof. There exist a sequence of stopping times {τn }n∈N which almost surely satisfy that
τn → ∞ as n → ∞ such that (Mt∧τn )t∈[0,∞) is a martingale. Therefore, for each n ∈ N
and s ≤ t ∈ [0, ∞),
E[E(M )t∧τn |Fs ] = E(Ms∧τn .
Hence, by positivity and Fatou’s lemma,
E[E(M )t |Fs ] = E[lim inf E(M )t∧τn |Fs ]
n→∞
≤ lim inf E[E(M )t∧τn |Fs ]
n→∞
= E[E(M )t∧τn |Fs ]Ms∧τn
= E(M )s .
And, since E(M )0 = 1 almost surely, for every t ∈ [0, ∞),
E[E(M )t ] ≤ E[E(M )0 ] = 1.
(c) Show that (E(M )t )t∈[0,∞) is a continuous martingale if and only if E[E(M )t )] = 1 for every
t ∈ [0, ∞).
Proof. Let s ≤ t ∈ [0, ∞). Since it holds almost surely that
E[E(M )t |Fs ] ≤ E(M )s ,
we have that
1 = E[E(M )t ] = E[E(M )t : E[E(M )t |Fs ] = Ms ] + E[E(M )t : E[E(M )t |Fs ] < Ms ]
≤ E[E(M )s ] = E[E(M )t : E[E(M )t |Fs ] = Ms ] + E[E(M )s : E[E(M )t |Fs ] < Ms ]
≤ E[E(M )s ] = 1,
with equality if and only if P[E[E(M )t |Fs ] < Ms ] = 0 since necessarily E(M )s > 0 on the
set {E[E(M )t |Fs ] < Ms }. We therefore conclude that
P(E[E(M )t |Fs ] = Ms ) = 1,
(xxvi) The following is Kazamaki’s criterion. Let (Lt )t∈[0,∞) be a continuous local martingale. Prove
that if (exp( 12 Lt ))t∈[0,∞) is a uniformly integrable submartingale, then the stochastic expo-
nential (E(L)t )t∈[0,∞) is a uniformly integrable martingale.
Proof. Let α ∈ (0, 1). Since the exponential function is invertiable and since(exp( 12 Lt ))t∈[0,∞)
is a uniformly integrable submartingale, there exists a measurable L∞ such that, as t → ∞,

1 1
exp Lt → exp L∞ almost surely and strongly in L1 (Ω).
2 2
It follows by definition of the stochastic exponential and the equality hαLit = α2 hLit that
α2

E(αL)t = exp αLt − hLit
2
2 2
= (E(L)t )α (Ztα )1−α ,
α
for Ztα = exp( 1+α Lt ). It follows from Hölder’s inequality that, for every stopping time T , for
every A ∈ F,
h 2 2
i 2 2
E [E(αL)T 1A ] = E (E(L)T )α (ZTα )1−α 1A ≤ E [E(L)T ]α E [ZTα 1A ]1−α .
By Question (i) since (E(L)t )t∈[0,∞) is a supermartingale the optional stopping theorem proves
that E[E(L)T ] ≤ 1 and therefore that
2
E [E(αL)T 1A ] ≤ E [ZTα ]1−α .
Since α/1+α ≤ 1/2 for every α ∈ (0, 1), it follows from every α ∈ (0, 1) that the family
{ZTα : T is a stopping time} is uniformly integrable,
and therefore for every α ∈ (0, 1) the family
{E(αL)T : T is a stopping time} is uniformly integrable.
Therefore, since for each α ∈ (0, 1) Question (i) proved that (E(αL)t )t∈[0,∞) is a local martin-
gale, it follows by uniform integrability that (E(αL)t )t∈[0,∞) is a uniformly integrable martin-
gale. Therefore, for each α ∈ (0, 1),
2 2
1 = E[E(αL)0 ] = E[E(αL)∞ ] ≤ E [E(L)∞ ]α E [Z∞
α 1−α
] .
Since (exp(1/2Lt ))t∈[0,∞) is a uniformly integrable submartingale by assumption, it follows
from the dominated convergence theorem that

α 1
lim E [Zt ] = E exp Lt ,
α→1 2
and therefore by Question (i) that, for each t ∈ [0, ∞)
1 ≤ E[E(L)∞ ] ≤ E[E(L)t ] ≤ E[E(L)0 ] = 1.
By Question (i) and the martingale convergence theorem we conclude that (E(L)t )t∈[0,∞) is a
uniformly integrable martingale.
(xxvii) The following is Novikov’s criterion. Prove that if (Lt )t∈[0,∞) is a continuous local martingale
which satisfies
1
E exp hLi∞ < ∞,
2
then (E(L))t∈[0,∞) is a uniformly integrable martingale.
Proof. It follows by Taylor expansion that, for every p ∈ N,

p p 1
E [hLi∞ ] ≤ 2 p!E exp hLi∞ .
2
This implies in particular that the martingale (Lt )t∈[0,∞) is L2 -bounded and hence uniformly
integrable. The uniform integrability proves that (Lt )t∈[0,∞) is a martingale and that
L∞ = lim Lt exists almost surely and strongly in Lp (Ω).
t→∞
Therefore, since we have

1 1 1 1 1 1
exp L∞ = exp L∞ − hLi∞ + hLi∞ = E(L)∞ exp
2
hLi∞ ,
2 2 4 4 4
and since it follows from Question (i) that (E(L)t )t∈[0,∞) is a supermartingale, and since the
optional stopping theorem proves thatE[E(L)∞ ] ≤ 1, we have from Hölder’s inequality and
the assumption that
1
1 1 1 2
E exp L∞ ≤ E [E(L)∞ ] E exp
2 hLi∞ < ∞.
2 2
We conclude that (exp( 12 Lt ))t∈[0,∞) is a submartingale because it is the convex function of
a martingale. Furthermore, it follows that (exp( 21 Lt ))t∈[0,∞) is uniformly integrable because,
for each K ∈ (0, ∞) and t ∈ [0, ∞),

1 1 1 1
E exp Lt : exp Lt ≥ K ≤ E exp L∞ : exp Lt ≥ K
2 2 2 2
and , for each t ∈ [0, ∞),

1 1
E exp Lt ≤ E exp L∞ ,
2 2
and Chebyshev’s inequality and the continuity of the Lebesgue integral. We therefore conclude
that (exp( 12 Lt ))t∈[0,∞) is a uniformly integrable submartingale, and therefore that Kazamaki’s
criterion from Question (ii) applies.
(xxviii) Let (Bt )t∈[0,∞) be a standard one-dimensional Brownian motion defined on a probability
space (Ω, F, P) with respect to a filtration (Ft )t∈[0,∞) . Suppose that b : R → R is a bounded,
measurable function. Define measures {QT }T ∈(0,∞) on {(Ω, FT , P)}T ∈(0,∞) which satisfy the
following three properties.
(a) The measures are compatible in the sense that QT1 = QT2 on FT1 for every T1 ≤ T2 ∈
[0, ∞).
(b) For every T ∈ [0, ∞) the measure QT is mutually absolutely continuous with respect to P
on (Ω, FT , P).
(c) The process (B̃t )t∈[0,∞) defined by
Z t
B̃t = Bt − b(Bs ) dBs ,
0
is for every T ∈ (0, ∞) a standard Brownian motion with respect to (QT , FT ) on [0, T ].
Prove that for every T ∈ (0, ∞) the pair (Bt , B̃t )t∈[0,T ] is a weak solution to the equation
(
dBt = b(Bt ) dt + dB̃t in (0, T ),
(8.19)
B0 = 0,
with respect to (QT , FT ). Deduce that uniqueness in law holds for (8.19).
Proof. Using the boundedness of b, define the martingale (Lt )t∈[0,∞) by
Z t
Lt = b(Bs )dBs ,
0
and for each T ∈ (0, ∞) define the measure QT by
dQT |Ft = E(L)t∧T dP|Ft .
The measures are compatible by definition, and they are mutually absolutely continuous
because E(L)t∧T is almost surely positive for each t ≤ T ∈ (0, ∞). Then, since either
the Kazamaki criterion or Novikov’s criterion prove that, for each T ∈ (0, ∞) the process
(E(L)t∧T )t∈[0,∞) is a martingale, the Girsanov theorem implies that, for each T ∈ (0, ∞), the
process (B̃t )t∈[0,∞) defined by
Z t
B̃t = Bt − hB, Lit = Bt − b(Bs ) ds
0
is a Brownian motion on [0, T ] with respect to the measure QT . Therefore, for each T ∈ (0, ∞),
dBt = dB̃t + b(Bt ) dt with B0 = 0
is a weak solution on [0, T ] with respect to QT . Since T ∈ (0, ∞) is arbitrary, this completes
the proof.
(xxix) (Skorokhod’s Lemma) Let y : [0, ∞) → R be a real-valued function that satisfies y(0) ≥ 0.
Prove that there exist unique functions a, z : [0, ∞) → R which satisfy the following three
properties.
(a) We have the decomposition
z = y + a.
(b) The function z is nonnegative
z ≥ 0.
(c) The function a is increasing, continuous, and vanishes at zero and the corresponding
Riemann-Stieltjes measure da is supported on the set {s : z(s) = 0}.
Furthermore, the function a is given by
" #
a(t) = sup (−y(s)) ∨ 0.
s∈[0,t]
Proof. It follows by definition of a, the continuity of y, and the fact that y(0) ≥ 0 that a
vanishes at zero, is continuous, and is nondecreasing. Furthermore, if we define z : [0, ∞) → R
by
z(t) = y(t) + a(t),
then it follows by definition that z ≥ 0. Finally, it follows by continuity that a is constant on
the set {t ∈ [0, ∞) : z(t) > 0} since for t ∈ [0, ∞) we have z(t) > 0 if and only if.
y(t) + sup (−y(s)) > 0.
s∈[0,t]
That is, z(t) > 0 if and only if

y(t) > − sup (−y(s)) = inf y(s).
s∈[0,t] s∈[0,t]
This implies that for every t0 ∈ {t ∈ [0, ∞) : z(t) > 0} there exists δ0 ∈ (0, ∞) such that, for
every δ ∈ (0, δ0 ),
a(t + δ) = a(t).
This implies that da vanishes on the set {t ∈ [0, ∞) : z(t) > 0} and therefore that da is
supported on the set {t ∈ [0, ∞) : z(t) = 0}.
To prove uniqueness, suppose that a1 , z1 and a2 , z2 are two such decompositions. Then,
z1 − a1 = z2 − a2 ,
and therefore
z1 − z2 (t) = a1 − a2 is a process of bounded variation.
Therefore, by properties (b) and (c), for each t ∈ [0, ∞),

Z t
2
(z1 (t) − z2 (t)) = 2 (z1 (s) − z2 (s)) d(a1 − a2 )( ds)
0
Z t Z 2
=− z1 (s) da2 ( ds) − z2 (s) da1 (s)
0 0
≤ 0,
which proves that z1 = z2 and therefore that a1 = a2 .
(xxx) Let (Bt )t∈[0,∞) be a standard one-dimensional Brownian motion. Define the process (Xt )t∈[0,∞)
by
Z t
Xt = sgn(Bs ) dBs .
0
(a) Show that (Xt )t∈[0,∞) is a standard Brownian motion with respect to the filtration
|B|
Ft = σ(|Bs | : s ∈ [0, t]).
|B|
Furthermore, observe that Ft ( FtB .
Proof. Levy’s theorem proves that (Xt )t∈[0,∞) is a Brownian motion. Since it follows by
Tanaka’s formula that
Z t
|Bt | = (Bs ) dBs + L0t (B) = Xt + L0t (B),
0
and since it follows by Question (v) that

L0t (B) = sup (−Xs ),
s∈[0,T ]
|B|
it follows for each t ∈ [0, ∞) that Ft ⊆ FtX . Similarly, since it follows almost surely that
Z t Z t
0 1 1
Lt (B) = lim 1−ε≤|Bs |≤ε ds = lim 1|Bs |≤ε ds,
ε→0 2ε 0 ε→0 2ε 0
L0 (B) |B|
it follows for each t ∈ [0, ∞) that Ft ⊆ Ft and therefore since Xt = |Bt | − L0t (B)
|B|
it follows for every t ∈ [0, ∞) that FtX ⊆ Ft . And therefore we have that (Xt )t∈[0,∞) is
|B| |B|
Ft -adapted. Finally, for each t ∈ (0, ∞) we have Ft ( FtB since
|B|
{Bt > 0} ∈ FtB but {Bt > 0} ∈
/ Ft .

(b) Let (L0t )t∈[0,∞) denote the local time of (Bt )t∈[0,∞) at zero. Prove that
L0t = sup (−Xs ) .

s∈[0,t]
Proof. This was shown in the previous problem, and is a consequence of Question (v) and
Tanaka’s formula.
(c) Let (St )t∈[0,∞) be defined by St = sups∈[0,t] |Bs |. Show that the two-dimensional processes
(St − Bt , St ) and (|Bt | , L0t ) have the same law.
Proof. By Tanaka’s formula, |B|t = Xt + L0t (B) and St − Bt = −Bt + St . Therefore, by

Question (v), the process (St −Bt , St )t∈[0,∞) is obtained from (−Bt )t∈[0,∞) and the process
(|Bt | , L0t (B))t∈[0,∞) is obtained from (Xt )t∈[0,∞) using the same deterministic procedure.
Since (−Bt )t∈[0,∞) and (Xt )t∈[0,∞) have the same law, since both are Brownian motions, we
conclude that the two-dimensional processes (St − Bt , St )t∈[0,∞) and (|Bt | , L0t (B))t∈[0,∞)
have the same law.
(d) For every a ∈ R let (Lat )t∈[0,∞) denote the local time of (Bt )t∈[0,∞) at a ∈ R. Deduce for
every a ∈ R that P[La∞ = ∞] = 1.
Proof. By definition, for each a ∈ R,

La∞ = lim Lat .
t→∞
Since properties of Brownian motion prove that

P[S∞ = ∞] = 1,
it follows from part (iii) that
P[L0∞ = ∞] = 1.
Let a ∈ R and define the stopping time Ta = inf{t ∈ [0, ∞) : Bt = a}. Then, since (BTa +t −
a)t∈[0,∞) is again a Brownian motion by Levy’s theorem, and since La∞ (B) = L0∞ (BTa +· − a),
we have
P[L0∞ (BTa +· − a)] = P[La∞ (B)] = 1.

(xxxi) Let (Bt )t∈[0,∞) be a standard one-dimensional Brownian motion. Prove that the stochastic
differential equation
dXt = sgn(Xt ) dBt in (0, ∞),

(8.20)
X0 = 0,
had a weak solution but no strong solution. Deduce that uniqueness in law holds for equation
(8.20) but that pathwise uniqueness does not hold.
Proof. Let (Bt )t∈[0,∞) be a standard Brownian motion, and define the process (B̃t )t∈[0,∞) by
Z t
B̃t = sgn(Bs ) dBs .
0
It follows from Levy’s theorem that (B̃t )t∈[0,∞) is a Brownian motion, and therefore since
dB̃t = sgn(Bt ) dBt ,

we have that
dBt = sgn(Bt ) dB̃t with B0 = 0.
We therefore conclude that (Bt , B̃t )t∈[0,∞) is a weak solution of the equation. Since Levy’s
theorem implies that every weak solution of the equation is a Brownian motion, we conclude
that uniqueness in law must hold. However, pathwise uniqueness does not hold, since given any
weak solution (Bt , B̃t )t∈[0,∞) the process (−Bt , B̃t )t∈[0,∞) is again weak solution and (Bt )t∈[0,∞)
is nonzero since in law it is a Brownian motion.
(xxxii) Let α ∈ (0, 1/2), let σ : R → R be defined by

σ(x) = |x|α ∧ 1,
and let (Bt )t∈[0,∞) be a standard one-dimensional Brownian motion. Show that the map
Z t
t ∈ [0, ∞) → σ −2 (Bs ) ds,
0
is almost surely finite. Let (τt )t∈[0,∞) denote the associated time-changes
Z s
−2
τt = inf s ∈ [0, ∞) : σ (Br ) dr = t .
0
Show that Xt = Bτt and Xt = 0 are two solutions of the equation

X0 = 0.
Conclude that uniqueness in law does not hold for this equation, despite the fact that the
second of these solutions is a strong solution.
Proof. We first observe that, for each t ∈ [0, ∞), since α ∈ (0, 12 ),
!
Z t Z 2
1 |x|
E σ −2 (Bs ) dBs = (2πt)− 2 |x|−2α exp − dx
0 R 2t
1 Z ∞ 2
2π 2 r
= r−2α exp − dr < ∞.
t 0 2t
Therefore, since the process
Z t
σ −2 (Bs ) dBs is increasing,
0 t∈[0,∞)
we conclude that for every t ∈ [0, ∞) it is almost surely finite. Define the process (B̃t )t∈[0,∞)
by Z τt
B̃t = σ −1 (Bs ) dBs .
0
Levy’s theorem and the definition of (τt )t∈[0,∞) proves that (B̃t )t∈[0,∞) is a Brownian motion
and that (Xt )t∈[0,∞) defined by Xt = Bτt satisfies
dXt = σ(Xt ) dB̃t for t ∈ (0, ∞) with X0 = 0.
We therefore conclude that (Xt , B̃t )t∈[0,∞) is a nonzero weak solution of the equation. Since
the zero process is also a weak solution, we conclude that pathwise uniqueness does not hold,
despite the fact that the zero solution is a strong solution.

SDENotes 2

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

SDENotes 2

Uploaded by

Copyright:

Available Formats

STOCHASTIC DIFFERENTIAL EQUATIONS

We can also take conditional expectations with respect to a sigma-algebra.

Proposition 1.10. Let {Yn : Ω → R}n∈N and Y : Ω → R be integrable random variables on a

(iii) For every s ≤ t ∈ I,

is a Fn -martingale. If in addition (Hn )n∈N is nonnegative, then ((H · M )n )n∈N0 is a (sub/super)

Observe that, for every n ∈ N,

(2.31) sup E[Mn2 ] < ∞.

Therefore, after passing to the limit n → ∞, it follows from (2.31) that

Since the same computation proves that, for every n ≤ m ∈ N0 ,

infinite sequence of stopping times {τk }k∈N by

We first observe that, by continuity of the process (Mt )t∈[0,∞) ,

We therefore conclude from (2.34) that, for every K ∈ N,

Then, there exists M∞ ∈ L1 (Ω) such that, for almost every ω ∈ Ω, as t → ∞,

suppose by contradiction that there exist ω ∈ Ω0 and a, b ∈ [−∞, ∞] such that

which proves that M∞ ∈ L1 (Ω). This completes the proof.

Then, there exists M∞ ∈ L1 (Ω) such that, for almost every ω ∈ Ω, as t → ∞,

Then there exists M∞ ∈ Lp (Ω) such that

and since we have for every t ∈ [0, ∞) that

it follows from (2.44) and the dominated convergence theorem that

Therefore, on the interval [0, 1], for a partition

As the mesh |∆| → 0, we expect by stochastic integration that

Motivated by the quadratic variation, we expect that, as |∆| → 0,

In particular, if (Bt )t∈[0,∞) is a one-dimensional Brownian motion,

For the variance, for every s ≤ t ∈ [0, ∞),

= E[Mt2 |Fs ] − E[Mt2n |Fs ] + E[Mt2n |Fs ] − Ms2 + Ts∆ (M ).

It then follows by Hölder’s inequality that

|Hs | |Ks | d |hM, N i|s ≤ Hk2 hM ittkk−1 Kk2 hN ittkk−1

We therefore define an inner product and norm on H for M, N ∈ H by

Therefore, there exists a subsequence {nk }k∈N such that, as k → ∞,

and therefore, for every t ∈ [0, ∞),

Let (Mt )t∈[0,∞) ∈ H, and suppose that

∆ = {0 = t0 < t1 < t2 < t3 < . . .},

∆ = {0 = t0 < t1 < t2 < t3 < . . .},

it follows by definition and the martingale property that

and such that, for every n ∈ N,

3.3. The integration-by-parts formula. In this section, we will prove an integration-by-parts

Returning to (3.6), it follows from the identities

4. Applications of stochastic calculus

defined in Proposition 4.1 by

is a martingale. Define the process of bounded variation

which completes the proof.

and therefore that

is a process of bounded variation,

In particular, if we define the L2 -bounded martingale (Mt )t∈[0,∞) by

Then, for every f ∈ S, let (Mtf )t∈[0,∞) denote the martingale

and the stochastic exponential (E(M f )t )t∈[0,∞) defined by

which takes the form (4.12) for Ht = E(M f )t ft .

and therefore that

Define the measure dQ = Y dP and define the measurable map F : Ω → Rn by

We therefore conclude that {(Hsn )}s∈[0,∞) is a Cauchy sequence in L2 (B).

and such that, as n → ∞,

Since by definition, for every k ∈ N ,

and almost surely satisfies, for every t ∈ [0, ∞),

which completes the proof.

Define the convex function f N : R → R by

and observe by a direct computation using the compact support of µf N that

Jensen’s inequality and Fubini’s theorem then prove that

If for some k ∈ (1, ∞) we have that

We observe in particular that the righthand side of (5.2) is independent of x ∈ R. Therefore, if we

We observe that (L̂at )a∈R,t∈[0,∞)

t ∈ [0, ∞), the dominated convergence theorem proves that

Equations (5.3) and (5.4) prove that (L̂at )a∈R,t∈[0,∞)

6. The Girsanov theorem