Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Revised: 06/06/2019
© 2019 - WorldQuant University – All rights reserved.
1
MScFE 620 Discrete-time Stochastic Processes – Table of Contents
Part I .............................................................................................................................9
Part II ......................................................................................................................... 15
The study of probability theory begins with what is called a random experiment: an experiment
whose outcome cannot be pre-determined with certainty. An example of such a random
experiment is the tossing of three fair coins, with the (uncertain) outcome being the face shown by
each of them.
We associate with every random experiment a set Ω of all possible outcomes of the random
experiment. An outcome 𝜔 is simply an element of Ω.
We define an event to be a subset of Ω and we say that an event 𝐴 occurs if the outcome 𝜔 belongs
to 𝐴, i.e. 𝜔 ∈ 𝐴.
In our example of tossing three coins, the sample space can be chosen to be:
where H means “heads" and T means “tails". For instance, the event 𝐴 that exactly two heads come
up corresponds to the set
For complex models, it turns out that we have to restrict the class of events to a sub-collection
ℱ of subsets of Ω that satisfies the following properties:
𝒜1. ∅ ∈ ℱ;
⋃ 𝐴𝑛 ∈ ℱ.
𝑛=1
A collection of subsets of 𝛺 that satisfies 𝒜1, 𝒜2 and 𝒜3 is called an algebra in 𝛺, and an algebra
that also satisfies 𝜎𝒜4 is called a 𝜎 − 𝑎𝑙𝑔𝑒𝑏𝑟𝑎1. Thus, the collection of admissible events ℱ should
be a 𝜎-𝑎𝑙𝑔𝑒𝑏𝑟𝑎.1
Once we have the collection of events ℱ, we want to assign probabilities to each one of them. We
call a set function ℙ: ℱ → ℝ
̅ = ℝ ∪ {∞, −∞} a probability measure if ℙ satisfies the following:
ℙ1. ℙ(∅) = 0;
∞ ∞
ℙ (⋃ 𝐴𝑛 ) = ∑ ℙ(𝐴𝑛 ).
𝑛=1 𝑛=1
ℙ4. ℙ(Ω) = 1.
Upon careful observation, we see that there are several other set functions that behave like a
probability. Here are some examples of set functions that satisfy conditions ℙ1 − ℙ3:
o Length of subsets of the real line. This is the function 𝜆1 that assigns to certain subsets of
ℝ the “length” of that subset. If 𝐴 = (𝑎, 𝑏) is a bounded interval, then 𝜆1 (𝐴) = 𝑏 − 𝑎.
However, 𝜆1 is defined on a larger collection than intervals (the intervals by themselves
do not form a 𝜎-algebra). This measure, as we will refer to it, satisfies conditions ℙ1 −
ℙ3, but not ℙ4. Also, 𝜆1 (𝛺) = 𝜆1 (ℝ) = ∞, hence 𝜆1 takes values on the extended real
numbers ℝ̅ ≔ ℝ ∪ {−∞, ∞}. This measure is called the Lebesgue measure, after the
French mathematician Henri Lebesgue, who was instrumental in laying the foundations of
the field of Measure Theory.
1
Note that for a 𝜎-algebra, condition 𝒜3 can be removed since it is implied by condition 𝜎𝒜4
Given an arbitrary set Ω and a 𝜎-algebra ℱ of subsets of Ω, we say that a set function 𝜇: ℱ → ℝ
̅ is a
positive measure (or measure) on (Ω, ℱ) if it satisfies ℙ1, ℙ2, and ℙ3 (with 𝜇 replacing ℙ, of course).
If, in addition, the measure 𝜇 also satisfies ℙ4, then 𝜇 is a probability measure. This means that the
study of measure theory contains probability theory as a special case! The pair (Ω, ℱ) is called a
measurable space and the triple (Ω, ℱ, 𝜇) is called a measure space. If 𝜇 is a probability measure,
then (Ω, ℱ, 𝜇) is called a probability space. Thus, on our road to understanding probability theory
we are going to take a detour through abstract measure theory. We will prove many important
results that hold in general measure spaces and later apply them to probability spaces.
Here are some examples of algebras and 𝜎 -algebras. Check that these are indeed algebras or 𝜎-
algebras.
1 The “smallest” 𝜎-algebra on any set Ω is 𝒜 ≔ {∅, Ω}. Any other algebra (or 𝜎-algebra) must
necessarily contain it.
2 On the other hand, the “largest” 𝜎-algebra on a set Ω is the power set of Ω, 2Ω . (A power set
is a set of all subsets of a set.) Every algebra (or 𝜎-algebra) is contained in 2Ω .
Notice that in all the examples above the number of elements in each algebra is a power of 2. In
general, we say that a 𝜎-algebra ℱ is generated by blocks if there exists a countable partition
{ℬ𝓃 : 𝑛 = 0,1,2, … } of Ω such that
ℱ = {⋃ ℬ𝓃 : 𝐽 ⊆ ℕ}.
𝑛∈𝐽
to be subsets of ℝ that are finite unions of (open, closed, bounded, empty, etc.) intervals in
ℝ. Then 𝒜 is an algebra but not a 𝜎-algebra, since (for instance) 𝐴𝑛 : = {𝑛} = [𝑛, 𝑛] ∈ 𝐴 for
each 𝑛 ∈ ℕ, yet ∪𝑛∈ℕ 𝐴𝑛 = 𝑁 ∉ 𝒜.
Then ℱ is a 𝜎-algebra.
1 The Dirac Measure. Let Ω be any non-empty set and pick 𝑎 ∉ Ω. If we let ℱ = 2Ω and define
𝛿𝑎 : ℱ → ℝ by
1 if 𝑎 ∈ 𝐴
δ𝑎 (𝐴) ≔ {
0 otherwise,
|𝐴| if 𝐴 is finite
#(𝐴) ≔ {
∞ otherwise,
where |𝐴| represents the cardinality (number of elements) of 𝐴. Then this is a measure on
(Ω, 2Ω ), called the counting measure.
3 The Lebesgue Measure on ℝ. This measure 𝜆1 is defined on a 𝜎-algebra that includes all
the intervals and it extends the notion of length of an interval to such sets. If 𝐼 = (𝑎, 𝑏) is an
interval, then
λ1 (𝐼) = 𝑏 − 𝑎.
A property 𝑃 is said to be true almost everywhere with respect to a measure 𝜇 if the set of points
where 𝑃 is false is contained in a set of measure zero. We write 𝑃 − 𝜇 a.e., and we say 𝜇 − a.s. if 𝜇 is
a probability measure.
A special 𝜎-algebra is the Borel 𝜎-algebra on ℝ denoted by ℬ(ℝ). It is the 𝜎-algebra generated by
intervals in ℝ, i.e.,
The elements of ℬ(ℝ) are called Borel sets. We also define the Borel 𝜎-algebra of a subset 𝐴 ⊆ ℝ
as the restriction of ℬ(ℝ) to 𝐴; i.e.,
ℬ(𝐴) ≔ {𝐵 ∩ 𝐴: 𝐵 ∈ ℬ(ℝ)}.
Part I
Intuitively, a random variable is a numerical variable whose value is determined by the outcome of
a random experiment. This description suggests that we define a random variable 𝑋 as a function
𝑋: Ω → ℝ. Unfortunately, defining a random variable this way is not sufficient in calculating
probabilities about 𝑋. Indeed, to calculate ℙ({𝑋 ≤ 2}) for instance, we need to make sure that the
set 𝑋 −1 ((−∞, 2]) is in the 𝜎-algebra ℱ.
This definition of a random variable does not require a measure to be defined on (Ω, ℱ); all we
need is Ω and the 𝜎-algebra ℱ . We now extend this definition to general measurable spaces.
Let (Ω, ℱ) and (𝐸, ℰ) be measurable spaces and 𝑓: Ω → 𝐸 be a function. (We will write
𝑓: (Ω, ℱ) → (𝐸, ℰ).) We say that 𝑓 is ℱ/ℰ -measurable if 𝑓 −1 (𝐵) ∈ ℱ for every 𝐵 ∈ ℰ.
1 If (𝐸, ℰ) = (ℝ, ℬ(ℝ)), we say that 𝑓 is Borel measurable, or, simply, measurable.
2 If (Ω, ℱ, ℙ) is a probability space and (𝐸, ℰ) = (ℝ, ℬ(ℝ)), then 𝑓 is called a random variable
and we usually use the letter 𝑋 instead of 𝑓.
3 If (Ω, ℱ, ℙ) is a probability space and (𝐸, ℰ) = (ℝ𝑛 , ℬ(ℝ𝑛 )), then 𝑓 is called a random vector
and we also use the letter 𝑋 instead of 𝑓.
Whenever we say 𝑓 is measurable without mentioning (𝐸, ℰ), we will assume that we mean Borel
measurable – i.e. (𝐸, ℰ) = (ℝ, ℬ(ℝ)).
1 Let (Ω, ℱ) be any measurable space and 𝐴 ⊆ 𝛺. We define the indicator function 𝐼𝐴 : 𝛺 → ℝ
of 𝐴 as
1 𝜔∈𝐴
𝐼𝐴 (𝜔) = { .
0 𝜔∉𝐴
Then 𝐼𝐴 is measurable (Borel measurable) if and only if 𝐴 ∈ ℱ. To see this, note that for any
𝐵 ∈ ℬ(ℝ),
∅ 0,1 ∉ 𝐵
𝐴 1 ∈ 𝐵, 0 ∉ 𝐵
𝐼𝐴−1 𝐵 = { 𝑐
𝐴 0 ∈ 𝐵, 1 ∉ 𝐵
𝛺 0,1 ∈ 𝐵,
3 On (𝛺, {∅, 𝛺}), the only (Borel) measurable functions 𝑓: 𝛺 → ℝ are constants.
4 On (𝛺, {∅, 𝛺, 𝐴, 𝐴𝑐 })𝐴 ⊆ 𝛺 , the only measurable functions 𝑓: 𝛺 → ℝ are constant on 𝐴 and
constant on 𝐴𝑐 . That is, 𝑓: 𝛺 → ℝ is measurable if and only if 𝑓 = 𝛼𝐼𝐴 + 𝛽𝐼𝐴𝑐 for some
constants 𝛼, 𝛽 ∈ ℝ.
Let us now look at a concrete example. Consider the example where two fair coins are tossed and
only the outcome of the first coin is recorded. Let 𝐴 be the event that the outcome of the first coin
is a head and 𝐵 be the event that the outcome of the first coin is a tail. Then 𝛺 =
{𝐻𝐻, 𝐻𝑇, 𝑇𝐻, 𝑇𝑇} and
ℱ ≔ 𝜎({𝐴, 𝐵}) = 𝜎({{𝐻𝐻, 𝐻𝑇}, {𝑇𝐻, 𝑇𝑇}}) = {{𝐻𝐻, 𝐻𝑇}, {𝑇𝐻, 𝑇𝑇}, ∅, 𝛺}.
Notice that since 𝐴𝑐 = 𝐵 , then ℬ = {𝐴, 𝐵} are the blocks that generate ℱ . Both 𝐼𝐴 and 𝐼𝐵 are
random variables. In fact, the only random variables are linear combinations of the two. In other
words, 𝑋 is a random variable if it is constant on both 𝐴 and 𝐵. This simply means that 𝑋 is only
dependent on the outcome of the first toss.
The preceding example suggests that if ℱ is generated by (countably many) blocks, then the only
ℱ -measurable functions are the ones that are constant on the blocks. The following result makes
this more precise.
Theorem 1
Let 𝛺 = ℝ and choose ℱ = 𝜎({{𝜔}: 𝜔 ∈ ℤ}). Note that ℱ is generated by the following
(countably many) blocks:
𝑔(𝜔) 𝜔∈ℤ
𝑓(𝜔) = { .
𝑐 𝜔 ∈ ℝ\ℤ
If the 𝜎-algebra ℱ is not generated by countably many blocks, it is generally difficult to prove that
a function 𝑓 is measurable. The difficulty is caused by the fact that we don't have a general form of
an arbitrary Borel set 𝐵. (𝐵 could be an interval, a countably infinite set, or even the irrational
numbers.) Thus, checking the pre-image condition for arbitrary Borel sets might be challenging.
The next theorem tells us that we do not need to check this condition for every Borel set, but it is
enough to check the condition for a sub-collection that generates ℬ(ℝ).
Theorem 2
Let (𝛺, ℱ) and (𝐸, ℰ) be measurable spaces and 𝒜 ⊆ ℰ be a collection of subsets of 𝐸 such that
𝜎(𝒜) = ℰ . A function 𝑓: 𝛺, ℱ → (𝐸, ℰ) is ℱ/ℰ -measurable if and only if for each 𝐵 ∈ 𝒜,
𝑓 −1 (𝐵) ∈ ℱ .
1 𝑓 is Borel measurable
2 {𝑓 < 𝑐} ≔ 𝑓 −1 ((−∞, 𝑐)) ∈ ℱ for every 𝑐 ∈ ℝ
3 {𝑓 ≤ 𝑐} ≔ 𝑓 −1 ((−∞, 𝑐]) ∈ ℱ for every 𝑐 ∈ ℝ
4 {𝑓 > 𝑐} ≔ 𝑓 −1 ((𝑐, ∞)) ∈ ℱ for every 𝑐 ∈ ℝ
5 {𝑓 ≥ 𝑐} ≔ 𝑓 −1 ([𝑐, ∞)) ∈ ℱ for every 𝑐 ∈ ℝ
6 {𝑎 < 𝑓 < 𝑏} ≔ 𝑓 −1 ((𝑎, 𝑏)) ∈ ℱ for every 𝑎, 𝑏 ∈ ℝ
7 {𝑎 ≤ 𝑓 < 𝑏} ≔ 𝑓 −1 ([𝑎, 𝑏)) ∈ ℱ for every 𝑎, 𝑏 ∈ ℝ
8 {𝑎 < 𝑓 ≤ 𝑏} ≔ 𝑓 −1 ((𝑎, 𝑏]) ∈ ℱ for every 𝑎, 𝑏 ∈ ℝ
9 {𝑎 ≤ 𝑓 ≤ 𝑏} ≔ 𝑓 −1 ([𝑎, 𝑏]) ∈ ℱ for every 𝑎, 𝑏 ∈ ℝ
Let 𝛺 = [0,1], ℱ = ℬ([0,1]) ≔ 𝜎({(𝑎, 𝑏): 0 ≤ 𝑎 < 𝑏 ≤ 1}) = Borel subsets of [0,1] and ℙ(𝐴) = 𝜆1 (𝐴)
for each 𝐴 ∈ ℱ. Define 𝑋: 𝛺 → ℝ as
1 1
0≤𝜔<
𝑋(𝜔) ≔ 4 2.
1
𝜔 ≤𝜔≤1
2
Then 𝑋 is a random variable. For each 𝑐 ∈ ℝ, we have
𝛺 𝑐≥1
1
[0, 𝑐] ≤𝑐<1
{𝑋 ≤ 𝑐} = 2
1 1,
[0, ] 0≤𝑐<
2 2
{ ∅ 𝑐<0
Let (𝛺, ℱ), (𝐸, ℰ) and (𝐺, 𝒢) be measurable spaces and 𝑓: (𝛺, ℱ) → (𝐸, ℰ), 𝑔: (𝐸, ℰ) → (𝐺, 𝒢) be
measurable functions. Then ℎ ≔ 𝑔 ∘ 𝑓: (𝛺, ℱ) → (𝐺, 𝒢) is ℱ/𝒢-measurable.
2
Strictly speaking, these are extended random variables. See the next set
of notes.
𝛺 𝑐≥2
1
[ , 1] 1≤𝑐<2
2
{𝑋 ≤ 𝑐} = 1 1
[ , √𝑐] ≤𝑐≤1
2 4
1
{ ∅ 𝑐< .
4
Thus
1 1 1
𝜎(𝑋) = 𝜎({{𝑋 ≤ 𝑐} : 𝑐 ∈ ℝ}) = 𝜎 ({𝛺, [ , 1] , ∅} ∪ {[ , 𝑏] : ≤ 𝑏 ≤ 1})
2 2 2
1 1
= 𝜎 {{[ , 𝑏] : ≤ 𝑏 ≤ 1}}
2 2
1 1 1
= {𝐵 ⊆ [0,1]: 𝐵 ∈ ℬ ([ , 1]) or 𝐵 = [0, ) ∪ 𝐵′ where 𝐵′ ∈ ℬ ([ , 1])}
2 2 2
1
The set [0, ) is a block of 𝜎(𝑋) since 𝑋 does not distinguish between elements of this set, implying
2
1
that if a function 𝑌: 𝛺 → ℝ is 𝜎(𝑋)-measurable, then 𝑌 must also be constant on [0, ).
2
Theorem 3:
Doob-Dynkin Lemma
Let 𝛺 be a set and 𝑋, 𝑌: 𝛺 → ℝ be functions. Then 𝑌 is 𝜎(𝑋)-measurable if and only if there exists a
Borel measurable function 𝑔: (ℝ, ℬ(ℝ)) → (ℝ, ℬ(ℝ)) such that 𝑌 = 𝑔(𝑋).
Part II
Let (Ω, ℱ, ℙ) be a probability space, 𝑋: (Ω, ℱ) → (ℝ, ℬ(ℝ)) be a random variable (i.e. 𝑋 ∈ 𝑚ℱ) and
𝐵 ∈ ℬ(ℝ). To calculate the probability that 𝑋 ∈ 𝐵 , we first find 𝑋 −1 (𝐵) and then evaluate ℙ
(𝑋 −1 (𝐵)). This means that we can define a set function ℙ𝑋 : ℬ(ℝ) → [0,1] by:
∞ ∞ ∞ ∞ ∞
−1
ℙ𝑥 (⋃ 𝐵𝑛 ) = ℙ (𝑋 (⋃ 𝐵𝑛 )) = ℙ (⋃ 𝑋 −1 (𝐵𝑛 )) = ∑ ℙ(𝑋 −1 (𝐵𝑛 )) = ∑ ℙ𝑋 (𝐵𝑛 ).
𝑛=1 𝑛=1 𝑛=1 𝑛=1 𝑛=1
Thus, each random variable induces a measure on ℬ(ℝ). We will call ℙ𝑋 the probability
distribution or Law of 𝑋.
Consider the random experiment of tossing 2 fair coins and let 𝑋 count the number of heads.
1
ℙ= ∑ 𝛿 .
4 𝜔
𝜔∈Ω
1 1 1
ℙ𝑋 = δ0 + δ1 + δ2 .
4 2 4
We will later see that if 𝑋 is a “discrete” random variable, then ℙ𝑋 is a linear combination of Dirac
measures.
We now calculate ℙ𝑋 .
ℙ𝑋 ((−∞, 𝑥]) = ℙ({𝑋 ≤ 𝑥}) = ℙ({ω ∈ (0,1]: 4ω ≤ 𝑥}) = ℙ((0,1] ∩ (−∞, 𝑥/4])
0 𝑥 ≤0
𝑥
={ 0<𝑥<4
4
1 𝑥 ≥4.
λ1 ((0,4] ∩ 𝐵)
ℙ𝑋 (𝐵) = , 𝐵 ∈ ℬ(ℝ).
4
Let 𝑋: (Ω, ℱ, ℙ) → ℝ be a random variable. The Cumulative Distribution Function (CDF) of 𝑋 is the
function 𝐹𝑋 : ℝ → ℝ defined by:
The CDF characterizes the distribution of a random variable. That is, if 𝑋, 𝑌: (Ω, ℱ) → (ℝ, ℬ(ℝ))
are random variables, then:
ℙ𝑋 = ℙ𝑌 ⟺ 𝐹𝑋 = 𝐹𝑌 .
The converse of the above is also true. That is, if you start with an arbitrary function that satisfies
the above properties, then that function is a CDF of some random variable.
∑|𝑏𝑖 − 𝑎𝑖 | < 𝛿,
𝑖=1
Then
ℙ𝑋 = ∑ 𝑝𝑋 (𝑥𝑛 )δ𝑥𝑛 ,
𝑛=1
where the function 𝑝𝑋 : ℝ → ℝ, called the probability mass function (PMF) of 𝑋, is defined by
𝑝𝑋 (𝑥) = ℙ𝑋 ({𝑥}), 𝑥 ∈ ℝ.
The coin toss random variable counting the number of heads is obviously discrete. In fact, every
random variable with a countable range is discrete.
Unit 3: Expectation
In this section, we develop the Lebesgue integral of a measurable function 𝑓 with respect to a
measure 𝜇 . In the special case when 𝜇 = ℙ is a probability measure and 𝑓 = 𝑋 is a random
variable, this integral is the expected value of 𝑋.
Our intuition about integration is that it represents “aggregation” or “summing up”. We will see
that by an appropriate choice of 𝜇 and Ω, the integral will cover the many forms of aggregation we
are used to in mathematics, including summation, expectation, and calculating area/volume using
the Riemann integral. The advantage now is that we can look at all these forms of aggregation as
essentially doing the same thing: integrating with respect to a measure. This point of view will help
us prove one theorem that applies to all these special cases. In applications to probability theory, it
will enable us to treat discrete, continuous, and other random variables the same way. There will
be no need, for instance, to prove a theorem for discrete random variables using summation and
then prove the same theorem again for absolutely continuous random variables using integrals.
Let (Ω, ℱ, 𝜇) be a measure space and 𝑓: Ω → ℝ be a measurable function. In this section, we want
to define the integral of 𝑓 with respect to the measure 𝜇 , denoted by
∫ 𝑓 𝑑𝜇.
𝛺
𝔼(𝑋) ≔ ∫ 𝑋 𝑑ℙ.
Ω
For reasons we will see later, it will be useful to define the integral for extended measurable
̅ ≔ ℝ ∪ {∞, −∞}, we define the Borel 𝜎-algebra
function. First, on the extended real numbers ℝ
ℬ(ℝ̅ ) to be
̅ ) ≔ {𝐵 ⊆ ℝ
ℬ(ℝ ̅ : 𝐵 ∩ ℝ ∈ ℬ(ℝ)}
̅ , ℬ(ℝ
A function 𝑓: (𝛺, ℱ) → (ℝ ̅ )) that is ℱ/ℬ(ℝ
̅ )-measurable will be called an extended
measurable function. If 𝑓 = 𝑋 is a random variable, we will call 𝑋 an extended random variable.
Throughout this chapter we will assume that all measurable functions are extended measurable
functions.
1 We first define the integral for the case when 𝑓 = 𝐼𝐴 is an indicator function of a
measurable set 𝐴 ∈ ℱ .
2 We will then extend the definition to when 𝑓 is positive3 and is a (finite) linear combination
of indicators. Such functions are called (positive) simple functions.
3 We then extend the definition to all positive measurable functions and prove useful results
that aid with the calculation of the integral in many concrete examples.
4 Finally, we extend the definition to general measurable (sign-changing) functions 𝑓 and
define the notion of integrability. This will be covered in the next section.
We begin with indicators. Throughout the rest of the chapter, (𝛺, ℱ, 𝜇) is a measure space and
̅ , ℬ(ℝ
𝑓: (𝛺, ℱ) → (ℝ ̅ )) is ℱ/ℬ(ℝ
̅ )-measurable.
If 𝑓 = 𝐼𝐴 for some 𝐴 ∈ ℱ , then we define the integral of 𝑓 with respect to the measure 𝜇 to be
There is nothing surprising about this definition; it is consistent with our intuition of aggregation.
We move on to simple functions.
𝑠 = ∑ 𝑠𝑘 𝐼𝐴𝑘 ,
𝑘=1
3
We should really be saying “non-negative”.
where 𝐴𝑘 = 𝑠 −1 {(𝑠𝑘 )} ∈ ℱ for 𝑘 = 1,2, … , 𝑛 are pairwise disjoint. We will call this the canonical
representation of a simple function and assume this representation when we talk about a simple
function.
𝑠 = ∑ 𝑠𝑘 𝐼𝐴𝑘 .
𝑘=1
Note that the value of this integral could be ∞, even if 𝑓 is finite. The probability space equivalent
of this definition is that if 𝑋 is a simple random variable with
𝑋 = ∑ 𝑠𝑘 𝐼𝐴𝑘 ,
𝑘=1
then
𝑠 = ∑ 𝑠𝑘 𝐼𝐴𝑘 ,
𝑘=1
then
When 𝜇 = ℙ is a probability measure and 𝑓 = 𝑋 is a random variable, then this definition means
If 𝑓 ∈ 𝒮 + , then this definition gives the same result as the previous definition of the integral for
positive simple functions. That is
even when 𝑓 ∈ 𝒮 + .
∫ 𝑓 𝑑𝜇 ≤ ∫ 𝑔 𝑑μ.
𝛺 𝛺
This definition is not user-friendly. We would like to avoid having to use it to calculate integrals in
concrete cases. The following two results will help us do just that. First, we need some definitions.
Let (𝑓𝑛 ) be a sequence of measurable functions. We say that (𝑓𝑛 ) is increasing if for each 𝜔 ∈ 𝛺
and 𝑛 ≥ 1, 𝑓𝑛 (𝜔) ≤ 𝑓𝑛+1 (𝜔). If 𝑓 ∈ 𝑚ℱ , we say that (𝑓𝑛 ) converges point-wise to 𝑓 if for each
𝜔 ∈ 𝛺, 𝑙𝑖𝑚 𝑓𝑛 (𝜔) = 𝑓(𝜔). We will simply write 𝑓𝑛 → 𝑓 to denote point-wise convergence of (𝑓𝑛 )
𝑛→∞
to 𝑓 .
Theorem 4:
Let 𝑚ℱ + and (𝑓𝑛 ) be an increasing sequence of positive measurable functions that converges to 𝑓
point-wise. Then
That is,
The probability space equivalent statement is that if (𝑋𝑛 ) is an increasing sequence of non-
negative random variables that converges point-wise to a non-negative random variable 𝑋, then
This result says that we can interchange between taking limits and evaluating integrals in the case
when the sequence of functions is increasing and positive. We will see other theorems of this
nature in the next section. The result will be helpful in calculating the integral of 𝑓 ∈ 𝑚ℱ + if we
already know how to calculate the integrals of the 𝑓𝑛 's. In other words, we need to be able to find a
sequence (𝑓𝑛 ) such that
1 (𝑓𝑛 ) is increasing,
2 𝑓𝑛 ∈ 𝑚ℱ + for each 𝑛 ≥ 1,
3 (𝑓𝑛 ) converges point-wise to 𝑓, and
4 ∫𝛺 𝑓𝑛 𝑑 is easy to calculate.
Theorem 5:
If 𝑓 ∈ 𝑚ℱ + , then there exists an increasing sequence of positive simple functions (𝑠𝑛 ) such that
(𝑠𝑛 ) converges point-wise to 𝑓 .
by (MON).
Let us see this theorem in action. Consider the measure space ([0,1], ℬ([0,1]), 𝜆1 ) and 𝑓: [0,1] →
̅ defined by 𝑓(𝑥) = 𝑥 . We have
ℝ
𝑛2𝑛 −1
𝑘
𝑠𝑛 = ∑ 𝐼 𝑘 𝑘+1 + 𝑛𝐼{𝑓≥𝑛} ,
2𝑛 {2𝑛 ≤𝑓< 2𝑛 }
𝑘=0
where
𝑘 𝑘+1 𝑘 𝑘+1 𝑘 𝑘+1
{ 𝑛
≤ 𝑓 < 𝑛 } = {𝑥 ∈ [0,1]: 𝑛 ≤ 𝑓(𝑥) < 𝑛 } = [0,1] ∩ [ 𝑛 , 𝑛 ),
2 2 2 2 2 2
2𝑛 −1 2𝑛 −1
𝑘 𝑘
𝑠𝑛 = ∑ 𝑛 𝐼 𝑘 𝑘+1 = ∑ 𝑛 𝐼 𝑘 𝑘+1
2 {2𝑛≤𝑓< 2𝑛 } 2 [2𝑛, 2𝑛 )
𝑘=0 𝑘=0
2𝑛 −1 2𝑛 −1
𝑘 𝑘 𝑘+1 𝑘 1 1
∫ 𝑠𝑛 𝑑𝜆1 = ∑ 𝑛 𝜆1 ([ 𝑛 , 𝑛 )) = ∑ 2𝑛 = (1 − 𝑛 ).
[0,1] 2 2 2 2 2 2
𝑘=0 𝑘=0
1 1 1
∫ 𝑓 𝑑𝜇 = 𝑙𝑖𝑚 ∫ 𝑠𝑛 𝑑𝜇 = 𝑙𝑖𝑚 (1 − 𝑛 ) = .
[0,1] 𝑛→∞ [0,1] 𝑛→∞ 2 2 2
𝑏
∫ 𝑓 𝑑𝜆1 = ∫ 𝑓(𝑥)𝑑𝑥 .
[𝑎,𝑏] 𝑎
That is, integrating with respect to the Lebesgue measure is the same as ordinary Riemann
integration when 𝑓 is Riemann integrable on a bounded interval.
Also note that ([0,1], ℬ([0,1]), 𝜆1 ) is actually a probability space, therefore 𝑓 is random variable.
It is easy to show that 𝑓 ∼ 𝑈(0,1), which is consistent with the fact the integral represents the
expected value of 𝑓 .
We now extend the definition of the integral to functions in 𝑚ℱ that may change sign. If 𝑓 ∈ 𝑚ℱ ,
we define the positive part of 𝑓 as the function 𝑓 + ≔ 𝑚𝑎𝑥{𝑓, 0}, and the negative part of 𝑓 as
𝑓 − ≔ 𝑚𝑎𝑥{−𝑓, 0}. It is easy to show that both 𝑓 + , 𝑓 − ∈ 𝑚ℱ + , 𝑓 = 𝑓 + 𝑓 − and |𝑓| = 𝑓 + + 𝑓 − .
provided at least one of the integrals on the right is finite. In that case, we say that the integral
exists, otherwise, we say that the integral does not exist. We will say that 𝑓 is integrable if both
integrals on the right are finite, i.e., if
We will denote by ℒ 1 ≔ ℒ 1 (𝛺, ℱ, 𝜇) the set of all integrable functions. In general, if 𝑝 ∈ [1, ∞),
we define
provided the integral exists. If 𝜇 = ℙ is a probability space and 𝑓 = 𝑋 is a random variable, we will
write this as
𝔼(𝑋; 𝐴) ≔ 𝔼(𝑋𝐼𝐴 ).
Theorem 6:
Let (𝑓𝑛 ) be a sequence of measurable functions such that there exists 𝑔 ∈ ℒ 1 ∩ 𝑚ℱ + such that
|𝑓𝑛 (𝜔)| ≤ 𝑔(𝜔) for every 𝜔 ∈ 𝛺 and 𝑛 ≥ 1. If 𝑓 ∈ 𝑚ℱ and 𝑙𝑖𝑚 𝑓𝑛 (𝜔) = 𝑓(𝜔) for every 𝜔 ∈ 𝛺,
𝑛→∞
then
1 𝑓 ∈ ℒ1,
3 ∫𝛺 𝑓 𝑑𝜇 = 𝑙𝑖𝑚 ∫𝛺 𝑓𝑛 .
𝑛→∞
In the language of probability, the theorem says that if (𝑋𝑛 ) is a sequence of random variables
such that there exists a random variable 𝑌 ∈ ℒ 1 ∩ 𝑚ℱ + such that |𝑋𝑛 (𝜔)| ≤ 𝑌(𝜔) for every 𝜔 ∈
𝛺 , 𝑛 ≥ 1, and 𝑋 ∈ 𝑚ℱ is such that 𝑙𝑖𝑚 𝑋𝑛 (𝜔) = 𝑋(𝜔) for every 𝜔 ∈ 𝛺, then
𝑛→∞
1 𝑋 ∈ ℒ1,
Theorem 7
∫ 𝑓 𝑑# = ∑ 𝑓(𝑛),
ℕ 𝑛∈ℕ
∫ 𝑓 𝑑𝛿𝑎 = 𝑓(𝑎).
𝛺
∫ 𝑓 𝑑𝜇 = ∑ 𝑎𝑘 ∫ 𝑓 𝑑𝜇𝑘 ,
𝛺 𝑘 𝛺
provided 𝑓 is 𝜇 -integrable.
μ
ν𝑓 (𝐴) = ∫ 𝑓 𝑑μ, 𝐴 ∈ ℱ.
𝐴
𝝁 𝝁
Then 𝝂𝒇 is a positive measure (exercise). Also, if 𝝁(𝑨) = 𝟎 then 𝝂𝒇 (𝑨) = 0. Likewise, if (Ω, ℱ, ℙ) is
a probability space and 𝑓 ∈ 𝑚ℱ + is a non-negative random variable with 𝔼(𝑋) = 1, then the
function Q : ℱ → [0,1] defined by
≥ 1.
𝜇
We have seen above that 𝜈𝑓 ≪ 𝜇. The Radon-Nikodym Theorem tells us that all measures that are
absolutely continuous with respect to 𝜇 arise in this way (at least in the 𝜎 −finite case).
Theorem 8:
Radon-Nikodym
Let (𝛺, ℱ) be a measurable space, and 𝜇, 𝜈 be two 𝜎 −finite measures on (𝛺, ℱ) such that 𝑣 ≪ 𝜇.
Then there exists 𝑓 ∈ 𝑚ℱ + such that 𝜈 = 𝜈𝑓𝜇 , in the sense that
ν(𝐴) = ∫ 𝑓 𝑑μ, 𝐴 ∈ ℱ.
𝐴
μ
Furthermore, if 𝑔 ∈ 𝑚ℱ + is such that ν = ν𝑔 , then 𝑓 = 𝑔 𝜇 − 𝑎. 𝑒.. Thus, 𝑓 is unique up to a null set.
𝜇
The function 𝒇 such that ν = 𝜈𝑓 is called the Radon-Nikodym derivative or density of 𝜈 with
respect to 𝜇 and is denoted by
𝑑ν
𝑓 =: .
𝑑μ
1 If (Ω, ℱ, 𝜇) is a measure space, then define (for any constant 𝑎 > 0), 𝜈: = 𝑎𝜇. We have that
𝑣 ≪ 𝜇 and
𝑑𝜈
= 𝑎, 𝜇 − 𝑎 .𝑒
𝑑𝜇
2 If (Ω, ℱ, 𝜇) is a measure space and 𝐹 ∈ ℱ, define 𝜈(𝐴) : = 𝜇(𝐴 ∩ 𝐹) for every 𝐴 ∈ ℱ. Then 𝜈
is also a measure on (Ω, ℱ) and 𝑣 ≪ 𝜇 with
𝑑𝜈
= 𝐼𝐹 , 𝜇 − 𝑎. 𝑒..
𝑑𝜇
Theorem 9
Let (𝛺, ℱ) be a measurable space and 𝜈, 𝜇 be measures with 𝑣 ≪ 𝜇. If 𝑔 ∈ ℒ 1 (𝛺, ℱ, 𝑣), then
𝑔𝑑𝜈
𝑑𝜇
∈ ℒ1 (𝛺, ℱ, 𝜇) and
𝑑𝜈
∫ 𝑔 𝑑𝜈 = ∫ 𝑔 𝑑𝜇 , for every 𝑨 ∈ 𝓕.
𝐴 𝐴 𝑑𝜇
𝑑(𝑎𝑣 + 𝑏𝜆) 𝑑𝑣 𝑑𝜆
=𝑎 +𝑏 , 𝜇 − 𝑎. 𝑒.
𝑑𝜇 𝑑𝜇 𝑑𝜇
𝑑𝜆 𝑑𝜆 𝑑𝑣
= .
𝑑𝜇 𝑑𝑣 𝑑𝜇
𝑑𝑣 𝑑𝜇
3 If 𝑣 ≡ 𝜇, then > 0 , 𝜇 − 𝑎. 𝑒. , > 0 𝑣 − 𝑎. 𝑒. , and
𝑑𝜇 𝑑𝑣
𝑑𝜇 1
= 𝑣 − 𝑎. 𝑒.
𝑑𝑣 𝑑𝑣
𝑑𝜇
Problem Set
Problem 1
𝑓(𝑥) = (𝑥 + 2)3 .
Solution:
𝑦 = 𝑓(𝑥) = (𝑥 + 2)3 .
𝑥 = 𝑦1/3 − 2.
Now, we get the value for the extremes of the given set, 𝑓 −1 ((−1,1)). For −1, we have 𝑦 1/3 −
2 = −1 − 2 = −3 and for 1 we get 𝑦1/3 − 2 = 1 − 2 = −1. Thus, the solution is the interval
(−3, −1).
Problem 2
Consider the measure space ([0,1], ℬ([0,1]), 𝜇 ≔ 𝜆1 + 2𝛿0 ) and the measurable function
𝑓(𝜔) = 2𝜔. Find
∫ 𝑓 𝑑𝜇.
[0,1]
Solution:
𝑏
∫ 𝑓(𝑥) 𝑑𝑥 = ∫ 𝑓 𝑑𝜆1 .
𝑎 [𝑎,𝑏]
We will need to apply this together with the following property: For any set 𝛺 and 𝑎 ∈ 𝛺 , we have
∫ 𝑓 𝑑𝛿𝑎 = 𝑓(𝑎).
𝛺
Thus,
1
∫ 2𝑥 𝑑𝜆1 = ∫ 2𝑥 𝑑𝑥 = 1.
[0,1] 0
∫ 4𝑥 𝑑𝛿0 = 𝑓(0) = 4 ∗ 0 = 0.
[0,1]
Problem 3
Consider the measurable space (ℝ, ℬ(ℝ)) and the random variable 𝑋(𝜔) = |𝜔|. Define the
measures ℙ and ℙ∗ as
𝑑 ℙ∗
ℙ ≔ 0.3𝛿0 + 0.4𝛿1 + 0.3𝛿2 , = 𝑋.
𝑑ℙ
Solution:
Thus, as we are interested in the interval ℙ∗ ([0,1]), we need to find the probability of 𝜔 = 0 and
𝜔 = 1.
Problem 4
Consider the measurable space (ℝ, ℬ(ℝ)) and the random variables 𝑋 and 𝑌 by 𝑋(𝜔) =
2𝐼(−∞,0] (𝜔) and 𝑌(𝜔) = 𝜔2 . Define the measures ℙ and ℙ∗ as
𝑑ℙ∗
ℙ ≔ 0.3𝛿0 + 0.5𝛿1 + 0.2𝛿−1 , ≔ 𝑋.
𝑑ℙ
Solution:
Thus, as we already know 𝑌(𝜔) (since it is given), we just need to compute 𝑃∗ (see Problem 3) as
follows:
Problem 5
Consider the measure space (ℕ, 2ℕ , 𝜇 = #) and the measurable function 𝑓(𝜔) = 𝑒 −𝜔 . Find
∫ 𝑓 𝑑𝜇.
ℕ
Solution:
From the lecture notes we know that on (ℕ, 2ℕ , #), integration reduces to summation:
∫ 𝑓 𝑑# = ∑ 𝑓(𝑛),
ℕ 𝑛∈ℕ
provided the series converges. Thus, our problem could be re-written as follows,
−𝑛
1 𝑛
∫ 𝑓𝑑# = ∑ 𝑓(𝑛) = ∑ 𝑒 = ∑( ) .
ℕ 𝑒
𝑛∈ℕ 𝑛∈ℕ 𝑛∈ℕ
The last expression above is just a geometric series with 𝑟 = 1/𝑒 and, thus the infinity sum is
computed as,
−𝑛
1 𝑛 1
∑𝑒 = ∑( ) = ,
𝑒 1 − 𝑒 −1
𝑛∈ℕ 𝑛∈ℕ
In this module, you are required to complete a collaborative review task, which is designed to test
your ability to apply and analyze the knowledge you have learned during the week.
Brief
𝑑ℙ𝑋 1 1 2
(𝑥) = 𝑒 −2 𝑥 , 𝑥 ∈ ℝ.
d𝜆1 √2𝜋
Construct (i.e., give an example of) a probability space (Ω, ℱ, ℙ) and a random variable 𝑋: Ω → ℝ on
(Ω, ℱ, ℙ) such that 𝑋 has a standard normal distribution. In your example, be sure to verify that 𝑋
does indeed have a standard normal distribution.