An Introduction to Stochastic Differential

Equations

Adair Antonio da Silva Neto

August 18, 2022

Contents

1 Introduction 1

2 Stochastic Integrals 2
2.1 Formalizing the ‘noise’ 2
2.2 Preparing the Terrain 3
2.3 Constructing the Itô Integral 5
2.4 Properties 9
2.5 Extensions 10
2.6 Stratonovich integral 10

3 The Ito Formula and the Martingale Representation Theorem 12

4 Stochastic Differential Equations 13

A Measure Theory 14
A.1 Measurable Spaces 14
A.2 Measures 14
A.3 Measurable Functions 17
A.4 Integration 19
A.5 Product Spaces 23
A.6 The Monotone Class Lemma 24
A.7 Modes of Convergence 25
A.8 Conditional Expectation 26

B Probability Theory 28
B.1 Probability Spaces 28
B.2 Distribution and Density of a Random Variable 30
B.3 Expected Value and Variance 31
B.4 Stochastic Processes 34
B.5 Brownian Motion 35
B.6 Riemann-Stieltjes Integral 36
B.7 Modes of Convergence 37

References 39

2
1 Introduction

An equation that models an evolution process that contains ‘noise’, with a

certain randomness, is an equation of the form:

dX
= b(t, Xt ) + σ(t, Xt ) · ‘white noise’ (1)
dt

Problems like this appears naturally in biology (populational growth mod-

els), physics (charge in electrical circuits), engineering (filtering problems, like
the Kalman’s Filter exihibted in the Figure 1) and finance (optimal stopping,
optimal portfolio and option pricing).

Figure 1: Kalman’s Filter: Observed data; Filtered process; Real data

[Wik22].

From the mathematical viewpoint, we need to make sense of this kind of

equations, since with the usual tools of calculus, it is not possible to treat them.
Our goal then, in this work, is to give the foundations that allow us to treat
this equations. In particular, firstly we’re going to define the Itô’s Integral and
then we’ll be ready to tackle stochastic differential equations.

1
2 Stochastic Integrals

2.1 Formalizing the ‘noise’

To study the equation (1), the first step is to give a formalization for the
‘noise’ that we are going to consider. In our case, we’ll consider the ‘white
noise’, which is a generalized stochastic process Wt with gaussian probability
distribution with mean zero, finite variance such that E[Ws Wt ] = 0 whenever
t ̸= s.
In other words, the ‘noise’ will be represented by some stochastic process
Wt and our goal is to satisfy the following properties:

• If t1 ̸= t2 , then Wt1 and Wt2 are independent.

• The collection {Wt } is stationary, i.e., the joint distribution of any collec-
tions of Wt does not depend on t.

• E[Wt ] = 0 for all t.

However, this process Wt does not have reasonable properties (does not
have continuous paths, and it is not measurable for finite time). So our first
goal is to replace Wt with some convenient stochastic process.
Let 0 = t0 < t1 < . . . < tm = t and consider the discrete version

Xk+1 – Xk = b(tk , Xk )∆tk + σ(tk , Xk )Wk ∆tk

Now, replace Wk ∆tk = ∆Vk = Vk+1 – Vk , where Vt is a stochastic process.

Since we want Vt to have stationary independent increments with mean zero,
the only suitable process with continuous paths is the Brownian motion, rep-
resented in the Figure 2.
Hence, we take Vt = Bt , thus obtaining
k–1
X k–1
X
Xk = X0 + b(tj , Xj )∆tj + σ(tj , Xj )∆Bj
j=0 j=0

Taking the limit as ∆tj goes to zero, we have

Zt Zt
Xk = X0 + b(s, Xs ) ds + σ(s, Xs ) dBs
0 0

2
 Notice that
Bt – Bs = Ws (t – s)
for all t ≥ s, i.e., the ‘white noise’ is the derivative, with respect to time, of the
Brownian motion.

0.5

–0.5

–1
0 50 100 150 200

Figure 2: Two realizations of the Brownian motion [Jak]

2.2 Preparing the Terrain

Now we can start treating the equation. In order to do that, notice that in
the determinist case, in which there isn’t a ‘noise’, we have an equation of the
form
dXt
= b(t, Xt ), X0 = x0 (2)
dt
In this case, a solution is a function Xt such that
ZT
Xt – X0 = b(s, Xs ) ds (3)
0

This function can be understood as a stochastic process over a space of proba-

bility given by a single point.
Using the formalism above, we propose as a solution to the equation (1)
an identity of the form
Zt Zt
Xt – X0 = b(s, Xs ) ds + σ(s, Xs ) dBs (4)
0 0

3
where the left integral is a Lebesgue-Stieltjes’ integral, but the right integral is
what still needs to be formalized.
As said earlier, our task is to understand the following integral
ZT
f (t, ω) dBt (ω) (5)
S

where f : [0, ∞] × Ω −→ R.
The idea will be the following: we’ll first define the integral for the most
elementary functions and, after that, using approximations and convergence
criteria, we’ll define it for the other classes of functions.
Since it is natural to approximate a given function f (t, ω) by
X
f (tj∗ , ω)χ[tj ,tj+1 ) (t)
j

where tj∗ ∈ [tj , tj+1 ], we could define the integral (5) as follows
ZT
f (t, ω) dBt (ω) = lim f (tj∗ , ω)[Btj+1 – Btj ](ω)
n→∞
S

However, the choice of points tj∗ make a difference here. If we take tj∗ = tj ,
the left end point, then we obtain the Itô integral, denoted by
ZT
f (t, ω) dBt (ω)
S

If, on the other hand, we take tj∗ = (tj + tj+1 )/2, the mid point, then we
obtain the Stratonovich integral, denoted by
ZT
f (t, ω) ◦ dBt (ω)
S

One of the advantages of Itô integral is its application in Finance. Intu-

itively, by taking the leftmost point, one does not have to know the future, like
whether a stock goes up or down.
Related to this intuition, and developing the Ito integral, we must restrict
ourselves to functions f that only depends on the behaviour of Bs (ω) up to the
time tj .

4
Definition 2.1. Let Bt (ω) be a Brownian motion. We define Ft as the σ-algebra
generated by the random variables Bs , where s ≤ t. I.e., Ft is the smallest σ-
algebra containing all sets

{ω : Bt1 (ω) ∈ F1 , . . . , Btk (ω)} ∈ Fk

where tj ≤ t and Fj ⊆ Rn are Borel sets, j ≤ k = 1, 2, . . ..

The intuition behind this definition is that Ft can be thought as the ‘history
of Bs up to the time t’. A function h(ω) is Ft -measurable if h depends only on
the values from B0 up to Bt . I.e., it does not depend on the ‘future’.

Definition 2.2 (Nt -Adapted). Let {Nt }t≥0 be an increasing family of σ-algebras
of subsets of Ω. A process g(t, ω) : [0, ∞)×Ω −→ Rn is Nt -adapted if for each
t ≥ 0 the function ω −→ g(t, ω) is Nt -measurable.

Example 2.1. The process h1 (t, ω) = Bt/2 (ω) is Ft -measurable and, hence,
Ft -adapted. Notice that this process only depends on previous information.
However, the process h2 (t, ω) = B2t (ω) is not Ft -measurable and, hence,
it is not Ft -adapted. That’s because the process h2 depends on future informa-
tion.

2.3 Constructing the Itô Integral

The class of functions for which the Itô integral will be first defined is the
following.

Definition 2.3. Let V(S, T) be the class of functions

f (t, ω) : [0, ∞) × Ω −→ R

such that

1. The function (t, ω) −→ f (t, ω) is B × F-measurable, where B denotes

the Borel σ-algebra on [0, ∞).

2. The function f (t, ω) is Ft -adapted.

”R T —
3. E S f (t, ω)2 dt < ∞.

5
 Given functions f ∈ V, in order to define the Ito integral
ZT
I[f ](ω) = f (t, ω) dBt (ω)
S

we are going to define I[ϕ] for a simple class of functions ϕ and then show
how each f ∈ V can be approximated by ϕ.

Definition 2.4 (Elementary functions). A function ϕ ∈ V is elementary if it

has the form X
ϕ(t, ω) = ej (ω)χ[tj ,tj+1 ) (t)
j

Since ϕ ∈ V, each function ej is Fj -measurable. And, as we did before, we

define ZT X
ϕ(t, ω) dBt (ω) = ej (ω)[Bj+1 – Bj ](ω)
S j≥0

To further our development of the Ito integral, we need the following result.

Theorem 2.1 (Itô isometry). If ϕ(t, ω) is bounded and elementary, then

 
‚Z T Œ2 –Z T ™
2
E ϕ(t, ω) dBt (ω)  = E ϕ(t, ω) dt
S S

Using this isometry, it is possible to extended the previous definition to

functions in V by the following steps.

1. Let g(ω) ∈ V be bounded and continuous for each ω. Then there exists
elementary functions ϕn ∈ V such that
–Z T ™
E (g – ϕn )2 dt −→ 0
S

as n → ∞.

2. Let h ∈ V be bounded, then there exists bounded functions gn ∈ V such

that gn (ω) are continuous for all ω and n, and
–Z T ™
E (h – gn )2 dt −→ 0
S

6
 3. Let f ∈ V. Then there exists a sequence of functions (hn ) ⊆ V such that
hn is bounded for each n and
–Z T ™
E (f – hn )2 dt −→ 0
S
as n → ∞.

Please note that we started from bounded and continuous functions, gen-
eralized into bounded functions and then generalized further into any function
in our class V.
These steps can be summarized in the following result, which guarantees
that the elementary functions are dense in V(S, T).
Theorem 2.2. If g ∈ V(S, T), then there exists a sequence of elementary func-
tions ϕn ∈ V(S, T) such that ϕn → g in L2 (P)

Finally, we can define the Itô integral as follows.

Definition 2.5 (The Itô integral). Let f ∈ V(S, T), then the Itô integral of f ,
from S to T is
ZT ZT
f (t, ω) dBt (ω) = lim ϕn (t, ω) dBt (ω)
n→∞
S S

where the limit is in L2 (P) and (ϕn ) is a sequence of elementary functions such
that
T
–Z ™
E (f (t, ω) – ϕn (t, ω))2 dt −→ 0
S
as n → ∞.

This new definition induces a generalized form of the Ito isometry.

Theorem 2.3 (Ito isometry). For all f ∈ V(S, T),
 
‚Z T Œ2 –Z T ™
E f (t, ω) dBt (ω)  = E f (t, ω)2 dt
S S

Corollary 2.4. If f (t, ω) ∈ V(S, T) and fn (t, ω) ∈ V(S, T) and

 
‚Z T Œ2
E fn (t, ω) – f (t, ω) dt −→ 0
S

7
as n → ∞, then
ZT ZT
fn (t, ω) dBt (ω) −→ f (t, ω) dB)t(ω)
S S

in L2 (P) as n → ∞.
Example 2.2. We wish to compute the integral
ZT
B(t) dB(t)
0

Step 1. Let
n–1
X
ϕn (t) = Bn (tin )[Bn (ti+1
n
) – Bn (tin )]
i=0
be a sequence of elementary functions. Then,
 
–Z T
X Z tj+1 X Z tj+1
™
2 2
E (ϕn – Bs ) ds = E  (Bj – Bs ) ds = (s – tj )2 ds
0 j tj j tj
X1
2
= tj+1 – tj −→ 0 when ∆tj → 0
2
j
n – tn ) → 0 by the conti-
I.e., ϕn (t) converges to B(t) almost surely as maxi (ti+1 i
nuity of B(t).
Step 2. Now notice that

Bn (ti )[Bn (ti+1 ) – Bn (ti )] = Bn (ti+1 )Bn (ti ) – B2n (ti ) + B2n (ti+1 ) – B2n (ti+1 )
1
= [B2n (ti+1 ) – B2n (ti ) – (Bn (ti+1 ) – Bn (ti ))2 ]
2

Step 3. With these two results, we can write the original integral as
ZT n–1 n–1
1 X 2  1X
Xn (t) dB(t) = Bn (ti+1 ) – B2n (ti ) – [Bn (ti+1 ) – Bn (ti )]2
0 2 2
i=0 i=0

Step 4. Since the first sum is a telescopic sum and the second one converges
in probability to T by the quadratic variation of Brownian motion, we have
ZT
1 1 1 1 1
Xn (t) dB(t) = B2 (t) – B2 (0) – T = B2 (t) – T
0 2 2 2 2 2

8
 Hence, the integral converges in L2 (P) to
ZT ZT
1 1
B(t) dB(t) = lim Xn (t) dB(t) = B2 (t) – T
0
n→∞
0 2 2

2.4 Properties

Before heading to more theoretical and important results, let us notice some
natural facts of the Itô integral.

Theorem 2.5. Let f , g ∈ V(0, T) and let 0 ≤ S < U < T. Then

RT RU RT
1. S
f dBt = S
f dBt + U
f dBt for a.a. ω.
RT RT RT
2. S (cf + g) dBt = c S f dBt + S g dBt where c ∈ R.
”R T —
3. E S f dBt = 0.
RT
4. S
f dBt is FT -measurable.

With the Doob’s martingale inequality, it can be proved that the Itô integral
can be chosen to depend continuously on t. A proof of this fact is on the third
chapter of [Oks13].
Now an important result is that the Itô integral is a martingale.

Theorem 2.6. Let f (t, ω) ∈ V(0, T) for all T. Then

Zt
Mt (ω) = f (s, ω) dBs
0

is a martingale w.r.t. Ft and for λ, T > 0

  –Z T ™
1 2
P sup |Mt | ≥ λ ≤ 2 E f (s, ω) ds
0≤t≤T λ 0

Summarizing, the Itô integral is continuous, adapted, linear, a martingale,

satisfies the Itô isometry and has Quadratic variation.

9
2.5 Extensions

Using the concept of martingales, we can generalize the Itô integral for a
larger class of functions than V.
Considering the Definition 2.3, we can relax the condition 2. into

2.’ There exists an increasing familiy of σ-algebras Ht such that Bt is a mar-

tingale w.r.t. Ht and ft is Ht -adapted.

The idea here is that Bt must remain a martingale with respect to the history
of fs .
Another way of extending the Itô integral definition is by weaking the con-
dition 3. into
”R T —
3.’ P f (s, ω) 2 ds < ∞ = 1
S

Too understand why this works, let WH (S, T) denote the class of processes
satisfying conditions 1, 2’ and 3’ above. Then, in the 1-dimensional Brownian
motion, for all t there exists step functions fn ∈ WH such that
Zt
|fn – f |2 ds −→ 0
0

in probability.
Rt
Now, since 0 fn (s, ω) dBs (ω) converges in probability to a random variable
and the limit depends only on f , we define
Zt Zt
f (s, ω) dBs (ω) = lim fn (s, ω) dBs (ω)
n→∞
0 0

where the limit is in probability and f ∈ WH .

2.6 Stratonovich integral

As we saw, the Itô integral is one possible interpretation of the integral

ZT
f (t, ω) dBt (ω)
S

10
 The Stratonovich integral is another possibility and, in general, leads to
different results (except when the integrating function has a derivative). In
some cases, the Stratonovich definition may be adequate.
In other cases, Itô’s feature of ‘not looking in the future’ (as [Oks13] puts
it), justifies its use in biology and finance, for example. Moreover, Itô integral is
a martingale, and Stratonovich’s is not. This gives an important computational
advantage to Itô’s definition.
And, to quote [S+ 04],

However, it [Stratonovich’s integral] is inappropriate for finance.

In finance, the integrand represents a position in an asset and the
integrator represents the price of that asset. We cannot decide at
1:00 p.m. which position we took at 9:00 a.m. We must decide the
position at the beginning of each time interval, and the Ito integral
is the limit of the gain achieved by that kind of trading as the time
between trades approaches zero.

However, Stratonovich integral is used in manifolds and is of particular

interest to physicists for obeying rules of classical calculus, such as the chain
rule.
In this section, we presented a formalism for modeling evolution process
with noise using the Brownian motion. We saw that the classical notion of
solution doesn’t work and we proposed a new interpretation. As a result, it
was necessary to construct a new theory of integration and, with that, the Itô
and Stratonovich integrals. We hence studied the properties of these integrals.
Given that, we can advance on more elaborated questions, for example:
does the equation (1) has a solution? Under which hypothesis? How can we
study the behavior of the solution?

11
3 The Ito Formula and the Martingale Representation
Theorem

12
4 Stochastic Differential Equations

13
A Measure Theory

A.1 Measurable Spaces

Given a set S, on what collection S of subsets of S are suitable to be a

domain of measure?

Definition A.1. Let S be a set and S be a family of subsets of S. Then S is

called a σ-algebra on S if

1. ;, S ∈ S.

2. If A ∈ S, then Ac ∈ S.
S∞
3. If A1 , A2 , . . . ∈ S, then i=1 Ai ∈ S.

The pair (S, S) is said to be a measurable space, and any subset A ⊆ S is

called an S-measurable set.

For any set S and any collection A of subsets of S there is at least one σ-
algebra containing A: the family of all subsets of S. Taking the intersection of
all the σ-algebras containing A, we obtain the smallest σ-algebra containing
A, which is called the σ-algebra generated by A.
Particularlly, the smallest σ-algebra containing all of the open sets of R is
called the Boreal algebra for R and is denoted by B. Any set in B is called a
Borel set.

A.2 Measures

Now, how can we assign a size (or a probability) to all the sets in S?

Definition A.2. Let (S, S) be a measurable space. A measure is an extended

real-valued function µ : S −→ R such that

1. µ(;) = 0;

2. µ(A) ≥ 0, for all A ∈ S;

S∞
3. If (An )∞ is a countable, disjoint sequence of subsets in S, then µ( n=1 An ) =
P∞ n=1
n=1 µ(An ).

14
 The triple (S, S, µ) is called a measure space.

A measure is nonnegative, assigns zero to the null set, and is countably

additive. If µ(S) < ∞, then µ is finite.
We say that a proposition holds almost everywhere (a.e.) if there exists a
set A ∈ S with µ(A) = 0, such that the proposition holds on the complement
of A. Intuitively, the proposition holds everywhere except on sets of measure
zero.
For example, a sequence of functions (fn ) on S converges a.e. to a function
f if there exists A ∈ S with µ(A) = 0 such that limn→∞ fn (x) = f (x) for all
x ∈ Ac .
If µ(S) = 1, then µ is a probability measure and (S, S, µ) is called a prob-
ability space. Any measurable set A ∈ S is called and event and µ(A) is the
probability of the event A. In a probability space, the phrase almost surely
(a.s.) is used interchangeably with ‘almost everywhere’.
Notice that if A, B ∈ S and A ⊆ B, then µ(A) ≤ µ(B). If µ(A) < ∞, then
µ(B \ A) = µ(B) – µ(A).

Theorem A.1. Let (S, S, µ) be a measure space.

1. If (An )∞
n=1 is an increasing sequence in S (i.e. if An ⊆ An+1 for all n),
then
∞
‚ Œ
[
µ An = lim µ(An )
n→∞
n=1

2. If (Bn )∞
n=1 is a decreasing sequence in S (i.e. if Bn+1 ⊆ Bn for all n) and
if µ(Bm ) < ∞ for some m, then
∞
‚ Œ
\
µ Bn = lim µ(Bn )
n→∞
n=1

First, we define a measure on a small family of sets and then present an

extension theorem.

Definition A.3. Let S be a set and A be a family of subsets of S. Then A is

called an algebra if

1. ;, S ∈ A;

2. A ∈ A implies Ac ∈ A;

15
 Sn
3. A1 , A2 , . . . , An ∈ A implies i=1 Ai ∈ A.

I.e., an algebra is closed under complementation and finite union. On an

algebra, the idea of measure is very much similar to the definition on a σ-
algebra.
Definition A.4. Let S be a set, and let A be an algebra of subsets of S. A
measure is a real-valued function µ : A −→ R such that

1. µ(;) = 0;

2. µ(A) ≥ 0, for all A ∈ A;

S∞
3. If (An )∞ is a disjoint sequence of sets in A with n=1 An ∈ A, then
S∞n=1
P∞
µ n=1 An = n=1 µ(An ).

Notice that last item is different from the definition on a σ-algebra. Here,
we require that the union is contained in the algebra.
Although it is easier to define measures on algebras, it is more convenient
to work with σ-algebras.
Theorem A.2 (Caratheodory Extension Theorem). Let S be a set, A an algebra
of its subsets, and µ a measure on A. Let S be the smallest σ-algebra contain-
ing A. Then there exists a measure µ∗ on S such that µ(A) = µ∗ (A), for all
A ∈ A.

To rule out the possibility of more than one extension of µ to S, we’ll need
the following definition
Definition A.5. Let S be a set, A an algebra of its subsets, and µ a measure on
A. If there is a countable sequence of sets (Ai )∞
i=1
∈ A with µ(Ai ) < ∞ for all
S∞
i, and S = i=1 Ai , then µ is called σ-finite.

By definition, any probability measure is σ-finite. And the next theorem

shows that the extension of a σ-finite measure is unique.
Theorem A.3 (Hahn Extension Theorem). Let S, A, µ and S be as specified in
Caratheodory Extension Theorem. If µ is σ-finite, then the extension µ∗ to S
is unique.
Definition A.6. Let (S, S, µ) be a measure space, A ∈ S be any set with mea-
sure zero, and let C ⊆ A. Denoting by C the family of such sets C, i.e.,

C = {C ⊂ S : C ⊆ A for some A ∈ S with µ(A) = 0}

16
 The completion of S is the family S′ constructed by starting with any set
B ∈ S, and then adding and subtracting from it sets in C. That is

S′ = {B′ ⊆ S : B′ = (B ∪ C1 ) \ C2 , B ∈ S, C1 , C2 ∈ C}

Intuitively, S′ consists of all the subsetes of S that differ from a set in S

by a setset of a set of µ-measure zero. Using this definition, a measure µ on
(S, S) can be extended to (S, S′ ).

A.3 Measurable Functions

Definition A.7. Given a measurable space (S, S), a real-valued function f :

S −→ R is measurable with respect to S (or S-measurable) if

{s ∈ S : f (s) ≤ a} ∈ S, ∀a ∈ R

If the space is a probability space, then f is called a random variable.

Some cases in which measurability can be easily verified:

• Any monotone or continuous function f : R −→ R is measurable with

respect to Borel sets.

• If S is a countable set and S is a complete σ-algebra for S (i.e. contains

all subsets of S), then all functions f : S −→ R are S-measurable.

In more general cases, the measurability of a function is established by

showing that it is the limit of a sequence of ‘simpler’ functions. We’ll do this
progressively.
First, let (S, S) be a measurable space and consider the indicator function
χA : S −→ R of the from
¨
1, if s ∈ A
χA (s) =
/A
0, if s ∈

Clearly, χA is S-measurable iff. A ∈ S.

Now, consider the finite weighted sums of indicator functions:
n
X
ϕ(s) = ai χAi (s)
i=1

17
where (Ai )ni=1 is a sequence of subsets of S, and (ai )ni=1 is a sequence of real
numbers. Functions like this are called simple functions.
If the sets (Ai ) form a partition of S and if all of the values ai are distinct,
then the previous equation is the standard representation of the function. If
this is the case, then ϕ is measurable iff. each Ai ∈ S.
As we’ll see in the next two theorems, the set of all measurable functions
consists of those that are pointwise limits of measurable simple functions.
First, any function f that is pointwise limit of a sequence (fn ) of measurable
function is itself measurable.
Theorem A.4 (Pointwise convergence preserves measurability). Let (S, S) be
a measurable space, and let (fn ) be a sequence of S-measurable functions con-
verging pointwise to f , i.e.,
lim f (s) = f (s), ∀s ∈ S
n→∞ n

Then f is also S-measurable.

Second, any measurable function f can be expressed as the pointwise limit

of a sequence (ϕn ) of measurable simple functions. If f is nonnegative, the
sequence can be chosen to be strictly increasing. And if f is bounded, the
sequence can be chosen to converge uniformly.
The idea is that we can construct a class of measurable functions by taking
the class of measurable simple function and then closing this under pointwise
convergence. The next result shows that this set contains all the measurable
functions.
Theorem A.5 (Approximation of measurable functions by simple functions).
Let (S, S) be a measurable space. If f : S −→ R is S-measurable, then there
exists a sequence of measurable simple functions (ϕn ) such that ϕn −→ f point-
wise. If 0 ≤ f , then the sequence can be chosen so that
0 ≤ ϕn ≤ ϕn+1 ≤ f
for all n. If f is bounded, then the sequence can be chosen so that ϕn −→ f
uniformly.

With these two theorems, we see that a function is measurable iff. it is the
pointwise limit of a sequence of measurable simple functions. The standard
way to prove that a function is measurable is to find such an approximation.
Some important properties follow. Suppose that f and g are S-measurable
functions on S and c ∈ R. Then,

18
 • The functions f + g, f · g, |f | and c · f are S-measurable.

• If (fn ) is a sequence of S-measurable functions, then inf fn , sup fn , lim inf fn

and lim sup fn are all S-measurable.

• All continuous functions on Rl are Bl -measurable.

• The composition of Borel measurable functions are Borel measurable.

However, this is not true for Lebesgue measurable functions.

Definition A.8. Let (S, S) and (T, T) be measurable spaces. Then the function
f : S −→ T is measurable if the inverse image of every measurable set is
measurable, i.e.,
{s ∈ S : f (s) ∈ A} ∈ S, ∀A ∈ T

Hence, if (S, S), (T, T), and (U, U) are measurable spaces, and f : S −→ T
and g : T −→ U are measurable functions, then h : S −→ U defined by h(s) =
g ◦ f (s) is a measurable function.

Definition A.9 (Measurable selection). Let (S, S) and (T, T) be measurable

spaces, and let Γ be a correspondence of S into T. Then the function h : S −→ T
is a measurable selection from Γ if h is measurable and h(s) ∈ Γ (s), for all
s ∈ S.

Theorem A.6 (Measurable Selection). Let S ⊆ Rl and T ⊆ Rm be Borel sets,

with their Borel subsets S and T. Let Γ : S −→ T be a nonempty compact-
valued and upper hemi-continuous (UHC) correspondence. Then there exists
a measurable selection from Γ .

A.4 Integration

Let (S, S, µ) be a fixed measure space and M(S, S) be the space of measur-
able, extended real-valued functions on S.
For a nonnegative, measurable and simple function, we have the following
definition.

Definition A.10. Let ϕ ∈ M+ (S, S

P)nbe a measurable simple function, with the
standard representation ϕ(s) = i=1 ai χAi (s). Then the integral of ϕ with
respect to µ is
Z Xn
ϕ(s) µ(ds) = ai µ(Ai )
S i=1

19
 To extended from simple functions to all M+ (S, S), we define
Definition A.11. For f ∈ M+ (S, S), the integral of f with respect to µ is
Z Z
f (s) µ(ds) = f (s)χA (s)µ(ds)
A S
R R
We also denote the integrals above as f dµ and A f dµ.

Since any function f ∈ M+ (S, S) can be expressed as the limit of an in-

creasing sequence (ϕn ) of simple functions in M+ (S, S). However, we cannot
define the integral as the limit of such a sequence because it would leave open
the possibility that the limit depends on the particular sequence chosen. The
Monotone Convergence Theorem shows that the limit value is unique.
Before that, the following result shows that a simple function on a measure
space can be used to define a new measure on the space.
Lemma A.7. If ϕ ∈ M+ (S, S) is a simple function and λ : S −→ R is defined
by Z
λ(A) = ϕ dµ
A
for all A ∈ S, then λ is a measure on S.
Theorem A.8 (Monotone Convergence Theorem). If (fn ) is a monotone in-
creasing sequence of functions in M+ (S, S) converging pointwise to f then
Z Z
f dµ = lim fn dµ
n→∞

Some important properties:

1. If f , g ∈ M+ (S, S) and c ≥ 0, then

Z Z Z
(f + g) dµ = f dµ + g dµ

and Z Z
c · f dµ = c f dµ

2. If f , g ∈ M+ (S, S) and f ≤ g, then

Z Z
f dµ ≤ g dµ

20
 3. If f ∈ M+ (S, S) and A, B ∈ S with A ⊆ B, then
Z Z
f dµ ≤ f dµ
A B

4. If f ∈ M+ (S, S) and A ∈ S with µ(A) = 0, then

Z
f dµ = 0
A

5. If f ∈ M+ (S, S), S f dµ < ∞, and A = {s ∈ S : f (s) = +∞}, then

R

µ(A) = 0.

6. If f ∈ M+ (S, S) and λ : S −→ R+ , defined as λ(A) = A f dµ for all

R

A ∈ S, then λ is a measure on (S, S).

7. Let (gi ) be a sequence of functions in M+ (S, S). Then

Z ‚X n n Z
Œ
X
gi dµ = gi dµ
i=1 i=1

Lemma A.9 (Fatou’s Lemma). If (fn ) is a sequence of functions in M+ (S, S),

then Z Z
(lim inf fn ) dµ ≤ lim inf fn dµ

To generalize to functions that take on both positive and negative values,

we begin by defining the positive and negative parts of a function f + and f – .
¨
f (s), if f (s) ≥ 0
f + (s) =
0, if f (s) < 0
and ¨
–f (s), if f (s) ≤ 0
f – (s) =
0, if f (s) > 0

Hence, if f is measurable, f + and f – are both in M+ (S, S) and f = f + – f – .

This motivates the following generalization.
Definition A.12. Let (S, S, µ) be a measure space, and let f be a measurable
real-valued function on S. If f + and f – both have finite integrals with respect
to µ, then f is integrable and the integral of f with respect to µ is
Z Z Z
f dµ = f + dµ – f – dµ

21
 In probability, if (S, S, µ) is a probability space and f is a measurable func-
tion, then f is a random variable. And if f is integrable, the integral of f with
respect to µ is called the expected value of f and is denoted by
Z
E[f ] = f dµ

Now let L(S, S, µ) denote the set of all S-measurable, real-valued functions
on S that are integrable with respect to µ. How can we determine whether a
function is integrable?

1. If f is bounded, measurable, real-valued function on S, and µ(S) < ∞,

then f is µ-integrable.
R R
2. f is µ-integrable iff. |f | is µ-integrable. In that case, | f dµ| ≤ |f | dµ.

3. If f isRS-measurable,
R g is µ-integrable, and |f | ≤ |g|, then f is µ-integrable
and |f | dµ ≤ |g| dµ.

4. If f and g are µ-integrable and α ∈ R, then αf and f + g are µ-integrable,

satisfying the standard rules of calculus.

Theorem A.10 (Lebesgue Dominated Convergence Theorem). Let (S, S, µ)

be a measure space, and let (fn ) be a sequence of integrable functions that
converges almost everywhere to a measurable function f .If there exists an in-
tegrable function g such that |fn | ≤ g for all n, then f is integrable and
Z Z
f dµ = lim fn dµ

Definition A.13. Let λ and µ be finite measures on (S, S). If

µ(A) = 0 =⇒ λ(A) = 0 ∀A ∈ S

then λ is absolutely continuous with respect to µ, written λ << µ.

If there is A ∈ S such that λ(B) = λ(A ∩ B), for all B ∈ S, then λ is
concentrated on A.
If there are disjoint sets A, B ∈ S such that λ is concentrated on A and µ is
concentrated on B, then λ and µ are mutually singular, written λ ⊥ µ.

22
Theorem A.11 (Radon-Nikodym Theorem). Let λ and µ be σ-finite positive
measures on (S, S) with λ << µ. Then there exists an integrable function h
such that Z
λ(A) = h(s) µ(ds), all A ∈ S
A
And the function h is called the Radon-Nikodym derivative of λ with respect
to µ.

Lemma A.12. Let λ1 and λ2 be finite measures on (S, S). Then there exists a
triple of measures γ, α1 , α2 such that

λi = γ + α1

where i = 1, 2, and α1 ⊥ α2 .

A.5 Product Spaces

Let (X, X) and (Y, Y) be measurable spaces, and let Z = X × Y.

Our first task is to define a σ-algebra of subsets of Z that is a natural product
of X and Y.

Definition A.14 (Measurable rectangle). A set C = A × B ⊆ Z is a measurable

rectangle if A ∈ X and B ∈ Y.

We denote by C the set of all measurable rectangles and E the set of all finite
unions of measurable rectangles. Notice that E is an algebra and that every set
in E can be written as the finite union of disjoint measurable rectangles.

Definition A.15 (Product Space). Let F = X × Y be the σ-algebra generated

by E. The measurable space (Z, F) is called the product space.

Theorem A.13. Let (X, X), (Y, Y), C, and E be as specified above. Let µ : C −→
R+ have the following properties:

• µ(;) = 0;
S∞
• If (Ci ) = ((Ai × Bi ))∞ is a sequence of disjoint sets in C and i=1 Ci is
S∞ i=1
P∞
in C, then µ i=1 Ci = i=1 µ(Ci ).

Then there is a measure on E that coincides with µ on C.

23
 This theorem can be naturally extended to any space that is the product of
a finite number of measurable spaces.

Definition A.16. Let (X, X) and (Y, Y) be measurable spaces, and let (Z, F) be
the product space. And let E ⊆ Z and x ∈ X. Then the x-section of E is the set
Ex = y ∈ Y : (x, y) ∈ E. And the y-section of E is the set Ey = x ∈ X : (x, y) ∈ E.
Let f : Z −→ R and let x ∈ X. The x-section of f is the function fx : Y −→ R
defined by fx (y) = f (x, y). And the y-section of f is the function fy : X −→ R
defined by fy (x) = f (x, y).

Intuitively, the x-section of a function is found by fixing x and viewing f

only as a function of y.

Theorem A.14. Let (X, X) and (Y, Y) be measurable spaces, and let (Z, F) be
the product space.

1. If the set E in Z is F-measurable, then every section of E is measurable.

2. If the function f : Z −→ R is F-measurable, then every section of f is

measurable.

A.6 The Monotone Class Lemma

The main question that this section aims to answer is whether a propery P
holds at a given σ-algebra.

Definition A.17 (Monotone Class). A monotone class is a nonempty collec-

tion M of sets such that it contains

1. The union of every nested increasing sequence A1 ⊆ A2 ⊆ . . . of sets in

M.

2. The intersection of every nested decreasing sequence A1 ⊇ A2 ⊇ . . . of

sets in M.

Some important facts:

1. Every σ-algebra is a monotone class.

2. Given a nonempty collection of subsets of S, denoted by A, there exists

a smallest monotone class containing A, which is called the monotone
class generated by A.

24
 3. The σ-algebra generated by A contains the monotone class generated by
A.

4. If a monotone class is an algebra, then it is a σ-algebra.

Lemma A.15 (Monotone class). Leet S be a set and let A be an algebra of
subsets of S. Then the monotone class M generated by A is the same as the
σ-algebra S generated by A.

Using this lemma, our original task in this section is solved. To establish
that P holds for all sets in a product σ-algebra, it suffices to show that

1. P holds for all finite unions of disjoint measurable rectangles.

2. The family of sets E for which P holds is a monotone class.

A.7 Modes of Convergence

To the traditional ideas of convergence pointwise and uniformly, learned

in an undergraduate course in Analysis, we’ll add new forms of convergence,
as the functions fn acquire an infinite and inequivalent number of ways to
approach their limit f .
Let (X, B, µ) be a measure space, and the functions fn (and their limit f )
measurable with respect to this space. Then, we say that fn converges to f

1. pointwise almost everywhere if, for (µ–)almost everywhere x ∈ X, fn (x)

converges to f (x).

2. uniformly almost everywhere, essentially uniformly , or in L∞ norm

if, for every ϵ > 0, there exists N such that for every n ≥ N,

|fn (x) – f (x)| ≤ ϵ

for µ-almost every x ∈ X.

3. almost uniformly if, for every ϵ > 0, there exists a set E ∈ B of measure
µ(E) ≤ ϵ such that fn converges uniformly to f on the complement of E.

4. in L1 norm if the quantity

Z
∥fn – f ∥L1 (µ) = |fn (x) – f (x)| dµ
x
converges to 0 as n → ∞.

25
 5. in measure if, for every ϵ > 0, the measures

µ({x ∈ X : |fn (x) – f (x)| ≥ ϵ})

converge to zero as n → ∞.

In probability theory, if fn and f are random variables, convergence in L1

is usually referred to as convergence in mean, pointwise convergence almost
everywhere is referred to as almost sure convergence, and convergence in
measure is often referred to as convergence in probability.
Recall that a property P(x) is said to hold almost everywhere if the set of
x for which P(x) fails has Lebesgue measure zero.

A.8 Conditional Expectation

Let (Ω, F, µ) be a probability space.

Definition A.18 (Conditional Probability). For a measurable set A satisfying

µ(A) > 0 and for any measurable set B we define the conditional probability
of B given A as
µ(B ∩ A)
µA (B) :=
µ(A)
We’ll also use P(B|A) to denote this probability.

Now notice that µA : F −→ [0, 1], i.e., it is also a probability measure.

Definition A.19 (Conditional Expectation). The conditional expectation of f

given A is defined as Z
E[f |A] = f dµA

Definition A.20 (Measurable Partition). A family Aη η∈H of subsets of Ω is a

measurable partition of Ω if the following conditions hold:

1. Aη ∈ F, ∀η ∈ H.

2. η∈H Aη = H.
S

3. Aη ∩ Aη′ = 0, ∀η ̸= η′ .

If the index set H is countable, then we call the measurable partition countable.

26
 To formalize it better, extending the concept to σ-algebras not generated
by a countable partition, we’ll introduce the following.
Definition A.21 (Conditional Expectation). Let A ⊂ F be a σ-algebra and let
f : Ω −→ R be an integrable function. Then the conditional expectation of f
relative to A is an A-measurable function E[f |A] : Ω −→ R such that
Z Z
E[f |A](ω)µ (dω) = f (ω)µ (dω), ∀C ∈ A
C C

It can be proved that this function exists and is unique in the sense that if
g also satisties the condition above, then E[f |A] = g.
Notice that to compute conditional probability P(B|Aη ) we may take f to
be the indicator function χB . Then
Z
E[χB |A](ω̂)µ(Aη ) = χb dµ = µ(B ∩ Aη )
Aη

for all ω̂ ∈ Aη , all η ∈ H.

If µ(Aη ) > 0, then
µ(B ∩ Aη )
E[χB |A](ω̂) =
µ(Aη )

Summarizing this fact,

P(B|Aη ) = E[χB |A](ω̂)
for all ω̂ ∈ Aη , all η ∈ H.
We end this section with an important property of conditional expectations.
Let A1 ⊆ A2 ⊆ F. Then we have
Z Z
E[E(f |A2 )|A1 ] dµ = E(f |A2 )(ω) dµ
C c
Z
= f (ω) dµ
C
Z
= E[f |A1 ](ω) dµ, ∀C ∈ A1
C

This is known as the law of the iterated expectation. To state it in another

way, Z
E[f – E(f |A2 )|A1 ] dµ = 0, ∀C ∈ A1
C

27
B Probability Theory

B.1 Probability Spaces

A probability space starts with a set Ω called the sample space, which is,
intuitively, a list of all possible outcomes of an experiment. Each ω ∈ Ω is a
sample point, and each subset A ⊆ Ω is an event.
To filter ‘well-behaved’ subsets, where it will be possible to measure a prob-
ability, the following definition is necessary.

Definition B.1 (σ-algebra). If Ω is a set, then a σ-algebra F on Ω is a family

of subsets of Ω satisfying:

1. ; ∈ F.

2. If A ∈ F, then Ac ∈ F.
S∞
3. If A1 , A2 , . . . ∈ F, then i=1 Ai ∈ F.

The pair (Ω, F) is said to be a measurable space, and any subset B ⊆ Ω that
also belongs to F is called a measurable set.

Example B.1 (Examples of σ-algebras). The following are σ-algebras.

1. The family {;, Ω} is called the trivial σ-algebra, and is the smallest one
possible.

2. The power set P(Ω) is called the discrete σ-algebra, and is the largest
one possible.

3. The family {;, Ω, A, Ac } is the σ-algebra generated by the set A and is

usually denoted by FA .

But the most important σ-algebra for probability theory is the Borel σ-
algebra, which will be denoted by B. Taking Ω = R, the Borel σ-algebra is
generated by the intersection of all σ-algebras containing the real line inter-
vals.
Notice that the Borel σ-algebra contains all open sets, closed sets, and all
their countable operations with union ∪, intersection ∩, and their complements
c . This is the smallest σ-algebra containing all open subsets.

28
 More generally, if U is the collection of all open subsets of a topological
space Ω, then B is called the Borel σ-algebra on Ω. The elements B ∈ B are
called Borel sets.
Given a measurable space, it’s possible to assign each outcome to a proba-
bility.

Definition B.2 (Probability Measure). Let (Ω, F) be a measurable space. A

probability measure P is the function

P : F −→ [0, 1]

satisfying

1. P(;) = 0 and P(Ω) = 1.

2. (σ-additivity). If A1 , A2 , . . . ∈ F and Ai ∩Aj = ;, for i ̸= j (i.e. are mutually

exclusive), then Œ ∞
∞
‚
[ X
P Ai = P(Ai )
i=1 i=1

And the triple (Ω, F, P) is called a probability space.

In order to restrict the functions to sets in the σ-algebra, the following

definition will be needed.

Definition B.3 (F-measurable function). Given a probability space (Ω, F, P), a

function Y from the sample space Ω to Rn is called F-measurable if

Y–1 (U) = {ω ∈ Ω : Y(ω) ∈ U} ∈ F

for all open sets U ∈ Rn (i.e. for all Borel sets U ⊆ Rn ). In other words, the
inverse image of U is in the σ-algebra.

And to attach numerical values to each ω ∈ Ω, define the following func-

tion.

Definition B.4 (Random variable). An F-measurable function X : Ω −→ Rn is

called a random variable on a complete probability space (Ω, F, P).

This definition means that if we know which event U in the F has occurred,
then we know which value of X has occurred.

29
 Consider the measurable space (Ω, P(Ω)), and let X be a random variable
with values xi , i = 1, 2, . . . , k. The sets

Ai = {ω ∈ Ω : X(ω) = xi } ⊆ Ω

form a partition of Ω, and the σ-algebra generated by this partition is called

the σ-algebra generated by X.
Notice that this is the smallest σ-algebra that contains all the sets Ai . This
σ-algebra is often denoted by FX . Intuitively, this σ-algebra represents all
information available about the sample point ω by observing X.

B.2 Distribution and Density of a Random Variable

However, given only the probability measure P, it is not always immedi-

ate how to compute the probability of a given interval, set or value. For that
purpose, the following three definitions will be useful.

Definition B.5 (Distribution). The distribution of a random variable X is a

function µX : Ω −→ Rn defined as

µX (B) = P[X–1 (B)] = P[X ∈ B] = P({ω : X(ω) ∈ B})

Definition B.6 (Distribution Function). The cumulative distribution func-

tion (CDF) of a random variable X is defined as

FX (x) = P[X ≤ x] = P({ω : X(ω) ≤ x})

where x ∈ R. Notice that F is non-decreasing, and right-continuous, and ap-

proaches 0 at –∞ and 1 at +∞.

Definition B.7 (Joint Distribution). If X and Y are random variables, then their
joint distribution function is

FX,Y (x, y) = P[X ≤ x, Y ≤ y]

where x, y ∈ R.

The distribution function gives the probability that the random variable X
is on the interval (–∞, x]. This also allows to look for an interval [a, b] using
FX (b) – FX (a).
While the distribution of X returns the probability of an event {ω : X(ω) ∈
B}.

30
Definition B.8 (Density function). A random variable X has a probability den-
sity function (PDF) f : Rn −→ R if f is a measurable function and
Zx
FX (x) = P[X ≤ x] = f (y) dy
–∞

B.3 Expected Value and Variance

Another useful tool for any given random variable is to know its mean value
and how much it varies. This motivates the following.

Definition B.9 (Expectation).

R Let (Ω, F, P) be a probability space and X a ran-
dom variable. If Ω |X(ω)| dP(ω) < ∞, then the expected value (or mean
value of X with respect to P is
Z Z
E[X] = X(ω) dP(ω) = x dµX (x)
Ω Rn

Notice that the expectation is linear. I.e., if X and Y are integrable and a
and b are constants, then

E[aX + bY] = aE[X] + bE[Y]

R
If f is a Borel measurable function and Ω |f (X)| dP(ω) < ∞, then
Z Z
E[f (X)] = f (X(ω)) dP(ω) = f (X) dµX (x)
Ω Rn

Theorem B.1 (Chebychev’s inequality).

1
P[|X| ≥ λ] ≤ p E[|X|p ]
λ
for all λ ≥ 0.

Definition B.10 (Variance). Let µ := E[X], i.e., the expected value of a random
variable X. Then the variance of X is given by
Z
Var[X] = E[∥X – µ∥2 ] = |X – µ|2 dP(ω)
Ω

31
Definition B.11 (Covariance). Let X and Y be integral random variables. Let
µX = E[X] and µY = E[Y]. If XY is integrable, then the covariance of X and Y
is:
Cov[X, Y] = E[(X – µX )(Y – µY )] = E[XY] – µX µY
Notice that Var[X] = Cov[X, X].

The variance of X can be computed more simply using the following theo-
rem.

Theorem B.2. Let X be a random variable. Then,

1. Var[X] = E[X2 ] – E[X]2 .

2. Var[aX + b] = a2 Var[X].

3. Cov[X, Y] = Cov[Y, X].

4. Cov[aX + bY, Z] = aCov[X, Z] + bCov[Y, Z].

5. Var[X + Y] = Var[X] + Var[Y] + 2Cov[X, Y].

Definition B.12 (Independence). Two events A, B ∈ F are independent if

P[A ∩ B] = P[A] · P[B]

More generally, any collection of events Ai , i = 1, 2, . . ., is called indepen-

dent if for any n ∈ N and any choice of indices ik , k = 1, 2, . . .,
n n
‚ Œ
\ Y

P A ik = P A ik
k=1 k=1

Definition B.13 (Conditional Expectation). Let X and Y be random variables.

Then, given that fY (y) > 0, the conditional distribution of X given Y = y is
given by the conditional density

f (x, y)
f (x|y) =
fY (y)

And the conditional expectation of X given Y = y is

Z∞
E[X|Y = y] = xf (x|y) dx
–∞

32
 The intuition behind this definition is to build an estimate of the random
variable X given the information available in Y.
Some important and very useful properties of conditional expectation are
listed below.

Theorem B.3. Let X, Y : Ω −→ Rn be random variables with E[|X|] < ∞ and

E[|Y|] < ∞, H ⊆ F a σ-algebra, and let a, b ∈ R.

1. E[aX + bY|H] = aE[X|H] + bE[Y|H].

2. E[E[X|H]] = E[X].

3. E[X|H] = X if X is H-measurable.

4. E[X|H] = E[X] if X is independent of H.

5. E[Y · X|H] = Y · E[X|H] if Y is H-measurable.

6. If G ⊆ H is a σ-algebra, then

E[X|G] = E[E[X|H]|G]

7. If ϕ : R −→ R is convex and E[|ϕ(X)|] < ∞, then

ϕ(E[X|H]) ≤ E[ϕ(X)|H]

Definition B.14 (Infinitely often). If A1 , A2 , . . . , An , . . . are events in the prob-

ability space, then the event
∞ [
\ ∞
Am = {ω ∈ Ω : ω belongs to infinitely many of the An }
n=1 m=n

is called An infinitely often, or simply ‘An i.o.’.

The next lemma helps us check if some sequence of events occurs infinitely
often.
P∞
Lemma B.4 (Borel-Cantelli). If n=1 P(An ) < ∞, then
∞ [∞
‚ Œ
\
P(An i.o.) = P Am = 0
n=1 m=n

The most important distribution in this text is the normal (or gaussian)
distribution.

33
Definition B.15 (Normal Distribution). If random variable X has mean µ, vari-
ance σ2 and a density function of the form

–(x – µ)2
 
1
f (x) = p exp
2πσ 2σ2

then X is said to have a normal distribution, which we denote by X ∼ N(µ, σ2 ).

Notice that X can be normalized (i.e. transformed into a distribution of the

X–µ
form N(0, 1)) by taking Z = σ .
Definition B.16 (Lipschitz and Hölder Conditions). A function f is said to be
Hölder continuous of order α, 0 < α ≤ 1, on an interval [a, b] ⊆ R if there
exists a constant K > 0, such that for all x, y ∈ [a, b] we have

|f (x) – f (y)| ≤ K|x – y|α

If α = 1, then the function f is said to be Lipschitz continuous.

B.4 Stochastic Processes

Definition B.17 (Stochastic Process). A stochastic process is a collection of

random variables {X(t)} parametrized by time t ∈ T.
For each point ω ∈ Ω, the mapping t −→ Xt (ω) is the respective sample
path, also called realization or trajectory.
Definition B.18 (Filtration). A filtration F is a collection of σ-algebras

F = {F0 , F1 , . . . , Ft , . . . , FT }

such that Ft ⊆ Ft+1 .

The idea behind this definition is to model the flow of information. As

the time t passes, the observer has more information, and does not lose any
previous data. Notice that this implies finer partitions of the sample space Ω.
Given a stochastic process {X(t)}, let Ft be the σ-algebra generated by the
random variables Xs , s = 0, . . . , t. Since Ft ⊆ Ft+1 , these σ-algebras form a
filtration called the natural filtration of the process {X(t)}, and contain all
available information of the process up to the time t.
Definition B.19 (Martingale). A stochastic process {Mt } on (Ω, F, P) is called
a martingale with respect to a filtration Mt if

34
 1. Mt is Mt -measurable for all t.

2. E[|Mt |] < ∞ for all t.

3. E[Ms |Mt ] = Mt for all s ≥ t.

Intuitively, a martingale is a stochastic process in which the future has no

tendency to go up or down. In other words, the expected value of any time in
the future is equal to the value of the process at the present time.

Theorem B.5 (Doob’s martingale inequality). If Mt is a martingale such that

the mapping t −→ Mt (ω) is continuous a.s., then for all p ≥ 1, T ≥ 0 and
λ > 0,  
1
P sup |Mt | ≥ λ ≤ p E[|MT |p ]
0≤t≤T λ

B.5 Brownian Motion

In 1828 the Scottish botanist Robert Brown described the irregular motion
of pollen grains suspended in fluid, that is now known as the Brownian Motion.
To describe this mathematically, we use a stochastic process Bt (ω), which is the
position at time t of the particle ω.
This can also be understood as the model for the cumulative effect of ‘noise’.
And is also called Wiener process, after N. Wiener, who formalized mathemat-
ically the Brownian motion.

Definition B.20 (Brownian Motion). A Brownian motion {B(t)} is a stochastic

process with the following properties:

1. (Independence of increments). B(t) – B(s), for t > s, are independent.

That means that the direction that the particle will go does not depend
on the past.

2. (Normal increments). B(t) – B(s) has normal distribution with mean 0

and variance t–s. Notice that by taking s = 0 we have B(t)–B(0) ∼ N(0, t).

3. (Continuity of paths). B(t), t ≥ 0, are continuous functions of t.

The existence and continuity of Brownian motion is proved using Kolmogorov’s

extension and continuity theorem [Oks13].

35
Theorem B.6 (Quadratic variation). If t ≥ s,
E[(Bt – Bs )2 ] = t – s

More generally (in Rn ):

E[(Bt – Bs )2 ] = n(t – s)
Theorem B.7. Z
1 x2
E[f (Bt )] = p f (x)e– 2t dx
2πt R
Theorem B.8. The expected value of odd moments of Bt (other than one) are
zero. And the even moments are given by
(2k)! k
E[B2k
t ]= t
2k k!
Last but not least, two important identities, useful for further calculations
are:
1
Bj (Bj+1 – Bj ) = [B2j+1 – B2j – (Bj+1 – Bj )2 ]
2
1 1
B2j (Bj+1 – Bj ) = (B3j+1 – B3j ) – Bj (Bj+1 – Bj )2 – (Bj+1 – Bj )3
3 3

B.6 Riemann-Stieltjes Integral

Definition B.21 (Riemann-Stieltjes Integral). If ϕ is a continuous function on

[a, b] and F is a distribution function, we define the Riemann-Stieltjes Integral
of ϕ on [a, b] in relation to F (or weighted by F) as
Xn Zb
lim ϕ(yi )[F(xi+1 ) – F(xi ))] = ϕ(x) dF(x)
∥∆∥→0 a
i=1
where a = x1 < x2 < . . . < xn+1 – b, yi is an arbitrary point of [xi , xi+1 ], and
∥∆∥ = max1≤i≤n (xi+1 – xi ).

If F is a discrete random variable, then

Zb X
ϕ dF = ϕ(xi )p(xi )
a i:a<xi ≤b

And if F is a continuous random variable, then

Zb Zb
ϕ dF = ϕ(x)f (x) dx
a a

36
B.7 Modes of Convergence

Definition B.22. A space Lp (Ω, F, P) is the space of measurable functions in

which the p-th power of the absolute value is Lebesgue integrable. I.e., func-
tions of the form Z  1/p
∥f ∥p := |f |p dµ <∞
Ω

We can also write this as

Æ
p
∥X∥p := E[|X|p ] < ∞

A sequence of random variable can converge in different ways. The follow-

ing definitions are given in an increasing order of strength.

Definition B.23 (Convergence in Distribution). A sequence of random vari-

ables {Xn } converges in distribution to X if their distribution functions FXn (x)
converge to the distribution function FX (x) at any point of continuity of FX .
In other words, for any bounded function f : R −→ R we have

lim E[f (Xn )] = E[f (X)]

n→∞

I.e., Z Z
lim f (Xn ) dPn = f (X) dP
n→∞
Ωn Ω

Definition B.24 (Convergence in Probability). A sequence of random variables

{Xn } converges in probability to X if for any ϵ > 0, we have

lim P[|Xn (ω) – X(ω)| > ϵ] = 0

n→∞

Definition B.25 (Convergence Almost Surely). A sequence of random vari-

ables {Xn } converges almost surely (a.s.) to X if for any ω outside a set of
probability zero we have that

lim X (ω) = X(ω)

n→∞ n

Alternatively, we can write

P[Xn (ω) → X(ω)] = 1

37
Definition B.26 (Lp -Convergence). A sequence of random variables {Xn } con-
verges in Lp to X if {Xn } ⊆ Lp , p ∈ [1, ∞), and

lim ∥Xn – X∥p = 0

n→∞

I.e., E [|Xn |p ] < ∞ and

lim E |Xn – X|p = 0

 
n→∞

38
References

[Dur19] Rick Durrett. Probability: theory and examples, volume 49. Cam-
bridge university press, 2019.

[Eva12] Lawrence C Evans. An introduction to stochastic differential equations,

volume 82. American Mathematical Soc., 2012.

[Jak] Jake. How to draw brownian motions in tikz/pgf. [Link]

[Link]/a/59934. [Acesso em: 07/07/2022]. 3
[Kle12] Fima C Klebaner. Introduction to stochastic calculus with applications.
World Scientific Publishing Company, 2012.

[Oks13] Bernt Oksendal. Stochastic differential equations: an introduction with

applications. Springer Science & Business Media, 2013. 9, 11, 35

[S+ 04] Steven E Shreve et al. Stochastic calculus for finance II: Continuous-
time models, volume 11. Springer, 2004. 11

[Wik22] Wikipedia. Kalman filter — Wikipedia, the free encyclopedia.

[Link]
20filter&oldid=1091166052, 2022. [Online; accessed 06-July-
2022]. 1

39