Math 5010

Introduction to Probability
Based on D. Stirzaker’s book

Cambridge University Press

and

R.B. Ash’s book

Dover Publications

Davar Khoshnevisan
University of Utah
Firas Rassoul-Agha
University of Utah

Last Scribed October 3, 2014

 Acknowledgments: We thank Daniel Conus for considerably expanding
the list of exercises. We thank Ryan Viertel for catching many typos. We
thank Daniel Conus, Stewart Ethier, Nicos Georgiou, Kunwoo Kim, and
Ryan Viertel for valuable comments.
 Lecture 1

Example 1.1 (The Monty Hall problem, Steve Selvin, 1975). Three doors:
behind one is a nice prize; behind the other two lie goats. You choose a
door. The host (Monty Hall) knows where the prize is and opens a door
that has a goat. He gives you the option of changing your choice to the
remaining unopened door. Should you take his offer?
The answer is “yes.” Indeed, if you do not change your mind, then
to win you must choose the prize right from the start. This is 1 in 3. If
you do change your mind, then you win if you choose a goat right from
the start (for then the host opens the other door with the goat and when
you switch you have the prize). This is 2 in 3. Your chances double if you
switch.

1. The sample space, events, and outcomes

We need a math model for describing “random” events that result from
performing an “experiment.”
We cannot use “frequency of occurrence” as a model because it does
not have the power of “prediction.” For instance, if our definition of a fair
coin is that the frequency of heads has to converge to 1/2 as the number
of tosses grows to infinity, then we have done things backwards: to predict Model is used to predict the
behavior of a experiment
how a fair coin would behave, we would first have to toss the coin infinitely without have to do the
many times to verify it is a fair coin. This beats the whole purpose of a experiment.

model, which is to predict the behavior without having to toss the coin
that many times. What we should do is first define a model, then draw
from it the prediction about the frequency of heads.

1
2 1

Here is how we will do things. First, we define a state space (or sample
space) that we will denote by Ω. We think of the elements of Ω as outcomes
of the experiment.
Then we specify a collection F of subsets of Ω. Each of these subsets
Space: All possible results
is called an event. We will “only be allowed” to talk about the probability Events: Subgroups of the
results in the space.
of these events; i.e. it shall be illegal to ask about the probability of an
event not in F . Probability only for Events,
When Ω is finite, F can be taken to be the collection of all its subsets. there is no probability of Space
(It will be always 1)
In this case, we are allowed to talk about the probability of any event.
Thus the next step is to assign a probability P(A) to every A ∈ F . We
will talk about this after the following examples.

Example 1.2. Toss a coin. A natural sample space is

Ω = {H, T }.

Since Ω is finite we let F be all subsets of Ω:

F = ∅, {H}, {T }, {H, T } .

Note that the event {H, T } reads: “we tossed the coin and got heads or
tails” NOT “heads and tails”!

Example 1.3. Roll a six-sided die; what is the probability of rolling a six?
First, write a sample space. Here is a natural one:

Ω = {1, 2, 3, 4, 5, 6}.

In this case, Ω is finite and we want F to be the collection of all subsets

of Ω. That is,

F = ∅ , {1} , . . . , {6} , {1, 2} , . . . , {1, 6} , . . . , {1, 2, . . . , 6} .

For example, the event that we rolled and got an even outcome is the event
{2, 4, 6}. The event that we got an odd number or a 6 is {1, 3, 5, 6}. And so
on.

Example 1.4. Toss two coins; what is the probability that we get two
heads? A natural sample space is

Ω = (H1 , H2 ) , (H1 , T2 ) , (T1 , H2 ) , (T1 , T2 ) .

Once we have readied a sample space Ω and an event-space F , we

need to assign a probability to every event. This assignment cannot be
made at whim; it has to satisfy some properties.
3. Algebra of events 3

2. Rules of probability
Rule 1: Probability always
Rule 1. 0 6 P(A) 6 1 for every event A. between 0 and 1

Rule 2. P(Ω) = 1. “Something will happen with probability one.” Rule 2: Probability of the Space
is 1

Rule 3 (Addition rule). If A and B are disjoint events [i.e., A ∩ B = ∅], Rule 3: P( A or B) = P(A)+P(B)
P(Space)=P(Space)+P(empty)
then the probability that at least one of the two occurs is the sum of the -> P(empty)=0
individual probabilities. More precisely put,
P(A ∪ B) = P(A) + P(B).

Note that Ω ∩ ∅ = ∅ and hence these two events are disjoint. Further-
more, Ω ∪ ∅ = Ω. So Rule 3, when applied to the two disjoint events Ω
and ∅, implies the following:
P(Ω) = P(Ω) + P(∅).
Canceling P(Ω) on both sides gives that P(∅) = 0. This makes sense: the
probability that nothing happens is zero.
Example 1.5 (Coin toss model). We have seen that to model a coin toss
we set Ω = {H, T } and let F be all subsets of Ω. Now, we can assign
probabilities to the events in F . We know P{H, T } = 1 by Rule 2. Also, we
have just seen that P(∅) = 0. To complete the model we just need to assign
a probability to each of {H} and {T }. The numbers have to be between 0
and 1 by Rule 1. So pick p ∈ [0, 1] and let P{H} = p. By Rule 3 we have
P{H} + P{T } = P{H, T } = 1. Rule 3: The sum of two unlinked events is the Probability of one or other
happening. In this case, the entire space. By Rule 2, P(Space) = 1

Thus, P{T } = 1 − p.
This is our first probability model. It models a coin with heads loaded
to come out with probability p and tails with probability 1 − p.

Let us recall some set-theoretical notation.

3. Algebra of events
Given two sets A and B that are subsets of some bigger set Ω:
União é a probabilidade de um OU outro ocorrer. União soma os dois conjuntos
• A∪B is the “union” of the two and consists of elements belonging
to either set; i.e. x ∈ A ∪ B is equivalent to x ∈ A or x ∈ B.
• A ∩ B is the “intersection” of the two and consists of elements
Intersecção é AND, quando os
shared by the two sets; i.e. x ∈ A ∩ B is equivalent to x ∈ A and dois ocorrer ao mesmo tempo.
x ∈ B.
• Ac is the “complement” of A and consists of elements in Ω that Complement é tudo que não faz
are not in A. parte daquele evento.
 4 Elements of A not in B 1
Toda intersecção indo

de A1 com A2 até A1

com An vai ser o próprio We write A r B for A ∩ Bc ; i.e. elements in A but not in B.
A1 porque para ser Clearly, A∪B = B∪A and A∩B = B∩A. Also, A∪(B∪C) = (A∪B)∪C,
which we simply write as A ∪ B ∪ C. Thus, it is clear what is meant by
elemento de qualquer
A1 ∪ · · · ∪ An . Similarly for intersections.
intersecção com A1 tem
We write A ⊆ B when A is inside B; i.e. x ∈ A implies x ∈ B. It is clear
que pertencer ao A1 that if A ⊆ B, then A ∩ B = A and A ∪ B = B. Thus, if A1 ⊆ A2 ⊆ · · · ⊆ An , Se A está dentro de B, então A
interseção com B Vai ser sempre A
then ∩n n
i=1 Ai = A1 and ∪i=1 Ai = An .
It is clear that A ∩ Ac = ∅ and A ∪ Ac = Ω. It is also not very hard to
see that (A ∪ B)c = Ac ∩ Bc . (Not being in A or B is the same thing as not
being in A and not being in B.) Similarly, (A ∩ B)c = Ac ∪ Bc .
Toda união (ser elemento
deum OU outro - Soma os Todo elemento que não pertencer a A mesma coisa aqui: dizer que o elemento que não pertence a A e B (interseção
elemtentos dos dois conjuntos) A ou B é o mesmo que dizer que o ocorrendo ao mesmo tempo) é o mesmo que dizer que o elemento não pertence a A ou
de A1 ou A2 até A1 ou An vai elemento não pertence a A ou não não pertence a B (para não fazer parte da intersecção). Ou ele A ou ele é B
ser o próprio An porque ele é o pertence a B
maior conjunto de eventos e
todos os são subconjuntos de
An
 3. Algebra of events 5

Homework Problems

Exercise 1.1. You ask a friend to choose an integer N between 0 and 9. Let
A = {N 6 5}, B = {3 6 N 6 7} and C = {N is even and > 0}. List the
points that belong to the following events:
(a) A ∩ B ∩ C
(b) A ∪ (B ∩ Cc )
(c) (A ∪ B) ∩ Cc
(d) (A ∩ B) ∩ ((A ∪ C)c )
Exercise 1.2. Let A, B and C be events in a sample space Ω. Prove the
following identities:
(a) (A ∪ B) ∪ C = A ∪ (B ∪ C)
(b) A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C)
(c) (A ∪ B)c = Ac ∩ Bc
(d) (A ∩ B)c = Ac ∪ Bc
Exercise 1.3. Let A, B and C be arbitrary events in a sample space Ω.
Express each of the following events in terms of A, B and C using inter-
sections, unions and complements.
(a) A and B occur, but not C;
(b) A is the only one to occur;
(c) at least two of the events A, B, C occur;
(d) at least one of the events A, B, C occurs;
(e) exactly two of the events A, B, C occur;
(f) exactly one of the events A, B, C occurs;
(g) not more than one of the events A, B, C occur.
Disjoint sets: Prop. of SETS Exercise 1.4. Two sets are disjoint if their intersection is empty. If A and B
(sample space)
are disjoint events in a sample space Ω, are Ac and Bc disjoint? Are A ∩ C
Mutually exclusives: Prop.
of Events (subsets)
and B ∩ C disjoint ? What about A ∪ C and B ∪ C?
Exercise 1.5. We roll a die 3 times. Give a sample space Ω and a set of
events F for this experiment.
Exercise 1.6. An urn contains three chips: one black, one green, and one
red. We draw one chip at random. Give a sample space Ω and a collection
of events F for this experiment.
Exercise 1.7. If An ⊂ An−1 ⊂ · · · ⊂ A1 , show that ∩n
i=1 Ai = An and
∪n
i=1 A i = A 1 .
 Lecture 2

1. Algebra of events, continued

We say that A1 , · · · , An are disjoint if ∩n
i=1 Ai = ∅. We say they are pair-
wise disjoint if Ai ∩ Aj = ∅, for all i 6= j.

Example 2.1. The sets {1, 2}, {2, 3}, and {1, 3} are disjoint but not pair-wise
disjoint.

Example 2.2. If A and B are disjoint, then A ∪ C and B ∪ C are disjoint only
when C = ∅. To see this, we write (A∪C)∩(B∪C) = (A∩B)∪C = ∅∪C = C.
On the other hand, A ∩ C and B ∩ C are obviously disjoint.

Example 2.3. If A, B, C, and D are some events, then the event “B and
at least A or C, but not D” is written as B ∩ (A ∪ C) r D or, equivalently,
B ∩ (A ∪ C) ∩ Dc . Similarly, the event “A but not B, or C and D” is written
(A ∩ Bc ) ∪ (C ∩ D).

Example 2.4. Now, to be more concrete, let A = {1, 3, 7, 13}, B = {2, 3, 4, 13, 15},
C = {1, 2, 3, 4, 17}, D = {13, 17, 30}. Then, A ∪ C = {1, 2, 3, 4, 7, 13, 17},
B ∩ (A ∪ C) = {2, 3, 4, 13}, and B ∩ (A ∪ C) r D = {2, 3, 4}. Similarly,
A ∩ Bc = {1, 7}, C ∩ D = {17}, and (A ∩ Bc ) ∪ (C ∩ D) = {1, 7, 17}.

Example 2.5. We want to write the solutions to |x−5|+|x−3| > |x| as a union
of disjoint intervals. For this, we first need to figure out what the absolute
values are equal to. There are four cases. If x 6 0, then the inequality
becomes 5 − x + 3 − x > −x, that is 8 > x, which is always satisfied (when
x 6 0). Next, is the case 0 6 x 6 3, and then we have 5 − x + 3 − x > x,
which means 8 > 3x, and so 8/3 < x 6 3 is not allowed. The next case is
3 6 x 6 5, which gives 5 − x + x − 3 > x and thus 2 > x, which cannot

7
8 2

happen (when 3 6 x 6 5). Finally, x > 5 implies x−5+x−3 > x and x > 8,
which rules out 5 6 x < 8. In short, the solutions to the above equation
are the whole real line except the three intervals (8/3, 3], [3, 5], and [5, 8).
This is really the whole real line except the one interval (8/3, 8). In other
words, the solutions are the points in (−∞, 8/3] ∪ [8, ∞).

We have the following distributive relation.

Lemma 2.6. A∪(B∩C) = (A∪B)∩(A∪C) and A∩(B∪C) = (A∩B)∪(A∩C).

Proof. First, we show that A ∪ (B ∩ C) ⊆ (A ∪ B) ∩ (A ∪ C). Indeed, if

x ∈ A ∪ (B ∩ C), then either x is in A or it is in both B and C. Either way, x
is in A ∪ B and in A ∪ C.
Next, we show that (A ∪ B) ∩ (A ∪ C) ⊆ A ∪ (B ∩ C). Here too, if x is
in A ∪ B and in A ∪ C, then either x ∈ A, or x is not in A and hence it is in
both B and C. Either way, it is in A ∪ (B ∩ C).
To prove the second equality either proceed similarly to the above
proof, or apply the first equality to Ac , Bc , and Cc , and take complements
of both side to get
A∩(B∪C) = (Ac ∪(Bc ∩Cc ))c = ((Ac ∪Bc )∩(Ac ∪Cc ))c = (A∩B)∪(A∩C).


Recall that we say a set I is countable if there is a bijective (1-to-1

and onto) function from N = {1, 2, 3, . . . } onto I; in other words if we can
“count” I. Examples of countable sets are N, Z, Z2 , Z3 , . . . , and Q. An ex-
ample of an uncountable set is the interval [0, 1), or any nonempty interval
for that matter.
One can form countable unions of sets by defining ∪i>1 Ai to be the
set of elements that are in at least one of the sets Ai . Similarly, ∩i>1 Ai is
the set of elements that are in all of the Ai ’s simultaneously. (If there are
no such elements, then the intersection is simply the empty set.)
Example 2.7. Let a < b − 1. Then, ∪n>1 (a, b − 1/n) = (a, b). It is clear that
(a, b − 1/n) ⊂ (a, b), for all n > 1. Thus, ∪n>1 (a, b − 1/n) ⊆ (a, b). On the
other hand, if x ∈ (a, b), then there exists an n > 1 such that x < b − 1/n.
For otherwise, x > b − 1/n for all n > 1 and thus taking n → ∞ we have
x > b. We just proved that if x ∈ (a, b), then there is an n > 1 such that
x ∈ (a, b − 1/n). Thus, (a, b) ⊆ ∪n>1 (a, b − 1/n).
Example 2.8. Similarly to the above we can show that if a < b, then
∩n>1 (a, b + 1/n) = (a, b]. In particular, ∩n>1 (0, 1/n) = ∅. (Even though
this is a sequence of nonempty decreasing sets, their intersection is empty!)
1. Algebra of events, continued 9

Homework Problems

Exercise 2.1. A public opinion poll (fictional) consists of the following

three questions:
(1) Are you a registered Democrat?
(2) Do you approve of President Obama’s performance in office?
(3) Do you favor the Health Care Bill?
A group of 1000 people is polled. Answers to the questions are either yes
or no. It is found that 550 people answer yes to the third question and 450
answer no. 325 people answer yes exactly twice (i.e. their answers contain
2 yeses and one no). 100 people answer yes to all three questions. 125 reg-
istered Democrats approve of Obama’s performance. How many of those
who favor the Health Care Bill do not approve of Obama’s performance
and in addition are not registered Democrats? (Hint: use a Venn diagram.)
Exercise 2.2. Let A and B be events in a sample space Ω. We remind that
A \ B = A ∩ Bc . Prove the following:
(a) A ∩ (B \ C) = (A ∩ B) \ (A ∩ C)
(b) A \ (B ∪ C) = (A \ B) \ C
(c) Is it true that (A \ B) ∪ C = (A ∪ C) \ B ?
Exercise 2.3. Let Ω be the reals. Establish
∞  ∞ h
[ 1i [ 1 
(a, b) = a, b − = a + ,b
n n
n=1 n=1
∞ h ∞ 
\ 1 \ 1 i
[a, b] = a, b + = a − ,b
n n
n=1 n=1
 Lecture 3

1. About the set of events F

You can think of F as the set of events A for which you are allowed to
ask the question: what is the probability of A? We can be very general
and choose it to be the set of all subsets of Ω, allowing ourselves to ask
about anything. This turns out to be OK if the space is finite [or even if it
is countably infinite]. However, this turns out to be too much to ask for if
the space is, say, Ω = [0, 1]. (We will see why shortly.)
In any case, the empty set must belong to F ; i.e. we should be able to
ask about the probability that nothing happens. Here is another obvious
property any choice of F must satisfy:

if A, B ∈ F , then Ac ∈ F and A ∪ B ∈ F .

In other words, if we can ask about the probability an event occurs, we

should be able to ask about the probability it does not occur. Furthermore,
if we can ask about the probabilities of two events, we should be able to
ask about the probability at least one of them occurs. (Note that A ∩ B =
(Ac ∪ Bc )c is then also in F .)

Example 3.1. We can take F = {∅, Ω}. This is the smallest possible F . In
this case, we are only allowed to ask about the probability of something
happening and that of nothing happening.

Example 3.2. If A is a subset of Ω, we can take F = {∅, A, Ac , Ω}. This

is the smallest possible F containing A. In this case, we are only allowed
to ask about the probability of something happening and that of nothing
happening, as well as about the probabilies of A occurring or not.

11
12 3

Figure 3.1. Félix Édouard Justin Émile Borel (Jan 7, 1871 – Feb 3, 1956, France)

Example 3.3. If A and B are subsets of Ω, the smallest F containing them

is
{∅, A, Ac , B, Bc ,A ∩ B, A ∩ Bc , Ac ∩ B, Ac ∩ Bc , A ∪ B, A ∪ Bc ,
Ac ∪ B, Ac ∪ Bc , (A ∩ Bc ) ∪ (Ac ∩ B), (Ac ∪ B) ∩ (A ∪ Bc ), Ω}.

Finally, if we have an infinite space, then F should also satisfy:

if Ai , i > 1, are in F , then so is ∪i>1 Ai .
(Note that, since Aci ∈ F , the above automatically implies that ∩i>1 Ai =
(∪i>1 Aci )c ∈ F . So we do not need an extra “requirement”.)
Example 3.4. It turns out that if Ω = [0, 1], then there is a “smallest” F
that satisfies the above requirements and contains all the intervals (a, b),
0 < a < b < 1. This set turns out to be much smaller than the set of all
subsets of [0, 1]. It is called the Borel σ-algebra and cannot be described
in a simpler way than just “the smallest F that contains all the intervals”!
Proving its existence requires quite a bit of mathematical analysis, which
we do not go into in this class. It is noteworthy that one can also prove
that any set in F can be written as intersections and unions of (countably
many) intervals and complements of intervals of the form (a, b].

2. Rules of probability, continued

Recall rules 1–3 (from Lecture 1).
Lemma 3.5. Choose and fix an integer n > 1. If A1 , A2 , . . . , An are pairwise
disjoint events, then
n
!
[
P Ai = P(A1 ) + · · · + P(An ).
i=1
3. Why F is not “everything” 13

Proof. The proof uses mathematical induction.

Claim. If the assertion is true for n − 1, then it is true for n.

The assertion is clearly true for n = 1, and it is true for n = 2 by Rule

3. Because it is true for n = 2, the Claim shows that the assertion holds
for n = 3. Because it holds for n = 3, the Claim implies that it holds for
n = 4, etc.

Proof of Claim. We can write A1 ∪· · ·∪An as A1 ∪B, where B = A2 ∪· · ·∪An .

Evidently, A1 and B are disjoint. Therefore, Rule 3 implies that P(A) =
P(A1 ∪ B) = P(A1 ) + P(B). But B itself is a disjoint union of n − 1 events.
Therefore P(B) = P(A2 ) + · · · + P(An ), thanks to the assumption of the
Claim [“the induction hypothesis”]. This ends the proof. 

Rules 1–3 suffice if we want to study only finite sample spaces. But
infinite sample spaces are also interesting. This happens, for example, if
we want to write a model that answers, “what is the probability that we
toss a coin 12 times before we toss heads?” This leads us to the next, and
final, rule of probability.
Rule 4 (Extended addition rule). If A1 , A2 , . . . are (countably-many) paire-
wise disjoint events, then
∞ ∞
X
!
[
P Ai = P(Ai ).
i=1 i=1

This rule will be extremely important to us soon. It looks as if we

might be able to derive this as a consequence of Lemma 3.5, but that is not
the case. It does need to be assumed as part of our model of probability
theory.

3. Why F is not “everything”

Now we can learn why the set of events F cannot be taken as the set
of all subsets of Ω, if Ω is not countable; e.g. if Ω = [0, 1]. The reason
is simply because then there are too many sets that one has to take into
account and it is not clear if there is even one probability measure that
can satisfy rules 1-4. If one instead uses the smallest F that contains the
“sets of interest”, e.g. the intervals, then one can prove that there are lots
of probability measures.
In what follows, F will be in the background. We will not need it
explicitly in this course. Any events we ask about the probability of will
14 3

happen to be in this mysterious F . Thus, we will not talk about it anymore

(even though it is essential when doing more serious work in probability
theory).

4. Properties of probability
Rules 1–4 have other consequences as well.

Example 3.6. Let A ⊆ B. Note that A and B r A are disjoint. Because

B = A ∪ (B r A) is a disjoint union, Rule 3 implies then that
P(B) = P(A ∪ (B r A))
= P(A) + P(B r A).
Thus, we obtain the statement that
A ⊆ B =⇒ P(B r A) = P(B) − P(A).
As a special case, taking B = Ω and using Rule 2, we have the physically–
appealing statement that
P(Ac ) = 1 − P(A).
For instance, this yields P(∅) = 1 − P(Ω) = 0. “Chances are zero that
nothing happens.”

Example 3.7. Since P(B r A) > 0, the above also shows another physically-
appealing property:
A ⊆ B =⇒ P(A) 6 P(B).

5. Equally-likely outcomes
Suppose Ω = {ω1 , . . . , ωN } has N distinct elements (“N distinct outcomes
of the experiment”). One way of assigning probabilities to every subset of
Ω is to just let
|A| |A|
P(A) = = ,
|Ω| N
where |E| denotes the number of elements of E. Let us check that this prob-
ability assignment satisfies Rules 1–4. Rules 1 and 2 are easy to verify, and
Rule 4 holds vacuously because Ω does not have infinitely-many disjoint
subsets. It remains to verify Rule 3. If A and B are disjoint subsets of Ω,
then |A ∪ B| = |A| + |B|. Rule 3 follows from this. In this example, each
outcome ωi has probability 1/N. Thus, this is the special case of “equally
likely outcomes.”
5. Equally-likely outcomes 15

Example 3.8. Let

Ω = (H1 , H2 ) , (H1 , T2 ) , (T1 , H2 ) , (T1 , T2 ) .
There are four possible outcomes. Suppose that they are equally likely.
Then, by Rule 3,
 
P({H1 }) = P {H1 , H2 } ∪ {H1 , T2 }
= P({H1 , H2 }) + P({H1 , T2 })
1 1 1
= + = .
4 4 2
In fact, in this model for equally-likely outcomes, P({H1 }) = P({H2 }) =
P({T1 }) = P({T2 }) = 1/2. Thus, we are modeling two fair tosses of two fair
coins.
Example 3.9. Let us continue with the sample space of the previous ex-
ample, but assign probabilities differently. Here, we define P({H1 , H2 }) =
P({T1 , T2 }) = 1/2 and P({H1 , T2 }) = P({T1 , H2 }) = 0. We compute, as we did
before, to find that P({H1 }) = P({H2 }) = P({T1 }) = P({T2 }) = 1/2. But now
the coins are not tossed fairly. In fact, the results of the two coin tosses
are the same in this model; i.e. the first coin is a fair coin and once it is
tossed and the result is known the second coin is simply flipped to match
the result of the first coin. Thus, each of the two coins seems fair, but the
second toss depends on the first one and is not hence a fair toss.
16 3

Homework Problems

Exercise 3.1. We toss a coin twice. We consider three steps in this experi-
ment: 1. before the first toss; 2. after the first toss, but before the second
toss; 3. after the two tosses.
(a) Give a sample space Ω for this experiment.
(b) Give the collection F3 of observable events at step 3.
(c) Give the collection F2 of observable events at step 2.
(d) Give the collection F1 of observable events at step 1.
Exercise 3.2. Aaron and Bill toss a coin one after the other until one of
them gets a head. Aaron starts and the first one to get a head wins.
(a) Give a sample space for this experiment.
(b) Describe the events that correspond to ”Aaron wins”, ”Bill wins”
and ”no one wins” ?
Exercise 3.3. Give an example to show that P(A\B) does not need to equal
P(A) − P(B).
 Lecture 4

1. Properties of probability, continued

The following generalizes Rule 3, because P(A ∩ B) = 0 when A and B are
disjoint.
Lemma 4.1 (Another addition rule). If A and B are events (not necessarily
disjoint), then
P(A ∪ B) = P(A) + P(B) − P(A ∩ B). (4.1)

Proof. We can write A ∪ B as a disjoint union of three events:

A ∪ B = (A ∩ Bc ) ∪ (Ac ∩ B) ∪ (A ∩ B).
By Rule 3,
P(A ∪ B) = P(A ∩ Bc ) + P(Ac ∩ B) + P(A ∩ B). (4.2)
Similarly, write A = (A ∩ Bc ) ∪ (A ∩ B), as a disjoint union, to find that
P(A) = P(A ∩ Bc ) + P(A ∩ B). (4.3)
There is a third identity that is proved the same way. Namely,
P(B) = P(Ac ∩ B) + P(A ∩ B). (4.4)
Add (4.3) and (4.4) and solve to find that
P(A ∩ Bc ) + P(Ac ∩ B) = P(A) + P(B) − 2P(A ∩ B).
Plug this into the right-hand side of (4.2) to finish the proof. 

As a corollary we have the following useful fact.

17
18 4

Brown eyes or hair Brown eyes

0.80 0.45

Brown eyes and hair

Brown hair 0.6+0.45-0.8=0.25
0.60

0.25-0.1=0.15 0.1
Brown eyes and hair Brown eyes and hair
and not math major and math major

Figure 4.1. Venn diagram for Example 4.3.

Lemma 4.2. If Ai , i > 1, are [countably many] events (not necessarily disjoint),
then X
P(∪i>1 Ai ) 6 P(Ai ).
i>1

For finitely many such events, A1 , · · · , An , the proof of the lemma goes
by induction using the previous lemma. The proof of the general case of
infinitely many events uses rule 4 and is omitted.
Example 4.3. The probability a student has brown hair is 0.6, the probabil-
ity a student has brown eyes is 0.45, the probability a student has brown
hair and eyes and is a math major is 0.1, and the probability a student
has brown eyes or brown hair is 0.8. What is the probability of a student
having brown eyes and hair, but not being a math major? We know that
P{brown eyes or hair}
= P{brown eyes} + P{brown hair} − P{brown eyes and hair}.
Thus, the probability of having brown eyes and hair is 0.45 + 0.6 − 0.8 =
0.25. But then,
P{brown eyes and hair} = P{brown eyes and hair and math major}
+ P{brown eyes and hair and not math major}.
Therefore, the probability we are seeking equals 0.25 − 0.1 = 0.15. See
Figure 4.1.

Formula 4.1 has a generalization. The following is called the “inclusion-

exclusion” rule.
X
n X
P(A1 ∪ · · · ∪ An ) = (−1)i−1 P(Aj1 ∩ · · · ∩ Aji ).
i=1 16j1 ,...,ji 6n
j1 ,...,ji all different
3. Rolling dice 19

For example,
P(A ∪ B ∪ C) = P(A) + P(B) + P(C) − P(A ∩ B) − P(A ∩ C) − P(B ∩ C) + P(A ∩ B ∩ C).
(4.5)
Proving the inclusion-exclusion formula is deferred to Exercise 33.2.

2. Word of caution
One has to be careful when working out the state space. Consider, for
example, tossing two identical fair coins and asking about the probability
of the two coins landing with different faces; i.e. one heads and one tails.
Since the two coins are identical and one cannot tell which is which, the
state space can be taken as
Ω = {“two heads”,“two tails”,“one heads and one tails”}.
A common mistake, however, is to assume these outcomes to be equally
likely. This would be a perfectly fine mathematical model. But it would
not be modeling the toss of two identical fair coins. For example, if we do
the tossing a large number of times and observe the fraction of time we
got two different faces, this fraction will not be close to 1/3. It will in fact
be close to 1/2.
To resolve the issue, let us paint one coin in red. Then, we can tell
which coin is which and a natural state space is
Ω = {(H1 , H2 ), (T1 , T2 ), (H1 , T2 ), (T1 , H2 )}.
Now, these outcomes are equally likely. Since coins do not behave differ-
ently when they are painted, the probabilities assigned to the state space
in the previous case of identical coins must be
P{two heads} = P{two tails} = 1/4 and P{one heads and one tails} = 1/2.
This matches what an empirical experiment would give, and hence is the
more accurate model of a toss of two fair coins.

3. Rolling dice
Roll two fair dice fairly; all possible outcomes are equally likely.

3.1. A good sample space is

 

(1, 1) (1, 2) · · · (1, 6)
Ω= .. .. .. ..
 . . . . 
 
(6, 1) (6, 2) · · · (6, 6)
We have already seen we can assign P(A) = |A|/|Ω| for any event A. There-
fore, the first question we address is, “how many items are in Ω?” We can
20 4

think of Ω as a 6-by-6 table; so |Ω| = 6 × 6 = 36, by second-grade arith-

metic.
Before we proceed with our example, let us document this observation
more abstractly.

Proposition 4.4 (The second principle of counting). If we have m distinct

forks and n distinct knives, then mn distinct knife–fork combinations are possible.

. . . not to be mistaken with . . .

Proposition 4.5 (The first principle of counting). If we have m distinct forks

and n distinct knives, then there are m + n utensils.

. . . back to our problem . . .

3.2. What is the probability that we roll doubles? Let

A = {(1, 1) , (2, 2) , . . . , (6, 6)}.
We are asking to find P(A) = |A|/36. But there are 6 items in A; hence,
P(A) = 6/36 = 1/6.

3.3. What are the chances that we roll a total of five pips? Let
A = {(1, 4) , (2, 3) , (3, 2) , (4, 1)}.
We need to find P(A) = |A|/36 = 4/36 = 1/9.

3.4. What is the probability that we roll somewhere between two and five
pips (inclusive)? Let
sum = 2 sum =4

 z }| { z }| { 
A = (1, 1) , (1, 2) , (2, 1) , (1, 3) , (2, 2) , (3, 1) , (1, 4) , (4, 1) , (2, 3) , (3, 2) .
 | {z } | {z }
sum =3 sum=5

We are asking to find P(A) = 10/36.

3.5. What are the odds that the product of the number of pips thus rolls
is an odd number? The event in question is
 
(1, 1), (1, 3), (1, 5)
A := (3, 1), (3, 3), (3, 5) .
 
(5, 1), (5, 3), (5, 5)
And P(A) = 9/36 = 1/4.
4. Easy cards 21

4. Easy cards
There are 52 cards in a deck. You deal two cards, all pairs equally likely.
Math model: Ω is the collection of all pairs [drawn without replace-
ment from an ordinary deck]. What is |Ω|? To answer this note that 2|Ω|
is the number of all possible ways to give a pair out; i.e., 2|Ω| = 52 × 51,
by the principle of counting. Therefore,
52 × 51
|Ω| = = 1326.
2
• The probability that exactly one card is an ace is 4 × 48 = 192
divided by 1326. This probability is ' 0.1448
• The probability that both cards are aces is (4 × 3)/2 = 6 divided
by 1326, which is ' 0.0045.
• The probability that both cards are the same is P{ace and ace} +
· · · + P{king and king} = 13 × 6/1326 ' 0.0588.
22 4

Homework Problems

Exercise 4.1. A fair die is rolled 5 times and the sequence of scores recorded.
(a) How many outcomes are there?
(b) Find the probability that first and last rolls are 6.
Exercise 4.2. If a 3-digit number (000 to 999) is chosen at random, find the
probability that exactly one digit will be larger than 5.
Exercise 4.3. A license plate is made of 3 numbers followed by 3 letters.
(a) What is the total number of possible license plates?
(b) What is the number of license plates with the alphabetical part
starting with an A?
Exercise 4.4. An urn contains 3 red, 8 yellow and 13 green balls; another
urn contains 5 red, 7 yellow and 6 green balls. We pick one ball from each
urn at random. Find the probability that both balls are of the same color.
 Lecture 5

1. The birthday problem

n people in a room; all birthdays are equally likely, and assigned at ran-
dom. What are the chances that no two people in the room are born on
the same day? You may assume that there are 365 days a years, and that
there are no leap years.
Let p(n) denote the probability in question.
To understand this consider finding p(2) first. There are two people in
the room.
The sample space is the collection of all pairs of the form (D1 , D2 ),
where D1 and D2 are birthdays. Note that |Ω| = 3652 [principle of count-
ing].
In general, Ω is the collection of all “n-tuples” of the form (D1 , . . . , Dn )
where the Di ’s are birthdays; |Ω| = 365n . Let A denote the collection of
all elements (D1 , . . . , Dn ) of Ω such that all the Di ’s are distinct. We need
to find |A|.
To understand what is going on, we start with n = 2. In order to
list all the elements of A, we observe that we have to assign two separate
birthdays. [Forks = first birthday; knives = second birthday]. There are
therefore 365 × 364 outcomes in A when n = 2. Similarly, when n = 3,
there are 365 × 364 × 363, and in general, |A| = 365 × · · · × (365 − n + 1).
Check this with induction!
Thus,

|A| 365 × · · · × (365 − n + 1)

p(n) = = .
|Ω| 365n

23
24 5

For example, check that p(10) ' 0.88 while p(50) ' 0.03. In fact, if n > 23,
then p(n) < 0.5.

2. An urn problem
n purple and n orange balls are in an urn. You select two balls at ran-
dom [without replacement]. What are the chances that they have different
colors?
Let us number the purple balls 1 through n and the orange balls n + 1
through 2n. This is only for convenience, so that we can define a sample
space and compute probabilities. The balls of course do not know they are
numbered!
The sample space Ω is then the collection of all pairs of distinct num-
bers 1 through 2n. Note that |Ω| = 2n(2n − 1) [principle of counting].
P{two different colors} = 1 − P{the same color}.
Also,
P{the same color} = P(P1 ∩ P2 ) + P(O1 ∩ O2 ),
where Oj denotes the event that the jth ball is orange, and Pk the event
that the kth ball is purple. The number of elements of P1 ∩ P2 is n(n − 1);
the same holds for O1 ∩ O2 . Therefore,
 
n(n − 1) n(n − 1)
P{different colors} = 1 − +
2n(2n − 1) 2n(2n − 1)
n
= .
2n − 1
In particular, regardless of the value of n, we always have
1
P{different colors} > .
2

3. Ordered selection with replacement

Theorem 5.1. Let n > 1 and k > 0 be integers. There are nk ways to pick k
balls from a bag containing n distinct (numbered 1 through n) balls, replacing
the ball each time back in the bag.

Proof. To prove this think of the case k = 2. Let B be the set of balls.
Then, B2 = B × B is the state space corresponding to picking two balls
with replacement. The second principle of counting says |B2 | = |B|2 = n2 .
More generally, when picking k balls we have |Bk | = |B|k = nk ways. 
3. Ordered selection with replacement 25

Note that the above theorem implies that the number of functions from
a set A to a set B is |B||A| . (Think of A = {1, . . . , k} and B being the set of
balls. Each function from A to B corresponds to exactly one way of picking
k balls from B, and vice-versa.)
Example 5.2. What is the probability that 10 people, picked at random,
are all born in May? Let us assume the year has 365 days and ignore leap
years. There are 31 days in May and thus 3110 ways to pick 10 birthdays in
May. In total, there are 36510 ways to pick 10 days. Thus, the probability
3110
in question is 365 10 .

Example 5.3. A PIN number is a four-symbol code word in which each

entry is either a letter (A-Z) or a digit (0-9). Let A be the event that exactly
one symbol is a letter. What is P(A) if a PIN is chosen at random and all
outcomes are equally likely? To get an outcome in A, one has to choose
which symbol was the letter (4 ways), then choose that letter (26 ways),
then choose the other three digits (10 × 10 × 10 ways). Thus,
4 × 26 × 10 × 10 × 10
P(A) = ' 0.0619.
36 × 36 × 36 × 36
Example 5.4. An experiment consists of rolling a fair die, drawing a card
from a standard deck, and tossing a coin. Then, the probability that the
die score is even, the card is a heart, and the coin is heads is equal to
3×13×1
6×52×2 = 1/16.

Example 5.5. We roll a fair die then toss a coin the number of times shown
on the die. What is the probability of the event A that all coin tosses result
in heads? One could use the state space
Ω = {(1, H), (1, T ), (2, H, H), (2, T , T ), (2, T , H), (2, H, T ), · · · }.
However, the outcomes are then not all equally likely. Instead, we continue
tossing the coin up to 6 times regardless of the outcome of the die. Now,
the state space is Ω = {1, · · · , 6} × {H, T }6 and the outcomes are equally
likely. Then, the event of interest is A = A1 ∪ A2 ∪ A3 ∪ A4 ∪ A5 ∪ A6 , where
Ai is the event that the die came up i and the first i tosses of the coin came
up heads. There is one way the die can come up i and 26−i ways the first
i tosses come up heads. Then,
26−i 1
P(Ai ) = 6
= .
6×2 6 × 2i
These events are clearly disjoint and
11 1 1 1 1 1 21
P(A) = + + + + + = .
6 2 4 8 16 32 64 128
26 5

Homework Problems

Read the examples from the lecture that were not covered in class.
Exercise 5.1. Suppose that there are 5 duck hunters, each a perfect shot.
A flock of 10 ducks fly over, and each hunter selects one duck at random
and shoots. Find the probability that 5 ducks are killed.
 Lecture 6

1. Ordered selection without replacement: Permutations

The following follows directly from the second principle of counting.
Theorem 6.1. Let 1 6 k 6 n be integers. There are n(n−1) · · · (n−k+1) ways
to pick k balls out of a bag of n distinct (numbered 1 through n) balls, without
replacing the balls back in the bag.

As a special case one concludes that there are n(n − 1) · · · (2)(1) ways
to put n objects in order. (This corresponds to picking n balls out of a bag
of n balls, without replacement.)
Definition 6.2. If n > 1 is an integer, then we define “n factorial” as the
following integer:
n! = n · (n − 1) · (n − 2) · · · 2 · 1.
For consistency of future formulas, we define also
0! = 1.

Note that the number in the above theorem can be written as

n!
n(n − 1) · · · (n − k + 1) = .
(n − k)!
Example 6.3. 6 dice are rolled. What is the probability that they all show
different faces?
Ω =?
|Ω| = 66 .
If A is the event in question, then |A| = 6 × 5 × 4 × 3 × 2 × 1.

27
28 6

Example 6.4. Five rolls of a fair die. What is P(A), where A is the event
that all five show different faces? Note that |A| is equal to 6 [which face is
left out] times 5!. Thus,
6 · 5! 6!
P(A) = 5 = 5 .
6 6
Example 6.5. The number of permutations of cards in a regular 52-card
deck is 52! > 8 × 1068 . If each person on earth shuffles a deck per second
and even if each of the new shuffled decks gives a completely new per-
mutation, it would still require more than 3 × 1050 years to see all possible
decks! The currently accepted theory says Earth is no more than 5 × 109
years old and our Sun will collapse in about 7 × 109 years. The Heat Death
theory places 3 × 1050 years from now in the Black Hole era. The matter
that stars and life was built of no longer exists.
Example 6.6. Eight persons, consisting of four couples are to be seated in
a row of eight chairs. What is the probability that significant others in each
couple sit together? Since we have 4 couples, there are 4! ways to arrange
them. Then, there are 2 ways to arrange each couple. Thus, there are 4!×24
4
ways to seat couples together. The probability is thus 4!×28! = 1/105.

2. Unordered selection without replacement: Combinations

Theorem 6.7. The number of ways to choose k balls from a bag of n identical
(unnumbered) balls is “n choose k.” Its numerical value is
 
n n!
= .
k k!(n − k)!
More generally, let k1 , . . . , kr > 0 be integers such that k1 + · · · + kr = n. Then,
the number of ways we can choose k1 balls, mark them 1, k2 balls, mark them 2,
. . . , kr balls, mark them r, out of a bag of n identical balls, is equal to
 
n n!
= .
k1 , . . . , kr k1 ! · · · kr !

Before we give the proof, let us do an example that may shed a bit of
light on the situation.
Example 6.8. If there are n people in a room, then they can shake hands
in n2 many different ways. Indeed, the number of possible hand shakes

is the same as
the number of ways we can list all pairs of people, which
is clearly n2 . Here is another, equivalent, interpretation. If there are n
n


vertices in a “graph,” then there are 2 many different possible “edges”
that can be formed between distinct vertices. The reasoning is the same.
Another way to reason is to say that there are n ways to pick the first
2. Unordered selection without replacement: Combinations 29

vertex of the edge and n − 1 ways to pick the second one. But then we
would count each edge twice (once from the point of view
of each end of
the edge) and thus the number of edges is n(n − 1)/2 = n2 .

Proof of Theorem 6.7. Let us first consider the case of n distinct balls.
Then, there is no difference between, on the one hand, ordered choices
of k1 balls, k2 balls, etc, and on the other hand, putting n balls in order.
There are n! ways to do so. Now, each choice of k1 balls out of n identical
balls corresponds to k1 ! possible choices of k1 balls out of n distinct balls.
Hence, if the number of ways of choosing k1 balls, marking them 1, then
k2 balls, marking them 2, etc, out of n identical balls is N, we can write
k1 ! · · · kr !N = n!. Solve to finish. 
Example 6.9. Roll 4 dice; let A denote the event that all faces are different.
Then,  
6 6! 6!
|A| = 4! = = .
4 2! 2
The 6-choose-4 is there because that is how many ways we can choose the
different faces. Note that another way to count is via permutations. We
are choosing 4 distinct faces out of 6. In any case,
6!
P(A) = .
2 × 64
Example 6.10. A poker hand consists of 5 cards dealt without replacement
and without regard to order from a standard 52-cards deck. There are
 
52
= 2, 598, 960
5
different standard poker hands possible.
Example 6.11. The number of different “pairs” {a, a, b, c, d} in a poker
hand is
   
4 12
13
|{z} × × × |{z} 43 .
2 3
choose the a |{z} | {z } deal b, c, d
deal the two a’s choose the b, c, and d

The last 43 corresponds to an ordered choice because once b, c, and d are

chosen, they are distinct and the order in which the suites are assigned
does matter. (That is, it matters if b is a heart and c is a diamond or if it
is the other way around.) Also, it is a choice with replacement because in
each case all 4 suites are possible.
From the above we conclude that
13 × 42 × 12 3



3 ×4
P(pairs) = 52

≈ 0.42.
5
30 6

We also can compute this probability by imposing order on the posi-

tion of the card. (So now, the dealer is giving the cards one at a time, and
we are taking into account which card came first, which came second, and
so on.) Then, the number of ways to get one pair is
 
5
13
|{z} × × 3}
4| {z × × 12 | × 11{z × 10} × |{z} 43 .
2
choose the a deal the two a’s |{z} choose the b, c, and d deal b, c, d
choose where the a’s go

Then
13 × 4 × 3 × 52 × 12 × 11 × 10 × 43


P(pairs) = .
52 × 51 × 50 × 49 × 48
Check this is exactly the same as the above answer.
Example 6.12. Let A denote the event that we get two pairs [a, a, b, b, c] in
a poker hand. Then,
   2
13 4
|A| = × × |{z}
11 × |{z}4 .
2 2
| {z } | {z } choose c deal c
choose a, b deal the a, b

Another way to compute this (which some may find more intuitive) is as:
13


3 is to pick the face values, times 3 to pick which face value is the single

2
card and which are the two pairs, and then times 42 × 4 to deal the cards.
Check that this gives the same answer as above.
In any case,
13 4 2



2 × 2 × 11 × 4
P(two pairs) = 52

≈ 0.06.
5
2. Unordered selection without replacement: Combinations 31

Homework Problems

Exercise 6.1. Suppose that 8 rooks are randomly placed on a chessboard.

Show that the probability that no rook can capture another is 8!8!/(64 ×
63 × · · · × 57).
Exercise 6.2. A conference room contains m men and w women. These
people seat at random in m + w seats arranged in a row. Find the proba-
bility that all the women will be adjacent.
Exercise 6.3. If a box contains 75 good light bulbs and 25 defective bulbs
and 15 bulbs are removed, find the probability that at least one will be
defective.
Exercise 6.4. A lottery is played as follows: the player picks six numbers
out of {1,2,. . . ,54}. Then, six numbers are drawn at random out of the 54.
You win the first prize of you have the 6 correct numbers and the second
prize if you get 5 of them.
(a) What is the probability to win the first prize ?
(b) What is the probability to win the second prize ?
Exercise 6.5. Another lottery is played as follows: the player picks five
numbers out of {1,2,. . . ,50} and two other numbers from the list {1,. . . ,9}.
Then, five numbers are drawn at random from the first list and two from
the random list.
(a) You win the first prize if all numbers are correct. What is the
probability to win the first prize ?
(b) Which lottery would you choose to play between this one and the
one from the previous problem ?
Exercise 6.6. Find the probability that a five-card poker hand (i.e. 5 out of
a 52-card deck) will be :
(a) Four of a kind, that is four cards of the same value and one other
card of a different value (xxxxy shape).
(b) Three of a kind, that is three cards of the same value and two other
cards of different values (xxxyz shape).
(c) A straight flush, that is five cards in a row, of the same suit (ace
may be high or low).
(d) A flush, that is five cards of the same suit, but not a straight flush.
(e) A straight, that is five cards in a row, but not a straight flush (ace
may be high or low).
32 6

Exercise 6.7. Suppose that n people are to be seated at a round table. Show
that there are (n − 1)! distinct seating arrangements. Hint: the mathemat-
ical significance of a round table is that there is no dedicated first chair.
Exercise 6.8. An experiment consists of drawing 10 cards from an ordinary
52-card deck.
(a) If the drawing is made with replacement, find the probability that
no two cards have the same face value.
(b) If the drawing is made without replacament, find the probability
that at least 9 cards will have the same suit.
Exercise 6.9. An urn contains 10 balls numbered from 1 to 10. We draw
five balls from the urn, without replacement. Find the probability that the
second largest number drawn is 8.
Exercise 6.10. Eight cards are drawn without replacement from an ordi-
nary deck. Find the probability of obtaining exactly three aces or exactly
three kings (or both).
Exercise 6.11. How many possible ways are there to seat 8 people (A,B,C,D,E,F,G
and H) in a row, if:
(a) No restrictions are enforced;
(b) A and B want to be seated together;
(c) assuming there are four men and four women, men should be
only seated between women and the other way around;
(d) assuming there are five men, they must be seated together;
(e) assuming these people are four married couples, each couple has
to be seated together.
Exercise 6.12. John owns six discs: 3 of classical music, 2 of jazz and one
of rock (all of them different). How many possible ways does John have if
he wants to store these discs on a shelf, if:
(a) No restrictions are enforced;
(b) The classical discs and the jazz discs have to be stored together;
(c) The classical discs have to be stored together, but the jazz discs
have to be separated.
Exercise 6.13. How many (not necessarily meaningful) words can you form
by shuffling the letters of the following words: (a) bike; (b) paper; (c) letter;
(d) minimum.
 Lecture 7

1. Properties of combinations
Clearly, n choose 0 and n choose n are both equal to 1. The following is
also clear from the definition.

Lemma 7.1. For any integers 0 6 k 6 n

   
n n
= .
k n−k

Recall that n choose k is the number of ways one can choose k elements
out of a set of n elements. Thus, the above formula is obvious: choosing
which k balls we remove from a bag is equivalent to choosing which n − k
balls we keep in the basket. This is called a combinatorial proof.

Lemma 7.2. For 1 6 k 6 n − 1 integers we have

     
n n−1 n−1
= + .
k k−1 k

Proof. We leave the algebraic proof to the student and give instead the
combinatorial proof. Consider a set of n identical balls and mark one of
them, say with a different color. Any choice of k balls out of the n will
either include or exclude the marked ball. There are n − 1 choose k ways
to choose k elements that exclude the ball and n − 1 choose k − 1 ways
to choose k elements that include the ball. The formula now follows from
the first principle of counting. 

33
34 7

Figure 7.1. Blaise Pascal (Jun 19, 1623 – Aug 19, 1662, France)

This allows to easily generate the so-called Pascal’s triangle [Chan-

das Shastra (5th?-2nd? century BC), Al-Karaji (953-1029), Omar Khayyám
(1048-1131), Yang Hui (1238-1298), Petrus Apianus (1495-1552), Niccolò
Fontana Tartaglia (1500-1577), Blaise Pascal (1653)]:

Example 7.3. How many subsets does {1 , . . . , n} have? Assign to each

element of {1 , . . . , n} a zero [“not in the subset”] or a one [“in the subset”].
Thus, the number of subsets of a set with n distinct elements is 2n .
Example 7.4. Choose and fix an integer
r ∈ {1 , . . . , n}. The number of
subsets of {1 , . . . , n} that have size r is nr . This, and the preceding proves
the following amusing combinatorial identity:
Xn  
n
= 2n .
r
r=0
You may need to also recall the first principle of counting.

The preceding example has a powerful generalization.

Theorem 7.5 (The binomial theorem). For all integers n > 0 and all real
numbers x and y,
X
n  
n j n−j
(x + y)n = xy .
j
j=0
1. Properties of combinations 35

Remark 7.6. When n = 2, this yields the familiar algebraic identity

(x + y)2 = x2 + 2xy + y2 .
For n = 3 we obtain
       
3 3 0 3 3 1 2 3 2 1 3 3 0
(x + y) = x y + x y + x y + x y
0 1 2 3
= y3 + 3xy2 + 3x2 y + x3 .

Proof. This is obviously correct for n = 0, 1, 2. We use induction. Induc-

tion hypothesis: True for n − 1.
(x + y)n = (x + y) · (x + y)n−1
X n − 1
n−1
= (x + y) xj yn−j−1
j
j=0

X
n−1
n − 1 j+1 n−(j+1) X n − 1 j n−j
 n−1  
= x y + xy .
j j
j=0 j=0

Change variables [k = j + 1 for the first sum, and k = j for the second] to
deduce that
Xn 
n − 1 k n−k X n − 1 k n−k
 n−1  
n
(x + y) = x y + x y
k−1 k
k=1 k=0
X n − 1 n − 1
n−1
= + xk yn−k + xn + yn
k−1 k
k=1
X n
n−1
= xk yn−k + xn + yn .
k
k=1
The binomial theorem follows. 
Remark 7.7. A combinatorial proof of the above theorem consists of writ-
ing
(x + y)n = (x + y)(x + y) · · · (x + y) .
| {z }
n-times
Then, one observes that to get the term xk yn−k one has to choose k of the
above n multiplicands and pick x from them, then pick y from the n − k
remaining multiplicands. There are n choose k ways to do that.
Example 7.8. The coefficient in front of x3 y4 in (2x − 4y)7 is 73 23 (−4)4 =

71680.

One can similarly work out the coefficients in the multinomial theo-
rem.
36 7

Theorem 7.9 (The multinomial theorem). For all integers n > 0 and r > 2,
and all real numbers x1 , . . . , xr ,
X 
n

n
(x1 + · · · + xr ) = xk1 · · · xk
r ,
r
k 1 , · · · , kr 1
06k1 ,...,kr 6n
k1 +···+kr =n
n


where k1 ,··· ,kr was defined in Theorem 6.7.
The sum in the above display is over r-tuples (k1 , . . . , kr ) such that each
ki is an integer between 0 and n, and the ki ’s add up to n. In the case
r = 2, these are simply (k, n − k) where k runs from 0 to n. So there are
n + 1 terms. The following theorem gives the number of terms for more
general r.
Theorem 7.10. The number of terms in the expansion of (x1 + · · · + xr )n is
n+r−1


r−1 .

For example, the number of terms in the expansion of (a + b + c)5 is

5+3−1
= 72 = 21 terms.



3−1

Proof of Theorem 7.10. To prove the above theorem imagine we have a

collection of n indistinguishable balls that we want to split among r friends.
(Friend number 1 then gets k1 balls, etc.) We want to compute the number
of ways we can do this.
To split the balls among the friend, put the n balls in a row and insert
r−1 indistinguishable stones in between the balls. This will break the balls
into exactly r groups. The first group goes to friend number 1, and so on.
Now we see that the problem amounts to just putting n + r − 1 objects
(the balls and the stones) in a row, in any order. However, (as was done
in the proof of Theorem 6.7) since the balls are indistinguishable, we need
to divide by n!. Similarly, since all stones are indistinguishable, we need
to divide by (r − 1)!. Hence, the number of ways we can split n identical
balls into r groups is
 
(n + r − 1)! n+r−1
= . 
n!(r − 1)! r−1

2. Conditional Probabilities
Example 7.11. There are 5 women and 10 men in a room. Three of the
women and 9 of the men are employed. You select a person at random
from the room, all people being equally likely to be chosen. Clearly, Ω is
the collection of all 15 people, and
2 1 4
P{male} = , P{female} = , P{employed} = .
3 3 5
2. Conditional Probabilities 37

Also,
9 1
P{male and employed} = , P{female and employed} = .
15 5
Someone has looked at the result of the sample and tells us that the person
sampled is employed. Let P(female | employed) denote the conditional
probability of “female” given this piece of information. Then,
|female among employed| 3 1
P(female | employed) = = = .
|employed| 12 4
Definition 7.12. If A and B are events and P(B) > 0, then the conditional
probability of A given B is
P(A ∩ B)
P(A | B) = .
P(B)
For the previous example, this amounts to writing
|female and employed|/|Ω| 1
P(Female | employed) = = .
|employed|/|Ω| 4
The above definition is consistent with the frequentist intuition about
probability. Indeed, if we run an experiment n times and observe that an
event B occurred nB times, then probabilistic intuition tells us that P(B) '
nB /n. If among these nB times an event A occurred nAB times, then
P(A | B) should be about nAB /nB . Dividing through by n one recovers the
above definition of conditional probability.
Example 7.13. If we deal two cards fairly from a standard deck, the prob-
ability of K1 ∩ K2 [Kj = {King on the j draw}] is
4 3
P(K1 ∩ K2 ) = P(K1 )P(K2 | K1 ) = × .
52 51
This agrees with direct counting: |K1 ∩ K2 | = 4 × 3, whereas |Ω| = 52 × 51.
Similarly,
P(K1 ∩ K2 ) P(K3 ∩ K1 ∩ K2 )
P(K1 ∩ K2 ∩ K3 ) = P(K1 ) × ×
P(K1 ) P(K1 ∩ K2 )
= P(K1 )P(K2 | K1 )P(K3 | K1 ∩ K2 )
4 3 2
= × × .
52 51 50
Or for that matter,
4 3 2 1
P(K1 ∩ K2 ∩ K3 ∩ K4 ) = × × × . (Check!)
52 51 50 49
38 7

Homework Problems

Exercise 7.1. Find the coefficient of x5 in (2 + 3x)8 .

Exercise 7.2 (The game of rencontre). An urn contains n tickets numbered
1, 2, . . . , n. The tickets are shuffled thoroughly and then drawn one by one
without replacement. If the ticket numbered r appears in the r-th drawing,
this is denoted as a match (French: rencontre). Show that the probability of
at least one match is
1 1 (−1)n−1
1 − + − ··· + → 1 − e−1 as n → ∞.
2! 3! n!
Exercise 7.3. Show that
          
n+m n m n m n m
= + + ··· + ,
r 0 r 1 r−1 r 0
where 0 6 r 6 min(n, m), r, m, n ∈ N. Try to find a combinatorial proof
and an algebraic proof.
Exercise 7.4. (a) Prove the equality
Xn  
n
k = n 2n−1
k
k=1
by computing in two different ways the number of possible ways
to form a team with a captain out of n people. (The size of the
team can be anything.)
(b) Similarly as in (a), prove that
Xn  
n
k(k − 1) = n(n − 1) 2n−2 .
k
k=1
(c) Find again the results of (a) and (b) by applying the binomial
theorem to (1 + x)n and taking derivatives with respect to x.
Exercise 7.5. We are interested in 4-digit numbers. (The number 0013 is a
2-digit number, not a 4-digit number.)
(a) How many of them have 4 identical digits?
(b) How many of them are made of two pairs of 2 identical digits?
(c) How many of them have 4 different digits?
(d) How many of them have 4 different digits, in increasing order
(from left to right)?
(e) What are the answers to (a), (c) and (d) if we replace 4 by n?
 Lecture 8

1. Conditional Probabilities, continued

Sometimes, to compute the probability of some event A, it turns out to be
helpful if one knew something about another event B. This then can be
used as follows.

Theorem 8.1 (Law of total probability). For all events A and B,

P(A) = P(A ∩ B) + P(A ∩ Bc ).
If, in addition, 0 < P(B) < 1, then
P(A) = P(A | B)P(B) + P(A | Bc )P(Bc ).

Proof. For the first statement, note that A = (A ∩ B) ∪ (A ∩ Bc ) is a disjoint

union. For the second, write P(A ∩ B) = P(A | B)P(B) and P(A ∩ Bc ) =
P(A | Bc )P(Bc ). 

Example 8.2. Once again, we draw two cards from a standard deck. The
probability P(K2 ) (second draw is a king, regardless of the first) is best
computed by splitting it into the two disjoint cases: K1 ∩ K2 and Kc1 ∩ K2 .
Thus,
P(K2 ) = P(K2 ∩ K1 ) + P(K2 ∩ Kc1 ) = P(K1 )P(K2 | K1 ) + P(Kc1 )P(K2 | Kc1 )
4 3 48 4
= × + × .
52 51 52 51
In the above theorem what mattered was that B and Bc partitioned the
space Ω into two disjoint parts. The same holds if we partition the space
into any other number of disjoint parts (even countably many).

39
40 8

Figure 8.1. Thomas Bayes (1702 – Apr 17, 1761, England)

Example 8.3. There are three types of people: 10% are poor (π), 30% have
middle-income (µ), and the rest are rich (ρ). 40% of all π, 45% of µ, and
60% of ρ are over 25 years old (Θ). Find P(Θ). The result of Theorem 8.1
gets replaced with
P(Θ) = P(Θ ∩ π) + P(Θ ∩ µ) + P(Θ ∩ ρ)
= P(Θ | π)P(π) + P(Θ | µ)P(µ) + P(Θ | ρ)P(ρ)
= 0.4P(π) + 0.45P(µ) + 0.6P(ρ).
We know that P(ρ) = 0.6 (why?), and thus
P(Θ) = (0.4 × 0.1) + (0.45 × 0.3) + (0.6 × 0.6) = 0.535.
Example 8.4. Let us recall the setting of Example 5.4. We can now use the
state space
{(D1, H1 ), (D1, T1 ), (D2, T1 , T2 ), (D2, T1 , H2 ), (D2, H1 , T2 ), (D2, H1 , H2 ), · · · },
even though we know the outcomes are not equally likely. We can then
compute
P(A) = P{(D1, H1 )} + P{(D2 , H1 , H2 )} + · · · + P{(D6, H1 , H2 , H3 , H4 , H5 , H6 )}
= P(D1)P(H1 | D1) + P(D2 )P{(H1 , H2 ) | D2 } + · · ·
= P(D1)P(H1 | D1) + P(D2)P(H1 | D2)P(H2 | D1, H1 ) + · · · .
We will finish this computation once we learn about independence in the
next lecture.

2. Bayes’ Theorem
The following question arises from time to time: Suppose A and B are two
events of positive probability. If we know P(B | A) but want P(A | B), then
2. Bayes’ Theorem 41

1/2 H

1/2
1/2 T

1/2 1/3 H

2/3
T

Figure 8.2. Boldface arrows indicated paths giving heads. The path go-
ing to the boldface circle corresponds to choosing the first coin and get-
ting heads. Probabilities multiply along paths by Bayes’ formula.

we can proceed as follows:

P(A ∩ B) P(B | A)P(A)
P(A | B) = = .
P(B) P(B)
If we know only the conditional probabilities, then we can write P(B), in
turn, using Theorem 8.1, and obtain
Theorem 8.5 (Bayes’s Rule). If A, Ac and B are events of positive probability,
then
P(B | A)P(A)
P(A | B) = .
P(B | A)P(A) + P(B | Ac )P(Ac )
Example 8.6. As before, deal two cards from a standard deck. Then,
P(K1 | K2 ) seems complicated to compute. But Bayes’ rule says:
P(K1 ∩ K2 ) P(K1 )P(K2 | K1 )
P(K1 | K2 ) = =
P(K2 ) P(K1 )P(K2 | K1 ) + P(Kc1 )P(K2 | Kc1 )
4 3
52 × 51
= 4 3 48 4
.
52 × 51 + 52 × 51

Example 8.7. There are two coins on a table. The first tosses heads with
probability 1/2, whereas the second tosses heads with probability 1/3. You
select one at random (equally likely) and toss it. Say you got heads. What
are the odds that it was the first coin that was chosen?
Let C denote the event that you selected the first coin. Let H denote
the event that you tossed heads. We know: P(C) = 1/2, P(H | C) = 1/2,
and P(H | Cc ) = 1/3. By Bayes’s formula (see Figure 8.2),
1 1
P(H | C)P(C) × 3
P(C | H) = = 2 2

= .
P(H | C)P(C) + P(H | Cc )P(Cc ) 1 1 1 1


2 × 2 + 3 ×2
5
Remark 8.8. The denominator in Bayes’ rule simply computes P(B) using
the law of total probability. Sometimes, partitioning the space Ω into A
and Ac is not the best way to go (e.g. when the event Ac is complicated).
42 8

In that case, one can apply the law of total probability by partitioning the
space Ω into more than just two parts (as was done in Example 8.3 to
compute the probability P(Θ)). The corresponding diagram (analogous to
Figure 8.2) could then have more than two branches out of each node. But
the methodology is the same. See Exercise 8.10 for an example of this.

3. Conditional probabilities as probabilities

Suppose B is an event of positive probability. Consider the conditional
probability distribution, Q( · · · ) = P( · · · | B).
Theorem 8.9. Q is a probability on the new sample space B. [It is also a proba-
bility on the larger sample space Ω, why?]

Proof. Rule 1 is easy to verify: For all events A,

P(A ∩ B) P(B)
0 6 Q(A) = 6 = 1,
P(B) P(B)
because A ∩ B ⊆ B and hence P(A ∩ B) 6 P(B).
For Rule 2 we check that
P(B ∩ B)
Q(B) = P(B | B) = = 1.
P(B)
Next suppose A1 , A2 , . . . are disjoint events. Then,
∞ ∞
! !
[ 1 [
Q An = P An ∩ B .
P(B)
n=1 n=1
Note that ∪∞
n=1 An ∩B = ∪∞
n=1 (An ∩ B), and (A1 ∩ B), (A2 ∩ B), . . . are
disjoint events. Therefore,
∞ ∞ ∞
1 X X
!
[
Q An = P (An ∩ B) = Q(An ).
P(B)
n=1 n=1 n=1
This verifies Rule 4, and hence Rule 3. 
3. Conditional probabilities as probabilities 43

Homework Problems

Exercise 8.1. We have a coin that gives heads with probability p. In 10

independent tosses of the coin, find the conditional probability that all
successes will occur consecutively (i.e., no two successes will be separated
by one or more failures), given that the number of successes is between
four and six.
Exercise 8.2. We toss a fair coin n times. What is the probability that we
get at least 3 heads given that we get at least one.
Exercise 8.3. A fair die is rolled. If the outcome is odd, a fair coin is tossed
repeatedly. If the outcome is even, a biased coin (with probability of heads
p 6= 12 ) is tossed repeatedly. If the first n throws result in heads, what is
the probability that the fair coin is being used?
Exercise 8.4. We select a positive integer I with P{I = n} = 21n . If I = n,
we toss a coin with probability of heads p = e−n . What is the probability
that the result is heads?
Exercise 8.5. A bridge player and his partner are known to have six spades
between them. Find the probability that the spades will be split (a) 3-3, (b)
4-2 or 2-4, (c) 5-1 or 1-5, (d) 6-0 or 0-6.
Exercise 8.6. An urn contains 30 white and 15 black balls. If 10 balls are
drawn with (respectively without) replacement, find the probability that
the first two balls will be white, given that the sample contains exactly six
white balls.
Exercise 8.7. In a certain village, 20% of the population has some disease.
A test is administered which has the property that if a person is sick, the
test will be positive 90% of the time and if the person is not sick, then
the test will still be positive 30% of the time. All people tested positive
are prescribed a drug which always cures the disease but produces a rash
25% of the time. Given that a random person has the rash, what is the
probability that this person had the disease to start with?
Exercise 8.8. An insurance company considers that people can be split in
two groups : those who are likely to have accidents and those who are
not. Statistics show that a person who is likely to have an accident has
probability 0.4 to have one over a year; this probability is only 0.2 for a
person who is not likely to have an accident. We assume that 30% of the
population is likely to have an accident.
(a) What is the probability that a new customer has an accident over
the first year of his contract?
44 8

(b) A new customer has an accident during the first year of his con-
tract. What is the probability that he belongs to the group likely
to have an accident?
Exercise 8.9. A transmitting system transmits 0’s and 1’s. The probability
of a correct transmission of a 0 is 0.8, and it is 0.9 for a 1. We know that
45% of the transmitted symbols are 0’s.
(a) What is the probability that the receiver gets a 0?
(b) If the receiver gets a 0, what is the probability the the transmitting
system actually sent a 0?
Exercise 8.10. 46% of the electors of a town consider themselves as inde-
pendent, whereas 30% consider themselves democrats and 24% republi-
cans. In a recent election, 35% of the independents, 62% of the democrats
and 58% of the republicans voted.
(a) What proportion of the total population actually voted?
(b) A random voter is picked. Given that he voted, what is the prob-
ability that he is independent? democrat? republican?
Exercise 8.11. To go to the office, John sometimes drives - and he gets late
once every other time - and sometimes takes the train - and he gets late
only once every other four times. When he get on time, he always keeps
the same transportation the day after, whereas he always changes when
he gets late. Let p be the probability that John drives on the first day.
(a) What is the probability that John drives on the nth day?
(b) What is the probability that John gets late on the nth day?
(c) Find the limit as n → ∞ of the results in (a) and (b).
 Lecture 9

1. Independence
It is reasonable to say that A is independent of B if

P(A | B) = P(A), P(Ac | B) = P(Ac ), P(A | Bc ) = P(A), and P(Ac | Bc ) = P(Ac );

i.e. ”knowledge of B tells us nothing new about A.” It turns out that the
first equality above implies the other three. (Check!) It also is equivalent
to the definition that we will actually use: A and B are independent if and
only if
P(A ∩ B) = P(A)P(B).
Note that this is now a symmetric formula and thus B is also independent
of A. Note also that the last definition makes sense even if P(B) = 0 or
P(A) = 0.

Example 9.1. In fact, if P(A) = 0 or P(A) = 1, then A is independent of

any other event B. Indeed, if P(A) = 0 then P(A ∩ B) 6 P(A) implies that
P(A ∩ B) = 0 = P(A)P(B). Also, if P(A) = 1, then P(Ac ) = 0 and

P(Bc ) 6 P(Ac ∪ Bc ) 6 P(Ac ) + P(Bc )

implies that P(Ac ∪ Bc ) = P(Bc ) and thus P(A ∩ B) = P(B) = P(A)P(B).

Example 9.2. Conversely, if A is independent of any other event B (and so

in particular A is independent of itself!), then it must be the case that P(A)
is 0 or 1. To see this observe that P(A ∩ A) = P(A)P(A).

It is noteworthy that being independent and being disjoint have noth-

ing to do with each other.

45
46 9

Example 9.3. Roll a die and let A be the event of an even outcome and B
that of an odd outcome. The two are obviously dependent. Mathemati-
cally, P(A ∩ B) = 0 while P(A) = P(B) = 1/2. On the other hand the two
are disjoint. Conversely, let C be the event of getting a number less than or
equal to 2. Then, P(A ∩ C) = P{2} = 1/6 and P(A)P(C) = 1/2 × 1/3 = 1/6.
So even though A and C are not disjoint, they are independent.

Two experiments E1 and E2 are independent if A1 and A2 are inde-

pendent for all choices of events A1 and A2 of experiments E1 and E2 ,
respectively.
Example 9.4. Toss two fair coins; all possible outcomes are equally likely.
Let Hj denote the event that the jth coin landed on heads, and Tj = Hcj .
Then,
1
P(H1 ∩ H2 ) = = P(H1 )P(H2 ).
4
In fact, the two coins are independent because P(T1 ∩ T2 ) = P(T1 ∩ H2 ) =
P(H1 ∩ H2 ) = 1/4 also. Conversely, if two fair coins are tossed indepen-
dently, then all possible outcomes are equally likely to occur. What if the
coins are not fair, say P(H1 ) = P(H2 ) = 1/4?

Similarly to the above reasoning, three events A1 , A2 , and A3 are inde-

pendent if any combination of two is independent of both the third and
of its complement; e.g. A1 and Ac2 ∩ A3 are independent as well as are Ac2
and A1 ∩ A3 and so on. It turns out that all these relations follow simply
from saying that any two of the events are independent and that also
P(A1 ∩ A2 ∩ A3 ) = P(A1 )P(A2 )P(A3 ). (9.1)
For example, then
P(A1 ∩ Ac2 ∩ A3 ) = P(A1 ∩ A3 ) − P(A1 ∩ A2 ∩ A3 )
= P(A1 )P(A3 ) − P(A1 )P(A2 )P(A3 )
= P(A1 )(1 − P(A2 ))P(A3 )
= P(A1 )P(Ac2 )P(A3 ).
Note that (9.1) by itself is not enough for independence. It is essential
that on top of that every two events are independent.
Example 9.5. Roll two dice and let A be the event of getting a number less
than 3 on the first die, B the event of getting 3, 4, or 5, on the first die, and
C the event of the two faces adding up to 9. Then, P(A ∩ B ∩ C) = 1/36 =
P(A)P(B)P(C) but P(A ∩ B) = 1/6 6= 1/4 = P(A)P(B).

Also, it could happen that any two are independent but (9.1) does not
hold and hence A1 , A2 , and A3 are not independent.
1. Independence 47

Example 9.6. Roll two dice and let A be the event of getting a number less
than 3 on the first die, B the event of getting a number larger than 4 on the
second die, and C the event of the two faces adding up to 7. Then, each
two of these are independent (check), while
1
P(A ∩ B ∩ C) = P{(1, 6), (2, 5), (3, 4)} =
12
but P(A)P(B)P(C) = 1/24.

More generally, having defined independence of n − 1 events, then

A1 , A2 , A3 , . . . , An are independent if any n − 1 of them are and
P(A1 ∩ · · · ∩ An ) = P(A1 ) · · · P(An ).
n experiments are independent if A1 , · · · , An are, for any events Aj of
experiment j.
Example 9.7. In 10 fair tosses of a coin that comes up heads with proba-
bility p, the conditional probability that all heads will occur consecutively,
given that the number of heads is between four and six, is equal to the ra-
tio of the probability of getting exactly four, five, or six consecutive heads
(and the rest tails), by the probability of getting between four and six
heads. That is,
7p4 (1 − p)6 + 6p5 (1 − p)5 + 5p6 (1 − p)4
10 4

10 5

10 6

.
6 5 4
4 p (1 − p) + 5 p (1 − p) + 6 p (1 − p)
(7 ways to get 4 heads in a row and the rest tails, etc.)
Example 9.8. We can now finish the computation from Example 8.4 which
gives an alternative solution to Example 5.4. Indeed, the die and the coins
are independent. Hence,
P(A) = P(D1)P(H1 | D1) + P(D2)P(H1 | D2)P(H2 | D2, H1 ) + · · ·
= P(D1)P(H1 ) + P(D2)P(H1 )P(H2 ) + · · ·
11 1 1
= + + ··· + 6 .
6 2 4 2
48 9

Homework Problems

Read: section 1.7 of Ash’s book carefully.

Exercise 9.1. A single card is drawn from a standard 52-card deck. Give
examples of events A and B that are:
(a) Disjoint but not independent;
(b) Independent but not disjoint;
(c) Independent and disjoint;
(d) Neither independent nor disjoint.
Exercise 9.2. Six fair dice are rolled independently. Find the probability
that the number of 1’s minus the number of 2’s is equal to 3.
Exercise 9.3. Prove the following statements.
(a) If an event A is independent of itself, then P(A) = 0 or 1.
(b) If P(A) = 0 or 1, then A is independent of any event B.
Exercise 9.4. We toss a fair coin three times. Let G1 be the event “the
second and third tosses give the same outcome”, G2 the event “tosses 1
and 3 give the same outcome” and G3 the event “tosses 1 and 2 give the
same outcome”. Prove that these events are pairwise independent but not
independent.
Exercise 9.5. We assume that the gender of a child is independent of the
gender of the other children of the same couple and that the probability
to get a boy is 0.52. Compute, for a 4-child family, the probabilities of the
following events:
(a) all children have the same gender;
(b) the three oldest children are boys and the youngest is a girl;
(c) there are exactly 3 boys;
(d) the two oldest are boys;
(e) there is at least a girl.
 Lecture 10

Figure 10.1. Christiaan Huygens (Apr 14, 1629 – Jul 8, 1695, Netherlands)

1. Gambler’s ruin formula (Huygens)

You, the “Gambler,” are playing independent repetitions of a fair game
against the “House.” When you win, you gain a dollar; when you lose,
you lose a dollar. You start with k dollars, and the House starts with K
dollars. What is the probability that the House is ruined before you?
Observe that if you reach k+K dollars, the house is ruined, while if you
reach 0 dollars, you are ruined. In either case the game ends. Let us think
slightly more generally and define Pj to be the conditional probability that
when the game ends you have K + k dollars (i.e. you win), given that you
start with j dollars initially. We want to find Pk .
Two easy cases are: P0 = 0 and Pk+K = 1.
By direct use of the definitions we see that
P(A ∩ B ∩ C) P(A ∩ B ∩ C) P(B ∩ C)
P(A∩B | C) = = = P(A | B∩C)P(B | C).
P(C) P(B ∩ C) P(C)
[This is the conditional version of P(A ∩ B) = P(A | B)P(B).]
Let H be the event “the house is ruined”, W be the event we win the
next $1 and L be the event we lose our next $1. By Theorem 8.1 we then

49
50 10

have for j = 1, . . . , k + K − 1
Pj = P(H | start with $j)
= P(H ∩ W | start with $j) + P(H ∩ L | start with $j)
= P(W | start with $j)P(H | W and start with $j)
+ P(L | start with $j)P(H | L and start with $j).
Since winning or losing $1 is independent of how much we start with, and
the probability of each is just 1/2, and since starting with $j and winning
$1 results in us having $(j + 1), and similarly for losing $1, we have
1 1
Pj = P(H | start with $(j + 1)) + P(H | start with $(j − 1))
2 2
1 1
= Pj+1 + Pj−1 .
2 2
In order to solve this, write Pj = 12 Pj + 21 Pj , so that
1 1 1 1
Pj + Pj = Pj+1 + Pj−1 for 0 < j < k + K.
2 2 2 2
Multiply both side by two and solve:
Pj+1 − Pj = Pj − Pj−1 for 0 < j < k + K.
In other words,
Pj+1 − Pj = P1 for 0 < j < k + K.
This is the simplest of all possible “difference equations.” In order to solve
it you note that, since P0 = 0,
Pj+1 = (Pj+1 − Pj ) + (Pj − Pj−1 ) + · · · + (P1 − P0 ) for 0 < j < k + K
= (j + 1)P1 for 0 < j < k + K.
Apply this with j = k + K − 1 to find that
1
1 = Pk+K = (k + K)P1 , and hence P1 = .
k+K
Therefore,
j+1
Pj+1 = for 0 < j < k + K.
k+K
Set j = k − 1 to find the following:

Theorem 10.1 (Gambler’s ruin formula). If you start with k dollars, then the
probability that you end with k + K dollars before losing all of your initial fortune
is k/(k + K) for all k, K > 1.
2. Random Variables 51

2. Random Variables
We often want to measure certain characteristics of an outcome of an ex-
periment; e.g. we pick a student at random and measure their height.
Assigning a value to each possible outcome is what a random variable
does.
Definition 10.2. A D-valued random variable is a function X from Ω to D.
The set D is usually [for us] a subset of the real line R, or d-dimensional
space Rd .

We use capital letters (X, Y, Z, etc) for random variables.

Example 10.3. Define the sample space,

Ω=  , .
, , , ,

Then, the random variable X() = 1, X() = 2, . . . , X( ) = 6 models the

number of pips in a roll of a fair six-sided die.
  
The random variable Y( ) = Y( ) = 5, Y( ) = Y( ) = Y( ) = 2,

and Y( ) = −1 models the game where you roll a die and win $5 if you
get 1 or 3, win $2 if you get 2, 5 or 6, and lose $1 if you get 4.

Now if, say, we picked John and he was 6 feet tall, then there is nothing
random about 6 feet! What is random is how we picked the student; i.e.
the procedure that led to the 6 feet. Picking a different student is likely
to lead to a different value for the height. This is modeled by giving a
probability P on the state space Ω.
Example 10.4. In the previous example assume the die is fair; i.e. all out-
comes are equally likely. This corresponds to the probability P on Ω that
gives each outcome a probability of 1/6. As a result, for all k = 1, . . . , 6,
1
P ({ω ∈ Ω : X(ω) = k}) = P({k}) = . (10.1)
6
This probability is zero for other values of k, since X does not take such
values. Usually, we write {X ∈ A} in place of the set {ω ∈ Ω : X(ω) ∈ A}.
In this notation, we have

1
if k = 1, . . . , 6,
P{X = k} = 6 (10.2)
0 otherwise.
This is a math model for the result of rolling a fair die. Similarly,
1 1 1
P{Y = 5} = , P{Y = 2} = , and P{Y = −1} = . (10.3)
3 2 6
This is a math model for the the game mentioned in the previous example.
52 10

Observe that we could have chosen our state space as

Ω = a, l, m, F, ˇ, ˜ .

If we then define the random variable as Y(a) = Y(m) = 5, Y(l) =

Y(ˇ) = Y(˜) = 2, and Y(F) = −1, then we are still modeling the
same game and (10.3) still holds. In fact, if we change the weights of
our die so that a, l, m, F, ˇ, and ˜ come up with probabilities,
respectively, 1/12, 5/24, 3/12, 1/6, 1/24, and 1/4, then (10.3) still holds and
we are once again modeling the same game even though we are using a
different die! The point is that what matters are the values X takes and the
corresponding probabilities.
2. Random Variables 53

Homework Problems

Read: section 2.2 of Ash’s book.

Exercise 10.1. We toss a fair coin 3 times. Let X be the number of tails
we obtain. Give a sample space Ω, a probability measure P and a random
variable X : Ω → R corresponding to this experiment.
Exercise 10.2. We roll a fair die 3 times. Let X be the product of the
outcomes. Give a sample space Ω, a probability measure P and a random
variable X : Ω → R corresponding to this experiment.
 Lecture 11

1. Random Variables, continued

Suppose X is a random variable, defined on some state space Ω. Let P be
a probability on Ω (with events-set F ). By the distribution (or the law) of
X under P we mean the collection of probabilities P{X ∈ A}, as A ranges
over all Borel sets in R. The law of a random variable determines all its
statistical properties and hence characterizes the random variable.
Example 11.1. In the previous example 10.4, (10.2) gave the law of X and
(10.3) gave the law of Y.

If we define PX (A) = P{X ∈ A}, then one can check that this collection
satisfies the rules of probability and is thus a probability on the set D. In
other words, the law of a D-valued random variable is itself a probability
on D. [In fact, this leads to a subtle point. Since we are now talking
about a probability on D, we need an events set, say G . But then we
need {ω : X(ω) ∈ A} to be in F , for all A ∈ G . This is actually another
condition that we should require when defining a random variable, but
we are overlooking this (important) technical point in this course.]
From now on we will focus on the study of two types of random vari-
ables: discrete random variables and continuous random variables.

2. Discrete Random Variables

If X takes values in a finite, or countably-infinite set, then we say that X
is a discrete random variable. Its distribution is called a discrete distribution.
The function
f(x) = P{X = x}

55
56 11

Figure 11.1. Jacob Bernoulli (also known as James or Jacques) (Dec 27,
1654 – Aug 16, 1705, Switzerland)

is then called the mass function of X. The values x for which f(x) > 0 are
called the possible values of X.
Note that in this case knowledge of the mass function is sufficient to
determine the law of X. Indeed, for any subset A ⊂ D,
X
P{X ∈ A} = P{X = x}, (11.1)
x∈A

since A = ∪x∈A {x} and this is a countable union of disjoint sets.

Here are two important properties of mass functions:

• 0 6 f(x) 6 1 for all x. [Easy]

P
• x∈Ω f(x) = 1. [Use A = D in (11.1)]

In fact, given a countable set D and a function f satisfying the above

two properties, we can reverse engineer a random variable with mass func-
tion f. Just take Ω = D, P({x}) = f(x) for x ∈ D, and X(x) = x.
The upshot is that to describe a discrete random variable it is enough
to give a formula for its mass function.

3. The Bernoulli distribution

Suppose we perform a trial once. Let p ∈ [0, 1] be the probability of
“success”. So the state space is Ω = {success, failure}. Let X(success) = 1
and X(failure) = 0. Then, X is said to have a Bernoulli distribution with
4. The binomial distribution (Bernoulli) 57

parameter p [X ∼ Bernoulli(p)]. The mass function is simple:



1 − p if x = 0,
f(x) = P{X = x} = p if x = 1,


0 otherwise.

A nice and useful way to rewrite this is as f(x) = px (1 − p)1−x , if x ∈ {0, 1},
and f(x) = 0 otherwise.

4. The binomial distribution (Bernoulli)

Suppose we perform n independent trials; each trial leads to a “success”
or a “failure”; and the probability of success per trial is the same number
p ∈ [0 , 1]. This is like fairly tossing n coins that each give heads with
probability p, and calling heads a success.
Let X denote the total number of successes in this experiment. This is
a discrete random variable with possible values 0, . . . , n. We say then that
X is a binomial random variable [X ∼ Binomial(n , p)].
Math modelling questions:
• Construct an Ω: Ω = {0, 1}n (with 1 being a success).
P
• Construct X on this Ω: X(ω1 , . . . , ωn ) = ni=1 ωi .

Let us find the mass function of X. We seek to find f(x), where x =

0, . . . , n. For all other values of x, f(x) = 0.
Now suppose x is an integer between zero and n. Note that f(x) =
P{X = x} is the

probability of getting exactly x successes and n − x failures.
There are n x ways to choose which x trials were successes. Moreover,
by independence, the probability of getting any specific combination of x
successes (e.g. first x trials were successes and the rest were failures) is
px (1 − p)n−x . Therefore,


n x n−x if x = 0, . . . , n,
f(x) = P{X = x} = x p (1 − p)
0 otherwise.
P
Note that x f(x) = (p + 1 − p)n = 1 by the binomial theorem.
Remark 11.2. Observe that if B1 , . . . , Bn are independent Bernoulli(p)
P ran-
dom variables (i.e. outcomes of n fair coin tosses), then X = n k=1 k is
B
Binomial(n, p).
Example 11.3. Ten percent of a certain (large) population smoke. If we take
a random sample [without replacement] of 5 people from this population,
what are the chances that at least 2 people smoke in the sample?
58 11

Let X denote the number of smokers in the sample. Then X ∼ Binomial(n , p),
with p = 0.1 and n = 5 [“success” = “smoker”]. Therefore,
P{X > 2} = 1 − P{X 6 1} = 1 − [f(0) + f(1)]
    
5 0 5−0 5 1 5−1
=1− (0.1) (1 − 0.1) + (0.1) (1 − 0.1)
0 1
= 1 − (0.9)5 − 5(0.1)(0.9)4 .
Alternatively, we can follow the longer route and write
P{X > 2} = P ({X = 2} ∪ · · · {X = 5}) = f(2) + f(3) + f(4) + f(5).
4. The binomial distribution (Bernoulli) 59

Homework Problems

Exercise 11.1. Consider a sequence of five Bernoulli trials. Let X be the

number of times that a head is followed immediately by a tail. For exam-
ple, if the outcome is ω = HHT HT then X(ω) = 2 since a head is followed
directly by a tail at trials 2 and 3, and also at trials 4 and 5. Find the
probability mass function of X.
Exercise 11.2. We roll a fair die three times. Let X be the number of times
that we roll a 6. What is the probability mass function of X?
Exercise 11.3. We roll two fair dice.
(a) Let X be the product of the two outcomes. What is the probability
mass function of X?
(b) Let X be the maximum of the two outcomes. What is the proba-
bility mass function of X?
Exercise 11.4. Let Ω = {1, . . . , 6}2 = {(ω1 , ω2 ) : ω1 , ω2 ∈ {1, . . . , 6}} and P
1
the probability measure given by P{ω} = 36 , for all ω ∈ Ω. Let X : Ω → R
be the number of dice that rolled even. Give the probability mass function
of X.
Exercise 11.5. An urn contains 5 balls numbered from 1 to 5. We draw 3
of them at random without replacement.
(a) Let X be the largest number drawn. What is the probability mass
function of X?
(b) Let X be the smallest number drawn. What is the probability mass
function of X?
Exercise 11.6. Of the 100 people in a certain village, 50 always tell the
truth, 30 always lie, and 20 always refuse to answer. A sample of size 30 is
taken with replacement.
(a) Find the probability that the sample will contain 10 people of each
category.
(b) Find the probability that there will be exactly 12 liars.
 Lecture 12

1. The geometric distribution

Suppose we now do not fix the number of independent trials at n. Instead,
we keep running the trials until the first success. Another way to think
about this is as follows. A p-coin is a coin that tosses heads with probability
p and tails with probability 1 − p. Suppose we toss a p-coin until the first
time heads appears. Let X denote the number of tosses made. Then X is a
so-called geometric random variable [“X ∼ Geometric(p)”].
Evidently, if n is an integer greater than or equal to one, then P{X =
n} = (1 − p)n−1 p. Therefore, the mass function of X is given by

p(1 − p)x−1 if x = 1, 2, . . . ,
f(x) =
0 otherwise.

1.1. The tail of the distribution.

Example 12.1. A couple has children until their first son is born. Suppose
the genders of their children are independent from one another, and the
probability of girl is 0.6 every time. Let X denote the number of their
children to find then that X ∼ Geometric(0.4). In particular,

P{X 6 3} = f(1) + f(2) + f(3)

= p + p(1 − p) + p(1 − p)2
= p 1 + 1 − p + (1 − p)2
 

= p 3 − 3p + p2
 

= 0.784.

61
62 12

This gives P{X > 4} = 1 − 0.784 = 0.216.

More generally, consider the tail of the distribution of X ∼ Geometric(p)

(probability of large values). Namely, the probability that X > n. This is
the same as the probability of failing in all of the first n − 1 experiments.
That is,
P{X > n} = (1 − p)n−1 .

In the above couples example, P{X > 4} = 0.63 .

Example 12.2. Let X ∼ Geometric(p). Fix an integer k > 1. Then, the
conditional probability of X − k = x, given X > k + 1 equals
P{X = k + x and X > k + 1}
P{X − k = x | X > k + 1} =
P{X > k + 1}
P{X = k + x} p(1 − p)x+k−1
= = = p(1 − p)x−1 .
P{X > k + 1} (1 − p)k
This says that if we know we have not gotten heads by the k-th toss (i.e.
X > k + 1), then the distribution of when we will get the first head, from
that moment on (i.e. X − k), is again geometric with the same parameter.
This, of course, makes sense: we are still using the same coin, still waiting
for the first heads to come, and the future tosses are independent of the
first k we made so far; i.e. we might as well consider we are starting afresh!
This fact is usually stated as: “the geometric distribution forgets the past.”

2. The negative binomial (or Pascal) distribution

Suppose we are tossing a p-coin, where p ∈ [0 , 1] is fixed, until we obtain
r heads. Let X denote the number of tosses needed. Then, X is a discrete
random variable with possible values r, r + 1, r + 2, . . . . When r = 1, then
X is Geometric(p). In general,
 
 x − 1 pr (1 − p)x−r if x = r, r + 1, r + 2, . . . ,
f(x) = r−1

0 otherwise.
This X is said to have a negative binomial distribution with parameters r and
p.
We can think of the whole process as follows: first toss a p-coin until
the first head is obtained, then toss an independent p-coin, until the second
head appears, and so on. This shows that the negative binomial(r, p) is in
fact the sum of r independent Geometric(p). We will prove this rigorously
when we study moment generating functions.
3. The Poisson distribution (Poisson, 1838) 63

Figure 12.1. Left: Siméon Denis Poisson (Jun 21, 1781 – Apr 25, 1840,
France). Right: Sir Brook Taylor (Aug 18, 1685 – Nov 30, 1731, England)

3. The Poisson distribution (Poisson, 1838)

Choose and fix a number λ > 0. A random variable X is said to have the
Poisson distribution with parameter λ (X ∼ Poisson(λ)) if its mass function is
 −λ x
e λ
if x = 0, 1, . . . ,
f(x) = x! (12.1)

0 otherwise.
In
P order to make sure that this makes sense, it suffices to prove that
x f(x) = 1, but this is an immediate consequence of the Taylor expansion
of eλ , viz.,
X∞
λ λk
e = .
k!
k=0

Poisson random variables are often used to model the length of a wait-
ing list or a queue (e.g. the number of people ahead of you when you stand
in line at the supermarket). The reason this makes a good model is made
clear in the following section.

3.1. Law of rare events. Is there a physical manner in which Poisson(λ)

arises naturally? The answer is “yes.” Let X = Binomial(n , λ/n). For
instance, X could denote the total number of sampled people who have a
rare disease (population percentage = λ/n) in a large sample of size n. Or,
the total number of people, in a population of size n, who decide to stand
in line in the supermarket, with each of them making an independent
decision to join the queue with a small chance of λ/n [in order to make
the “average length” of the line about n × λ/n = λ]. Then, for all fixed
64 12

integers k = 0 , . . . , n,
   k 
λ n−k

n λ
fX (k) = 1− . (12.2)
k n n
Poisson’s “law of rare events” states that if n is large, then the distribu-
tion of X is approximately Poisson(λ). This explains why Poisson random
variables make good models for queue lengths.
In order to deduce this we need two computational lemmas.
Lemma 12.3. For all z ∈ R,
 z n
lim 1+ = ez .
n→∞ n
Proof. Because ex is continuous, it suffices to prove that
 z
lim n ln 1 + = z. (12.3)
n→∞ n
By Taylor’s expansion,
 z z θ2
ln 1 + = + ,
n n 2
where θ lies between 0 and z/n. Equivalently,
z  z z z2
6 ln 1 + 6 + 2.
n n n 2n
Multiply all sides by n and take limits to find (12.3), and thence the lemma.
Alternatively, one can set h = z/n and write (12.3) as
ln(1 + h) ln(1 + h) − ln(1)
z lim = z lim = z(ln x) 0 x=0
= z. 
h→0 h h→0 h
Lemma 12.4. If k > 0 is a fixed integer, then
nk
 
n
∼ as n → ∞.
k k!
where an ∼ bn means that limn→∞ (an /bn ) = 1.

Proof. If n > k, then

n! n(n − 1) · · · (n − k + 1)
=
nk (n
− k)! nk
n n−1 n−k+1
= × × ··· ×
n n n
→1 as n → ∞.

nk
The lemma follows upon writing out nk / k! and applying the above. 
3. The Poisson distribution (Poisson, 1838) 65

Thanks to Lemmas 12.3 and 12.4, and to (12.2),

nk λk −λ e−λ λk
fX (k) ∼ e = .
k! nk k!
That is, when n is large, X behaves like a Poisson(λ), and this proves our
assertion.
66 12

Homework Problems

Exercise 12.1. Solve the following.

(a) Let XPbe a geometric random variable with parameter p. Prove
that ∞n=1 P{X = n} = 1.
(b) P
Let Y be a Poisson random variable with parameter λ. Prove that
∞
n=0 P{Y = n} = 1.

Exercise 12.2. Some day, 10,000 cars are travelling across a city ; one car
out of 5 is gray. Suppose that the probability that a car has an accident
this day is 0.002. Using the approximation of a binomial distribution by a
Poisson distribution, compute:
(a) the probability that exactly 15 cars have an accident this day;
(b) the probability that exactly 3 gray cars have an accident this day.
 Lecture 13

1. (Cumulative) distribution functions

Let X be a [real-valued] random variable. The (cumulative) distribution
function (CDF) F of X under P is defined by
F(x) = P{X 6 x}.

Here are some basic properties distribution functions have to satisfy.

(a) F(x) 6 F(y) whenever x 6 y; i.e. F is non-decreasing.
(b) limb→∞ F(b) = 1 and lima→−∞ F(a) = 0.
(c) F is right-continuous. That is, limy&x F(y) = F(x) for all x.
Property (a) just follows from the fact that (−∞, x] ⊂ (−∞, y]. The
other two properties follow from the facts that ∪n>1 (−∞, n] = (−∞, ∞),
∩n>1 (−∞, −n] = ∅, ∪n>1 (−∞, x + 1/n] = (−∞, x], and the following
lemma.
Lemma 13.1. Let P be a probability. Let An be an increasing set of events:
A1 ⊂ A2 ⊂ A3 ⊂ · · · . Then,


lim P(An ) = P ∪n>1 An .
n→∞

Similarly, if An is decreasing, i.e. A1 ⊃ A2 ⊃ A3 ⊃ · · · , then

lim P(An ) = P ∩n>1 An .
n→∞

Proof. Let us start with the first statement. The proof uses Rule 4 of prob-
ability. To do this we write
∪n>1 An = A1 ∪ (A2 r A1 ) ∪ (A3 r A2 ) ∪ · · · .

67
68 13

Note that the sets on the right-hand-side are disjoint. Hence,

X
P ∪n>1 An = P(A1 ) + P(Ai r Ai−1 )
i>2
X
n
= P(A1 ) + lim (P(Ai ) − P(Ai−1 ))
n→∞
i=2
= lim P(An ).
n→∞

The other statement follows from taking complements. 

In fact, the converse is also true. Any function F that satisfies the
above properties (a)-(c) is a distribution function of some random variable
X. This is because of the following property. If X has distribution function
F under P, then
(d) F(b) − F(a) = P{a < X 6 b} for a < b.
Now, say we have a function F satisfying (a)-(c) and we want to re-
verse engineer a random variable X with distribution function F. Let
Ω = (−∞, ∞) and for a < b define

P((a, b]) = F(b) − F(a). (13.1)

Recall at this point the Borel sets from Example 3.4. It turns out that
properties (a)-(c) are exactly what is needed to be able to extend (13.1) to a
collection {P(B) : B is a Borel set} that satisfies the rules of probability. This
fact has a pretty sophisticated proof that we omit here. But then, consider
the random variable X(ω) = ω. Its distribution function under P is equal
to


P{X 6 x} = P((−∞, x]) = P ∩n>1 (−n, x] = lim (F(x) − F(−n)) = F(x).
n→∞

The upshot is that it is in general (whether X is discrete, continuous,

or neither) enough to specify the CDF in order to fully describe a random
variable.
Here are two more useful properties of distribution functions.
(e) P{X > x} = 1 − F(x).
(f) P{X = x} = F(x) − limy%x F(y) is the size of the jump [if any] at x.
The last property is proved again using Lemma 13.1. It shows that for
a discrete random variable the distribution function is a step function that
jumps precisely at the possible values of X. The size of the jump at x is
1. (Cumulative) distribution functions 69

F(x) F (x)
1 1

0.5

0
a x 0 1 x
(a) X = a (b) Bernoulli(0.5)

F (x) F (x)
..
1 ..1 ... .
0.75 .
0.5 0.58
0.44
0.25 0.25
0 0
1 2 3 4 x 1 2 3 4 5x
(c) Uniform(1, 2, 3, 4) (d) Geometric(0.25)

Figure 13.1. CDFs for some discrete distributions

exactly the mass function f(x). In fact, in this case

X
F(x) = f(y). (13.2)
y: y6x

In particular, the CDF of a discrete random variable is piecewise constant.

Example 13.2. Let X be nonrandom. That is, P{X = a} = 1 for some

number a. Such a random variable is called “deterministic.” Then (see
Figure 13.1(a)),

0 if x < a,
F(x) =
1 if x > a.

Example 13.3. Let X be Bernoulli with parameter p ∈ [0, 1]. Then (see
Figure 13.1(b)),


0 if x < 0,
F(x) = 1 − p if 0 6 x < 1,


1 if x > 1.

Example 13.4. Let Ω = {1, . . . , n} and let X(k) = k for all k ∈ Ω. Let P the
probability on Ω corresponding to choosing an element, equally likely;
70 13

P{k} = 1/n for all k ∈ Ω. Then (see Figure 13.1(c)),



0 if x < 1,
F(x) = n k
if k 6 x < k + 1, k ∈ {1, · · · , n − 1},


1 if x > n.
Example 13.5. Let X be binomial with parameters n and p. Then,


0P if x < 0,
k n j

n−j
F(x) = j=0 j p (1 − p) if k 6 x < k + 1, 0 6 k < n,


1 if x > n.
Example 13.6. Let X be geometric with parameter p. Then (see Figure
13.1(d)), 
0 if x < 1,
F(x) = n
1 − (1 − p) if n 6 x < n + 1, n > 1.
Here, we used the fact that
f(0) + f(1) + · · · + f(n) = p + (1 − p)p + · · · + (1 − p)n−1 p = 1 − (1 − p)n .
1. (Cumulative) distribution functions 71

Homework Problems

Exercise 13.1. Let F be the function defined by:



 0 if x < 0,



 x2
if 0 6 x < 1,
 3
1
F(x) = if 1 6 x < 2,


3

 1 1
if 2 6 x < 4,

 6x + 3
 1 if x > 4.
Let X be a random variable which corresponds to F.
(a) Verify that F is a cumulative distribution function.
(b) Compute P{X = 2}.
(c) Compute P{X < 2}.
(d) Compute P{X = 2 or 1
2 6 X < 32 }.
1
(e) Compute P{X = 2 or 2 6 X 6 3}.
 Lecture 14

1. Continuous Random Variables

We say that X is a continuous random variable with (probability) density
function (pdf) f if f is a piecewise continuous nonnegative function, and
for all real numbers x,
Zx
F(x) = P{X 6 x} = f(y) dy.
−∞

As a consequence, F is a continuous function when X is a continuous

random variable.
It is noteworthy at this point that if F is not continuous nor piecewise
constant then it is not the CDF of a discrete nor of a continuous random
variable. (Draw a CDF like that!)
Comparing the above to (13.2) shows that f is playing the role of the
mass function that was used in the discrete case. However, note that when
X is continuous,
P{X = x} = F(x) − lim F(y) = 0,
y%x

for all x. Hence, f(x) is not a mass function. A good way to think about it
is as the “likelihood” of getting outcome x, instead of as the probability of
getting x.
If f is continuous at x, then by the fundamental theorem of calculus,

F 0 (x) = f(x).

And since F is non-decreasing, we have that f(x) > 0, for all x where f is
continuous.

73
74 14

R∞
Conversely, any piecewise continuous f such that −∞ f(y) dy = 1 and
f(x) > 0R corresponds to a continuous random variable X. Simply define
x
F(x) = −∞ f(y) dy and check that properties (a)-(c) of distribution func-
tions are satisfied!
In fact, if X has pdf f, then
Z
P{X ∈ A} = f(x) dx.
A
In particular,
Zb
P{a 6 X 6 b} = P{a < X 6 b} = P{a 6 X < b} = P{a < X < b} = f(x) dx.
a
Example 14.1. Say X is a continuous random variable with probability
density function
1
f(x) = 2 , if |x| > 1/2.
4x
Then, to find P{X4 − 2X3 − X2 + 2X > 0} we need to write the set in question
as a union of disjoint intervals and then integrate f over each interval and
add up the results. So we observe that
X4 − 2X3 − X2 + 2X = (X + 1)X(X − 1)(X − 2)
and thus the region in question is (−∞, −1) ∪ (0, 1) ∪ (2, ∞). Note that X
is never in (0, 1/2), since fX vanishes there. The probability of X being in
this region is then
Z −1 Z1 Z∞
1 1 1 5
2
dx + 2
dx + 2
dx = .
−∞ 4x 1/2 4x 2 4x 8
Example 14.2. Say X is a continuous random variable with probability
density function f(x) = 14 min(1, x12 ). Then f(x) = 41 for −1 6 x 6 1 and
f(x) = 4x1 2 , for |x| > 1. Thus,
Z −1 Z1 Z4
1 1 1 13
P{−2 6 X 6 4} = 2
dx + dx + 2
dx = .
−2 4x −1 4 1 4x 16
1. Continuous Random Variables 75

Homework Problems

Exercise 14.1. Is the random variable X from Exercise 13.1 discrete, con-
tinuous, or neither?
Exercise 14.2. Let X be a random variable with probability density function
given by
c(4 − x2 ) if − 2 < x < 2,
f(x) =
0 otherwise.
(a) What is the value of c?
(b) Find the cumulative distribution function of X.
Exercise 14.3. Let X be a random variable with probability density function
given by
c cos2 (x) if 0 < x < π2 ,
f(x) =
0 otherwise.
(a) What is the value of c?
(b) Find the cumulative distribution function of X.
Exercise 14.4. Let X be a random variable with probability density function
given by
1
f(x) = exp(−|x|).
2
Compute the probabilities of the following events:
(a) {|X| 6 2},
(b) {|X| 6 2 or X > 0},
(c) {|X| 6 2 or X 6 −1},
(d) {|X| + |X − 3| 6 3},
(e) {X3 − X2 − X − 2 > 0},
(f) {esin(πX) > 1},
(g) {X ∈ N}.
Exercise 14.5. Solve the following.
(a) Let f : R → R be defined by
 c
√ if x > 1,
f(x) = x
0 if x < 1.
Does there exist a value of c such that f becomes a probability
density function?
76 14

(b) Let F : R → R be defined by

1
e− x if x > 0,
F(x) =
0 if x 6 0.
Is F a cumulative distribution function? If yes, what is the associ-
ated probability density function?
Exercise 14.6. (a) Let f : R → R be defined by
 c
 if x > 0,
f(x) = 1 + x2

0 if x < 0.
Does there exist a value of c such that f becomes a probability
density function?
(b) Let F : R → R be defined by
 
1 x
F(x) = 1+ √ , x ∈ R.
2 1 + x2
Is F a cumulative distribution function ? If yes, what is the asso-
ciated probability density function?
 Lecture 15

1. Continuous Random Variables, continued

Here are some standard examples of continuous random variables.

Example 15.1 (Uniform density). If a < b are fixed, then the uniform
density on (a , b) is the function

 1 if a 6 x 6 b,
f(x) = b − a
0 otherwise;

see Figure 15.1(a). In this case, we can compute the distribution function
as follows:


 0 if x < a,
x − a
F(x) = if a 6 x 6 b,

 b−a

1 if x > b.

A random variable with this density (X ∼ Uniform(a, b)) takes any value in
[a, b] “equally likely” and has 0 likelihood of taking values outside [a, b].
Note that if a < c < d < b, then
d−c
P{c 6 X 6 d} = F(d) − F(c) = .
b−a
This says that the probability we will pick a number in [c, d] is equal to
the ratio of d − c ”the number of desired outcomes” by b − a ”the total
number of outcomes.”

77
78 15

f (x) f (x)

1 1

0 1 2 x 0 1 2 3 x
(a) Uniform(1, 2) (b) Exponential(1)

Figure 15.1. pdf for certain continuous distributions

Example 15.2 (Exponential densities). Let λ > 0 be fixed. Then


λe−λx if x > 0,
f(x) =
0 if x < 0
is a density, and is called the exponential density with parameter λ. See Figure
15.1(b). It is not hard to see that

1 − e−λx if x > 0,
F(x) =
0 if x < 0.
We write X ∼ Exponential(λ) to say that X is distributed exponentially with
parameter λ.

The exponential distribution is the continuous analogue of the geo-

metric distribution. In fact, just as Poisson’s law of rare events explains
how binomial random variables approximate Poisson random variables,
there is also a sense in which geometric random variables “approximate”
exponential ones. This explains why the latter are used to model waiting
times; e.g. the time it takes to be served when you are first in line at the
supermarket.
To see this, imagine the cashier operates as follows: they flip a coin
every 1/n seconds and serve you only when the coin falls heads. The
coin, however, is balanced to give heads with a small probability of λ/n.
So on average, it will take about n/λ coin flips until you get heads, and
you will be served in about 1/λ seconds.
Now, let n be large (i.e. decisions whether to serve you or not are made
very often). Let X be the time when you get served. Then, the probability
you get served by time x (P{X 6 x}) is the same as the probability the
coin lands heads in the first nx tosses. This is equal to the distribution
function at nx of a geometric variable with parameter λ/n. That is, 1 −
(1 − λ/n)nx . By Lemma 12.3 this converges to 1 − e−λx , which is the
distribution function of an exponential random variable with parameter λ.
1. Continuous Random Variables, continued 79

Remark 15.3. A familiar situation that may be helpful to have in mind

while making sense of the above is how continuously compound interest
arrises. Recall that if the interest is compounded n times a year, with
interest rate r, then an initial amount of A dollars becomes A(1 + r/n)n .
Compounding interest continuously simply means n → ∞ and thus the
amount becomes Aer . Note that if we compound n times, then the interest
rate for each time is r/n, not r. Do you see how this is similar to the above
derivation of the exponential distribution?
Example 15.4. Just as we have seen for a geometric random variable, an
exponential random variable does not recall history; see Example 12.2.
Indeed, if X ∼ Exponential(λ), then for a > 0 and x > 0 we have
P{a < X 6 a + x} e−λa − e−λ(a+x)
P{X − a 6 x | X > a} = = = 1 − e−λx ;
P{X > a} e−λa
i.e. given that you have not been served by time a, the distribution of the
remaining waiting time is again exponential with the same parameter λ.
Makes sense, no?
 Lecture 16

Figure 16.1. Baron Augustin-Louis Cauchy (Aug 21, 1789 – May 23,
1857, France)

1. Continuous Random Variables, continued

Example 16.1 (The Cauchy density (Cauchy, 1827)). Define for all real
numbers x,
1 1
f(x) = .
π 1 + x2
Because
d 1
arctan x = ,
dx 1 + x2
we have
Z∞ Z
1 ∞ 1 1
f(y) dy = 2
dy = [arctan(∞) − arctan(−∞)] = 1.
−∞ π −∞ 1 + y π
Hence, f is a density; see Figure 16.2(a). Also,
Z
1 x 1
F(x) = f(y) dy = [arctan(x) − arctan(−∞)]
π −∞ π
1 1
= arctan(x) + for all real x.
π 2
Note that f decays rather slowly as |x| → ∞ (as opposed to an exponential
distribution, for example). This means that a Cauchy distributed random
variable has a “good chance” of taking large values. For example, it turns

81
82 16

f (x) f (x)
f (x)

x −3 −2 −1 0 1 2 3 x µ − 2σ µ µ + 2σ x
(a) Cauchy (b) Standard Normal (c) Normal(µ,σ2 )

Figure 16.2. pdf for certain continuous distributions

out that it is a good model of the distance for which a certain type of
squirrels carries a nut before burring it. The fat tails of the distribution
then explain the vast spread of certain types of trees in a relatively short
time period!
Example 16.2 (Standard normal density). I claim that
1 x2
φ(x) = √ e− 2
2π
defines a density function; see Figure 2(b). Clearly, φ(x) > 0 and is con-
tinuous at all points x. So it suffices to show that the area under φ is one.
Define Z∞
A= φ(x) dx.
−∞
Then,
Z∞ Z∞
2 1 x2 +y2
A = e− 2 dx dy.
2π −∞ −∞
Changing to polar coordinates (x = r cos θ, y = r sin θ gives a Jacobian of
r) one has
Z Z
2 1 2π ∞ − r2
A = e 2 r dr dθ.
2π 0 0
R∞
Let s = r2 /2 to find that the inner integral is 0 e−s ds = 1. Therefore,
A2 = 1 and hence A = 1, as desired. [Why is A not −1?]
The distribution function of φ is
Zz
1 2
Φ(z) = √ e−x /2 dx.
2π −∞
Of course, we know that Φ(z) → 0 as z → −∞ and Φ(z) → 1 as z → ∞.
Due to symmetry, we also know that Φ(0) = 1/2. (Check that!) Unfor-
tunately, a theorem of Liouville tells us that Φ(z) cannot be computed (in
terms of other “nice” functions). In other words, Φ(z) cannot be computed
1. Continuous Random Variables, continued 83

Figure 16.3. Johann Carl Friedrich Gauss (Apr 30, 1777 – Feb 23, 1855, Germany)

exactly for any value of z other than z = 0, ±∞. Therefore, people have
approximated and tabulated Φ(z) for various choices of z, using standard
methods used for approximating integrals; see the table in Appendix C.
Here are some consequences of that table [check!!]:

Φ(0.09) ≈ 0.5359, Φ(0.90) ≈ 0.8159, Φ(3.35) ≈ 0.9996.

And because φ is symmetric, Φ(−z) = 1 − Φ(z). Therefore [check!!],

Φ(−0.09) = 1 − Φ(0.09) ≈ 1 − 0.5359 = 0.4641, etc.

Of course, nowadays one can also use software to compute Φ(z) very accu-
rately. For example, in Excel one can use the command NORMSDIST(0.09)
to compute Φ(0.09).

Example 16.3 (Normal or Gaussian density (Gauss, 1809)). Given two

numbers −∞ < µ < ∞ and σ > 0, the normal curve (Normal(µ,σ2 )) is
described by the density function

1 2 2
f(x) = √ e−(x−µ) /(2σ ) for −∞ < x < ∞;
2π σ

see Figure 16.2. Using a change of variables, one can relate this dis-
tribution to the standard normal one, denoted N(0,1). Indeed, for all
84 16

−∞ < a 6 b < ∞,
Zb Zb
1 2 2
f(x) dx = √ e−(x−µ) /(2σ ) dx
a a 2π σ
Z (b−µ)/σ
1 2
= √ e−z /2 dz [z = (x − µ)/σ]
(a−µ)/σ 2π
Z (b−µ)/σ (16.1)
= φ(z) dz
(a−µ)/σ
   
b−µ a−µ
=Φ −Φ .
σ σ
One can take a → −∞ or b → ∞ to compute, respectively,
Zb   Z∞  
b−µ a−µ
f(x)dx = Φ and f(x)dx = 1 − Φ .
−∞ σ a σ
Note at this point that taking both a → −∞ and b → ∞ proves that f
is indeed a density curve (i.e. has area 1 under it). The operation x 7→
z = (x − µ)/σ is called standardization. Thus, the above calculation shows
that the area between a and b under the Normal(µ,σ2 ) curve is the same
as the one between the standard scores of a and b but under the standard
Normal(0,1) curve. One can now use the standard normal table to estimate
these areas.
 Lecture 17

1. The binomial distribution, the golden theorem, and a normal

approximation
Consider n independent coin tosses, each giving heads with probability p
and tails with probability 1 − p. As was mentioned at the very beginning
of the course, one expects that as n becomes very large the proportion of
heads approaches p. While this cannot be used as the definition of the
probability of heads being equal to p, it certainly is a consequence of the
probability models we have been learning about. We will later see the
following fact.

Theorem 17.1 (Bernoulli’s golden theorem a.k.a. the law of large numbers,
1713). Suppose 0 6 p 6 1 is fixed. Then, with probability 1, as n → ∞,
Number of heads
≈ p.
n
In other words: in a large sample (n large), the probability is nearly one that
the percentage in the sample is quite close to the percentage the population (p);
i.e. with high probability, random sampling works well for large sample sizes.
Next, a natural question comes to mind: for a given a 6 b with 0 6
a, b 6 n, we know that
Xb  
n j
P {Number of heads is somewhere between a and b} = p (1 − p)n−j .
j
j=a

Can we estimate this sum, if n is large? The answer is “yes.” Another

remarkable fact we will see later on is the following:

85
86 17

Figure 17.1. Left: Abraham de Moivre (May 26, 1667 – Nov 27, 1754,
France). Right: Pierre-Simon, marquis de Laplace (Mar 23, 1749 – Mar
5, 1827, France).

Theorem 17.2 (The De Moivre–Laplace central limit theorem, 1733). Sup-

pose 0 < p < 1 is fixed. Then, as n → ∞,
! !
b − np a − np
P {Between a and b successes} ≈ Φ p −Φ p ,
np(1 − p) np(1 − p)
(17.1)
where Φ is the standard normal CDF (cumulative distribution function),
Zz
1 2
Φ(z) := √ e−x /2 dx for all −∞ < z < ∞.
−∞ 2π
Remark 17.3. Taking a → −∞ it follows that
!
b − np
P{Less than b successes} ≈ Φ p .
np(1 − p)
Similarly, taking b → ∞ we have
!
a − np
P{More than a successes} ≈ 1 − Φ p .
np(1 − p)

Next we learn to use this theorem; we will learn to understand its

actual meaning later on.
Recall that Φ(z) is equal to the area to the left of z and under the
standard normal curve and that, according to (16.1), the difference on the
right-hand side in (17.1) is the area between a and b under the normal
curve N(np,np(1 − p)). The De Moivre–Laplace central limit theorem tells
us then that if n is large, then the binomial probability of having between a and
b successes is approximately equal to theparea between a and b under the normal
curve with parameters µ = np and σ = np(1 − p).
1. The binomial distribution, the golden theorem, and a normal approximation 87

Example 17.4. The evening of a presidential election the ballots were opened
and it was revealed that the race was a tie between the democratic and the
republican candidates. In a random sample of 1963 voters what is the
chance that more than 1021 voted for the republican candidate?
The exact answer to this question is computed from a binomial distri-
bution with n = 1963 and p = 0.5. We are asked to compute
X 1963  1 j 
1963
1 1963−j

P {more than 1021 republican voters} = 1− .
j 2 2
j=1021
p
Because np = 981.5 and np(1 − p) = 22.15, the normal approximation
(Theorem 17.2) yields the following which turns out to be a quite good
approximation:
 
1021 − 981.5
P {more than 1021 republican voters} ≈ 1 − Φ
22.15
≈ 1 − Φ(1.78) ≈ 1 − 0.9625 = 0.0375.
In other words, the chances are approximately 3.75% that the number of
republican voters in the sample is more than 1021.
Example 17.5. A certain population is comprised of half men and half
women. In a random sample of 10,000 what is the chance that the percent-
age of the men in the sample is somewhere between 49% and 51%?
The exact answer to this question is computed from a binomial distri-
bution with n = 10, 000 and p = 0.5. We are asked to compute
X 10000  1 j 
5100
1 10000−j

P {between 4900 and 5, 100 men} = 1− .
j 2 2
j=4900
p
Because np = 5000 and np(1 − p) = 50, the normal approximation (The-
orem 17.2) yields the following which turns out to be a quite good approx-
imation:
   
5100 − 5000 4900 − 5000
P {between 4900 and 5100 men} ≈ Φ −Φ
50 50
= Φ(2) − Φ(−2) = Φ(2) − (1 − Φ(2))
= 2Φ(2) − 1
≈ (2 × 0.9772) − 1 = 0.9544.
In other words, the chances are approximately 95.44% that the percentage
of men in the sample is somewhere between 49% and 51%. This is con-
sistent with law of large numbers: in a large sample, the probability is
nearly one that the percentage of the men in the sample is quite close to
the percentage of men in the population.
88 17

Homework Problems

Exercise 17.1. Let X be the number of successes in a sequence of 10, 000

Bernoulli trials with probability of success 0.8. Estimate P{7940 6 X 6
8080}.
 Lecture 18

1. Continuous Random Variables, continued

Example 18.1 (Gamma densities). Choose and fix two numbers (parame-
ters) α, λ > 0. The gamma density with parameters α and λ is the probability
density function that is proportional to

xα−1 e−λx if x > 0,
0 if x < 0.
The above is nonnegative, but does not necessarily integrate to 1. Thus, to
make it into a density function we have to divide it by its integral (from 0
to ∞). Now, Z∞ Z
1 ∞ α−1 −y
xα−1 e−λx dx = α y e dy.
0 λ 0
Define the gamma function as
Z∞
Γ (α) = yα−1 e−y dy for all α > 0.
0
One can prove that there is “no nice formula” that “describes” Γ (α) for
all α (theorem of Liouville). Thus, the best we can do is to say that the
following is a Gamma density with parameters α, λ > 0:
 α
 λ xα−1 e−λx if x > 0,
f(x) = Γ (α)

0 if x < 0.
Rx
You can probably guess by now (and correctly!) that F(x) = −∞ f(y) dy
cannot be described by nice functions either. Nonetheless, let us finish
by making the observation that Γ (α) is computable for some reasonable

89
90 18

values of α > 0. The key to unraveling this remark is the following “re-
producing property”:

Γ (α + 1) = αΓ (α) for all α > 0. (18.1)

The proof uses integration by parts:

Z∞
Γ (α + 1) = xα e−x dx
Z0∞
= u(x)v 0 (x) dx,
0

where u(x) = xα and v 0 (x) = e−x . Integration by parts states that1

Z Z
uv = uv − v 0 u
0
for indefinite integrals.

Evidently, u 0 (x) = αxα−1 and v(x) = −e−x . Hence,

Z∞
Γ (α + 1) = xα e−x dx
0
∞
Z∞
= uv − v 0u
0 0
Z∞
α−1 −x

∞
= −αx e +α xα−1 e−x dx.
0 0

The first term is zero, and the Rsecond (the integral) is αΓ (α), as claimed.
∞
Now, it easy to see that Γ (1) = 0 e−x dx = 1. Therefore, Γ (2) = 1 × Γ (1) =
1, Γ (3) = 2 × Γ (2) = 2, . . . , and in general,

Γ (n) = (n − 1)! for all integers n > 1.

It is also not too hard to see that

Z∞ √ Z ∞ −y2 /2 √
√
−1/2 −x 2π √
Γ (1/2) = x e dx = 2 e dy = 2 × = π.
0 0 2
Thus,
√
Γ (n + 1/2) = (n − 1/2)(n − 3/2) · · · (1/2) π for all integers n > 1.

In other words, even though Γ (α) is usually hard to compute, for a general
α, it is quite easy to compute for α’s that are are half nonnegative integers.

1This follows immediately from integrating the product rule: (uv) 0 = u 0 v + uv 0 .

2. Functions of a discrete random variable 91

2. Functions of a discrete random variable

Example 18.2. Suppose X has the mass function



1
if x = −1,

 6
 1 if x = 0,
fX (x) = 13
 2 if x = 1,



0 otherwise.

Define a new random variable Y = 2X2 + 1. Then, Y takes the values 1 and
3. The mass function of Y is
fY (y) = P{Y = y} = P{2X2 + 1 = y} = P{X2 = (y − 1)/2}
p p
= P X = (y − 1)/2 or X = − (y − 1)/2 .
When y = 3 we have
1 1 2
fY (3) = P{X = 1 or X = −1} = fX (1) + fX (−1) = + = .
6 2 3
When y = 1 we get
1
fY (3) = P{X = 0 or X = 0} = fX (0) = .
3
The procedure of this example actually produces a theorem.
Theorem 18.3. Let X be a discrete random variable with the set of possible values
being D. If Y = g(X) for a function g, then the set of possible values of Y is g(D)
and P
x: g(x)=y fX (x) if y ∈ g(D),
fY (y) =
0 otherwise.

When g is one-to-one and has inverse h (i.e. x = h(y)) then the formula
simplifies to
fY (y) = fX (h(y)). (18.2)
In the above example, solving for x in terms of y gives

−1 or 1 if y = 3,
x=
0 if y = 1.
Thus, 
fX (−1) + fX (1) if y = 3,
fY (y) =
fX (0) if y = 1.
 Lecture 19

1. Functions of a continuous random variable

The basic problem: If Y = g(X), then how can we compute fY in terms of
fX ? One way is to first compute FY from FX and then take its derivative.
Example 19.1. Suppose X is uniform on (0 , 1), and Y = − ln X. Then, we
compute fY by first computing FY , and then using fY = FY0 . Here are the
details:
FY (y) = P{Y 6 y} = P {− ln X 6 y} = P {ln X > −y} .
Now, the exponential function is an increasing function. Therefore, ln X >
−y if and only if X > e−y . Recalling that FX (x) = x for x ∈ [0, 1] we have

FY (y) = P X > e−y = 1 − FX (e−y ) = 1 − e−y , for y > 0.
We know, of course, that Y does not take negative values and so FY (y) = 0
for y 6 0. Consequently, fY (y) = 0 for y < 0 and for y > 0 we have
fY (y) = FY0 (y) = (1 − e−y ) 0 = e−y .

Let us make the observation that X = e−Y and

fY (y) = FY0 (y) = (1 − FX (e−y )) 0 = −fX (e−y )(e−y ) 0 .

This is not a coincidence.

Theorem 19.2. Suppose X is a continuous random variable with density function
fX supported on a set D ⊂ R. Let g : D → R be a one-to-one function with
inverse h and let Y = g(X). Then,

fX (h(y)) |h 0 (y)| for y ∈ g(D),
fY (y) =
0 otherwise.

93
94 19

[Compare the above formula with the one for the discrete case (18.2)!]

Proof. We have two cases. If g is increasing, then so is h and we have

FY (y) = P{Y 6 y} = P{g(X) 6 y} = P{X 6 h(y)} = FX (h(y)).
Thus,
fY (y) = fX (h(y)) h 0 (y) = fX (h(y)) |h 0 (y)|.
If, on the other hand, g is decreasing, then so is h and we have
FY (y) = P{Y 6 y} = P{g(X) 6 y} = P{X > h(y)} = 1 − FX (h(y)).
[We have used the fact that X is continuous.] Thus,
fY (y) = −fX (h(y)) h 0 (y) = fX (h(y)) |h 0 (y)|. 
Example 19.3. Suppose µ ∈ R and σ > 0 are fixed constants, and de-
fine Y = µ + σX. Find the density of Y in terms of that of X. Since the
transformation is one-to-one and its inverse is x = (y − µ)/σ, we have
 
dx 1 y−µ
fY (y) = fX (x) = fX .
dy σ σ
For example, if X is standard normal, then
1 (y−µ)2
fµ+σX (y) = √ e− 2σ2 .
2πσ2
In other words, Y ∼N(µ , σ2 ).
Example 19.4. Let X ∼ N(µ, σ2 ) and Y = eX . Then, y = ex > 0, x = ln y,
and
1 −(ln y−µ)2
fY (y) = √ e 2σ2 , for y > 0.
y 2πσ2
This is called the log-normal distribution. It is often encountered in med-
ical and financial applications. By the central limit theorem, normally
distributed random variables appear when a huge number of small inde-
pendent errors are added. In chemistry, for example, concentrations are
multiplied. So in huge reactions the logarithms of concentrations add up
and give a normally distributed random variable. The concentration is
then the exponential of this variable and is, therefore, a log-normal ran-
dom variable.

Now what if g is not one-to-one?

The solution: First compute FY , by hand, in terms of FX , and then use the
fact that FY0 = fY and FX
0 =f .
X
1. Functions of a continuous random variable 95

Example 19.5. Suppose X has density fX . Then let us find the density
function of Y = X2 . Again, we seek to first compute FY . Now, for all y > 0,
√ √ √ √
FY (y) = P{X2 6 y} = P {− y 6 X 6 y} = FX ( y) − FX (− y) .
Differentiate [d/dy] to find that
√
√

fX y + fX − y
fY (y) = √
2 y
On the other hand, FY (y) = 0 if y 6 0 and so fY (y) = 0 as well.
For example, consider the case that X is standard normal. Then,
 −y
 √e if y > 0,
fX2 (y) = 2πy

0 if y 6 0.
Or if X is Cauchy, then

 √ 1 if y > 0,
fX2 (y) = π y(1 + y)

0 if y 6 0.

Example 19.6.√If X is uniform (0 , 1) and Y = X2 , then X2 = Y has one

solution: X = Y. Thus, we are in the 1:1 situation and

 √ 1
if 0 < y < 1,
fX2 (y) = 2 y

0 otherwise.

But what happens if X is uniform on (−1, 2) and Y = X2 ? Well, then

X2 = Y has two solutions when 0 < Y < 1 and only one solution when
1 < Y < 4. Repeat the above method to get that

 1 1 1

 √ + √ = √ if 0 < y < 1,

 6 y 6 y 3 y
fX2 (y) = 1
 √ if 1 < y < 4,

 6 y


0 otherwise.
(We leave this as an exercise but will show how it is done in the next
example.)
96 19

Homework Problems

Exercise 19.1. Let X be a uniform random variable on [−1, 1]. Let Y = e−X .
What is the probability density function of Y ?
Exercise 19.2. Let X be an exponential random variable with parameter
λ > 0. What is the probability density function of Y = X2 ?
Exercise 19.3. Solve the following.
(a) (Log-normal distribution) Let X be a standard normal random
variable. Find the probability density function of Y = eX .
(b) Let X be a standard normal random variable and Z a random
variable solution of Z3 + Z + 1 = X. Find the probability density
function of Z.
Exercise 19.4. Solve the following.
(a) Let X be an exponential random variable with parameter λ > 0.
Find the probability density function of Y = ln(X).
(b) Let X be a standard normal random variable and Z a random
π π


variable with values in − 2 , 2 solution of Z + tan(Z) = X. Find
the density function of Z.
Exercise 19.5. Let X be a continuous random variable with probability
density function given by fX (x) = x12 if x > 1 and 0 otherwise. A random
variable Y is given by 
2X if X > 2,
Y=
X2 if X < 2.
Find the probability density function of Y.
Exercise 19.6. Solve the following.
(a) Let f be the probability density function of a continuous random
variable X. Find the probability density function of Y = X2 .
(b) Let X be a standard normal random variable. Show that Y = X2
has a Gamma distribution and find the parameters.
Exercise 19.7. We throw a ball from the origin with velocity v0 and an
angle θ with respect to the x-axis. We assume v0 is fixed and θ is uniformly
distributed on [0, π2 ]. We denote by R the distance at which the object lands,
i.e. hits the x-axis again. Find the probability density function of R. Hint :
we remind you that the laws of mechanics show that the distance is given
v20 sin(2θ)
by R = g , where g is the gravity constant.
 Lecture 20

1. Functions of a continuous random variable, continued

Example 20.1. Suppose X is exponential with parameter λ = 3. Let Y =
(X − 1)2 . Then,
√ √
FY (y) = P{1 − y 6 X 6 1 + y}.
Now, one has to be careful. If 0 6 y 6 1, then
Z 1+√y
FY (y) = √
3e−3x dx
1− y

and √ √
3e−3(1+ y)+ 3e−3(1− y)
fY (y) = √ .
2 y
y) /√y
√
This formula cannot be true for y large. Indeed e−3(1− goes to ∞
as y → ∞, while fY integrates to 1.
In fact, if y > 1, then
Z 1+√y
FY (y) = 3e−3x dx
0

and √
3e−3(1+ y)
fY (y) = √ .
2 y
Another way to see the above is to write
 √ √
1 − y or 1 + y if 0 < y < 1,
x= √
1+ y if y > 1.

97
98 20

(The second solution is rejected when y > 1 because X is an exponential

and is thus always nonnegative.) Now,
 √ √ √ √
fX (1 − y)|(1 − y) 0 | + fX (1 + y)|(1 + y) 0 | if 0 < y < 1,
fY (y) = √ √
fX (1 + y)|(1 + y) 0 | if y > 1.

Finish the computation and check you get the same answer as before.

Example 20.2. Another common transformation is g(x) = |x|. In this case,

let Y = |X| and note that if y > 0, then

FY (y) = P{−y < X < y} = FX (y) − FX (−y).

Else, FY (y) = 0. Therefore,


fX (y) + fX (−y) if y > 0,
fY (y) =
0 if y 6 0.

For instance, if X is standard normal, then

r
 2 −y2 /2
e if y > 0,
f|X| (y) = π

0 if y 6 0.

Or if X is Cauchy, then

2 1
if y > 0,
f|X| (y) = π 1 + y2

0 otherwise.

Can you guess f|X| when X is uniform (−1 , 1)?

Example 20.3. As you can see, it is best to try to work on these problems on
a case-by-case basis. Here is another example where you need to do that.
Let Θ be uniformly distributed between −π/2 and π/2. Let Y = tan Θ.
Geometrically, Y is obtained by picking a line, in the xy-plane, passing
through the origin so that the angle of this line with the x-axis is uniformly
distributed. The y-coordinate of the intersection between this line and
the line x = 1 is our random variable Y. What is the pdf of Y? The
transformation is y = tan θ and thus the pdf of Y is

1
fY (y) = fΘ (arctan(y)) | arctan 0 (y)| = .
π(1 + y2 )

That is, Y is Cauchy distributed.

2. Generating random variables from Uniform(0, 1) 99

2. Generating random variables from Uniform(0, 1)

Theorem 20.4. If X is any random variable with a continuous CDF F(x), then
U = F(X) ∼ Uniform(0, 1).

Proof. Clearly, 0 6 U 6 1. So F(u) = 0 for u < 0 and F(u) = 1 for u > 1.

If F is one-to-one, then we can simply write, for 0 < u < 1,
FU (u) = P{U 6 u} = P{F(X) 6 u} = P{X 6 F−1 (u)} = F(F−1 (u)) = u.
This is the CDF of a Uniform(0,1). If F is not one-to-one, then one needs
to be more careful.
Fix u ∈ (0, 1). Fix ε > 0 such that ε < u and ε < 1 − u. Then,
(u, u + ε) ⊂ (0, 1) and, by continuity of F and the fact that F goes to 0 at
−∞ and to 1 at ∞, its graph must pass between u and u + ε; i.e. there
exists a b such that F(b) ∈ (u, u + ε]. Similarly, there exists an a such that
F(a) ∈ [u − ε, u].
Now note that X 6 a implies F(X) 6 F(a), because F is nondecreasing.
This then implies that F(X) 6 u. Thus,
FU (u) = P{U 6 u} = P{F(X) 6 u} > P{F(X) 6 F(a)}
> P{X 6 a} = F(a) > u − ε.
Moreover, since F(X) 6 u implies F(X) < F(b) which implies X < b, we
have
FU (u) = P{U 6 u} = P{F(X) 6 u} 6 P{F(X) < F(b)}
6 P{X < b} = P{X 6 b} = F(b) 6 u + ε.
We have thus shown that |FU (u) − u| 6 ε. Now take ε → 0 to conclude
that FU (u) = u. 

It is noteworthy that the above does not work if F is not continuous.

Take for example X ∼ Bernoulli(0.5). Then, F(X) = F(0) = 0.5, with proba-
bility 0.5, and F(X) = F(1) = 1 with probability 0.5. This is certainly not a
Uniform(0, 1)!
The converse to the above theorem is quite useful.
Theorem 20.5. Let F be a strictly increasing continuous function such that
limx→−∞ F(x) = 0 and limx→∞ F(x) = 1. Let U ∼ Uniform(0, 1). Then,
X = F−1 (U) has CDF F(x).

Proof. Simply write

P{X 6 x} = P{F−1 (U) 6 x} = P{U 6 F(x)} = F(x). 
100 20

Example 20.6. To generate an exponential random variable with parameter

λ we solve
u = F(x) = 1 − e−λx
log(1−u)
to get x = − λ . Thus, −λ−1 log(1−U) has an exponential distribution
with parameter λ, where U ∼ Uniform(0, 1). [In this special case, 1 − U is
also Uniform(0, 1), and thus we can use −λ−1 log U as well.]
 Lecture 21

1. Generating random variables from Uniform(0, 1), continued

In fact, we can prove a much more general version of Theorem 20.5.

Theorem 21.1. Let F be any nondecreasing right-continuous function such that

limx→−∞ F(x) = 0 and limx→∞ F(x) = 1; i.e. F is any candidate for a CDF.
Define
G(u) = inf{x : u 6 F(x)} = min{x : u 6 F(x)};
see Figure 21.1. [Note that G(u) = F−1 (u) wherever the latter exists.] Let
U ∼ Uniform(0, 1). Then, X = G(U) has CDF F(x).

1
u6

u5 u4 u3

u2
G(u 3 ) = G(u 4 )
u1 = G(u 5 )

0
G(u 1 )G(u 2 ) G(u 6 )

Figure 21.1. The function G(u).

101
102 21

Proof. First, let us explain why the infimum in the definition of G is

attained (and is thus a minimum). This is a consequence of the right-
continuity of F. Indeed, if x & G(u) with F(x) & u, then F(G(u)) = u.
One consequence of the above equation is that
P{X 6 a} = P{G(U) 6 a} 6 P{F(G(U)) 6 F(a)} = P{U 6 F(a)} = F(a).
Next, we observe that the definition of G implies that if u 6 F(a), then
G(u) 6 a. Thus,
P{X 6 a} = P{G(U) 6 a} > P{U 6 F(a)} = F(a).
We conclude that P{X 6 a} = F(a), which means that X has CDF F. 

This theorem allows us to generate any random variable we can com-

pute the CDF of, if we simply have a random number generator that gen-
erates numbers between 0 and 1 “equally likely.”
Example 21.2. How do we flip a coin that gives heads with probability
0.6, using the random number generator on our calculator? The intuitive
answer is: generate a number and call it tails if it is less than 0.4 and heads
otherwise. Does the above theorem give the same answer?
Since the CDF of a Bernoulli(0.6) is not one-to-one, we need to com-
pute G. This turns out not to be too hard. Recall that


0 if x < 0,
F(x) = 0.4 if 0 6 x < 1,


1 if x > 1.
Then, 
0 if 0 6 u 6 0.4,
G(u) =
1 if 0.4 < u 6 1.
Just as our intuition had indicated.

Notice, by the way, that the above shows that one can start with a con-
tinuous random variable and transform it into a discrete random variable!
Of course, the transformation G is not continuous.

2. Joint distributions
If X and Y are two discrete random variables, then their joint mass function
is
f(x , y) = P{X = x , Y = y}.
We might write fX,Y in place of f in order to emphasize the dependence
on the two random variables X and Y.
Here are some properties of fX,Y :
2. Joint distributions 103

• f(x , y) > 0 for all x, y;

P P
• x y f(x , y) = 1;
P
• (x,y)∈C f(x , y) = P{(X , Y) ∈ C}.

Example 21.3. You roll two fair dice. Let X be the number of 2s shown,
and Y the number of 4s. Then X and Y are discrete random variables, and
f(x , y) = P{X = x , Y = y}



1
if x = 2 and y = 0,

 36

 1
if x = 0 and y = 2,

 36


 2
 36 if x = y = 1,
= 36 8
if x = 0 and y = 1,



 8
if x = 1 and y = 0,

 36

 16

 36 if x = y = 0,


0 otherwise.
Some times it helps to draw up a table of “joint probabilities”:

x\y 0 1 2
0 16/36 8/36 1/36
1 8/36 2/36 0
2 1/36 0 0
From this we can also calculate fX and fY . For instance,
10
fX (1) = P{X = 1} = f(1 , 0) + f(1 , 1) =
.
36
In general, you compute the row sums (fX ) and put them in the margin;
you do the same with the column sums (fY ) and put them in the bottom
row. In this way, you obtain:

x\y 0 1 2 fX
0 16/36 8/36 1/36 25/36
1 8/36 2/36 0 10/36
2 1/36 0 0 1/36
fY 25/36 10/36 1/36 1
The “1” designates the right-most column sum (which should be one),
and/or the bottom-row sum (which should also be one). This is also the
sum of the elements of the table (which should also be one).

En route we have discovered the next result, as well.

104 21

Theorem 21.4. For all x, y:

P
(1) fX (x) = b f(x , b).
P
(2) fY (y) = a f(a , y).

3. Independence
Definition 21.5. Let X and Y be discrete with joint mass function f. We say
that X and Y are independent if for all x, y,
fX,Y (x , y) = fX (x)fY (y).
• Suppose A and B are two sets, and X and Y are independent.
Then,
X X
P{X ∈ A , Y ∈ B} = f(x , y)
x∈A y∈B
X X
= fX (x) fY (y)
x∈A y∈B
= P{X ∈ A}P{Y ∈ B}.
• Similarly, if h and g are functions, then h(X) and g(Y) are inde-
pendent as well.
• All of this makes sense for more than 2 random variables as well.
Example 21.6 (Example 21.3, continued). Note that in this example, X and
Y are not independent. For instance,
10 1
f(1 , 2) = 0 6= fX (1)fY (2) = × .
36 36
Example 21.7. Let X ∼ Geometric(p1 ) and Y ∼ Geometric(p2 ) be independent.
What is the mass function of Z = min(X , Y)?
Let q1 = 1 − p1 and q2 = 1 − p2 be the probabilities of failure. Recall
from Lecture 10 that P{X > n} = qn−1
1 and P{Y > n} = qn−1
2 for all integers
n > 1. Therefore,
P{Z > n} = P{X > n , Y > n} = P{X > n}P{Y > n}
= (q1 q2 )n−1 ,
as long as n > 1 is an integer. Because P{Z > n} = P{Z = n} + P{Z > n + 1},
for all integers n > 1,
P{Z = n} = P{Z > n} − P{Z > n + 1} = (q1 q2 )n−1 − (q1 q2 )n
= (q1 q2 )n−1 (1 − q1 q2 ) .
Else, P{Z = n} = 0. Thus, Z ∼ Geometric(p), where p = 1 − q1 q2 .
3. Independence 105

This makes sense: at each step we flip two coins and wait until the
first time one of them comes up heads. In other words, we keep flipping
as long as both coins land tails. Thus, this is the same as flipping one coin
and waiting for the first time it comes up heads, as long as the probability
of tails in this third coin is equal to the probability of both of the original
coins coming up tails.
106 21

Homework Problems

Exercise 21.1. Let X and Y be two discrete random variables with joint
mass function f(x, y) given by
x|y 1 2
1 0.4 0.3
2 0.2 0.1
and f(x, y) = 0 otherwise.
(a) Determine if X and Y are independent.
(b) Compute P(XY 6 2).
Exercise 21.2. We roll two fair dice. Let X1 (resp. X2 ) be the smallest (resp.
largest) of the two outcomes.
(a) What is the joint mass function of (X1 , X2 )?
(b) What are the probability mass functions of X1 and X2 ?
(c) Are X1 and X2 independent?
Exercise 21.3. We draw two balls with replacement out of an urn in which
there are three balls numbered 2,3,4. Let X1 be the sum of the outcomes
and X2 be the product of the outcomes.
(a) What is the joint mass function of (X1 , X2 )?
(b) What are the probability mass functions of X1 and X2 ?
(c) Are X1 and X2 independent?
Exercise 21.4. Let F be the function defined by:


 0 if x < 0,



 x2
if 0 6 x < 1,
 3
1
F(x) = if 1 6 x < 2,


3

 1 1
if 2 6 x < 4,

 6x + 3
 1 if x > 4.
Let U be a Uniform(0,1) random variable. Give a transformation G that
would make X = G(U) a random variable with CDF F.
 Lecture 22

1. Sums of independent random variables

Example 22.1. Suppose X ∼ Poisson(λ) and Y ∼ Poisson(γ) are indepen-
dent. Then, what kind of random variable is X + Y? We can directly com-
pute as follows: The possible values of X + Y are 0, 1, . . . . Let z = 0, 1, . . .
be a possible value for X + Y. Then, X can be one of 0, 1, . . .. Let x be one
of these possible values for X. Then, Y = z − x. The chances X = x and
Y = z − x equal fX (x)fY (z − x). In other words,

∞
X
fX+Y (z) = fX (x)fY (z − x)
x=0
X∞
e−λ λx
= fY (z − x)
x!
x=0
Xz
e−λ λx e−γ γz−x
=
x! (z − x)!
x=0

e−(λ+γ) X z x z−x
z  
= λ γ
z! x
x=0
e−(λ+γ)
= (λ + γ)z ,
z!

thanks to the binomial theorem. For other values of z, it is easy to see that
fX+Y (z) = 0. This computation shows us that X + Y ∼ Poisson(λ + γ). This
makes sense, doesn’t it?

107
108 22

Observe that the first line in the above computation is completely gen-
eral and in fact proves the following theorem.
Theorem 22.2. If X and Y are discrete and independent, then
X
fX+Y (z) = fX (x)fY (z − x).
x
Example 22.3. Suppose X = ±1 with probability 1/2 each; and Y = ±2
with probability 1/2 each. Then,

1/4 if z = 3, −3, 1, −1,
fX+Y (z) =
0 otherwise.
Example 22.4. Let X and Y denote two independent Geometric(p) random
variables with the same parameter p ∈ (0 , 1). What is the mass function
of X + Y? If z = 2, 3, . . . , then
X ∞
X
fX+Y (z) = fX (x)fY (z − x) = pqx−1 fY (z − x)
x x=1
X
z−1 X
z−1
2
= pq x−1
pq z−x−1
=p qz−2 = (z − 1)p2 qz−2 .
x=1 x=1
Else, fX+Y (z) = 0. This shows that X + Y is a negative binomial. This again
makes sense, right?
Example 22.5. If X ∼ Binomial(n , p) and Y ∼ Binomial(m , p) for the
same parameter p ∈ (0 , 1), then what is the distribution of X + Y? If
z = 0, 1, . . . , n + m, then
X Xn  
n x n−x
fX+Y (z) = fX (x)fY (z − x) = p q fY (z − x)
x
x
x=0
X n 
m

x n−x
= p q pz−x qm−(z−x)
x z−x
06x6n
06z−x6m
X  
n m

z m+n−z
=p q .
x z−x
06x6n
z−m6x6z

[The sum is over all integers x such that x is between 0 and n, and x is also
between z − m and m.] For other values of z, fX+Y (z) = 0.
Equivalently, we can write for all z = 0, . . . , n + m,
  
n m
 
n + m z m+n−z X x z−x
fX+Y (z) = p q   .
z n+m
06x6n
z−m6x6z z
1. Sums of independent random variables 109

Thus, if we showed that the sum is one, then X + Y ∼ Binomial(n + m , p).

In order to show that the sum is one consider an urn that has n white balls
and m black balls. We choose z balls at random, without replacement. The
probability that we obtain exactly x white and z − x black is precisely,
  
n m
x z−x
  .
n+m
z
Therefore, if we add this probability over all possible values of x we should
get one. This does the job. [Can you find an algebraic proof? Hint: expand
the identity (a + b)n+m = (a + b)n (a + b)m and match the coefficients of
terms az bn+m−z .]
In particular, we have shown that if we add two independent Bernoulli(p)
random variables, then we get a Binomial(2, p) and that if we add to that
another independent Bernoulli(p), then we get a Binomial(3, p). Repeating
this inductively proves the fact we have already observed: the sum of n
independent Bernoulli(p) is a Binomial(n, p).
 Lecture 23

1. Jointly distributed continuous random variables

Definition 23.1. We say that (X, Y) is jointly distributed with joint density
function f if f is piecewise continuous, and for all “nice” two-dimensional
sets A,
ZZ
P{(X, Y) ∈ A} = f(x , y) dx dy.
A

If (X, Y) has a joint density function f, then:

(1) f(x , y) > 0 for all x and y;

R∞ R∞
(2) −∞ −∞ f(x , y) dx dy = 1.

For any function f of two variables that satisfies these properties, one
can reverse engineer two random variables that will have f as their joint
density function.

Example 23.2 (Uniform joint density). Suppose E is a subset of the plane

that has a well-defined finite area |E| > 0. Define

1 if (x , y) ∈ E,
f(x , y) = |E|

0 otherwise.

Then, f is a joint density function, and the corresponding random vector

(X, Y) is said to be distributed uniformly on E. Moreover, for all planar sets

111
112 23

Figure 23.1. Region of integration in Example 23.2.

E with well-defined areas,

ZZ
1 |E ∩ A|
P{(X, Y) ∈ A} = dx dy = .
|E| |E|
E∩A
See Figure 23.1. Thus, if the areas can be computed geometrically,
RR then, in
the case of a uniform distribution, there is no need to compute A f(x , y) dx dy.
Example 23.3. Let (X, Y) be uniformly distributed on [−1, 1]2 . That is,

1
if −1 6 x 6 1 and −1 6 y 6 1,
fX,Y (x , y) = 4
0 otherwise.

We want to find P{|X+Y| 6 1/2}. In this case, the areas are easy to compute
geometrically; see Figure 23.2. The area of the square is 22 = 4. The
shaded area is the sum of the areas of two identical trapezoids and a
parallelogram. It is thus equal to 2 × 12 × (1 + 21 )/2 + 1 × 1 = 7/4. Or,
alternatively, the non-shaded area is that of two triangles. The shaded
area is thus equal to 4 − 2 × 12 × 23 × 32 = 74 . Then, P{|X + Y| 6 1/2} = 7/16.
We could have used the definition of joint density functions and written
ZZ
P{|X + Y| 6 1/2} = fX,Y (x, y) dx dy
|x+y|61/2
Z −1/2 Z 1 Z 1/2 Z −x+1/2 Z 1 Z −x+1/2
1 1 1
= dy dx + dy dx + dy dx
−1 −x−1/2 4 −1/2 −x−1/2 4 1/2 −1 4
7
= .
16
1. Jointly distributed continuous random variables 113

y y

1 1

1/2 1/2
-1 1/2 1 -1 1/2 1
x x
-1/2 -1/2
-1/2
-1/2

-1 -1

Figure 23.2. Regions of integration for Example 23.3. Left: |x + y| 6 1/2.

Right: xy 6 1/2.

Next, we want to compute P{XY 6 1/2}. This area is not easy to com-
pute geometrically, in contrast to |x + y| 6 1/2; see Figure 23.2. Thus, we
need to compute it using the definition of joint density functions.
ZZ
P{XY 6 1/2} = fX,Y (x, y) dx dy
xy61/2
Z −1/2 Z 1 Z 1/2 Z 1 Z 1 Z 1/2x
1 1 1
= dy dx + dy dx + dy dx
1/2x 4 −1/2 −1 4 1/2 −1 4
−1
| {z } | {z } | {z }
(1/4−1/8x) 2/4 (1/8x+1/4)
x ln |x|  −1/2 1  ln |x| x 1 3 ln 2
= − + + + = + .
4 8 −1 2 8 4 1/2 4 4
Note that we could have computed the middle term geometrically: the
area of the rectangle is 2 × 1 = 2 and thus the probability corresponding to
it is 2/4 = 1/2. An alternative way to compute the above probability is by
computing one minus the integral over the non-shaded region in the right
Figure 23.2. If, on top of that, one observes that both the pdf and the two
non-shaded parts are symmetric relative to exchanging x and y, one can
quickly compute
Z1  Z1 Z1
1  1 1 3 ln 2
P{XY 6 1/2} = 1 − 2 dy dx = 1 − 2 ( − ) dx = + .
1/2 1/2x 4 1/2 4 8x 4 4
114 23

y=x

y=x/2

0 1 x

Figure 23.3. Region of integration in Example 23.4.

Example 23.4. Suppose (X, Y) has joint density


Cxy if 0 < y < x < 1,
f(x , y) =
0 otherwise.
Let us first find C, and then P{X 6 2Y}. To find C:
Z∞ Z∞ Z1 Zx
1= f(x , y) dx dy = Cxy dy dx
−∞ −∞ 0 0
Z 1 Z x  Z1
C C
=C x y dy dx = x3 dx = .
2 8
0
| 0
{z } 0
1 2
2x

Therefore, C = 8, and hence


8xy if 0 < y < x < 1,
f(x , y) =
0 otherwise.
Now ZZ
P{X 6 2Y} = P{(X, Y) ∈ A} = f(x , y) dx dy,
A
where A denotes the collection of all points (x , y) in the plane such that
x 6 2y. Therefore,
Z1 Zx
3
P{X 6 2Y} = 8xy dy dx = .
0 x/2 4
See Figure 23.3. (Graphing a figure always helps!)
1. Jointly distributed continuous random variables 115

Homework Problems

Exercise 23.1. Let X and Y be two continuous random variables with joint
density given by
1
4 if − 1 6 x 6 1 and − 1 6 y 6 1,
f(x, y) =
0 otherwise.
Compute the following probabilities:
(a) P{X + Y 6 21 },
(b) P{X − Y 6 12 },
(c) P{XY 6 41 },
Y
(d) P X 61 ,
 Y 21
(e) P X 6 2 ,
(f) P{|X| + |Y| 6 1},
(g) P{|Y| 6 eX }.
 Lecture 24

1. Marginals, distribution functions, etc.

If (X, Y) has joint density f, then
FX (a) = P{X 6 a} = P{(X, Y) ∈ A},
where A = {(x, y) : x 6 a}. Thus,
Z a Z ∞ 
FX (a) = f(x , y) dy dx.
−∞ −∞

Differentiate, and apply the fundamental theorem of calculus, to find that

Z∞
fX (a) = f(a , y) dy.
−∞

Similarly,
Z∞
fY (b) = f(x , b) dx.
−∞

Example 24.1 (Example 23.4, continued). Let


8xy if 0 < y < x < 1,
f(x , y) =
0 otherwise.
Then,
R x
08xy dy if 0 < x < 1,
fX (x) =
0 otherwise.

4x3 if 0 < x < 1,
=
0 otherwise.

117
118 24

[Note the typo in Stirzaker’s text, page 341.] Similarly,

R 1
y 8xy dx if 0 < y < 1,
fY (y) =
0 otherwise.

4y(1 − y2 ) if 0 < y < 1,
=
0 otherwise.
Example 24.2. Suppose (X, Y) is distributed uniformly in the circle of ra-
dius one about (0 , 0). That is,

1
if x2 + y2 6 1,
f(x , y) = π
0 otherwise.

Then,
Z √ 2


1−x
1
√ dy if −1 < x < 1,
fX (x) = 2 π

0
− 1−x
otherwise.
 p
 2 1 − x2 if −1 < x < 1,
= π
0 otherwise.
By symmetry, fY is the same function.

2. Independence
Just as in the discrete case, two continuous random variables are said to be
independent if fX,Y (x, y) = fX (x)fY (y), for all x and y. As a consequence,
one has
Z Z
P{X ∈ A, Y ∈ B} = fX,Y (x, y) dx dy = fX (x)fY (y) dx dy
A×B A×B
Z Z
= fX (x) dx fY (y) dy = P{X ∈ A}P{Y ∈ B}.
A B
This actually implies that if X and Y are independent, then f(X) and g(Y)
are also independent, for any functions f and g. We omit the short proof.
Example 24.3. Let X ∼ Exponential(λ1 ) and Y ∼ Exponential(λ2 ). What is
Z = min(X, Y)?
Let us compute
FZ (z) = P{min(X, Y) 6 z} = 1 − P{X > z, Y > z}
= 1 − P{X > z}P{Y > z} = 1 − (1 − FX (z))(1 − FY (z))
= 1 − e−λ1 z e−λ2 z = 1 − e−(λ1 +λ2 )z .
2. Independence 119

Thus, Z ∼ Exponential(λ1 + λ2 ). This makes sense: say you have two sta-
tions, with the first serving about λ1 people per unit time and the second
serving about λ2 people per unit time. Then, being served by these sta-
tions in a row is equivalent to being served by one station that serves about
λ1 + λ2 people per unit time.

It is noteworthy that X and Y are independent as soon as one can write

fX,Y (x, y) as the product of a function of x and a function of y. That is,
if and only if fX,Y (x, y) = h(x)g(y), for some functions h and g. This is
because we then have
 Z∞   Z∞ 
fX (x) = h(x) g(y) dy and fY (y) = g(y) h(x) dx
−∞ −∞
and
 Z∞  Z ∞ 
h(x) dx g(y) dy = 1
−∞ −∞
so that fX,Y (x, y) = fX (x)fY (y). In other words, the functions h and g are
really the same as the marginal density functions fX and fY , up to the
multiplicative constants that would make them integrate to one.
Example 24.4. Suppose (X, Y) is distributed uniformly on the square that
joins the origin to the points (1 , 0), (1 , 1), and (0 , 1). Then,

1 if 0 < x < 1 and 0 < y < 1,
fX,Y (x , y) =
0 otherwise.
Here, we see that fX,Y (x, y) does split into a product of a function of x and
a function of y. Indeed, both 1 = 1 × 1 and 0 = 0 × 0. Furthermore, the set
0 < x < 1 and 0 < y < 1 is a set that involves two independent conditions
on x and y. In fact, the marginals are equal to

1 if 0 < x < 1,
fX (x) =
0 otherwise,
and 
1 if 0 < y < 1,
fY (y) =
0 otherwise,
and thus we see clearly that fX,Y (x, y) = fX (x)fY (y). Note that we have
just shown that X and Y are both uniformly distributed on (0, 1).
Example 24.5. Let X and Y have joint density fX,Y (x, y) = 14 (1 + xy), for
−1 6 x 6 1 and −1 6 y 6 1. Then, the marginals are
Z1 Z
1 x 1 1
fX (x) = dy + y dy = ,
−1 4 4 −1 2
120 24

for −1 6 x 6 1, and similarly fY (y) = 12 , for −1 6 y 6 1. However, clearly

fX,Y (x, y) 6= fX (x)fY (y). This shows that X and Y are not independent.
To confirm that this is consistent with intuition we compute P{X >
0 and Y > 0} and P{X > 0}P{Y > 0}. First,
Z1 Z1
1 1 1 1 1 5
P{X > 0, Y > 0} = (1 + xy) dx dy = + × × = .
0 0 4 4 4 2 2 16
On the other hand, P{X > 0} = P{Y > 0} = 1/2, since both X and Y
R1 R1
are Uniform(−1, 1). (Alternatively, compute −1 0 41 (1 + xy) dx dy = 1/2.)
Thus
1 5
P{X > 0}P{Y > 0} = 6= = P{X > 0 and Y > 0}.
4 16
Example 24.6 (Example 24.1, continued). Let

8xy if 0 < y < x < 1,
f(x , y) =
0 otherwise.
It is tempting to say that X and Y are then independent, since f(x, y) seems
to be a product of two functions, one of x and one of y. However, one has
to be careful with the set: 0 < y < x < 1. This is where the dependence
occurs. Indeed, if we know that x = 1/2, then we know that y cannot
be larger than 1/2. This is made clear once we compute the marginals, in
Example 24.1, and observe that indeed fX,Y (x, y) is not equal to fX (x)fY (y).

The same caution needs to be applied to Example 24.2.

Example 24.7 (Order Statistic). Let X1 , . . . , Xn be independent random
variables with the same CDF F(x). We want to compute the CDF of
S = min(X1 , . . . , Xn ), the smallest of them. Then,
FS (s) = P{S 6 s} = 1 − P{S > s} = 1 − P{X1 > s, . . . , Xn > s}
= 1 − P{X1 > s}P{X2 > s} · · · P{Xn > s}
by independence. Because the variables have the same CDF, P{X1 > s} =
· · · = P{Xn > s} = 1 − F(s). Thus,
FS (s) = 1 − (1 − F(s))n .
Similarly, if T = max(X1 , . . . , Xn ) is the largest of the variables, then
FT (t) = P{T 6 t} = P{X1 6 t, . . . , Xn 6 t}
= P{X1 6 t}P{X2 6 t} · · · P{Xn 6 t}
= P{X1 6 t}n ,
and
FT (t) = (F(t))n .
2. Independence 121

Homework Problems

Exercise 24.1. Let X and Y be two continuous random variables with joint
density given by
c(x + y) if 0 6 x 6 1 and 0 6 y 6 1,
f(x, y) =
0 otherwise.
(a) Find c.
(b) Compute P{X < Y}.
(c) Find the marginal densities of X and Y.
(d) Compute P{X = Y}.
Exercise 24.2. Let X and Y be two continuous random variables with joint
density given by

 4xy if 0 6 x 6 1, 0 6 y 6 1 and x > y
f(x, y) = 6x2 if 0 6 x 6 1, 0 6 y 6 1 and x < y

0 otherwise.
(a) Find the marginal densities of X and Y.
(b) Let A = {X 6 12 } and B = {Y 6 12 }. Find P(A ∪ B).
Exercise 24.3. Let X and Y be two continuous random variables with joint
density given by
2e−(x+y) if 0 6 y 6 x,
f(x, y) =
0 otherwise.
Find the marginal densities of X and Y.
Exercise 24.4. Let (X, Y) be uniformly distributed over the parallelogram
with vertices (−1, 0), (1, 0), (2, 1), and (0, 1).
(a) Find and sketch the density functions of X and Y.
(b) A new random variable Z is defined by Z = X + Y. Show that Z is
a continuous random variable and find and sketch its probability
density function.
Exercise 24.5. Let (X, Y) be continuous random variables with joint density
f(x, y) = (x + y)/8, 0 6 x 6 2, 0 6 y 6 2; f(x, y) = 0 elsewhere.
(a) Find the probability that X2 + Y 6 1.
(b) Find the conditional probability that exactly one of the random
variables X and Y is 6 1, given that at least one of the random
variables is 6 1.
(c) Determine whether or not X and Y are independent.
 Lecture 25

1. Functions of a random vector

Basic problem: If (X, Y) has joint density f, then what, if any, is the joint
density of (U, V), where U = u(X, Y) and V = v(X, Y)? Or equivalently,
(U, V) = T (X, Y), where
 
u(x , y)
T (x , y) = .
v(x , y)
Example 25.1. Let (X, Y) be distributed uniformly in the disk of radius
ρ > 0 about the origin in the plane. Thus,

 1 if x2 + y2 6 ρ2 ,
fX,Y (x , y) = πρ2

0 otherwise.
√
We wish to write (X, Y), in polar coordinates, as (R, Θ), where R = X2 + Y 2
and Θ = arctan(Y/X). Then, we compute first the joint distribution function
FR,Θ of (R, Θ):
FR,Θ (r , θ) = P{R 6 r , Θ 6 θ}
= P{(X, Y) ∈ A},

where A is the “partial cone” {(x , y) : x2 + y2 6 r2 , arctan(y/x) 6 θ}. If

0 < r < ρ and −π < θ < π, then
ZZ
FR,Θ (r , θ) = fX,Y (x , y) dx dy
A
Zr Zθ
1
= s ds dϕ,
0 0 πρ2

123
124 25

p
after the change of variables s = x2 + y2 and ϕ = arctan(y/x). There-
fore, for all r ∈ (0 , ρ) and θ ∈ (−π , π),
r2 θ
FR,Θ (r , θ) = .
2πρ2
Rr Rθ
Since, by definition, FR,Θ (r, θ) = −∞ −∞ fR,Θ (s, ϕ) ds dϕ, we see that
∂2 FR,Θ
fR,Θ (r , θ) = (r , θ).
∂r∂θ
It is also clear that fR,Θ (r, θ) = 0 if r 6∈ (0, ρ) or θ 6∈ (−π, π). Therefore,

 r if 0 < r < ρ and −π < θ < π,
fR,Θ (r , θ) = πρ2
0 otherwise.
1
Observe that the above yields fΘ (θ) = 2π , if −π < θ < π, which implies
that Θ is Uniform(−π, π). On the other hand, fR (r) = ρ2r2 , if 0 < r < ρ,
which implies that R is not Uniform(0, ρ). Indeed, it is more likely to pick
a point with a larger radius (since there are more of them!).

The previous example can be generalized. Suppose T is invertible with

inverse function  
−1 x(u , v)
T (u , v) = .
y(u , v)
The Jacobian of this transformation is
∂(x , y) ∂x ∂y ∂x ∂y
= − .
∂(u , v) ∂u ∂v ∂v ∂u
Theorem 25.2. If T is “nice,” then
∂(x , y)
fU,V (u , v) = fX,Y (x(u , v) , y(u , v)) .
∂(u , v)
Compare the above to Theorem 19.2. The Jacobian is really what comes
up when doing change of variables in calculus, which is what the above
theorem is all about. We thus omit the proof.
Example 25.3. In the polar coordinates example (r = u, θ = v),
p
r(x , y) = x2 + y2 , θ(x , y) = arctan(y/x) = θ,
x(r , θ) = r cos θ, y(r , θ) = r sin θ.
Therefore, for all r > 0 and θ ∈ (−π , π), the Jacobian equals
(cos(θ) × r cos(θ)) − (−r sin(θ) × sin(θ)) = r.
Hence,
fR,Θ (r , θ) = r fX,Y (r cos θ , r sin θ).
1. Functions of a random vector 125

y=1
x=v
y=1

x=0
x=y u=v x²=y²=v
v=u²

Figure 25.1. Domains for pdfs in Example 25.4. Left: domain transfor-
mation. Right: integration area for CDF calculation.

You should check that this yields Example 25.1, for instance.
Example 25.4. Let fX,Y (x, y) = 2(x + y), if 0 < x < y < 1, and 0 other-
wise. We want to find fXY . In this case, we will first find fX,XY , and then
integrate the first coordinate out. This means we will use the transforma-
tion (u, v) =√(x, xy). Solving for x and y we get (x, y) = (u, v/u), with
0 < v < u < v < 1; see Figure 25.1. The Jacobian is then equal to
1 −v 1
1× −0× 2 = .
u u u
√
As a result, fU,V (u, v) = 2(u + v/u)/u, with 0 < v < u < v < 1, and
Z √v
fV (v) = 2 (1 + v/u2 ) du = 2(1 − v), for 0 < v < 1.
v

Alternatively, we could have computed the CDF of XY and then took

its derivative to find the pdf. Clearly, 0 < XY < 1 and thus FXY (v) = 0 for
v 6 0 and FXY (v) = 1 for v > 1. For 0 < v < 1 we have (see Figure 25.1)
ZZ
FXY (v) = P{XY 6 v} = fX,Y (x, y) dx dy
xy6v
Zv Z1 Z √v Z v/x
= 2(x + y) dy dx + 2(x + y) dy dx
0 x v x
Zv Z √v
= (2x − 2x2 + 1 − x2 ) dx + (2v − 2x2 + v2 /x2 − x2 ) dx
0 v
= (v2 − 2v3 /3 + v − v3 /3)
√ √
+ (2v v − 2v2 − 2v3/2 /3 + 2v3 /3 − v2 / v + v2 /v − v3/2 /3 + v3 /3)
= 2v − v2 .
0 (v) = 2 − 2v, for 0 < v < 1 (and 0 otherwise).
And fXY (v) = FXY
126 25

Homework Problems

Exercise 25.1. Let X and Y be independent and uniformly distributed be-

tween 0 and 1. Find and sketch the distribution and density functions of
the random variable Z = Y/X2 .
Exercise 25.2. Let X and Y be two continuous independent random vari-
ables with densities given by
e−x if x > 0,
f(x) =
0 otherwise.
(a) Find the probability density function of Z = X + Y.
Y
(b) Find the probability density function of W = X.

Exercise 25.3. Let X and Y be two independent random variables both with
Y
distribution N(0, 1). Find the probability density function of Z = X .
Exercise 25.4. A point-size worm is inside an apple in the form of the
sphere x2 + y2 + z2 = 4a2 . (Its position is uniformly distributed.) If the
apple is eaten down to a core determined by the intersection of the sphere
and the cylinder x2 + y2 = a2 , find the probability that the worm will be
eaten.
Exercise 25.5. A point (X, Y, Z) is uniformly distributed over the region
described by x2 + y2 6 4, 0 6 z 6 3x. Find the probability that Z 6 2X.
Exercise 25.6. Let T1 , . . . , Tn be the order statistics of X1 , . . . , Xn . That is, T1
is the smallest of the X’s, T2 is the second smallest, and so on. Tn is the
largest of the X’s. Assume X1 , . . . , Xn are independent, each with density
f. Show that the joint density of T1 , . . . , Tn is given by g(t1 , . . . , tn ) =
n! f(t1 , . . . , tn ) if t1 < t2 < · · · < tn and 0 otherwise.
Hint: First find P{T1 6 t1 , . . . , Tn 6 tn , X1 < X2 < . . . < Xn }.
Exercise 25.7. Let X, Y and Z be three continuous independent random
variables with densities given by
e−x if x > 0,
f(x) =
0 otherwise.
Compute P{X > 2Y > 3Z}.
Exercise 25.8. A man and a woman agree to meet at a certain place some
time between 10 am and 11 am. They agree that the first one to arrive will
wait 10 minutes for the other to arrive and then leave. If the arrival times
are independent and uniformly distributed, what is the probability that
they will meet?
1. Functions of a random vector 127

Exercise 25.9. When commuting to work, John can take public transporta-
tion (first a bus and then a train) or walk. Buses ride every 20 minutes and
trains ride every 10 minutes. John arrives at the bus stop at 8 am precisely,
but he doesn’t know the exact schedule of buses, nor the exact schedule of
trains. The total travel time on foot (resp. by public transportation) is 27
minutes (resp. 12 minutes).
(a) What is the probability that taking public transportation will take
more time than walking?
(b) If buses are systematically 2 minutes late, how does it change the
probability in (a)?
Exercise 25.10. Let X and Y be two independent random variable, both
with distribution N(0, σ2 ) for some σ > 0. Let R and Θ be two random
variables defined by
X = R cos(Θ),
Y = R sin(Θ),
where R > 0. Prove that R and Θ are independent and find their density
functions.
Exercise 25.11. A chamber consists of the inside of the cylinder x2 + y2 =
1. A particle at the origin is given initial velocity components vx = U
and vy = V, where (U, V) are independent random variables, each with
standard normal density. There is no motion in the z-direction and no
force acting on the particle after the initial push at time t = 0. If T is
the time at which the particle strikes the wall of the chamber, find the
distribution and density functions of T .
 Lecture 26

1. Functions of a random vector, continued

Example 26.1. Let us compute the joint density of U = X and V = X + Y.
Here,
u(x , y) = x, v(x , y) = x + y
x(u , v) = u, y(u , v) = v − u.
Therefore, the Jacobian equals
(1 × 1) − (0 × −1) = 1.
Consequently,
fU,V (u , v) = fX,Y (u , v − u).
This has an interesting by-product: The density function of V = X + Y is
Z∞ Z∞
fV (v) = fU,V (u , v) du = fX,Y (u , v − u) du.
−∞ −∞
Compare with Theorem 22.2.
Example 26.2. Let X and Y be two independent standard normal random
variables. We want to find the joint density of U = X + Y and V = X − Y.
Solving for x and y we get x = (u + v)/2 and y = (u − v)/2. The Jacobian
is then equal to
1 −1 1 1 1
× − × =− .
2 2 2 2 2
1 −(x2 +y2 )/2
The joint pdf of X and Y is fX,Y (x, y) = fX (x)fY (y) = 2π e . Thus,
1 1 1 −u2 /4 −v2 /4
− (u + v)2 + (u − v)2 /4 × =


fU,V (u, v) = exp e e .
2π 2 2π

129
130 26

This in fact shows that U and V are independent, even though they are
both mixtures of both X and Y. It also shows that they are both normal
random variables with parameters (mean) 0 and (variance) 2; i.e. N(0, 2).

Now, we will start building up the necessary material to make the link
between the mathematical definition of probability (state space, function
on events, etc) and the intuitive one (relative frequency). The starting point
is the notion of mathematical expectation.

2. Mathematical Expectation: Discrete random variables

The mathematical expectation (or just the expectation, or mean, or average)
E[X] of a discrete random variable X with mass function f is defined for-
mally as the average of the possible values of X, weighted by their corre-
sponding probabilities:
X
E[X] = x f(x). (26.1)
x

When X has finitely many possible values the above sum is well de-
fined. It corresponds to the physical notion of center of gravity of point
masses placed at positions x with weights f(x).
Example 26.3. We toss a fair coin and win $1 for heads and lose $1 for tails.
This is a fair game since the average winnings equal $0. Mathematically,
if X equals the amount we won, then E[X] = 1 × 12 + (−1) × 21 = 0.
Example 26.4. We roll a die that is loaded as follows: it comes up 6 with
probability 0.4, 1 with probability 0.2, and the rest of the outcomes come
up with probability 0.1 each. Say we lose $2 if the die shows a 2, 3, 4, or 5,
while we win $1 if it shows a 1 and $2 if it shows a 6. On average we win
−2 × 4 × .1 + 1 × 0.2 + 2 × 0.4 = 0.2;
that is we win 20 cents. In a simple case like this one, where X has a finite
amount of possible values, one can use a table:
x −2 1 2
f(x) = P{X = x} 4 × 0.1 0.2 0.4
xf(x) −0.8 0.2 0.8
E[X] = 0.2 is then the sum of the elements in the last row. Intuitively,
this means that if we play, say, 1000 times, we expect to win about $200.
Making this idea more precise is what we mean by “connecting the math-
ematical and the intuitive definitions of probability.” This also gives a fair
price to the game: 20 cents is a fair participation fee for each attempt.
2. Mathematical Expectation: Discrete random variables 131

Example 26.5. You role a fair die and lose as many dollars as pips shown
on the die. Then, you fairly toss an independent fair coin a number of
times equal to the outcome of the die. Each head wins you $2 and each
tail loses you $1. Is this a winning or a losing game? Let X be the amount
of dollars you win after having played the game. Let us compute the
average winning. First, we make a table of all the outcomes.

Outcome 1H 1T 2H 1H1T 2T
x −1 + 2 −1 − 1 −2 + 4 −2 + 2 − 1 −2 − 2
1 1 1 1 1 1
f(x) 6 × 2 6 × 2 6 × 4 2 × 61 × 14 1
6 × 4
1

Outcome 3H 2H1T 1H2T 3T 4H

x −3 + 6 −3 + 4 − 1 −3 + 2 − 2 −3 − 3 −4 + 8
1 1
f(x) 6 × 8 3 × 61 × 18 3 × 61 × 81 1
6 × 8
1 1 1
6 × 16
Outcome 3H1T 2H2T 1H3T 4T 5H
x −4 + 6 − 1 −4 + 4 − 2 −4 + 2 − 3 −4 − 4 −5 + 10
f(x) 4 × 16 × 161
6 × 61 × 161
4 × 16 × 16
1 1
6 × 16
1 1 1
6 × 32
Outcome 4H1T 3H2T 2H3T 1H4T 5T
x −5 + 8 − 1 −5 + 6 − 2 −5 + 4 − 3 −5 + 2 − 4 −5 − 5
f(x) 5 × 16 × 321
10 × 16 × 321
10 × 16 × 321
5 × 16 × 321 1 1
6 × 32
Outcome 6H 5H1T 4H2T 3H3T 2H4T
x −6 + 12 −6 + 10 − 1 −6 + 8 − 2 −6 + 6 − 3 −6 + 4 − 4
1 1
f(x) 6 × 64 6 × 61 × 641
15 × 16 × 641
20 × 61 × 641 1
15 × 61 × 64
Outcome 1H5T 6T
x −6 + 2 − 5 −6 − 6
f(x) 6 × 16 × 641 1
6 × 64
1

Then,

X 7
E[X] = x f(x) = − = −1.75.
4

In conclusion, the game is a losing game. In fact, I would only play if they
pay me a dollar and 75 cents each time!

Example 26.6. If X ∼ Bernoulli(p), then E[X] = p × 1 + (1 − p) × 0 = p.

More generally, if X ∼ Binomial(n , p), then I claim that E[X] = np. Here is
132 26

why:
f(k)
z  }| {
X
n
n k n−k
E[X] = k p (1 − p)
k
k=0
Xn
n!
= pk (1 − p)n−k
(k − 1)!(n − k)!
k=1
Xn  
n − 1 k−1
= np p (1 − p)(n−1)−(k−1)
k−1
k=1
X n − 1
n−1
= np pj (1 − p)(n−1)−j
j
j=0

= np(p + (1 − p))n−1 = np,

thanks to the binomial theorem.
 Lecture 27

1. Mathematical Expectation: Discrete random variables,

continued
If X has infinitely-many possible values, then the sum in (26.1) must be
defined. If P{X > 0} = 1 (i.e. all possible values of X are nonnegative), then
the sum in question is that of nonnegative numbers and is thus always
defined [though could be ∞]. Similarly, if P{X 6 0} = 1, then the sum is
that of nonpositive numbers and is always defined [though could be −∞].

Example 27.1. Suppose X ∼ Poisson(λ). Then, I claim that E[X] = λ. In-

deed,

∞
X ∞
X
e−λ λk e−λ λk−1
E[X] = k =λ
k! (k − 1)!
k=0 k=1
X∞
e−λ λj
=λ = λ,
j!
j=0

P
because eλ = ∞ j
j=0 λ /j!, thanks to Taylor’s expansion. So when modeling
the length of a waiting line, the parameter λ is the average length of the
line.

133
134 27

Example 27.2. Suppose X is negative binomial with parameters r and p.

Then, E[X] = r/p because

∞
X  
k−1 r
E[X] = k p (1 − p)k−r
r−1
k=r
X∞
k!
= pr (1 − p)k−r
(r − 1)!(k − r)!
k=r
X∞  
k r
=r p (1 − p)k−r
r
k=r
∞  
r X k r+1
= p (1 − p)(k+1)−(r+1)
p r
k=r
∞ 
r X

j−1
= pr+1 (1 − p)j−(r+1)
p (r + 1) − 1
j=r+1 | {z }
P{Negative binomial (r+1 ,p)=j}
r
= .
p

Thus, for example, E[Geometric(p)] = 1/p.

Example 27.3 (St.-Petersbourg paradox). Here is an example of a random

variable with infinite expectation. Let us toss a fair coin until we get
heads. The first toss wins us $2, and then each consecutive toss dou-
bles the winnings. So if X is the amount we win, then it has the mass
function f(2n ) = 1/2n , for n > 1; i.e. X = 2n with probability 1/2n . This is
a nonnegative random variable and thus the expectation must be defined.
However, 2n f(2n ) = 1 for all n > 1. Thus, the sum of these terms is indeed
infinite. This means that the game has an infinite price and you should be
willing to play regardless of the fee. The paradox is that this contradicts
our instincts. For example, if you are asked to pay $4 to play the game,
then you will probably agree since all you need is to get tails on your first
toss, which you assess as being quite likely. On the other hand, if you are
asked to pay $32, then you may hesitate. In this case, you need to get 4
tails in a row to break even, which you estimate as being quite unlikely.
But what if you get 5 tails in a row? Then you get the $32 back and get $32
more! This is what is hard to grasp. The unrealistic part of this paradox
is that it assumes the bank has infinite supplies and that in the unlikely
event of you getting 265 tails in a row, they will have $2266 to pay you!
(This is more than 1080 , which is the estimated number of atoms in the
observable universe!)
1. Mathematical Expectation: Discrete random variables, continued 135

If X has infinitely-many possible values but can take both positive and
negativePvalues, then we have to be careful with the definition of the sum
E[X] = x f(x). We can always add the positive and negative parts sepa-
rately. So, formally, we can write
X X
E[X] = x f(x) + x f(x).
x>0 x<0
Now, we see that if one of these two sums is finite then, even if the other
were infinite, E[X] would be well defined. Moreover, E[X] is finite if, and
only if, both sums are finite; i.e. if
X
|x| f(x) < ∞.

Example 27.4. Say X has the mass function f(2n ) = f(−2n ) = 1/2n , for
P
n > 2. (Note that the probabilities do add up to one: 2 n>2 21n = 1.)
Then, the positive part of the sum gives
X 1
2n × n = ∞,
2
n>2

and the negative part of the sum gives

X 1
(−2n ) × n = −∞.
2
n>2

This implies that E[X] is not defined. In fact, if we compute

X
N
1 X
M
1
2n × n
= N − 1 and (−2n ) × n = −M + 1,
2 2
n=2 n=2
then, in principle, to get E[X] we need to add the two and take N and M
to infinity. But we now see that the sum equals N − M and so depending
on how we take N and M to infinity, we get any value we want for E[X].
Indeed, if we take N = 2M, then we get ∞. If we take M = 2N we get
−∞. And if we take N = M + a, we get a, for any integer a.
136 27

Homework Problems

Exercise 27.1. In Las Vegas, a roulette is made of 38 boxes, namely 18

black boxes, 18 red boxes, a box ’0’ and a box ’00’. If you bet $1 on ’black’,
you get $2 if the ball stops in a black box and $0 otherwise. Let X be your
profit. Compute E[X].
Exercise 27.2. In the game Wheel of Fortune, you have 52 possible outcomes:
one “0”, one “00”, two “20”, four “10”, seven “5”, fifteen “2” and twenty-
two “1”. If you bet $1 on some number, you receive this amount of money
if the wheel stops on this number. (In particular, you do NOT lose the $1
you bet.) If the wheel stops at a different number, you lose the $1 you bet.
If you bet $1 on “0” or “00”, you receive $40 if the wheel stops on this
number (and in this case you do not lose the $1 you bet). For example, say
you bet $1 on 10. If the wheel stops on 10, your profit is $10. If it stops
on something other than 10, your profit is -$1 because you lose the $1 you
bet.
(a) Assume you bet $1 on each of the seven possible numbers or sym-
bols (for a total of $7), what is the expectation of your profit?
(b) Assume you want to bet $1 on only one of these numbers or sym-
bols, which has the best (resp. worst) profit expectation?

Remark: Try to redo the exercise with the assumption that you always
lose the $1 you bet. See how part (b) changes drastically, with just this
small change in the rules of the game!
Exercise 27.3. Let X be a Geometric random variable with parameter p ∈
[0, 1]. Compute E[X].
 Lecture 28

1. Mathematical Expectation: Continuous random variables

When X is a continuous random variable with density f(x), we can re-
peat the same reasoning as for discrete random variables and obtain the
formula

Z∞
E[X] = x f(x) dx.
−∞

R∞
The same issues as before arise: if −∞ |x| f(x) dx < ∞, then the above
R∞
integral is well defined and finite. If, on the other hand, 0 x f(x) dx < ∞
R0
but −∞ x f(x) dx = −∞, then the integral is again defined but equals −∞.
R∞ R0
Conversely, if 0 x f(x) dx = ∞ but −∞ x f(x) dx > −∞, then E[X] = ∞.
Finally, if both integrals are infinite, then E[X] is not defined.

Example 28.1 (Uniform). Suppose X is uniform on (a , b). Then,

Zb
1 1 b 2 − a2 1 (b − a)(b + a) b+a
E[X] = x dx = = = .
a b−a 2 b−a 2 b−a 2

N.B.: The formula of the first example on page 303 of Stirzaker’s text is
wrong.

137
138 28

Example 28.2 (Gamma). If X is Gamma(α , λ), then for all positive values
α
of x we have f(x) = Γλ(α) xα−1 e−λx , and f(x) = 0 for x < 0. Therefore,
Z
λα ∞ α −λx
E[X] = x e dx
Γ (α) 0
Z∞
1
= zα e−z dz (z = λx)
λΓ (α) 0
Γ (α + 1)
=
λΓ (α)
α
= .
λ
In the special case that α = 1, λ1 is the expectation of an exponential
random variable with parameter λ. So when modeling a waiting time,
the parameter of the exponential is one over the average waiting time. The
parameter λ is thus equal to the serving rate: the number of people served
per unit time. Now, you should understand a bit better the derivation of
the exponential distribution that came after Exercise 15.2. Namely, if we
recall from Exercise 27.2 that the average of a geometric with parameter p
is 1/p, we see that if we use p = λ/n, the average will be n/λ. If each coin
flip takes 1/n seconds, then the average serving time is 1/λ, as desired.
Another observation is that E[Gamma(α, λ)] = α/λ the same way as
E[Negative Binomial(r, p)] = r/p. This is not a coincidence and one can
derive the Gamma distribution from the negative binomial similarly to
how the exponential was derived from a geometric.
(x−µ)2
Example 28.3 (Normal). Suppose X ∼ N(µ , σ2 ); i.e. f(x) = σ√12π e− 2σ2 .
Then,
Z∞
(x − µ)2
 
1
E[X] = √ x exp − dx
σ 2π −∞ 2σ2
Z∞
1 2
=√ (µ + σz)e−z /2 dz (z = (x − µ)/σ)
2π −∞
Z ∞ −z2 /2 Z
e σ ∞ 2
=µ √ dz + √ ze−z /2 dz
2π 2π −∞
|−∞ {z } | {z }
1 0, by symmetry
= µ.
Example 28.4 (Cauchy). In this example, f(x) = π−1 (1 + x2 )−1 . Note that
the expectation is defined only if the following limit exists regardless of
how we let n and m tend to ∞:
Z
1 n y
dy.
π −m 1 + y2
2
1. Mathematical Expectation: Continuous random variables 139

Now I argue that the limit does not exist; I do so by showing two different
choices of (n , m) which give rise to different limiting “integrals.”
2
Suppose m = eπ a n, for some fixed number a. Then,
Z Z Z −π2 a n
1 n y 1 n y 1 e y
dy = 2 dy − 2 dy
π2 −e−π2 a n 1 + y2 π 0 1 + y2 π 0 1 + y2
Z 2 Z −2π2 a n2
1 1+n dz 1 1+e dz
= 2 − 2 (z = 1 + y2 )
2π 1 z 2π 1 z
1 + n2
 
1
= 2 ln 2
2π 1 + e−2π a n2
1 2
→ 2 ln e2π a = a as n → ∞.
2π
Thus, we can make the limit converge to any number a we want. In fact,
taking m = n2 and repeating the above calculation √ allows us to make the
limit converge to −∞, while taking m = n makes the limit equal to
∞. The upshot is that the Cauchy density does not have a well-defined
expectation. [That is not to say that the expectation is well defined, but
infinite.] In particular, we conclude that E[|X|] = ∞.
Theorem 28.5. If X is a positive random variable with density f, then
Z∞ Z∞
E[X] = P{X > x} dx = (1 − F(x)) dx.
0 0

Proof. The second identity is a consequence of the fact that 1 − F(x) =

P{X
R∞ > x}. In order to prove the first identity note that P{X > x} =
x f(y) dy. Therefore, if A = {(x, y) : y > x > 0} then
Z∞ Z∞  Z∞  ZZ
P{X > x} dx = f(y) dy dx = f(y) dx dy
0 0 x A
Z∞  Zy  Z∞
= f(y) dx dy = y f(y) dy
0 0 0
= E[X]. 

If X is a negative random variable, then −X is positive and we have

Z∞ Z0
E[X] = −E[(−X)] = − P{X < −x} dx = − P{X < x} dx.
0 −∞
R∞
If X takes negative and positive values, and at least one of 0 P{X > x} dx
R0
and −∞ P{X < x} dx is finite, then E[X] equals their difference:
Z∞ Z0
E[X] = P{X > x} dx − P{X < x} dx.
0 −∞
140 28

Note that the above formula does not involve the density function f.
It turns out (and we omit the math) that we can define the expectation of
any positive random variable (discrete, continuous, or other) using that
formula. That is to say the notion of mathematical expectation (or average
value) is general and applies to any real-valued random variable.
 Lecture 29

1. Some properties of expectations

The following theorem is useful when computing averages of transforma-
tions of random variables.
P
Theorem
P 29.1. If X has massP
function f(x) and x g(x) f(x) is well defined, i.e.
x:g(x)>0 g(x) f(x) < ∞ or x:g(x)60 g(x) f(x) > −∞, then
X
E[g(X)] = g(x)f(x).
x

Proof. Let Y = g(X). Then, by definition

X X X
E[g(X)] = E[Y] = y P{Y = y} = y P{X = x}
y y x:g(x)=y
X X X
= g(x) P{X = x} = g(x) P{X = x}
y x:g(x)=y x

as desired. 

The above can be generalized to the case of two random variables.

Theorem 29.2. If (X, Y) have joint mass function f(x, y) and g(x, y) is some
function, then
X
E[g(X, Y)] = g(x, y) f(x, y),
x,y

provided the sum is well defined.

The same holds in the continuous case.

141
142 29

Theorem 29.3. If (X, Y) have joint density function f(x, y) and g(x, y) is some
function, then
Z∞ Z∞
E[g(X, Y)] = g(x, y) f(x, y) dx,
−∞ −∞

provided the integral is well defined. In particular, if X has density f(x) and g(x)
is some function, then
Z∞
E[g(X)] = g(x) f(x) dx,
−∞

provided the integral is well defined.

Now, we can prove the following natural properties.

Theorem 29.4. Let X and Y be any random variables (discrete, continuous, or

other) with well defined expectations E[X] and E[Y]. Let a be any (nonrandom)
number. Then:

(1) E[aX] = aE[X];

(2) If either E[X] or E[Y] is finite, or if they are both ∞ or both −∞, then
E[X + Y] = E[X] + E[Y].
(3) E[a] = a;
(4) If P{X 6 Y} = 1, then E[X] 6 E[Y];
(5) If P{X > 0} = 1 and E[X] = 0, then P{X = 0} = 1;
(6) If X and Y are independent and are either both nonnegative, both non-
positive, or both have finite expectations, then E[XY] = E[X]E[Y].

Proof. We show the proofs in the discrete case. The proofs in the contin-
uous case are similar, and the proofs in the general case are omitted. To
prove (1) let x1 , x2 , . . . be the possible values of X. Then, ax1 , ax2 , . . . are
the possible values of aX. Moreover,
X X
E[aX] = axi f(xi ) = a xi f(xi ) = aE[X].
i i

Let us now prove (2). We will only treat the case where both variables are
nonnegative. Let x1 , x2 , . . . be the possible (nonnegative) values of X and
y1 , y2 , . . . the possible (nonnegative) values of Y. Then, the possible values
1. Some properties of expectations 143

of X + Y are {xi + yj : i = 1, 2, . . . , j = 1, 2, . . . } and are nonnegative. Thus,

X
E[X + Y] = (xi + yj )P{X = xi , Y = yj }
i,j
X X
= xi P{X = xi , Y = yj } + yj P{X = xi , Y = yj }
i,j i,j
X X X X
= xi P{X = xi , Y = yj } + yj P{X = xi , Y = yj }
i j j i
X X
= xi P{X = xi } + yj P{Y = yj }
i j
= E[X] + E[Y].
In the second-to-last equality we used the fact that the sets {X = xi } are
disjoint and their union is everything, and the same for the sets {Y = yj }.
If now X and Y are both nonpositive, then −X and −Y are nonnegative
and we can use property (1) to write
E[X] + E[Y] = −(E[−X] + E[−Y]) = −E[−(X + Y)] = E[X + Y].

If at least one of the variables takes positive and negative values, then
one needs to use slightly more involved arguments requiring facts about
infinite series. We omit the proof in this case.
Next, we prove (3). The only value the random variable a takes is
a and it takes it with probability 1. Thus, its mathematical expectation
simply equals a itself. To prove (4) observe that Y − X is a nonnegative
random variable; i.e. its possible values are all nonnegative. Thus, it has
a nonnegative average. But by (1) and (2) we have 0 6 E[Y − X] = E[Y] +
E[−X] = E[Y] − E[X]. Property (5) is obvious since if there
P existed an x0 > 0
for which f(x0 ) > 0, then we would have had E[X] = xf(x) > x0 f(x0 ) >
0, since the sum is over x > 0.
Finally, we prove (6). Again, let xi and yj be the possible values of X
and Y, respectively. Then, the possible values of XY are given by the set
{xi yj : i = 1, 2, . . . , j = 1, 2, . . . }. Thus,
X
E[XY] = xi yj P{X = xi , Y = yj }
i,j
X
= xi yj P{X = xi }P{Y = yj } (by independence)
i,j
X X
= xi P{X = xi } yj P{Y = yj }
i j
= E[X]E[Y].
144 29

The third equality was simply the result of summing over j first and then
over i. We can sum in any order because the terms are either of the same
sign, or are summable (if the expectations of X and Y are finite). 

As a consequence of the above theorem we have that E[aX + b] =

aE[X] + b for all constants a and b. Also, if P{X = Y} = 1 (i.e. X and Y are
almost-surely equal), then E[X] = E[Y].

Example 29.5. If X ∼ Binomial(n, p), then we found in Example 26.6 that

E[X] = np. Here is a quick way to recover this. Recall that if B1 , . . . , Bn
are independent Bernoulli random variables with parameter p, then X =
B1 + · · · + Bn ∼ Binomial(n, p). Now, recall that E[Bernoulli(p)] = p and
apply property (2) in Theorem 29.4 n times to get that E[X] = E[B1 ] + · · · +
E[Bn ] = np.

Example 29.6 (Bernoulli random variables). Suppose X ∼ Bernoulli(p).

Recall that E[X] = (1 − p) × 0 + p × 1 = p. Now let us compute E[X2 ]:

E[X2 ] = (1 − p) × 02 + p × 12 = p.

Two observations:

(1) This is obvious because X = X2 in this particular example; and

(2) E[X2 ] 6= (E[X])2 . In fact, the difference between E(X2 ) and (EX)2 is
an important quantity, called the variance of X. We will return to
this topic later.

Example 29.7. If X = Binomial(n , p), then what is E[X2 ]? It may help to

recall that E[X] = np. We have

X
n  
n k X
n
n!
2 2 n−k
E[X ] = k p (1 − p) = k pk (1 − p)n−k .
k (k − 1)!(n − k)!
k=0 k=1

The question is, “how do we reduce the factor k further”? If we had k − 1

instead of k, then this would be easy to answer. So let us first solve a
1. Some properties of expectations 145

related problem.
X
n  
n k
E[X(X − 1)] = k(k − 1) p (1 − p)n−k
k
k=0
Xn
n!
= k(k − 1) pk (1 − p)n−k
k!(n − k)!
k=2
X
n
(n − 2)!
= n(n − 1)
pk qn−k
k=2
(k − 2)! [n − 2] − [k − 2] !
Xn 
n−2 k

= n(n − 1) p (1 − p)n−k
k−2
k=2
Xn  
n − 2 k−2
2
= n(n − 1)p p (1 − p)[n−2]−[k−2]
k−2
k=2
X n − 2
n−2
2
= n(n − 1)p p` (1 − p)[n−2]−` .
`
`=0
The summand is the probability that Binomial(n − 2 , p) is equal to `. Since
that probability is added over all of its possible values, the sum is one.
Thus, we obtain E[X(X−1)] = n(n−1)p2 . But X(X−1) = X2 −X. Therefore,
we can apply Theorem 29.4 to find that
E[X2 ] = E[X(X − 1)] + E[X] = n(n − 1)p2 + np = (np)2 + np(1 − p).
Example 29.8. Suppose X ∼ Poisson(λ). We saw in Example 27.1 that
E[X] = λ. In order to compute E[X2 ], we first compute E[X(X − 1)] and find
that
∞
X ∞
e−λ λk X e−λ λk
E[X(X − 1)] = k(k − 1) =
k! (k − 2)!
k=0 k=2
∞
X
2 e−λ λk−2
=λ .
(k − 2)!
k=2
The sum is equal to one; change variables (j = k − 2) and recognize the jth
term as the probability that Poisson(λ) = j. Therefore,
E[X(X − 1)] = λ2 .
Because X(X − 1) = X2 − X, the left-hand side is E[X2 ] − E[X] = E[X2 ] − λ.
Therefore,
E[X2 ] = λ2 + λ.
146 29

Homework Problems

Exercise 29.1. Let X be an Exponential r.v. with parameter λ > 0. Compute

E[X] and E[X2 ].
Exercise 29.2. Let X be a random variable with N(0, 1) distribution. Show
that
0 if n is odd,
E[Xn ] =
(n − 1)(n − 3) · · · 3 · 1 if n is even.
Exercise 29.3. We assume that the length of a telephone call is given by a
random variable X with probability density function
xe−x if x > 0
f(x) =
0 otherwise.
The cost of a call is given as a function of the length by
2 if 0 < X 6 3
c(X) =
2 + 6(X − 3) if X > 3.
Find the average cost of a call.
 Lecture 30

1. Properties of expectation, continued

Example 30.1. Let X ∼ N(0, 1). We have seen that E[X] = 0. What is E[X2 ]?
We need to compute
Z∞
1 2
√ x2 e−x /2 dx.
2π −∞
2 /2
We use integration by parts. Letting u = x and v = −e−x we have
2 R R
uv 0 = x2 e−x /2 and since uv 0 dx = uv − vu 0 dx we get
Z∞ ∞
Z∞
1 2 −x2 /2 1 2 1 2 /2
√ x e dx = − √ xe−x /2 +√ e−x dx.
2π −∞ 2π −∞ 2π −∞

The first term is 0 and the second is 1 (why?). Thus, E[X2 ] = 1. One can
similarly compute E[Xn ] for integers n > 3.

Example 30.2. Let X and Y be independent Exponential(λ) random vari-

ables. Can you see real quick why E[XY] = 1/λ2 and E[X − Y] = 0? We
now want to compute E[|X − Y|]. Then,
ZZ ZZ
E[|X − Y|] = (x − y)f(x, y) dx dy + (y − x)f(x, y) dx dy
x>y x<y
Z∞  Zx 
2
= 2λ (x − y)e−λx e−λy dy dx
Z ∞0 0
Z∞  Zx 
= 2λ xe−λx (1 − e−λx ) dx − 2λ2 e−λx ye−λy dy dx.
0 0 0

147
148 30

We integrate by parts to compute

Zx Z
1 x
ye −λy
dy = − y(e−λy ) 0 dy
0 λ 0
Z
1 −λy x 1 x −λy
= − ye + e dy
λ 0 λ 0
1 1
= − xe−λx + 2 (1 − e−λx ).
λ λ
R∞ −λx
Now observe thatR λ 0 xe dx = E[Exponential(λ)] = 1/λ. The same
∞ −2λx dx = 1/(2λ). Also, we already know that
way
R∞ we have 2λ 0 xe R∞
λ 0 e−λx dx = 1 and 2λ 0 e−2λx dx = 1. Putting all this together we
get
2 1 1 2 2 1
E[|X − Y|] = − + − + = .
λ 2λ 2λ λ 2λ λ

2. Variance
When E[X] is well-defined, the variance of X is defined as
Var(X) = E (X − E[X])2 .
 

If E[X] = ∞ or −∞ the above is just infinite and does not carry any
information. Thus, the variance is a useful notion when E[X] is finite. The
next theorem says that this is the same as asking for E[|X|] to be finite.
(Think of absolute summability or absolute integrability in calculus.)
Theorem 30.3 (Triangle inequality). E[X] is well defined and finite if, and only
if, E[|X|] < ∞. In that case,
|E[X]| 6 E[|X|].

Proof. Observe that −|X| 6 X 6 |X| and apply (4) of Theorem 29.4. 

This of course makes sense: the average of X must be smaller than the
average of |X|, since there are no cancellations when averaging the latter.
Note that the triangle inequality that we know is a special case of the
above: |a + b| 6 |a| + |b|. Indeed, let X equal a or b, equally likely. Now
apply the above theorem and see what happens!
Thus, when E[|X|] is finite:
(1) We predict the as-yet-unseen value of X by the nonrandom num-
ber E[X] (its average value);
(2) Var(X) is the expected squared-error in this prediction. Note that
Var(X) is also a nonrandom number.
2. Variance 149

The variance measures the amount of variation in the random variable.

It vanishes if, and only if, there is no variation at all.
Theorem 30.4. If Var(X) = 0, then X is almost-surely constant. That is, there
exists a constant m such that P{X = m} = 1.

Proof. The constant m has to be the average of X. So we will prove that

if the variance vanishes, then P{X = E[X]} = 1. But this follows from
property (5) in Theorem 29.4. Indeed, 0 = Var(X) = E[(X − E[X])2 ] implies
that P{(X − E[X])2 = 0} = 1. This is what we wanted to prove. 

Here are some useful (and natural) properties of the variance.

Theorem 30.5. Let X be such that E[X2 ] < ∞ and let a be a nonrandom number.
(1) Var(X) > 0;
(2) Var(a) = 0;
(3) Var(aX) = a2 Var(X);
(4) Var(X + a) = Var(X).

The proofs go by direct computation and are left to the student. Note
that (2) says that nonrandom quantities have no variation. (4) says that
shifting by a nonrandom amount does not change the amount of variation
in the random variable.
Let us now compute the variance of a few random variables. But first,
here is another useful way to write the variance
E (X − E[X])2 = E X2 − 2XE[X] + (E[X])2 = E[X2 ] − 2E[X]E[X] + (E[X])2
   

= E[X2 ] − (E[X])2 .
Example 30.6. We have seen in the previous lecture that if X ∼ Poisson(λ),
then E[X2 ] = λ2 + λ. We have also seen in Example 27.1 that E[X] = λ.
Thus, in this case, Var(X) = λ.
Example 30.7. Suppose X ∼ Bernoulli(p). Then, X2 = X and E[X2 ] = E[X] =
p. But then, Var(X) = p − p2 = p(1 − p).
Example 30.8. If X = Binomial(n , p), then what is Var(X)? We have seen
that E[X] = np and E[X2 ] = (np)2 + np(1 − p). Therefore, Var(X) = np(1 −
p).

It is not a coincidence that the variance of Binomial(n, p) is n times the

variance of Bernoulli(p). It is a consequence of the following fact.
Theorem 30.9. Let X and Y be two independent random variables with both
E[|X|] < ∞ and E[|Y|] < ∞. Then Var(X + Y) = Var(X) + Var(Y).
150 30

Proof. Observe first that E[|XY|] = E[|X|]E[|Y|] < ∞. Thus by the triangle
inequality E[XY] is well defined and finite. Now, the proof of the theorem
follows by direct computation:
Var(X + Y) = E[(X + Y)2 ] − (E[X + Y])2
= E[X2 ] + 2E[XY] + E[Y 2 ] − (E[X])2 − 2E[X]E[Y] − (E[Y])2
= E[X2 ] + 2E[X]E[Y] + E[Y 2 ] − (E[X])2 − 2E[X]E[Y] − (E[Y])2
= E[X2 ] − (E[X]2 ) + E[Y 2 ] − (E[Y])2
= Var(X) + Var(Y).
In the third equality we used property (6) in Theorem 29.4. 
Example 30.10. Since a Binomial(n, p) is the sum of n independent Bernoulli(p),
each of which has variance p(1−p), the variance of a Binomial(n, p) is sim-
ply np(1 − p), as already observed by direct computation.
2. Variance 151

Homework Problems

Exercise 30.1. Let X and Y be two independent random variables, each

exponentially distributed with parameter λ = 1.
(a) Compute E[XY].
(b) Compute E[X − Y].
(c) Compute E[|X − Y|].
Exercise 30.2. Let X and Y be two random variables, each uniformly dis-
tributed on [−1, 1]. Compute E[max(X, Y)].
Exercise 30.3. Let X be a Binomial random variable with parameters n and
p. Compute E[X2 ] and Var(X).
Exercise 30.4. Let X be a Geometric random variable with parameter p.
Compute E[X2 ] and Var(X).
Exercise 30.5. Let X be uniformly distributed on [0, 2π]. Let Y = cos(X) and
Z = sin(X). Prove that E[YZ] = E[Y]E[Z] and Var(Y + Z) = Var(Y) + Var(Z).
Then prove that Y and Z are not independent. This shows that the two
equalities above do not imply independence.
Exercise 30.6. If X has the Poisson distribution with parameter λ, show
that for any integer k > 1
E[X(X − 1)(X − 2) · · · (X − k + 1)] = λk .
Conclude that E[X] = Var(X) = λ.
Exercise 30.7. If E[X] exists, show that |E[X]| 6 E[|X|].
 Lecture 31

1. Variance, continued
Example 31.1. Suppose X ∼ Geometric(p) distribution. We have seen al-
ready that E[X] = 1/p (Example 27.2). Let us find a new computation for
this fact, and then go on and find also the variance.

∞
X ∞
X
E[X] = kp(1 − p)k−1 = p k(1 − p)k−1
k=1 k=1
∞
X
!  
d k d 1 p 1
=p − (1 − p) =p − = 2 = .
dp dp p p p
k=0

In the above computation, we used that the derivative of the sum is the
sum of the derivatives. This is OK when we have finitely many terms.
Since we have infinitely many terms, one does need a justification that
comes from facts in real analysis. We will overlook this issue...
Next we compute E[X2 ] by first finding

∞
X X∞
p
E[X(X − 1)] = k(k − 1)p(1 − p)k−1 = k(k − 1)(1 − p)k−2
(1 − p)
k=1 k=1
∞
d2 X
!
d2
 
k p 1
= p(1 − p) (1 − p) =
dp2 (1 − p) dp2 p
k=0
 
d 1 2 2(1 − p)
= p(1 − p) − 2 = p(1 − p) 3 = .
dp p p p2

153
154 31

Because E[X(X − 1)] = E[X2 ] − E[X] = E[X2 ] − (1/p), this proves that
2(1 − p) 1 2−p
E[X2 ] = 2
+ = .
p p p2
Consequently,
2−p 1 1−p
Var(X) = − 2 = .
p2 p p2
For a different solution, see Example (13) on page 124 of Stirzaker’s text.
As a consequence of Theorem 30.9 we have the following.
Example 31.2. Let X be a negative binomial with parameters n and p.
Then, we know that X is a sum of n independent Geometric(p) random
variables. We conclude that Var(X) = n(1 − p)/p2 . Can you do a direct
computation to verify this?
Example 31.3 (Variance of Uniform(a , b)). If X is Uniform(a , b), then
E[X] = a+b
2 and
Zb
1 b2 + ab + a2
E[X2 ] = x2 dx = .
b−a a 3
(b−a)2
In particular, Var(X) = 12 .
Example 31.4 (Moments of N(0 , 1)). Compute E[Xn ], where X ∼ N(0 , 1)
and n > 1 is an integer:
Z∞
n 1 2
E[X ] = √ xn e−x /2 dx
2π −∞
= 0 if n is odd, by symmetry.
If n is even (or even when n is odd but we are computing E[|X|n ] instead
of E[Xn ]), then
Z∞ r Z
n 2 n −x2 /2 2 ∞ n −x2 /2
E[X ] = √ x e dx = x e dx
2π 0 π 0
r Z
2 ∞    √ 
= (2z)n/2 e−z (2z)−1/2 dz z = x2 /2 ⇔ x = 2z
π 0 | {z }
dx
Z∞
2n/2
= √ z(n−1)/2 e−z dz
π 0
2n/2
 
n 1
= √ Γ +
π 2 2
2n/2  n 1  n 3   3  1 
= √ − − ··· Γ (1/2) (Exercise 18.1)
π 2 2 2 2 2 2
= (n − 1)(n − 3) · · · (5)(3)(1).
1. Variance, continued 155

Example 31.5. We can now compute the variance of a normal random

variable with parameters µ and σ2 . Indeed,
Z∞ (x−µ)2
1
Var(X) = E[(X − E[X])2 ] = E[(X − µ)2 ] = √ (x − µ)2 e− 2σ2 dx.
σ 2π −∞
Use the change of variable z = (x − µ)/σ to get dx = σdx and
Z
σ2 ∞ 2 −z2 /2
Var(X) = √ z e dz = σ2 .
2π −∞
In the last step we used the √
previous exercise with n = 2 and recalled from
Exercise 18.1 that Γ (3/2) = π/2.
This is why one usually says that X is a normal random variable with
mean µ and variance σ2 .
Example 31.6. Let X ∼ Gamma(α, λ). Then, we know E[X] = α/λ. Now
compute
Z
2 λα ∞ α+1 −λx
E[X ] = x e dx
Γ (α) 0
Z∞
1
= 2 zα+1 e−z dz (z = λx)
λ Γ (α) 0
Γ (α + 2)
= 2
λ Γ (α)
(α + 1)α
= .
λ2
Thus,
α2 + α α2 α
Var(X) = 2
− 2 = 2.
λ λ λ
In particular, when α = 1, we see that the variance of an exponential
random variable with parameter λ equals 1/λ2 .
Example 31.7. If X is a Cauchy random variable, then we have seen that its
mean is not well defined (Exercise 28.4). Thus, it does not make sense to
talk about its variance. Furthermore, the first moment is infinite: E[|X|] =
∞. (Otherwise, the mean would be defined and finite.) In fact, a direct
computation shows that all moments are infinite: E[|X|n ] = ∞ for all n > 1.
Indeed,
Z Z Z Z
1 ∞ |x|n 2 ∞ xn 2 ∞ xn 1 ∞ n−2
dx = dx > dx > x dx = ∞.
π −∞ 1 + x2 π 0 1 + x2 π 1 1 + x2 π 1
(In the second inequality we used the fact that if x > 1 then 1 + x2 6 2x2 .)
 Lecture 32

1. Covariance
Theorem 32.1. If E[X2 ] < ∞ and E[Y 2 ] < ∞ then E[X], E[Y], and E[XY] are all
well-defined and finite.

Proof. First observe that if |X| > 1 then |X| 6 X2 and thus also |X| 6 X2 + 1.
If, on the other hand, |X| 6 1 then also |X| 6 X2 + 1. So in any case,
|X| 6 X2 + 1. This implies that E[|X|] < ∞ and by the triangle inequality
E[X] is well-defined and finite. The same reasoning goes for E[Y]. Lastly,
observe that (X + Y)2 > 0 and (X − Y)2 > 0 imply
−X2 − Y 2 6 2XY 6 X2 + Y 2
and thus |XY| 6 (X2 + Y 2 )/2 and E[XY] is well-defined and finite. 

From the above theorem we see that if E[X2 ] and E[Y 2 ] are finite then
we can define the covariance between X and Y to be
Cov(X, Y) = E [(X − E[X])(Y − E[Y])] . (32.1)
Because (X − E[X])(Y − E[Y]) = XY − XE[Y] − YE[X] + E[X]E[Y], we obtain the
following, which is the computationally useful formula for covariance:
Cov(X, Y) = E[XY] − E[X]E[Y]. (32.2)

Here are some properties of the covariance.

Theorem 32.2. Suppose E[X2 ], E[Y 2 ], and E[Z2 ] are finite and let a be a nonran-
dom number.
(1) Cov(X, X) = Var(X);
(2) Cov(X, Y) = Cov(Y, X);

157
158 32

(3) Cov(X, a) = 0 (and thus also Cov(a, Y) = 0);

(4) Cov(aX, Y) = a Cov(X, Y) (and thus also Cov(X, aY) = a Cov(X, Y));
(5) Cov(X + Z, Y) = Cov(X, Y) + Cov(Z, Y)
(and thus also Cov(X, Y + Z) = Cov(X, Y) + Cov(X, Z));
(6) Var(X + Y) = Var(X) + Var(Y) + 2Cov(X, Y).

The proofs go by directly applying the definition of covariance. Try

them as an exercise! The above shows that covariance is bilinear. So if a,
b, c, and d are nonrandom numbers and E[X2 ] < ∞, E[Y 2 ] < ∞, E[Z2 ] < ∞,
and E[U2 ] < ∞, then
Cov(aX+bY, cZ+dU) = ac Cov(X, Z)+ad Cov(X, U)+bc Cov(Y, Z)+bd Cov(Y, U).
Example 32.3 (Example 21.3, continued). Observe that the only nonzero
value XY takes with positive probability is 1 × 1. (For example, 2 × 1 and
2 × 2 have 0 probability.) Thus,
2 2
E[XY] = 1 × 1 × = .
36 36
Also,
25 10 1 12
E[X] = E[Y] = 0 × +1× +2× = .
36 36 36 36
Therefore,
2 12 12 72 1
Cov(X, Y) = − × =− =− .
36 36 36 1296 18

2. Correlation
The correlation between X and Y is the quantity,
Cov(X, Y)
ρ(X, Y) = p . (32.3)
Var(X) Var(Y)
Example 32.4 (Example 21.3, continued). Note that
25 10 1 14
E[X2 ] = E[Y 2 ] = 02 ×+ 12 × + 22 × = .
36 36 36 36
Therefore, the correlation between X and Y is
1/18 1
ρ(X, Y) = − q
=− .
5 5

5
18 18

We say that X and Y are negatively correlated. But what does this mean?
The following few sections will help explain this.

Correlation is always a number between −1 and 1.

Theorem 32.5. If E[X2 ] and E[Y 2 ] are positive and finite, then −1 6 ρ(X, Y) 6 1.
2. Correlation 159

Figure 32.1. Left: Karl Hermann Amandus Schwarz (Jan 25, 1843 – Nov
30, 1921, Hermsdorf, Silesia [now Jerzmanowa, Poland]). Right: Victor
Yakovlevich Bunyakovsky (Dec 16, 1804 – Dec 12, 1889, Bar, Ukraine,
Russian Empire)

This is a straightforward variant of the following inequality. [How?]

Theorem 32.6 (Cauchy–Bunyakovsky–Schwarz inequality). If E[X2 ] and E[Y 2 ]

are finite, then
q
|E[XY]| 6 E[X2 ] E[Y 2 ].

Proof. Note that

2
2
X2 E[Y 2 ] + Y 2 (E[XY])2 − 2XY E[Y 2 ]E[XY] = XE[Y 2 ] − YE[XY] > 0.

Therefore, taking expectation, we find

2
E[X2 ] E[Y 2 ] + E[Y 2 ] (E[XY])2 − 2E[Y 2 ] (E[XY])2 > 0

which leads to
 
E[Y 2 ] E[X2 ]E[Y 2 ] − (E[XY])2 > 0.

If E[Y 2 ] > 0, then we get

E[X2 ]E[Y 2 ] > (E[XY])2 ,

which is the claim. Else, if E[Y 2 ] = 0, then P{Y = 0} = 1 and P{XY = 0} = 1.

In this case the the result is true because it says 0 6 0. 

Applying the above inequality to two special cases one deduces two
very useful inequalities in mathematical analysis.
160 32

Example 32.7. Fix numbers a1 , . . . , an and b1 , . . . , bn . Let P{(X, Y) = (ai , bi )} =

1/n, for all i = 1, . . . , n. Then, P{X = ai } = 1/n and P{Y = bi } = 1/n. Ap-
plying the above theorem we get that
X
n 2 X
n X
n 
ai b i 6 a2i b2i .
i=1 i=1 i=1

Example 32.8. Let U ∼ Uniform(a, b) and let X = g(U) and Y = h(U) for
some functions g and h. Applying the above theorem we deduce that
 Zb 2  Z b  Z b 
g(u)h(u) du 6 |g(u)|2 du |h(u)|2 du .
a a a

3. Correlation and independence

We say that X and Y are uncorrelated if ρ(X, Y) = 0; equivalently, if Cov(X, Y) =
0. A significant property of uncorrelated random variables is that Var(X +
Y) = Var(X) + Var(Y); see Theorem 32.2(2). We saw that this also hap-
pens when the variables are independent; see Theorem 30.9. This is not a
coincidence.
Theorem 32.9. If X and Y are independent and E[X2 ] and E[Y 2 ] are finite, then
X and Y are uncorrelated.

Proof. It suffices to prove that E[XY] = E[X]E[Y]. But this is a consequence

of Theorem 29.4(6). 
Example 32.10 (A counter example). Sadly, it is only too common that
people some times think that the converse to Theorem 32.9 is also true. So
let us dispel this with a counterexample: Let Y and Z be two independent
random variables such that Z = ±1 with probability 1/2 each; and Y = 1
or 2 with probability 1/2 each. Define X = YZ. Then, I claim that X and Y
are uncorrelated but not independent.
First, note that X = ±1 and ±2, with probability 1/4 each. There-
fore, E[X] = 0. Also, XY = Y 2 Z = ±1 and ±4 with probability 1/4 each.
Therefore, again, E[XY] = 0. It follows that
Cov(X, Y) = E[XY] − E[X] E[Y] = 0.
| {z } |{z}
0 0

Thus, X and Y are uncorrelated. But they are not independent. Intuitively
speaking, this is clear because |X| = Y. Here is one way to logically justify
our claim:
1
P{X = 1 , Y = 2} = 0 6= = P{X = 1}P{Y = 2}.
8
4. Correlation and linear dependence 161

4. Correlation and linear dependence

Observe that if Y = aX + b for some nonrandom constants a 6= 0 and b,
then Cov(X, Y) = aCov(X, X) = aVar(X). Furthermore, Var(Y) = a2 Var(X).
Therefore, ρ(X, Y) = a/|a|, which equals 1 if a > 0 and −1 if a is negative.
In other words, if Y follows X linearly and goes up when X does, then
its correlation to X is +1. If it follows X linearly but goes down when X
goes up, then its correlation is −1. The converse is in fact true.

Theorem 32.11. Assume none of X and Y is constant (i.e. Var(X) > 0 and
Var(Y) > 0). If ρ(X, Y) = 1, then there exist constants b and a > 0 such that
P{Y = aX + b} = 1. Similarly, if ρ(X, Y) = −1, then there exist constants b and
a < 0 such that P{Y = aX + b} = 1.

Proof. Let a = Cov(X, Y)/Var(X). Note that a has the same sign as ρ(X, Y).
Recalling that ρ(X, Y) = 1 means (Cov(X, Y))2 = Var(X)Var(Y), we have
Var(Y − aX) = Var(Y) + Var(−aX) + 2Cov(−aX, Y)
= Var(Y) + a2 Var(X) − 2aCov(X, Y)
(Cov(X, Y))2
= Var(Y) −
Var(X)
= 0.
By Theorem 30.4 this implies the existence of a constant b such that
P{Y − aX = b} = 1. 

Consider now the function

f(a, b) = E[(Y − aX − b)2 ]
= E[X2 ] a2 + b2 + 2E[X] ab − 2E[XY] a − 2E[Y] b + E[Y 2 ].
This represents “how far” Y is from the line aX+b. Using some elementary
calculus one can find that the optimal a and b that minimize f (and thus
make Y as close as possible to aX + b) are the solutions to

E[X2 ] a + E[X] b − E[XY] = 0 and b + E[X] a − E[Y] = 0.

Var(X) a + E[X]E[Y] − E[X]2 a − E[XY] = 0 and b + E[X] a − E[Y] = 0.

Finding b in terms of a from the second equation and plugging back in
the first one one gets
Cov(X, Y)
a= and b = E[Y] − E[X] a.
Var(X)
162 32

Plugging back into f one has

h 2 i
f(a, b) = E (Y − aX) − E[Y − aX] = Var(Y − aX)
= Var(Y) + a2 Var(X) − 2a Cov(X, Y)
= Var(Y)(1 − ρ(X, Y)2 ).
So the closer |ρ(X, Y)| is to 1, the closer Y is to being a linear function of X.
4. Correlation and linear dependence 163

Homework Problems

Exercise 32.1. If X and Y are affinely dependent (i.e. there exist numbers a
and b such that Y = aX + b), show that |ρ(X, Y)| = 1.
Exercise 32.2. Show that equality occurs in the Cauchy-Bunyakovsky-Schwarz
inequality (i.e. E[XY]2 = E[X2 ]E[Y 2 ]) if and only if X and Y are linearly de-
pendent (i.e. there exists a number a such that Y = aX).
Exercise 32.3. Prove the following.
(a) For any real numbers a1 , . . . , an and b1 , . . . , bn ,
Xn 2 X n X
n
ai b i 6 a2i b2i .
i=1 i=1 i=1
Rb Rb Rb
(b) If a g2 (x)dx and a h2 (x)dx are finite, then so is a g(x)h(x)dx and
furthermore
 Zb 2 Z b Zb
g(x)h(x)dx 6 2
g (x)dx h2 (x)dx.
a a a

Exercise 32.4. Let X1 , . . . , Xn be a sequence of random variables with E[X2i ] <

∞ for all i = 1, . . . , n. Prove that
X
n X
n X
i−1
Var(X1 + · · · + Xn ) = Var(Xi ) + 2 Cov(Xi , Xj ).
i=1 i=1 j=1
 Lecture 33

Figure 33.1. Left: Pafnuty Lvovich Chebyshev (May 16, 1821 – Dec 8,
1894, Kaluga, Russia). Right: Andrei Andreyevich Markov (Jun 14, 1856
– Jul 20, 1922, Ryazan, Russia)

1. Indicator functions
Let A be an event. The indicator function of A is the random variable
defined by

1 if x ∈ A,
1IA (x) =
0 otherwise.
It indicates whether x is in A or not!
Example 33.1. If A and B are two events, then 1IA∩B = 1IA 1IB . This is
because 1IA (x)1IB (x) equals 1 when both indicators are 1, and equals 0
otherwise. But both indicators equal 1 only when x is in both A and B, i.e.
when x ∈ A ∩ B.

The following is a useful “trick.”

Lemma 33.2. If A is an event, then P(A) = E[1IA ].

Proof. The proof is simple. 1IA takes only two values: 0 and 1. Thus,
E[1IA ] = P(Ac ) × 0 + P(A) × 1 = P(A). 

Next, we prove a very useful inequality.

165
166 33

Lemma 33.3 (Chebyshev’s inequality). If h is a nonnegative function, then for

all λ > 0,
E[h(X)]
P{h(X) > λ} 6 .
λ

Proof. Let A be the event {x : h(x) > λ}. Then, because h is nonnegative,

h(x) > h(x)1IA (x) > λ1IA (x).

Thus,
E[h(X)] > λE[1IA ] = λP(A) = λP{h(X) > λ}.

Divide by λ to finish. 

Thus, for example,

E[|X|]
P {|X| > λ} 6 (Markov’s inequality)
λ

Var(X)
P {|X − E[X]| > λ} 6 , (33.1)
λ2

E[|X − E[X]|4 ]
P {|X − E[X]| > λ} 6 . (33.2)
λ4

To get Markov’s inequality, apply Lemma 33.3 with h(x) = |x|. To get the
second inequality, first note that |X−E[X]| > λ if and only if |X−E[X]|2 > λ2 .
Then, apply Lemma 33.3 with h(x) = |x − E[X]|2 and with λ2 in place of λ.
The third inequality is similar: use h(x) = |x − E[X]|4 and λ4 in place of λ.
In words:

• If E[|X|] < ∞, then the probability that |X| is large is small.

• If Var(X) = E[|X − E[X]|2 ] is small, then with high probability X ≈
E[X].
• If E[|X − E[X]|4 ] is small, then with even higher probability X ≈
E[X].

We are now ready for the link between the intuitive understanding
of probability (relative frequency) and the mathematical one (state space,
probability of an event, random variable, expectation, etc).
2. The law of large numbers 167

2. The law of large numbers

Theorem 33.4 (Weak Law of Large Numbers). Suppose X1 , X2 , . . . , Xn are
independent, all with the same (well defined) mean µ and (finite) variance σ2 <
∞. Then for all ε > 0, however small,

X1 + · · · + Xn
lim P −µ >ε = 0. (33.3)
n→∞ n

To see why this theorem is a step towards the connection with the
intuitive understanding of probability, think of the Xi ’s as being the results
of independent coin tosses: Xi = 1 if the i-th toss results in heads and
Xi = 0 otherwise. Then (X1 +· · ·+Xn )/n is precisely the relative frequency
of heads: the fraction of time we got heads, up to the n-th toss. On the
other hand, µ = E[X1 ] equals the probability of getting heads (since X1 is
really a Bernoulli random variable). Thus, the theorem says that if we toss
a coin a lot of times, the relative frequency of heads will, with very high
chance, be close to the probability the coin lands heads. If the coin is fair,
the relative frequency of heads will, with high probability, be close to 0.5.
The reason the theorem is called the weak law of large numbers is that
it does not say that the relative frequency will always converge, as n → ∞,
to the probability the coin lands heads. It only says that the odds the
relative frequency is far from the probability of getting heads (even by a
tiny, but fixed, amount) get smaller as n grows. We will later prove the
stronger version of this theorem, which then completes the link with the
intuitive understanding of a probability. But let us, for now, prove the
weak version.

Proof of Theorem 33.4. Let X = (X1 + · · · + Xn )/n. (This is simply the

sample mean.) Observe that

1
E[X] = E[X1 + · · · + Xn ]
n
1 
= E[X1 + · · · + Xn−1 ] + E[Xn ]
n
1 
= E[X1 + · · · + Xn−2 ] + E[Xn−1 ] + E[Xn ]
n
1 
= ··· = E[X1 ] + · · · + E[Xn ]
n
1
= (nµ) = µ.
n
168 33

Similarly, since Xi ’s are independent,

1
Var(X) = Var(X1 + · · · + Xn )
n2
1 
= 2 Var(X1 ) + · · · + Var(Xn )
n
1 σ2
= 2 (nσ2 ) = .
n n
Applying Chebyshev’s inequality, we find
X1 + · · · + Xn σ2
P −µ >ε 6 .
n nε2
Let n % ∞ to finish. 

Now, we will state and prove the stronger version of the law of large
numbers.
Theorem 33.5 (Strong Law of Large Numbers). Suppose X1 , X2 , . . . , Xn are
independent, all with the same (well defined) mean µ and finite fourth moment
β4 = E[X41 ] < ∞. Then,
X1 + · · · + Xn
P lim = µ = 1. (33.4)
n→∞ n
This theorem implies that if we flip a fair coin a lot of times and keep
track of the relative frequency of heads, then it will converge, as the num-
ber of tosses grows, to 0.5, the probability of the coin landing heads.
There is a subtle difference between the statements of the two versions
of the law of large numbers. This has to do with the different definitions
of convergence for a sequence of random variables.
Definition 33.6. A sequence Yn of random variables is said to converge
in probability to a random variable Y if for any ε > 0 (however small) the
quantity P{|Yn − Y| > ε} converges to 0 as n → ∞.
Convergence in probability means that the probability that Yn is far
from Y by more than the fixed amount ε gets small as n gets large. In
other words, Yn is very likely to be close to Y for large n.
Definition 33.7. A sequence Yn of random variables is said to converge
almost surely to a random variable Y, if
P Yn −→ Y = 1.
n→∞

Almost sure convergence means that the odds that Yn does not con-
verge to Y are nill. It is a fact that almost sure convergence implies con-
vergence in probability. We omit the proof. However, the converse is not
true, as the following example shows.
2. The law of large numbers 169

Example 33.8. Let Yn be a sequence of independent random variables such

that P{Yn = 3} = 1/n and P{Yn = 2} = 1 − 1/n. Then, for any ε ∈ (0, 1),
P{|Yn − 2| > ε} = P{Yn = 3} = 1/n converges to 0 as n → ∞. This proves
that Yn converges to the constant random variable Y = 2, in probability.
However, to say that Yn converges to 2 almost surely would mean to
say that Yn becomes equal to 2, for large n (since Yn takes only the values
2 and 3). If we fix two integers M > N, then the probability that Yn = 2
1
for all n between N and M is equal, by independence, to (1 − N )(1 −
1 1 −x
N+1 ) · · · (1 − M ). Observe now that if x ∈ (0, 1), then (1 − x) 6 e . Thus,
the above probability is smaller than
X
M
1
exp − ,
n
n=N

which goes to 0 as M → ∞, since the series with general term 1/n is

divergent. This proves that there is 0 probability that Yn = 2 for all n > N,
no matter what N is. In other words, Yn cannot converge to 2, almost
surely.
In words: as n grows, the odds that Yn is far from 2 decrease to 0.
Thus, Yn is converging to 2 in probability. However, with probability 1,
Yn will take the value 3 infinitely often and thus cannot be converging to
2 almost surely.

Back to the strong law of large numbers. It is noteworthy that the

theorem actually holds without the assumption of finiteness of the fourth
moment. In fact, one only needs that the mean of the Xi ’s is well defined
and finite. The proof, however, becomes quite harder.

Proof of Theorem 33.5. For n > 1 define the event

X1 + · · · + Xn 1
An = − µ > 1/8 .
n n
We start similarly to the proof of the weak version. By Chebyshev’s in-
equality (33.2), we have
X1 + · · · + Xn 4 1
P(An ) = P −µ >√
n n
= P{|(X1 − µ) + · · · + (Xn − µ)|4 > n7/2 }
E[|(X1 − µ) + · · · + (Xn − µ)|4 ]
6 .
n7/2
To compute E[|(X1 − µ) + · · · + (Xn − µ)|4 ] we expand the expression and
notice that if i 6= j then by independence we have E[(Xi − µ)(Xj − µ)3 ] =
170 33

E[Xi − µ]E[(Xj − µ)3 ], and this equals 0 because E[Xi ] = µ. Also, there are
n terms of the form
E[(Xi − µ)4 ] = E[X4i ] − 3E[X3i ]µ + 3E[Xi ]µ2 + µ4 .
Observe that by the Cauchy-Schwarz inequality (Theorem 32.6),
q
E[X2i ] = E[1 × X2i ] 6 1 × E[X4i ] = β2 < ∞
and q
E[|Xi |3 ] = E[|Xi | × X2i ] 6 E[X2i ]E[X4i ] 6 β3 < ∞.
Thus, E[(Xi − µ)4 ] 6 β4 + 3|µ|β3 + 3|µ|3 + µ4 = γ < ∞. Similarly, we have
n(n − 1) terms of the form E[(Xi − µ)2 (Xj − µ)2 ], with i 6= j. Here too the
Cauchy-Schwarz inequality gives
q
E[(Xi − µ)2 (Xj − µ)2 ] 6 E[(Xi − µ)4 ]E[(Xj − µ)4 ] 6 γ < ∞.

In conclusion, E[|(X1 − µ) + · · · + (Xn − µ)|4 ] 6 (n + n(n − 1))γ = n2 γ.

This gives us that
γ
P(An ) 6 3/2 .
n
But then
X X 1
P{∪n>N An } 6 P(An ) 6 γ 3/2
.
n>N n>N
n

Next, observe that the sets BN = ∪n>N An are decreasing; i.e. BN+1 ⊂
BN . Thus, by Lemma 13.1,
X 1
P{∩N>1 BN } = lim P(BN ) 6 γ lim .
N→∞ N→∞
n>N
n3/2

Because the series with general term 1/n3/2 is summable, the right-most
term above converges to 0 as N → ∞. Thus,
\ [
P {|(X1 + · · · + Xn )/n − µ| > 1/n1/8 } = 0.
N>1 n>N

Taking complements we have

[ \
P {|(X1 + · · · + Xn )/n − µ| < 1/n1/8 } = 1.
N>1 n>N

The event in question reads:

X1 + · · · + Xn 1
“There is an N > 1 such that for all n > N, − µ < 1/8 .”
n n
This has probability one, as we have shown, and implies that (X1 + · · · +
Xn )/n converges to µ. Thus the latter has probability one as well. 
2. The law of large numbers 171

Homework Problems

Exercise 33.1. Establish the following properties of indicator functions:

(a) 1IΩ = 1, 1I∅ = 0
(b) 1IA∩B = 1IA 1IB , 1IA∪B = 1IA + 1IB − 1IA∩B
P
(c) 1I∪∞
i=1 Ai
= ∞i=1 1IAi if the Ai are disjoint
(d) If A1 , A2 , . . . is an increasing sequence of events (An ⊂ An+1 for
all n) and ∪∞
n=1 An = A, or if A1 , A2 , . . . is a decreasing sequence of events
(An+1 ⊂ An for all n) and ∩∞ n=1 An = A, then 1IAn (ω) → 1IA (ω) for all ω.

Exercise 33.2. (a) Prove that

Xn X
1IA1 ∪···∪An = (−1)i−1 1IAj1 ∩···∩Aji .
i=1 16j1 ,...,ji 6n
j1 ,...,ji all different

(b) Deduce the inclusion-exclusion formula (4.5).

Exercise 33.3. 100 balls are tossed independently and at random into 50
boxes. Let X be the number of empty boxes. Find E[X].
Exercise 33.4. Let X have the exponential density f(x) = e−x , x > 0; f(x) =
0, x < 0. Let µ = E[X] and σ2 = Var(X). Evaluate P{|X − µ| > kσ} and
compare with Chebyshev’s inequality.
Exercise 33.5. Suppose that Xn is the amount you win on trial n in a game
of chance. Assume that the Xi are independent random variables, each
with finite mean µ and finite variance σ2 . Make the realistic assumption
that µ < 0. Show that
R1 + . . . + Rn
P < µ/2 −→ 1.
n n→∞
 Lecture 34

1. Conditioning
Say X and Y are two random variables. If they are independent, then
knowing something about Y does not say anything about X. So, for ex-
ample, if fX (x) were the pdf of X, then knowing that Y = 2 the pdf of X
is still fX (x). If, on the other hand, the two are dependent, then knowing
Y = 2 must change the pdf of X. For example, consider the case Y = |X|
and X ∼ N(0, 1). If we do not know anything about Y, then the pdf of X is
2
√1 e−x /2 . However, if we know Y = 2, then X can only take the values 2
2π
and −2 (with equal probability in this case). So knowing Y = 2 makes X a
discrete random variable with mass function f(2) = f(−2) = 1/2.

1.1. Conditional mass functions. We are given two discrete random vari-
ables X and Y with mass functions fX and fY , respectively. For all y, define
the conditional mass function of X given that Y = y as
 P{X = x , Y = y} fX,Y (x , y)
fX|Y (x | y) = P X = x Y = y = = ,
P{Y = y} fY (y)
provided that fY (y) > 0 (i.e. y is a possible value for Y).
As a function in x, fX|Y (x | y) is a probability mass function. That is:
(1) 0 6 fX|Y (x | y) 6 1;
P
(2) x fX|Y (x | y) = 1.

Example 34.1 (Example 21.3, continued). In this example, the joint mass
function of (X, Y), and the resulting marginal mass functions, were given
by the following:

173
174 34

x\y 0 1 2 fX
0 16/36 8/36 1/36 25/36
1 8/36 2/36 0 10/36
2 1/36 0 0 1/36
fY 25/36 10/36 1/36 1
Let us calculate the conditional mass function of X, given that Y = 0:
fX,Y (0 , 0) 16 fX,Y (1 , 0) 8
fX|Y (0 | 0) = = , fX|Y (1 | 0) = = ,
fY (0) 25 fY (0) 25
fX,Y (2 , 0) 1
fX|Y (2 | 0) = = , fX|Y (x | 0) = 0 for other values of x.
fY (0) 25
Similarly,
8 2
fX|Y (0 | 1) = , f (1 | 1) = , fX|Y (x | 1) = 0 for other values of x.
10 X|Y 10
and
fX|Y (0 | 2) = 1, fX|Y (x | 2) = 0 for other values of x.
These conditional mass functions are really just the relative frequencies in
each column of the above table. Similarly, fY|X (y | x) would be the relative
frequencies in each row.

Observe that if we know fX|Y and fY , then fX,Y (x, y) = fX|Y (x | y)fY (y).
This is really Bayes’ formula. The upshot is that one way to describe how
two random variables interact is by giving their joint mass function, and
another way is by giving the mass function of one and then the conditional
mass function of the other (i.e. describing how the second random variable
behaves, when the value of the first variable is known).
Example 34.2. Let X ∼ Poisson(λ) and if X = x then let Y ∼ Binomial(x, p).
By the above observation, the mass function for Y is
X X
fY (y) = fX,Y (x, y) = fX (x)fY|X (y | x)
x x
∞
X λx
 
−λ x y
= e p (1 − p)x−y
x=y
x! y
∞
py λy −λ X (λ(1 − p))x−y (λp)y
= e = e−λp .
y! x=y
(x − y)! y!

In other words, Y ∼ Poisson(λp). This in fact makes sense: after having

finished shopping you stand in line to pay. The length of the line (X) is a
Poisson random variable with average λ. But you decide now to use the
fact that you own the store and you give each person ahead of you a coin
1. Conditioning 175

to flip. The coin gives heads with probability p. If it comes up heads, the
person stays in line. But if it comes up tails, the person leaves the store!
Now, you still have a line of length Y in front of you. This is thus again
a Poisson random variable. Its average, though, is λp (since you had on
average λ people originally and then only a fraction of p of them stayed).

1.2. Conditional expectations. Define conditional expectations, as we did

ordinary expectations. But use conditional probabilities in place of ordi-
nary probabilities, viz.,
X
E[X | Y = y] = x fX|Y (x | y). (34.1)
x

Example 34.3 (Example 34.1, continued). Here,

   
8 2 2 1
E[X | Y = 1] = 0 × + 1× = = .
10 10 10 5
Similarly,
     
16 8 1 10 2
E[X | Y = 0] = 0 × + 1× + 2× = = ,
25 25 25 25 5
and
E[X | Y = 2] = 0.
Note that E[X] = 10/36 + 2 × 1/36 = 12/36 = 1/3, which is none of the
preceding. If you know, for example, that Y = 0, then your best bet for X
is 2/5. But if you have no extra knowledge, then your best bet for X is 1/3.
However, let us note Bayes’ formula in action:
E[X] = E[X | Y = 0]P{Y = 0} + E[X | Y = 1]P{Y = 1} + E[X | Y = 2]P{Y = 2}
     
2 25 1 10 1 12
= × + × + 0× = ,
5 36 5 36 36 36
as it should be.

The proof of Bayes’ formula is elementary:

X X X
E[X] = x P{X = x} = x P{X = x, Y = y}
x x y
X XX 
= x P{X = x | Y = y}P{Y = y} = x fX|Y (x | y) P{Y = y}
x,y y x
X
= E[X | Y = y]P{Y = y},
y

provided E[X] exists, of course, so that we can interchange summations

over x and y at will.
176 34

Example 34.4. Roll a fair die fairly n times. Let X be the number of 3’s and
Y the number of 6’s. We want to compute the conditional mass function
fX|Y (x | y). The possible values for Y are the integers from 0 to n. If we
know Y = y, for y = 0, . . . , n, then we know that the possible values for
X are the integers from 0 to n − y. If we know we got y 6’s, then the
probability of getting x 3’s is
 
n − y  1 x  4 n−y−x
fX|Y (x | y) = ,
x 5 5
for y = 0, . . . , n and x = 0, . . . , n − y (and it is 0 otherwise). In other words,
given that Y = y, X is a Binomial(n − y, 1/5). This makes sense, doesn’t it?
(You can also compute fX|Y using the definition.) Now, the expected value
of X, given Y = y, is clear: E[X | Y = y] = (n − y)/5, for y = 0, . . . , n.
Example 34.5 (Example 26.5, continued). Last time we computed the av-
erage amount one wins by considering a long table of all possible out-
comes and their corresponding probabilities. Now, we can do things much
cleaner. If we know the outcome of the die was x (an integer between 1
and 6), we lose x dollars right away. Then, we toss a fair coin x times and
the expected amount we win at each toss is 2 × 12 − 1 × 12 = 21 dollars. So
after x tosses the expected amount we win is x/2. Subtracting the amount
we already lost we have that, given the die rolls an x, the expected amount
we win is x/2 − x = −x/2. The probability of the die rolling x is 1/6.
Hence, Baye’s formula gives that the expected amount we win is
X
6 X
6 
x  1  7
E[W] = E[W | X = x]P{X = x} = − =− ,
2 6 4
x=1 x=1
as we found in the longer computation. Here, we wrote W for the amount
we win in this game and X for the outcome of the die.
1. Conditioning 177

Homework Problems

Exercise 34.1. If a single die is tossed independently n times, find the

average number of 2’s, given that the number of 1’s is k.
Exercise 34.2. Of the 100 people in a certain village, 50 always tell the
truth, 30 always lie, and 20 always refuse to answer. A single unbiased
die is tossed. If the result is 1, 2, 3, or 4, a sample of size 30 is taken with
replacement. If the result is 5 or 6, a samle of size 30 is taken without
replacement. A random variable X is defined then as follows:
X = 1 if the resulting sample contains 10 people of each category.
X = 2 if the sample is taken with replacement and contains 12 liars.
X = 3 otherwise.
Find E[X].
 Lecture 35

1. Conditioning, continued
1.1. Conditional density functions. We are now given two continuous
random variables X and Y with density functions fX and fY , respectively.
For all y, define the conditional density function of X given that Y = y as
f(x , y)
fX|Y (x | y) = , (35.1)
fY (y)
provided that fY (y) > 0.
As a function in x, fX|Y (x | y) is a probability density function. That is:
(1) fX|Y (x | y) > 0;
R∞
(2) −∞ fX|Y (x | y) dx = 1.
Example 35.1. Let X and Y have joint density fX,Y (x, y) = e−y , 0 < x < y.
If we do not have any information about Y, the pdf of X is
Z∞
fX (x) = e−y dy = e−x , x > 0
x
which means that X ∼ Exponential(1). But say we know that Y = y > 0.
We would like to find fX|Y (x | y). To this end, we first compute
Zy
fY (y) = e−y dx = ye−y , y > 0.
0
Then,
fX,Y (x, y) 1
fX|Y (x | y) = = , 0 < x < y.
fY (y) y
This means that given Y = y > 0, X ∼ Uniform(0, y).

179
180 35

Example 35.2. Now, say X is a random variable with pdf fX (x) = xe−x ,
x > 0. Given X = x > 0, Y is a uniform random variable on (0, x). This
means that Y has the conditional pdf fY|X (y | x) = x1 , 0 < y < x. Then,
fX,Y (x, y) = fX (x)fY|X (y | x) = e−x , 0 < y < x. This allows us to compute,
for example, the pdf of Y:
Z∞
fY (y) = e−x dx = e−y , y > 0.
y
So Y ∼ Exponential(1). We can also compute things like P{X + Y 6 2}.
First, we need to figure out the boundary of integration. We know that
0 < y < x and now we also have x + y 6 2. So x can go from 0 to 2
and then y can go from 0 to x or 2 − x, whichever is smaller. The switch
happens at x = 2 − x, and so at x = 1. Now we compute:
Z1  Zx  Z 2  Z 2−x 
−x
P{X + Y 6 2} = e dy dx + e−x dy dx
0 0 1 0
Z1 Z2
= xe−x dx + (2 − x)e−x dx
0 1
1 2 2
= −(x + 1)e−x − 2e−x + (x + 1)e−x = 1 + e−2 − 2e−1 .
0 1 1

1.2. Conditional expectations. Define conditional expectations, as we did

ordinary expectations. But use conditional probabilities in place of ordi-
nary probabilities, viz.,
Z∞
E[X | Y = y] = x fX|Y (x | y) dx.
−∞
Similarly, if g is a function of x, then
Z∞
E[g(X) | Y = y] = g(x) fX|Y (x | y) dx.
−∞

Example 35.3 (Example 35.2, continued). If we are given that X = x > 0,

then Y ∼ Uniform(0, x). This implies that E[Y | X = x] = x/2. Now, say we
are given that Y = y > 0. Then, to compute E[X | Y = y] we need to find
fX,Y (x, y) e−x
fX|Y (x | y) = = −y = e−(x−y) , 0 6 y < x.
fY (y) e
As a consequence, given Y = y > 0,
Z∞ Z∞
E[X | Y = y] = xe −(x−y)
dx = (z + y)e−z dz = 1 + y.
y 0
We can also compute, for y > 0,
Z∞ Z∞
E[e X/2
| Y = y] = x/2 −(x−y)
e e dx = e y
e−x/2 dx = 2ey e−y/2 = 2ey/2 .
y y
2. Conditioning on events 181

Let us note Bayes’ formula in action. On the one hand,

Z∞ Z∞
E[X] = x f(x) dx = x2 e−x dx = 2.
−∞ 0

(To see the last equality, either use integration by parts, or the fact that this
is the second moment of an Exponential(1), which is equal to its variance
plus the square of its mean: 1 + 12 = 2.) On the other hand,
Z∞ Z∞
E[X] = E[X | Y = y] fY (y) dy = (1 + y)e−y dy = 2,
−∞ 0

as it should be.

The proof of Bayes’ formula is similar to the discrete case. (Do it!)

2. Conditioning on events
So far, we have learned how to compute the conditional pdf and expecta-
tion of X given Y = y. But what about the same quantities, conditional on
knowing that Y 6 2, instead of a specific value for Y? This is quite simple
to answer in the discrete case. The mass function of X, given Y ∈ B, is:
P
P{X = x, Y ∈ B} y∈B fX,Y (x, y)
fX|Y∈B (x) = P{X = x | Y ∈ B} = = P .
P{Y ∈ B} y∈B fY (y)

The analogous formula in the continuous case is for the pdf of X, given
Y ∈ B: R
fX,Y (x, y) dy
fX|Y∈B (x) = BR . (35.2)
B fY (y) dy

Once we know the pdf (or mass function), formulas for expected val-
ues become clear:
P P
x x fX,Y (x, y)
E[X | Y ∈ B] = Py∈B ,
y∈B fY (y)

in the discrete case, and

R∞ R

−∞ x RB
f X,Y (x, y) dy dx
E[X | Y ∈ B] = , (35.3)
B fY (y) dy
in the continuous case. Observe that this can also be written as:
R R∞
−∞ x fX|Y (x | y) dx fY (y) dy


B
E[X | Y ∈ B] =
P{Y ∈ B}
R
E[X | Y = y] fY (y) dy
= B . (35.4)
P{Y ∈ B}
182 35

Example 35.4. Let (X, Y) have joint density function fX,Y (x, y) = e−x , 0 <
y < x. We want to find the expected value of Y, conditioned on X 6 5.
First, we find the conditional pdf. One part we need to compute is P{X 6
5}. The pdf of X is
Zx
fX (x) = e−x dy = xe−x , x > 0,
0
and, using integration by parts, we have
Z5
P{X 6 5} = xe−x dx = 1 − 6e−5 .
0
Now, we can go ahead with computing the conditional pdf using (35.3). If
X 6 5, then also Y < 5 (since Y < X) and
R5 R5 −x
f X,Y (x, y) dx ye dx e−y − e−5
fY|X65 (y) = 0 = = , 0 < y < 5.
1 − 6e−5 1 − 6e−5 1 − 6e−5
(Check that this pdf integrates to 1!) Finally, using integration by parts,
we can compute:
Z∞ R5
y(e−y − e−5 ) dy 1 − 18.5e−5
E[Y | X 6 5] = y fY|X65 (y) dy = 0 = ≈ 0.912.
−∞ 1 − 6e−5 1 − 6e−5
R∞
Remark 35.5. We have fY (y) = y e−x dx = e−y , y > 0. Thus, Y ∼
Exponential(1) and E[Y] = 1. Note now that the probability that X 6 5
is 1 − 6e−5 ≈ 0.96, which is very close to 1. So knowing that X 6 5 gives
very little information. This explains why E[Y | X 6 5] is very close to E[Y].
Try to compute E[Y | X 6 1] and see how it is not that close to E[Y] any-
more. Try also to compute E[Y | X 6 10] and see how it is even closer to
E[Y] than E[Y | X 6 5].

We could have done things in a different order to find E[Y | X 6 5].

First, we find the conditional expectationRE[Y | X = x]. To do so, we need
x
to find fY|X and thus to find first fX (x) = 0 e−x dy = xe−x , x > 0. Hence,
Rx
fY|X (y | x) = 1/x, for 0 < y < x. Now, we see that E[Y | X = x] = 0 y x1 dy =
x
2 . (This, of course, is not surprising since given X = x we found that
Y ∼ Uniform(0, x).) Finally, we can apply (35.4) and use integration by
parts to compute:
R5 x −x R
1 5 2 −x
0 2 xe dx 2 0x e dx 1 − 18.5e−5
E[Y | X 6 5] = = R5 = .
P{X 6 5} xe−x dx 1 − 6e−5
0

Example 35.6. Let X and Y be independent uniformly distributed on (0, 1).

Then, P{X + Y 6 1} = 1/2. (This is the area of the triangle that is half the
square [0, 1]2 .) Conditioned on knowing that X + Y 6 1, the pair (X, Y) is
2. Conditioning on events 183

still uniformly distributed but on the triangle {(x, y) ∈ [0, 1]2 : x + y 6 1}.
Consequently,
R1 R1−x
Z1
0x dy dx 1
E[X | X + Y 6 1] = 0
= 2 x(1 − x) dx = .
1/2 0 3
Alternatively, let U = X + Y. Using the transformation method we find
that fX,U (x, u) = 1, 0 < x < 1 and x < u < x + 1. (Do it!) This implies
Ru R1
that fU (u) = 0 dx = u, for 0 < u < 1, and fU (u) = u−1 dx = 2 − u, for
1 < u < 2. (This clearly integrates to 1. Use geometry to see that, rather
than doing the (easy) computation!) We can readily see that E[U] = 1/2.
(Again, use geometry rather than the (easy) computation.) Furthermore,
fX|U (x | u) = 1/u, for 0 < x < u < 1, and fX|U (x | u) = 1/(2 − u), for
0 < u − 1 < x < 1. Thus,
R u 1 u

 0 x u dx = 2 if 0 < u < 1,
E[X | U = u] =

R 1 1 1−(2−u)2
u−1 x 2−u dx = 2(2−u) if 1 < u < 2.
u
= .
2
Finally, using (35.4),
R1 R1 u
0 E[X | U = u]fU (u) du u du 1
E[X | X + Y 6 1] = E[X | U 6 1] = = 0 2 = .
P{U 6 1} 1/2 3
If instead we wanted to use (35.3), then we first write
R∞ R1
−∞ fX,U (x, u) du du
fX|U61 (x) = = x = 2(1 − x).
P{U 6 1} 1/2
Then, applying (35.3),
Z∞ Z1
1
E[X | X + Y 6 1] = E[X | U 6 1] = xfX|U61 (x) dx = 2x(1 − x) dx = .
−∞ 0 3
184 35

Homework Problems

Exercise 35.1. A number X is chosen with density fX (x) = 1/x2 , x > 1;

fX (x) = 0, x < 1. If X = x, let Y be uniformly distributed between 0 and x.
Find the distribution and density functions of Y.
Exercise 35.2. Let (X, Y) have density f(x, y) = e−y , 0 6 x 6 y, f(x, y) = 0
elsewhere. Find the conditional density of Y given X, and P{Y 6 y | X = x},
the conditional distribution function of Y given X = x.
Exercise 35.3. Let (X, Y) have density f(x, y) = k|x|, −1 6 y 6 x 6 1;
f(x, y) = 0 elsewhere. Find k; also find the individual densities of X and
Y, the conditional density of Y given X, and the conditional density of X
given Y.
Exercise 35.4. Let (X, Y) have density f(x, y) = e−y , 0 6 x 6 y, f(x, y) = 0
elsewhere. Let Z = Y − X. Find the conditional density of Z given X = x.
Also find P{1 6 Z 6 2 | X = x}.
Exercise 35.5. Let (X, Y) have density f(x, y) = 8xy, 0 6 y 6 x 6 1;
f(x, y) = 0 elsewhere.
(a) Find the conditional expectation of Y given X = x, and the condi-
tional expectation of X given Y = y.
(b) Find the conditional expectation of Y 2 given X = x.
(c) Find the conditional expectation of Y given A = {X 6 1/2}.
Exercise 35.6. Let (X, Y) be uniformly distributed over the parallelogram
with vertices (0, 0), (2, 0), (3, 1), (1, 1). Find E[Y|X = x].
Exercise 35.7. Let X and Y be independent random variables, each uni-
formly distributed between 0 and 2.
(a) Find the conditional probability that X > 1, given that X + Y 6 3.
(b) Find the conditional expectation of X, given that X + Y 6 3.
Exercise 35.8. The density for the time T required for the failure of a light
bulb is f(t) = λe−λt , t > 0. Find the conditional density function for T −t0 ,
given that T > t0 , and interpret the result intuitively.
Exercise 35.9. Let X and Y be independent random variables, each uni-
formly distributed between 0 and 1. Find the conditional expectation of
(X + Y)2 given X − Y. Hint: first find the joint density of (X + Y, X − Y).
Exercise 35.10. Let X and Y be independent random variables, each with
density f(x) = (1/2)e−x , x > 0; f(x) = 1/2, −1 6 x 6 0; f(x) = 0, x < −1.
Let Z = X2 + Y 2 . Find E[Z|X = x].
2. Conditioning on events 185

Exercise 35.11. Let X be the number of successes in n Bernoulli trials, with

probability p of success on a given trial. Find the conditional expectation
of X, given that X > 2.
Exercise 35.12. Let X be uniformly distributed between 0 and 10, and de-
fine Y by

X2 if 0 6 X 6 6,
Y=
3 if 6 < X 6 10.
Find the conditional expectation of Y given that 2 6 Y 6 4.
 Lecture 36

1. The moment generating function

The moment generating function (mgf) of a random variable X is the function
of t given by
P
etx f(x), in the discrete setting,
M(t) = E[e ] = R∞x tx
tX

−∞ e f(x) dx, in the continuous setting.

provided that the sum (or integral) exists. This is precisely the Laplace
transform of the mass function (or pdf).
Note that M(0) always equals 1 and M(t) is always nonnegative.
A related transformation is the characteristic function of a random vari-
able, given by
P
eitx f(x), in the discrete setting,
Φ(t) = E[eitX ] = R∞x itx
−∞ e f(x) dx, in the continuous setting.

While the moment generating function may be infinite at some (or even
all) nonzero values of t, the characteristic function is always defined and
finite. It is precisely the Fourier transform of the mass function (or the
pdf). In this course we will restrict attention to the moment generating
function. However, one can equally work with the characteristic function
instead, with the added advantage that it is always defined.

Example 36.1 (Bernoulli). If X ∼ Bernoulli(p), then its mgf is

M(t) = 1 − p + pet .

187
188 36

Example 36.2 (Uniform). If X ∼ Uniform(a , b), then

Zb
tX 1 ebt − eat
M(t) = E[e ] = etx dx = .
b−a a (b − a)t
Example 36.3 (Exponential). If X ∼ Exponential(λ), then
Z∞
tX
M(t) = E[e ] = λ etx e−λx dx.
0

This is infinite if t > λ and otherwise equals

λ
M(t) = , if t < λ.
λ−t
This is indeed a useful transformation, viz.,
Theorem 36.4 (Uniqueness). If X and Y are two random variables—discrete or
continuous—with moment generating functions MX and MY , and if there exists
δ > 0 such that MX (t) = MY (t) for all t ∈ (−δ , δ), then MX = MY and X and
Y have the same distribution. More precisely:
(1) X is discrete if and only if Y is, in which case their mass functions are
the same;
(2) X is continuous if and only if Y is, in which case their density functions
are the same.

We omit the proof. The theorem says that if we compute the mgf of
some random variable and recognize it to be the mgf of a distribution we
already knew, then that is precisely what the distribution of the random
variable is. In other words, there is only one distribution that corresponds
to any given mgf.
Example 36.5. If
1 1 1 √
M(t) = et + e−πt + e 2t ,
2 4 4
then M is the mgf of a random variable with mass function


1/2 if x = 1, √
f(x) = 1/4 if x = −π or x = 2,


0 otherwise.

2. Sums of independent random variables

Here is another reason why moment generating functions are a powerful
tool.
2. Sums of independent random variables 189

Theorem 36.6. If X1 , . . . , Xn are independent,

P with respective moment generat-
ing functions MX1 , . . . , MXn , then n X
i=1 i has the mgf,
M(t) = MX1 (t) × · · · × MXn (t).

Proof. By induction, it suffices to do this for n = 2 (why?). But then

h i
MX1 +X2 (t) = E et(X1 +X2 ) = E etX1 × etX2 .
 

By independence, this is equal to the product of E[etX1 ] and E[etX2 ], which

is the desired result. 
Example 36.7 (Binomial). Suppose X ∼ Binomial(n , p). Then we can write
X = X1 + · · · + Xn , where X1 , . . . , Xn are independent Bernoulli(p). We can
apply Theorem 36.6 then to find that
MX (t) = (1 − p + pet )n .
Example 36.8. If X ∼ Binomial(n, p) and Y ∼ Binomial(m , p) are indepen-
dent, then by the previous example and Theorem 36.6,
MX+Y (t) = (1 − p + pet )n (1 − p + pet )m = (1 − p + pet )n+m .
By the uniqueness theorem, X + Y ∼ Binomial(n + m , p). We found this
out earlier by applying much harder methods. See Example 22.5.
Example 36.9 (Poisson). If X ∼ Poisson(λ), then
∞
 X λk
M(t) = E etX = etk e−λ

k!
k=0
∞
X (λet )k
= e−λ .
k!
k=0
t
The sum gives the Taylor expansion of eλe . Therefore,
t −1)
M(t) = eλ(e .
Example 36.10. Now suppose X ∼ Poisson(λ) and Y ∼ Poisson(γ) are inde-
pendent. We apply the previous example and Theorem 36.6, in conjunc-
tion, to find that
t −1) t −1) t −1)
MX+Y (t) = eλ(e eγ(e = e(λ+γ)(e .
Thus, X + Y ∼ Poisson(γ + λ), thanks to the uniqueness theorem and Ex-
ample 36.9. For a harder derivation of the same fact see Example 22.1.
Example 36.11 (Geometric). Let X ∼ Geometric(p). Then,
∞
X ∞
kt k−1 p X
M(t) = e (1 − p) p= ((1 − p)et )k .
1−p
k=1 k=1
190 36

The sum converges only when (1−p)et < 1, and thus when t < − ln(1−p).
For example, the mgf of a Geometric(1/2) is only defined on the interval
(−∞, ln 2). So for a Geometric(p), the mgf is

pet
M(t) = , for t < − ln(1 − p).
1 − (1 − p)et

Example 36.12 (Negative Binomial). Since a negative binomial with pa-

rameters r and p is the sum of r independent geometrics with parameter
p, it has the mgf

 pet r
M(t) = , for t < − ln(1 − p).
1 − (1 − p)et

Example 36.13 (Gamma). If X ∼ Gamma(α , λ), then

Z∞ Z∞
tx λα α−1 −λx λα
M(t) = e x e dx = xα−1 e−(λ−t)x dx.
0 Γ (α) Γ (α) 0

If t > λ, then the integral is infinite. On the other hand, if t < λ, then

Z∞
λα zα−1 dz
M(t) = α−1
e−z (z = (λ − t)x)
Γ (α)
0 (λ − t) λ −t
α Z∞
λ
= zα−1 e−z dz
Γ (α) × (λ − t)α 0
| {z }
Γ (α)
λα
= .
(λ − t)α

Thus,
 α
λ
M(t) = , if t < λ.
λ−t

In particular, if α = 1 then we see (again) that the mgf of an Exponential(λ)

is

λ
M(t) = , if t < λ.
λ−t
2. Sums of independent random variables 191

Example 36.14 (Normal). If X = N(µ , σ2 ), then

Z∞
1 2 2
M(t) = √ etx e−(x−µ) /(2σ ) dx
2
2πσ −∞
2 t2 /2 Z ∞  2
µt+σ x − 2(σ2 t + µ)x + (σ2 t + µ)2

e
= √ exp − dx
2πσ2 −∞ 2σ2
Z
eµt+σ t /2 ∞
2 2
(x − σ2 t − µ)2
 
= √ exp − dx
2πσ2 −∞ 2σ2
Z
eµt+σ t /2 ∞ −u2 /2
2 2

= √ e du (u = (x − σ2 t − µ)/σ)
2π −∞
2 t2 /2
= eµt+σ .
In particular, the mgf of a standard normal N(0,1) is
2 /2
M(t) = et .
192 36

Homework Problems

Exercise 36.1. Let X ∼ Gamma(α, λ) and Y ∼ Gamma(β, λ). Assume X and

Y are independent. What is the distribution of X + Y?
Exercise 36.2. Let X1 , . . . , Xn be independent random variables with Xi ∼
N(µi , σ2i ). That is, each of them is normally distributed with its own mean
an variance. Show that X1 + · · · + Xn is again normally distributed, with
mean µ1 + · · · + µn and variance σ21 + · · · + σ2n .
Exercise 36.3. In each of the following, indicate whether or not the given
function can be a moment generating function. If it can, then find the mass
function or pdf of the corresponding random variable.
(a) M(t) = 1 − t.
(b) M(t) = 2e−t .
(c) M(t) = 1/(1 − t), for t < 1.
1
(d) M(t) = 3 + 12 e2t + 1 −2t
12 e + 1 13t
12 e .
Exercise 36.4. Show that if Y = aX + b, with nonrandom constants a and
b, then
MY (t) = ebt MX (at).
Exercise 36.5. Let X and Y take only the values 0, 1, or 2. Prove that if
MX (t) = MY (t) for all values of t, then X and Y have the same mass
function. Do not quote the Uniqueness Theorem 36.4.
 Lecture 37

1. Relation of MGF to moments

Suppose we know the function M(t) = E[etX ]. Then, we can compute
the moments of X from the function M by successive differentiation. For
instance, suppose X is a continuous random variable with moment gener-
ating function M and density function f, and note that
Z
0 d tX

d ∞ tx
M (t) = E[e ] = e f(x) dx.
dt dt −∞
Now, loosely speaking, if the integral of the derivative converges abso-
lutely, then a general fact states that we can take the derivative under the
integral sign. That is,
Z∞
M 0 (t) = xetx f(x) dx = E XetX .
 
−∞

The same end-result holds if X is discrete with mass function f, but this
time,
X
M 0 (t) = xetx f(x) = E XetX .
 
x
Therefore, in any event:
M 0 (0) = E[X].
In general, this procedure yields,

M(n) (t) = E Xn etX .

 

Therefore,
M(n) (0) = E [Xn ] .

193
194 37

Example 37.1 (Uniform). We saw earlier that if X is distributed uniformly

on (0 , 1), then for all real numbers t,
et − 1
M(t) = .
t
Therefore,
tet − et + 1 t2 et − 2tet + 2et − 2
M 0 (t) = , M 00 (t) = ,
t2 t3
whence
tet − et + 1 tet 1
E[X] = M 0 (0) = lim = lim = ,
t&0 t2 t&0 2t 2
by l’Hopital’s rule. Similarly,
t2 et − 2tet + 2et − 2 t2 e t 1
E X2 = lim
 
3
= lim 2
= .
t&0 t t&0 3t 3
Alternatively, these can be checked by direct computation, using the fact
R1
that E[Xn ] = 0 xn dx = 1/(n + 1).
Example 37.2 (Standard Normal). If X ∼ N(0, 1), then we have seen that
2 2
M(t) = et /2 . Thus, M 0 (t) = tet /2 and E[X] = M 0 (0) = 0. Also, M 00 (t) =
2
(t2 + 1)et /2 and E[X2 ] = M 00 (0) = 1.

2. The Central Limit Theorem

In this section we will address one of the most important theorems in
probability and statistics. In particular, we will answer the question “why
is the normal distribution so important?”
To start, let us consider a sequence X1 , . . . , Xn of independent iden-
tically distributed (iid) random variables. Assume µ = E[X1 ], the com-
mon average value of these random variables, exists and is finite. Then,
X1 − µ, . . . , Xn − µ represent the successive “measurement errors”. The law
of large numbers tells us that (X1 + · · · + Xn − nµ)/n converges to 0, with
probability 1. So the cumulative measurement error X1 + · · · + Xn − nµ
is not growing as fast as n. The question is then: how fast is it growing,
if it is growing at all? To get an idea of the answer, let us compute the
variance of this error. Let us assume σ2 = Var(X1 ), the common variance
of the random variables Xi , to be finite. Then,
Var(X1 +· · ·+Xn −nµ) = Var(X1 +· · ·+Xn ) = Var(X1 )+· · ·+Var(Xn ) = nσ2 .
So to have a quantity that has √ finite and positive variation we need to
consider
√ (X1 + · · · + X n − nµ)/ n. In fact, we will consider (X1 + · · · + Xn −
nµ)/(σ n), just to normalize things to have a variance of 1.
2. The Central Limit Theorem 195

What we are doing here is similar to the following simple idea: we

know an = 2n2 + 5n − 10 is growing to infinity as n grows, but how fast
is it growing? Note that an /n → ∞ and so an is growing faster than n
is. On the other hand, an /n3 → 0 means an is growing slower than n3 .
To find how fast an is growing we need to find just the right power α
for which an /nα will not grow to infinity nor will it go to 0. The correct
answer in this particular example is α = 2. In the case of X1 +· · ·+Xn −nµ,
the correct answer seems to be α = 1/2.
√
Now that we have an idea of how fast the error grows (like n), we
start wondering what the distribution of our cumulative error would √ look
like. In other words, if we draw a histogram of (X1 +· · ·+Xn −nµ)/(σ n),
then will it have some recognizable shape as n grows large? Note that
another way to write this cumulative error is as
X−µ
√ ,
σ/ n
where X is the sample mean (X1 + · · · + Xn )/n. So the question we are really
asking is: what does the distribution of the sample mean X look like, for
large samples (n large)?
Let us first find the answer to our question in the case when the ran-
dom variables are normally distributed.
Example 37.3. Let X1 , . . . , Xn be a sequence of independent random vari-
ables that are all N(µ, σ2 ). Then, E[X1 ] = µ and Var(X1 ) = σ2 . Let us com-
√
pute the moment generating function of Zn = (X1 + · · · + Xn − nµ)/(σ n):
h t i
MZn (t) = E[etZn ] = E exp √ (X1 + · · · + Xn − nµ)
σ n
√
− nµt/σ
√
n
=e MX1 (t/(σ n))
√ √ 2 2 2
n
= e− nµt/σ eµt/(σ n))+σ t /(2σ n)
2 /2
= et .
Thus, for any n > 1, Zn is a standard normal.

Motivated by the above use of the moment generating function, the

following theorem will be helpful in our quest.
Theorem 37.4 (Lévy’s continuity theorem). Let Xn be a random variables—
discrete or continuous—with moment generating functions Mn . Also, let X be a
random variable with moment generating function M. Suppose there exists δ > 0
such that:
(1) If −δ < t < δ, then Mn (t), M(t) < ∞ for all n > 1; and
196 37

(2) limn→∞ Mn (t) = M(t) for all t ∈ (−δ , δ), then

lim FXn (a) = lim P {Xn 6 a} = P{X 6 a} = FX (a),
n→∞ n→∞
for all numbers a at which FX is continuous.

The convergence of CDFs in (2) above roughly means that if we fix a

very large n and graph the cumulative histogram of the Xn , then it will
look like that of X. It will in fact look closer and closer to the CDF of X, as
we take n larger and larger. This is called convergence in distribution.
Example 37.5. Let Xn ∼ Uniform(0, 1/n). Since 0 6 Xn 6 1/n, it is clear
that P{Xn → 0} = 1; i.e. Xn → 0 almost-surely. It also converges to 0 in
t/n
distribution: MXn (t) = e t/n−1 → 1 as n → ∞. To see this, write h = t/n
and observe that we are looking for the limit of (eh − 1)/h as h → 0.
Now, either use the definition of derivative to see that the answer is the
derivative of ex at x = 0, or use de l’Hôpital’s rule. Now, since M(t) = 1
is the moment generating function of the random variable X = 0, Lévy’s
continuity theorem implies that the CDF of Xn must converge to that of X
at points of continuity of the latter. But FX (x) = 0 if x < 0 and 1 if x > 0.
(Recall that CDFs are right-continuous.) Note that FXn (x) = 0 if x < 0 and
thus converges in that case to FX (x). Similarly, FXn (x) = 1 if x > 1/n and
thus as n → ∞, FXn (x) → 1 for x > 0. However, FXn (0) = 0 which does
not converge to FX (0) = 1. This is not a problem, though, because x = 0 is
a point of discontinuity for FX .
Example 37.6 (Law of rare events). Suppose Xn ∼ Binomial(n , λ/n), where
λ > 0 is fixed, and n > λ. Then,
n
λ λe−t

−t n
→ exp −λ + λe−t .



MXn (t) = 1 − p + pe = 1− +
n n
Note that the right-most term is MX (t), where X = Poisson(λ). Therefore,
by Lévy’s continuity theorem,
lim P {Xn 6 a} = P {X 6 a} , (37.1)
n→∞
at all a where FX is continuous. But X is discrete and integer-valued.
Therefore, FX is continuous at a if and only if a is not a nonnegative
integer. If a is a nonnegative integer, then we can choose a non-integer
b ∈ (a , a + 1) to find that
lim P{Xn 6 b} = P{X 6 b}.
n→∞
Because Xn and X are both non-negative integers, Xn 6 b if and only if
Xn 6 a, and X 6 b if and only if X 6 a. Therefore, this time (37.1) holds
for all a, i.e. even at points where FX is discontinuous.
2. The Central Limit Theorem 197

Homework Problems

Exercise 37.1. Let X1 , . . . , Xn be independent Poisson(λ) random variables.

What is the distribution of X1 + · · √ · + Xn ? What is the moment generating
function of (X1 + · · · + Xn − nλ)/ nλ? Find the limit of this function as
n → ∞. Can you recognize the outcome as a moment generating function?
Exercise 37.2. Let X have pdf f(x) = e−(x+2) for x > −2, and f(x) = 0
otherwise. Find its mgf and use it to find E[X] and E[X2 ].
Exercise 37.3. Let Xn ∼ Geometric(λ/n). Show that Xn /n converges in
distribution to an Exponential(λ).
Hint: show that MXn /n (t) = MXn (t/n). Then, when taking n → ∞, write
h = 1/n and use de l’Hôpital’s rule.
Exercise 37.4. Let Xn ∼ Negative Binomial(r, λ/n). Show that Xn /n con-
verges in distribution to a Gamma(r, λ).
Remark 37.7. The above exercise shows that Gamma(α, λ) is a continuous
version of a Negative Binomial with a fractional r!
 Lecture 38

1. The Central Limit Theorem, continued

Let us now find the answer to our question about the distribution of the
sample mean in a few cases.
Example 38.1. Let X1 , . . . , Xn be a sequence of independent random vari-
ables that are all Poisson(λ). Then, E[X1 ] = λ, Var(X1 ) = λ, and MX1 (t) =
t √
eλ(e −1) . Let us compute the mgf of Zn = (X1 + · · · + Xn − nλ)/ nλ:
h t i
MZn (t) = E[etZn ] = E exp √ (X1 + · · · + Xn − nλ)
nλ
√
−t λn
√ √ √
=e MX1 (t/ λn) = exp − t λn + nλ(et/ λn − 1) .
n

According to the Taylor–MacLaurin expansion,

√ t t2
et/ λn
=1+ √ + + smaller terms.
λn 2λn
Thus,
√ √ t2 2 /2
MZn (t) = exp − t λn + t λn + + smaller terms −→ et .
2 n→∞
2
Since et /2 is the mgf of a standard normal, Lévy’s continuity theorem and
the fact that the CDF of a standard normal is continuous imply that
Za
1 2
P{Zn 6 a} −→ √ e−x /2 dx, for all a.
n→∞ 2π −∞
Example 38.2 (The de Moivre–Laplace central limit theorem). Suppose
Sn ∼ Binomial(n , p), where p ∈ (0 , 1) is fixed, and define Zn to be its

199
200 38

p
standardization. That is, Zn = (Sn − E[Sn ])/ Var(Sn ). Alternatively,
Sn − np
Zn = p .
np(1 − p)
Recall that Sn is really the sum of n independent Bernoulli(p) random
variables and that the mean of a Bernoulli(p) is p and its variance is p(1 −
p). Thus, the question of what the asymptotic distribution of Zn looks like
is precisely what we have been asking about in this section.

n
We know that for all real numbers t, MSn (t) = 1 − p + pet . We can
use this to compute MZn as follows:
" !#
Sn − np
MZn (t) = E exp t · p
np(1 − p)
√  t 
= e−npt/ np(1−p) MSn p
np(1 − p)
√  √ n
= e−t np/(1−p) 1 − p + pet/ np(1−p)
q q n
p
t 1−p
 −t n(1−p)
= (1 − p)e + pe np
.

According to the Taylor–MacLaurin expansion,

 s  s
1−p 1 − p t2 (1 − p)
exp t =1+t + + smaller terms,
np np 2np
t2 p
r r
p p
exp −t =1−t + + smaller terms.
n(1 − p) n(1 − p) 2n(1 − p)
Therefore,
 s  r
1−p p
p exp t + (1 − p) exp −t
np n(1 − p)
s !
1 − p t2 (1 − p) t2 p
 r 
p
=p 1+t + + · · · + (1 − p) 1 − t + + ···
np 2np n(1 − p) 2n(1 − p)
r r
p(1 − p) t2 (1 − p) p(1 − p) t2 p
=p+t + + · · · + (1 − p) − t + + ···
n 2n n 2n
t2
=1+ + smaller terms.
2n
Consequently,
n
t2

2
MZn (t) = 1 + + smaller terms → et /2 .
2n
1. The Central Limit Theorem, continued 201

We recognize the right-hand side as MZ (t), where Z ∼ N(0 , 1). Because

FZ is continuous, this prove the central limit theorem of de Moivre: For all
real numbers a,
Za
1 2
lim P{Zn 6 a} = √ e−x /2 dx.
n→∞ 2π −∞
The Central Limit Theroem (i.e. the limit theorem that is central in
probability theory), states that the above three results are not a coinci-
dence.
Theorem 38.3 (Central Limit Theorem). Let X1 , . . . , Xn , . . . be independent
random variables identically distributed. Assume σ2 = Var(X1 ) is finite. Then, if
we let µ = E[X1 ] (which exists because the variance is finite),
√
X1 + · · · + Xn − nµ (X − µ) n
Z= √ =
σ n σ
converges in distribution, as n → ∞, to a standard normal random variable.

If E[|X1 |3 ] < ∞, then the proof of the above theorem goes exactly the
same way as in the two examples above, i.e. through a Taylor–MacLaurin
expansion. We leave the details to the student.
A nice visualization of the Central Limit Theorem in action is done
using a Galton board. Look it up on Google and on YouTube.
One way to use this theorem is to approximately compute percentiles
of the sample mean X.
Example 38.4. The waiting time at a certain toll station is exponentially
distributed with an average waiting time of 30 seconds. If we use minutes
to compute things, then this average waiting time is µ = 0.5 a minute
and thus λ = 1/µ = 2. Consequently, the variance is σ2 = 1/λ2 = 1/4.
If 100 cars are in line, we know the average waiting time is 50 minutes.
This is only an estimate, however. So, for example, we want to estimate
of the probabilities they wait between 45 minutes and an hour. If Xi is
the waiting time of car number i, then we want to compute P{45 < X1 +
· · · + X100 < 60}. We can use the central limit theorem for this. The average
waiting time for the 100 cars is 50 minutes. The theorem tells us that the
distribution of
X1 + · · · + X100 − 50
Z= √
0.5 100
is approximately standard normal. Thus,
Z2
1 2
P{45 < X1 + · · · + X100 < 60} = P{−5/5 < Z < 10/5} ≈ √ e−z /2 dz,
2π −1
202 38

which we can find using the tables for the so-called error function
Zx
1 2
ϕ(x) = √ e−z /2 dz
2π −∞
(which is simply the CDF of a standard normal). The tables give that
ϕ(2) ≈ 0.9772. Most tables do not give ϕ(x) for negative numbers x.
This is because symmetry implies that ϕ(−x) = 1 − ϕ(x). Thus, ϕ(−1) =
1 − ϕ(1) ≈ 1 − 0.8413 = 0.1587. Hence, the probability we are looking for
is approximately equal to 0.9772 − 0.1587 = 0.8185, i.e. about 82%.
Example 38.5. In the 2004 presidential elections the National Election Pool
ran an exit poll. At 7:32 PM it was reported that 1963 voters from Ohio
responded to the poll, of which 941 said they voted for President Bush
and 1022 for Senator Kerry. It is safe to assume the sampling procedure
was done correctly without any biases (e.g. nonresponse, etc). We are
wondering if this data has significant evidence that President Bush had
lost the race in Ohio.
To answer this question, we assume the race resulted in a tie and com-
pute the odds that only 941 of the 1963 voters would vote for President
Bush. The de Moivre–Laplace central limit theorem tells us that
S1963 − 0.5 × 1963
Z= p
1963 × 0.5(1 − 0.5)
is approximately standard normal. Hence,
941 − 981.5
P{S1963 6 941} = P Z 6 √ ≈ ϕ(−1.8282) = 1−ϕ(1.8282) ≈ 0.03376.
490.75
In other words, had the result been a tie, there is chance of at most 3.4%
no more than 941 of the 1963 voters would have voted for President Bush.
1. The Central Limit Theorem, continued 203

Homework Problems

Exercise 38.1. A carton contains 144 baseballs, each of which has a mean
weight of 5 ounces and a standard deviation of 2/5 ounces. (Standard
deviation is the square root of the variance.) Find an approximate value
for the probability that the total weight of the baseballs in the carton is no
more than 725 ounces.
Exercise 38.2. Let Xi ∼ Uniform(0, 1), where X1 , . . . , X20 are independent.
Find normal approximations for each of the following:
X
20
(a) P Xi 6 12 .
i=1
X
20
(b) The 90-th percentile of Xi ; i.e. the number a for which
i=1

X
20
P Xi 6 a = 0.9.
i=1

Exercise 38.3. Let Xi be the weight of the i-th passenger’s luggage. As-
sume that the weights are independent, each with pdf
f(x) = 3x2 /803 , for 0 < x < 80,
X
100
and 0 otherwise. Approximate P Xi > 6025 .
i=1

Exercise 38.4. Let X be the number of baskets scored in a sequence of

10, 000 free throws attempts by some NBA player. This player’s rate of
scoring is 80%. Estimate the probability that he scores between 7940 and
8080 baskets.
 Appendix A

1. What next?
This course has covered some basics in probability theory. There are sev-
eral (not necessarily exclusive) directions to follow from here.
One direction is learning some statistics. For example, if you would
like to compute the average height of students at the university, one way
would be to run a census asking each student for their height. Thanks
to the law of the large numbers, a more efficient way would be to collect
a sample and compute the average height in that sample. Natural ques-
tions arise: how many students should be in the sample? How to take
the sample? Is the average height in the sample a good estimate of the
average height of all university students? If it is, then how large an error
are we making? These are very important practical issues. Example 38.4
touched on this matter. The same kind of questions arise, for example,
when designing exit polls. The main question is really about estimating
parameters of the distribution of the data; e.g. the mean, the variance, etc.
This is the main topic of Statistical Inference I (Math 5080).
Another situation where statistics is helpful is, for example, when
someone claims the average student at the university is more than 6 feet
tall. How would you collect data and check this claim? Clearly, the first
step is to use a sample to estimate the average height. But that would be
just an estimate and includes an error that is due to the randomness of
the sample. So if you find in your sample an average height of 6.2 feet, is
this large enough to conclude the average height of all university students
is indeed larger then 6 feet? What about if you find an average of 6.01
feet? Or 7.5 feet? Can one estimate the error due to random sampling

205
206 A

and thus guarantee that an average of 7.5 feet is not larger than 6 feet only
due to randomness but because the average height of all students is really
more than 6 feet? Inline with this, example 38.5 shows how one can use
probability theory to check the validity of certain claims. These issues are
addressed in Statistical Inference II (Math 5090).
Another direction is learning more probability theory. Here is a se-
lected subset of topics you would learn in Math 6040. The notion of al-
gebra of events is established more seriously, a very important topic that
we touched upon very lightly and then brushed under the rug for the rest
of this course. The two main theorems are proved properly: the strong
law of large numbers (with just one finite moment, instead of four as we
did in this course) and the central limit theorem (with just two moments
instead of three). The revolutionary object “Brownian motion” is intro-
duced and explored. Markov chains may also be covered, depending on
the instructor and time. And more... We will talk a little bit about Brow-
nian motion in the next two sections. Brownian motion (and simulations)
is also explored in Stochastic Processes and Simulation I & II (Math 5040
and 5050).

1.1. History of Brownian motion, as quoted from the Wiki. The Roman
Lucretius’s scientific poem On the Nature of Things (c. 60 BC) has a re-
markable description of Brownian motion of dust particles. He uses this
as a proof of the existence of atoms:
”Observe what happens when sunbeams are admitted into a build-
ing and shed light on its shadowy places. You will see a multitude of
tiny particles mingling in a multitude of ways... their dancing is an actual
indication of underlying movements of matter that are hidden from our
sight... It originates with the atoms which move of themselves [i.e. spon-
taneously]. Then those small compound bodies that are least removed
from the impetus of the atoms are set in motion by the impact of their
invisible blows and in turn cannon against slightly larger bodies. So the
movement mounts up from the atoms and gradually emerges to the level
of our senses, so that those bodies are in motion that we see in sunbeams,
moved by blows that remain invisible.”
Although the mingling motion of dust particles is caused largely by air
currents, the glittering, tumbling motion of small dust particles is, indeed,
caused chiefly by true Brownian dynamics.
Jan Ingenhousz had described the irregular motion of coal dust par-
ticles on the surface of alcohol in 1785. Nevertheless Brownian motion
is traditionally regarded as discovered by the botanist Robert Brown in
1827. It is believed that Brown was studying pollen particles floating in
1. What next? 207

water under the microscope. He then observed minute particles within the
vacuoles of the pollen grains executing a jittery motion. By repeating the
experiment with particles of dust, he was able to rule out that the motion
was due to pollen particles being ’alive’, although the origin of the motion
was yet to be explained.
The first person to describe the mathematics behind Brownian mo-
tion was Thorvald N. Thiele in 1880 in a paper on the method of least
squares. This was followed independently by Louis Bachelier in 1900 in
his PhD thesis ”The theory of speculation”, in which he presented a sto-
chastic analysis of the stock and option markets. However, it was Albert
Einstein’s (in his 1905 paper) and Marian Smoluchowski’s (1906) indepen-
dent research of the problem that brought the solution to the attention of
physicists, and presented it as a way to indirectly confirm the existence of
atoms and molecules.
However, at first the predictions of Einstein’s formula were refuted by
a series of experiments, by Svedberg in 1906 and 1907, which gave dis-
placements of the particles as 4 to 6 times the predicted value, and by
Henri in 1908 who found displacements 3 times greater than Einstein’s
formula predicted. But Einstein’s predictions were finally confirmed in a
series of experiments carried out by Chaidesaigues in 1908 and Perrin in
1909. The confirmation of Einstein’s theory constituted empirical progress
for the kinetic theory of heat. In essence, Einstein showed that the motion
can be predicted directly from the kinetic model of thermal equilibrium.
The importance of the theory lay in the fact that it confirmed the kinetic
theory’s account of the second law of thermodynamics as being an essen-
tially statistical law.

1.2. More history. Einstein predicted that the one-dimensional Brownian

motion is a random function of time, written as W(t) for “time” t > 0,
such that:
(a) At time 0, the random movement starts at the origin; i.e. W(0) = 0.
(b) At any given time t > 0, the position W(t) of the particle has the
normal distribution with mean 0 and variance t.
(c) If t > s > 0, then the displacement from time s to time t is inde-
pendent of the past until time s; i.e., W(t) − W(s) is independent
of all the values W(r); r 6 s.
(d) The displacement is time-homogeneous; i.e., the distribution of
W(t) − W(s) is the same as the distribution of W(t − s) which is
in turn normal with mean 0 and variance t − s.
(e) The random function W is continuous.
208 A

In 1923, Norbert Wiener (a professor at MIT and a child prodigy)

proved the existence of Brownian motion and set down a rm mathemati-
cal foundation for its further development and analysis. Wiener used the
recently-developed mathematics of É. Borel and H. Steinhaus (the subject
is called measure theory), and cleverly combined it with a nice idea from
a different mathematical discpline (harmonic analysis).
Finally, the classical development of Brownian motion was complete
in a 1939 work of Paul Lévy who proved the following remarkable fact:
If you replace the normal distribution by any other distribution in Ein-
steins predicates, then either there is no stochastic process that satises the
properties (a)–(d), or (e) fails to hold! Lévys work was closely related to
the concurrent and independent work of A. I. Khintchine in Russia, and is
nowadays called the Lévy-Khintchine Formula.
The work of Paul Lévy started the modern age of random processes,
and at its center, the theory of Brownian motion. The modern literature on
this is truly vast. But all probabilists would (or should) agree that a center-
piece of the classical literature is the 1942/1946 work of K. Itô who derived
a calculus – and thereby a theory of stochastic differential equations – that
is completely different from the ordinary nonstochastic theory. This theory
is nowadays at the very heart of the applications of probability theory to
mathematical nance, mathematical biology, turbulence, oceanography, etc.
For us, the final important step in the analysis of Brownian motion was
the 1951 work of Donsker who was a Professor of mathematics at The New
York University. Amongst other things, Donsker verified a 1949 conjecture
of the great American mathematician J. L. Doob by showing that once you
run them for a long time, all mean-zero variance-one random walks look
like Brownian motion! We will say more on this in the next section.

1.3. Random Walk and Brownian Motion. We describe the one-dimensional

case, but adding more dimensions is not too hard. In one dimension, imag-
ine the lattice of integer numbers. Say a particle starts at position 0. If the
particle is at position x, then it flips a fair coin and moves to x + 1 if the
coin lands Heads and to x − 1 if it lands Tails. The motion of the particle
is a random process called simple symmetric r andom walk. The increments
of the random walker are simply a sequence of independent random vari-
ables taking the value 1 with probability 1/2 and −1 with probability 1/2.
If we call the k-th increment Xk , then the position of the walker at time n
is Sn = X1 + · · · + Xn . This is a simplistic model of a dust particle moving
in discrete time and discrete space. As we have done a few times in this
course, we can try and derive a continuum model. The idea is to plot the
positions Sn against the times n and connect the dots. If we do this for
n very large and look from afar, so that the fine details are washed out,
1. What next? 209

but not too far, so that there is still a process going on and we do not just
see a straight line!, then a continuous curve emerges. This is the so-called
Brownian motion. This is not a trivial fact to work out mathematically and
is expressed in the following theorem. A picture explains it nicely, though;
see Figure 1.3.
Donsker’s Theorem. Let X1 , X2 , . . . denote independent, identically distributed
random variables with mean zero and variance one. The random walk is then
the random sequence √ Sn = X1 √ + · · · + Xn , and√for all n large, the random
graph (0, 0), (1, S1 / n), (2, S2 / n), · · · , (n, Sn/ n) (linearly interpolated in
between the values), is close to the graph of Brownian motion run until time one.

Once it is shown that the polygonal graph does have a limit, Einstein’s
predicates (a)–(d) are natural ((b) being the result of the central limit the-
orem). (e) is not trivial at all and is a big part of the hard work.
 Appendix B

Solutions

211
212 B. Solutions

Exercise 1.1
(a) {4}
(b) {0, 1, 2, 3, 4, 5, 7}
(c) {0, 1, 3, 5, 7}
(d) ∅
Exercise 1.2
(a) Let x ∈ (A ∪ B) ∪ C. Then we have the following equivalences:
x ∈ (A ∪ B) ∪ C ⇔ x ∈ A ∪ B or x ∈ C
⇔ x ∈ A or x ∈ B or x ∈ C
⇔ x ∈ A or x ∈ (B ∪ C)
⇔ x ∈ A ∪ (B ∪ C)
This proves the assertion.
(b) Let x ∈ A ∩ (B ∪ C). Then we have the following equivalences:
x ∈ A ∩ (B ∪ C) ⇔ x ∈ A and x ∈ B ∪ C
⇔ (x ∈ A and x ∈ B) or (x ∈ A and x ∈ C)
⇔ (x ∈ A ∩ B) or (x ∈ A ∩ C)
⇔ x ∈ (A ∩ B) ∪ (A ∩ C)
This proves the assertion.
(c) Let x ∈ (A ∪ B)c . Then we have the following equivalences:
x ∈ (A ∪ B)c ⇔ x 6∈ A ∪ B
⇔ (x 6∈ A and x 6∈ B)
⇔ x ∈ Ac and x ∈ Bc
⇔ x ∈ Ac ∩ Bc
This proves the assertion.
(d) Let x ∈ (A ∩ B)c . Then we have the following equivalences:
x ∈ (A ∩ B)c ⇔ x 6∈ A ∩ B
⇔ (x 6∈ A or x 6∈ B)
⇔ x ∈ Ac or x ∈ Bc
⇔ x ∈ Ac ∪ Bc
This proves the assertion.
Exercise 1.3
(a) A ∩ B ∩ Cc
(b) A ∩ Bc ∩ Cc
B. Solutions 213

(c) (A ∩ B ∩ Cc ) ∪ (A ∩ Bc ∩ C) ∪ (Ac ∩ B ∩ C) ∪ (A ∩ B ∩ C) =
(A ∩ B) ∪ (A ∩ C) ∩ (B ∩ C)
(d) A ∪ B ∪ C
(e) (A ∩ B ∩ Cc ) ∪ (A ∩ Bc ∩ C) ∪ (Ac ∩ B ∩ C)
(f) (A ∩ Bc ∩ Cc ) ∪ (Ac ∩ B ∩ Cc ) ∪ (Ac ∩ Bc ∩ C)
(g) (Ac ∩ Bc ∩ Cc ) ∪ (A ∩ Bc ∩ Cc ) ∪ (Ac ∩ B ∩ Cc ) ∪ (Ac ∩ Bc ∩ C)
Exercise 1.4 First of all, we can see that Ac and Bc are not disjoint: any
element that is not in A, nor in B will be in Ac ∩ Bc . Then, A ∩ C and B ∩ C
are disjoint as (A ∩ C) ∩ (B ∩ C) = A ∩ B ∩ C = ∅ ∩ C = ∅. Finally, A ∪ C
and B ∪ C are not disjoint as they both contain the elements of C (if this
one is not empty).
Exercise 1.5 The standard sample space for this experiment is to consider
Ω = {1, 2, 3, 4, 5, 6}3 , i.e. the set of all sequences of 3 elements chosen from
the set {1, 2, 3, 4, 5, 6}. In other words,
Ω = {(1, 1, 1), (1, 1, 2), . . . , (6, 6, 6)}.
There are 63 = 216 elements in Ω. As Ω is finite we can choose F to be the
set of all possible subsets of Ω.
Exercise 1.6 We can choose Ω = {B, G, R} where B denotes the black chip,
G the green one and R the red one. As the set is finite and the informa-
tion complete, we can choose F to be the set of all possible subsets of Ω.
Namely,
F = {∅, {B}, {G}, {R}, {B, G}, {B, R}, {R, G}, Ω} .
Exercise 1.7 See Ash’s exercise 1.2.7.
214 B. Solutions

Exercise 2.1 There are 150 people who favor the Health Care Bill, do not
approve of Obama’s performance and are not registered Democrats.
Exercise 2.2
(a) We have
(A ∩ B) \ (A ∩ C) = (A ∩ B) ∩ (A ∩ C)c = (A ∩ B) ∩ (Ac ∪ Cc )
= (A ∩ B ∩ Ac ) ∪ (A ∩ B ∩ Cc ) = A ∩ (B ∩ Cc ) = A ∩ (B \ C).
(b) We have
A \ (B ∪ C) = A ∩ (B ∪ C)c = A ∩ (Bc ∩ Cc ) = (A ∩ Bc ) ∩ Cc
= (A \ B) ∩ Cc = (A \ B) \ C.
(c) Let A = {1, 2, 3}, B = {2, 3, 4}, C = {2, 4, 6}. We have (A \ B) ∪ C =
{1, 2, 4, 6} and (A ∪ C) \ B = {1, 6}. Hence, the proposition is wrong.
Exercise 2.3 See Ash’s exercise 1.2.5.
B. Solutions 215

Exercise 3.1
(a) We can choose Ω = {HH, HT , T H, T T }, where H stands for heads
and T for tails. The first letter is the outcome of the first toss and
the second letter, the outcome of the second toss.
(b) The sample space Ω is finite and, at step 3, all information is
available, so we can choose F3 = P(Ω), the set of possible subsets
of Ω.
(c) At step 2, we do not know the result of the second toss. Hence,
if an observable subset contains xH, it has to contain xT , because
we would have no way to distinguish both. Hence, we have to
choose
F2 = {∅, {HH, HT }, {T H, T T }, Ω} .
Other subsets, such as {HT } cannot be observed at this step. In-
deed, one would need to know the outcome of the second toss to
decide if {HT } happens or not. As for the sets in F2 above, you do
not need to know the second outcome to decide if they happen or
not.
(d) At step 1, we know neither of the outcomes, so we can just decide
about the probability of the trivial events and we have to pick
F1 = {∅, Ω}.
Exercise 3.2
(a) We can choose Ω to be the set of all sequences of outcomes that
are made of only tails and one head at the end. That is
Ω = {H, T H, T T H, T T T H, . . .}.
Notice that we can also consider Ω = N, where ω = n means that
the game ended at toss number n. We also would like to point out
that it is customary to add an outcome ∆ called the cemetery out-
come which corresponds to the case where the experiment never
ends.
(b) With the different sample spaces above, we can describe the events
below this way.
{Aaron wins} = {H, T T H, T T T T H, . . .}, {Bill wins} = {T H, T T T H, T T T T T H, . . .}
and
{no one wins} = ∅.
If we consider the case where Ω = N, we obtain
{Aaron wins} = {odd integers}, {Bill wins} = {even integers}
216 B. Solutions

and
{no one wins} = ∅.
We notice that if you consider the cemetery outcome ∆, then {no
one wins} becomes {∆} rather than ∅. We will see later that with
a fair coin, the probability that no one wins is 0, hence modelling
this by ∅ is ok. To illustrate why ∆ could be useful, imagine
that the players play with a coin with two tails faces. Then the
only possible outcome is ∆ and this one can’t have probability 0.
Hence, modelling by ∅ would not be appropriate in this case.
Exercise 3.3 Let’s consider the case where we toss a coin twice and let
A = {we get H on the first toss} and B = {we get T on the second toss}.
Hence,
A = {HH, HT }, B = {HT , T T }, and A \ B = {HH}.
Hence,
1
P(A \ B) = P{HH} = ,
4
but
1 1
P(A) − P(B) = − = 0.
2 2
B. Solutions 217

Exercise 4.1 (a) There are 65 possible outcomes. (b) There are only 63
possible outcomes with the first and last rolls being 6. So the probability
in question is 63 /65 .
Exercise 4.2 There are 103 = 1000 ways to choose a 3-digit number at
random. Now, there are 3 ways to choose the position of the single digit
larger than 5, 4 ways to choose this digit (6 to 9) and 6 · 6 ways to choose
the two other digits (0 to 5). Hence, there are 3 · 4 · 6 · 6 ways to choose
a 3-digit number with only one digit larger than 5. The probability then
becomes:
3·4·6·6 432
p= = = 43.2%.
103 1000
Exercise 4.3
(a) We can apply the principles of counting and choosing each sym-
bol on the license plate in order. We obtain 26 × 26 × 26 × 10 ×
10 × 10 = 17, 576, 000 diffrent license plates.
(b) Similarly, we have 103 × 1 × 26 × 26 = 676, 000 license plates with
the alphabetical part starting with an A.

Exercise 4.4 Let A be the event that balls are of the same color, R, Y and G
the event that they are both red, yellow and green, respectively. Then, as
R, Y and G are disjoint,
3×5 8×7 13 × 6 149
P(A) = P(R∪Y∪G) = P(R)+P(Y)+P(G) = + + = = 0.345.
24 × 18 24 × 18 24 × 18 432
218 B. Solutions

Exercise 5.1 Each hunter has 10 choices: hunter 1 makes one of 10 choices,
then hunter 2 makes 1 of 10 choices, etc. So over all there are 105 possible
options. On the other hand, the number of ways to get 5 ducks shot is: 10
for hunter 1 then 9 for hunter 2, etc. So 10 × 9 × 8 × 7 × 6 ways. The answer
thus is the ratio of the two numbers: 10×9×8×7×6
105
.
B. Solutions 219

Exercise 6.1 There are 64

8 ways to place 8 rooks on a chessboard. If we
want the rooks not to check each other, we place the first rook on column
1 (8 ways) then the second rook on column 2 but not on the same row as
the first rook (7 ways) and so on. I total there are 8! ways to do this. So the
probability the 8 rooks are not checking each other is 8!/ 64


8 which equals
the claimed number. (Check this last claim yourself!)
Exercise 6.2 We put the women together in order to form one entity.
Hence, this problem comes back to seating m + 1 entity (m men and one
entity for the women). We have (m + 1)! ways to do it. Then, among the
women together, we have w! ways to seat them. As we have (m + w)! ways
to seat this people, the probability is
(m + 1)!w!
P(women together) = .
(m + w)!
Exercise 6.3 See Ash’s exercise 1.4.7.
Exercise 6.4
(a) There are 54


6 possible combinations of 6 numbers (the order
doesn’t matter). Only one of them will match the one you played.
Hence, the probability to win the first prize is
1 1
p= 54

= .
6
25, 827, 165

(b) We have 65 48



1 ways to choose a combination of 6 numbers that
shares 5 numbers with the one played (5 numbers out of the 6
played and 1 out of the 48 not played). Hence, the probability to
win the second prize is
6

48

· 6 · 48 288 3
p = 5 54
1 = 54
= = .
6 6
25, 827, 165 269, 033
Exercise 6.5
(a) There are 50


5 possible combinations of 5 numbers in the first list
and 92 combinations of 2 numbers in the second list. That makes


50

9

5 · 2 possible results for this lottery. Only one will match the
combination played. Hence, the probability to win the first prize
is
1 1
p = 50
9
=
5 2
76, 275, 360
(b) Based only on the probability to win the first prize, you would
definitely choose the first one which has a larger probability of
winning.
220 B. Solutions

52


Exercise 6.6 The number of possible poker hands is 5 as we have seen
in class.
(a) We have 13 ways to choose the value for the four cards. The suits
are all taken. Then, there are 48 ways left to choose the fifth card.
Hence, the probability to get four of a kind is
13 · 48
P(four of a kind) = 52

= 0.00024.
5

(b) We have 13 ways to choose the value for the three cards. The, 43

ways to choose the suits. Then, there are 12

2 · 4 · 4 ways left to
choose the last two cards (both of different values). Hence, the
probability to get three of a kind is
13 · 43 · 12


2
·4
P(three of a kind) = 52

2 = 0.0211.
5
(c) In order to make a straight flush, we have 10 ways to choose the
highest card of the straight and 4 ways to choose the suit. Hence,
the probability to get a straight flush is
10 · 4
P(straight flush) = 52

= 0.0000154.
5
(d) In order to make a flush, we have 4 ways to choose the suit and
then 13


5 ways to choose the five cards among the 13 of the suit
selected. Nevertheless, among those flushes, some of them are
straight flushes, so we need to subtract the number of straight
flushes obtained above. Hence, the probability to get a flush is
4 · 13


5 − 40
P(flush) = 52

= 0.00197.
5
(e) In order to make a straight, we have 10 ways to choose the highest
card of the straight and then 45 ways to choose the suits (4 for
each card). Nevertheless, among those straights, some of them
are straight flushes, so we need to subtract the number of straight
flushes obtained above. Hence, the probability to get a straight is
10 · 45 − 40
P(straight) = 52

= 0.00392.
5
Exercise 6.7 Observe first that once people are seated moving everyone
one seat to the right gives the same seating arrangement! So at a round
table the first person can sit anywhere. Then, the next person has n − 1
possible places to sit at, the next has n − 2 places, and so on. In total, there
are (n − 1)! ways to seat n people at a round table.
B. Solutions 221

Exercise 6.8

(a) For the draw with replacement, there are 5210 possible hands.
If we want no to cards to have the same face values, we have
13 · 12 · · · 4 ways to pick diffrent values and then, 410 ways to
choose the suits (4 for each card drawn). Hence, the probability
becomes
13 · · · · · 4 · 410
p= = 0.00753.
5210
(b) In the case of the draw without replacement, we have 52


10 possible
hands. The number of hands that have at least 9 card of the same
suit can have 9 of them or 10 of them. The first case corresponds
to 4 · 13


9 · 39 possibilities (4 possible suits, 9 cards out of this suit
and 1 additionnal card from the 39 remaining) and the second
case corresponds to 4 · 13


10 possibilities. Hence, the probability
becomes

4 · 13 13



9 · 39 + 4 · 10
p= 52

= 0.00000712.
10

Exercise 6.9 As the order doesn’t count, we have 10

5 possible ways to
draw the balls. If we want the second largest number to be 8, we need to
pick the 8, then pick one larger number among the two possible and pick
3 numbers among the 7 lower numbers. Hence, the probability becomes
2 7



1 · 3
p= 10

= 0.2778.
5

Exercise 6.10 See Ash’s exercise 1.4.8.

Exercise 6.11

(a) This comes back to counting the number of permutations of 8

different people. Hence, there are 8! (= 40320) possibles ways to
seat those 8 people in a row.
(b) People A and B want to be seated together. Hence, we will con-
sider them as one single entity that we will first treat as a single
person. Hence, we will assign one spot to each person and one
spot to the entity AB. There are 7 entities (6 people and the group
AB). There are 7! ways to seat them. For each of these ways, A
and B can be seated in 2 different ways in the group. As a con-
sequence, there are 2 · 7! (= 10080) possible ways to seat these 8
people with A and B seated together.
222 B. Solutions

(c) First of all, notice that there are two possible ways to sit men and
women in alternance, namely
wmwmwmwm or mwmwmwmw,
where w stands for a woman and m for a man. Then, for each
of the repartitions above, we have to choose the positions of the
women among themselves. There are 4! permutations. For each
repartition of the women, we need to choose the positions of the
men. There are 4! permutations as well. Hence, there are 2·4!·4! (=
1152) ways to seat 4 women and 4 men in alternance.
(d) Similarly as in (b), the 5 men form an entity that we will treat as
a single person. Then, there are 4 entities (3 women and 1 group
of men) to position. There are 4! ways to do it. For each of these
ways, the 5 men can be seated differently on the 5 consecutive
chairs they have. There are 5! to do it. Hence, there are 4! · 5! (=
2880) possible ways to seat those 8 people with the 5 men seated
together.
(e) We consider that each married couple forms an entity that we
will treat as a single person. There are then 4! ways to assign
seats to the couples. For each of these repartitions, there are two
ways to seat each person within the couple. Hence, there are
4! · 2 · 2 · 2 · 2 = 4! · 24 (= 384) possible ways to seat 4 couples.
Exercise 6.12
(a) There are 6 discs to store on the shelf. As they are all different,
there are 6! (= 720) ways to do it.
(b) Assume the classical discs, as well as the jazz discs form two en-
tities, that we will consider as a single disc. Then, there are 3 en-
tities to store and 3! ways to do it. For each of these repartitions,
the classical discs have 3! ways to be stored within the group and
the jazz discs hav 2 ways to be stored within the group. Globally,
there are 3! · 3! · 2 (= 72) ways to store the 6 discs respecting the
styles.
(c) If only the classical discs have to be stored together, we have 4
entities (the classical group, the three other discs). We have then
4! ways to assign their position. For each of their repartitions,
we have 3! to store the classical discs within the group. Hence,
we have 4! · 3! ways to store the discs with the classical together.
Nevertheless, among those repartitions, some of them have the
jazz discs together, which we don’t want. Hence, we subtract
from the number above, the number of ways to store the discs
according to the styles found in (b). Hence, there are (4! · 3!) −
B. Solutions 223

(3! · 3! · 2) (= 144 − 72 = 72) ways to store the discs with only the
classical together.
Exercise 6.13
(a) The 5 letters of the word “bikes” being different, there are 5! (=
120) ways to form a word.
(b) Among the 5 letters of the word “paper”, there are two p’s. First
5


choose their position, we have 2 ways to do it. Then, there are
3!
ways to position the other 3 different letters. Hence, we have
5 5!
2 · 3! = 2! = 60 possible words.
(c) First choose the positions of the e’s, then of the t’s and finally the
ones of the other letters. Hence, we have 62 42 · 2! = 2!2! 6!



= 180
possible words.
(d) Choose the poisition of the three m, then the ones of the two i’s
and finally the ones of the other different letters. Hence, we have
7 4 7!



3 2 · 2! = 3!2! = 420 possible words.
224 B. Solutions

8
Exercise 7.1 In (a + b)8 the coefficient of b5 is × a3 . Hence if a = 2 and


5
b = 3x, the coefficient of x5 is 85 × 23 × 35 .

Exercise 7.2 See Ash’s exercise 1.4.9.

Exercise 7.3 In order to prove that
          
n+m n m n m n m
= + + ··· + ,
r 0 r 1 r−1 r 0
we will consider a group of people made of m women and n men. Let
0 6 r 6 min(m, n). We will count the number of teams of r people that
we can form from
the m + n people available. On the one hand, this
number is m+nr . On another hand, we can count the number of teams

n of

m
r people with k women (and r − k men), with k 6 r. We have k r−k
of them. Summing the number of such teams over all possible values of k,
we obtain the total number of teams of r people, namely
Xr  
n m
     
n m n m
   
n m
= + + ··· + .
k r−k 0 r 1 r−1 r 0
k=0
The two results are solutions of a same combinatorial problem and have
to be equal. The result is proved.
Exercise 7.4
(a) The number of players in the team is not fixed. If we have to
form a k-player team, we have n


k ways to pick the players. Then,
there are k ways to choosePthe captain of this team. Summing
n n


from k = 1 to n, we obtain k=1 k k possible teams. On another
hand, we can first choose a captain (n ways), and then choose for
each of the n − 1 remaining people if they are part of the team or
not (2n−1 ways). That gives n 2n−1 possible teams.
(b) In that case, we proceed as in (a), but we choose a captain and
an assistant-captain. On the left hand-side, k(k − 1) represents
the number of ways to choose the captain and his assistant. On
the right-hand side, we first choose the captain and the assistant
(n(n−1)). There are n−2 people left to be part or not of the team.
That gives the result.
(c) The binomial theorem gives
X
n  
n k
(1 + x) = x .
k
k=0
Taking derivatives on both sides with respect to x, we have
Xn  
n k−1
n−1
n(1 + x) = k x .
k
k=1
B. Solutions 225

Taking x = 1 in the equation above gives (a). Differentiating a

second time, we get
∞
X  
n−2 n k−2
n(n − 1)(1 + x) = k(k − 1) x .
k
k=2
Taking x = 1 gives (b).
Exercise 7.5
(a) For each digit, except the zero, we can build a 4-digit number.
Hence, there are 9 possible numbers.
(b) Two cases can occur. First of all, the number of ways to build a 4-
digit number made of two pairs of different digits, different from
0 is 92 42 . Indeed, we choose two digits among 9 and two places

among four two place the first-type
digit. Secondly, if one of the
pairs is a pair of 0’s, we have 9 · 32 possible numbers. Indeed,
there are 9 ways to choose the second pair and we need to choose
two spots among three for the 0’s (we cannot put the 0 upfront).
Finally, there are 92 42 + 9 · 32 = 243 4-digit numbers made of

two pairs of two different digits.

(c) We again distinguish the cases with or without 0. There are 94 · 4!

numbers with 4 different digits without 0. Indeed, we choose 4

digits
among 9 that we can place in any order. Moreover, there
are 93 · 3 · 3! 4-digit numbers with 0. We choose the three other
numbers among 9, the position of 0 and
finally, we
can place the
others in any order. Hence, there are 94 · 4! + 93 · 3 · 3! = 4536
4-digit numbers with different digits.
(d) In the case where
the number have to be ordered in increasing
order, there are 94 ways to choose the 4 different digits (0 can’t
be chosen) and only one way to place them in order. Hence, there
are 94 4-digit ordered numbers.

(e) In (a), there are 9 possible numbers, for any value of n. In (d),
following

the same argument as for n = 4, we notice that there
are n9 n-digit ordered numbers for 1 6 n < 10. There are none
of them for n > 10. In (c), for 2 6 n 6 9, we have n9 · n! n-digit


9


numbers with different digits without 0 and n−1 ·(n−1)·(n−1)!
9 9



numbers with 0. Hence, we have n ·n!+ n−1 ·(n−1)·(n−1)! =
9·9!
(10−n)! n-digit numbers with different digits. There are 9 · 9! for
n = 10 and none for n > 10.
226 B. Solutions

Exercise 8.1 See Ash’s exercise 1.6.1.

Exercise 8.2 Let’s assume n > 3, otherwise the answer is 0. We will denote
by X the number of heads that we obtain. We want to find P{X > 3|X > 1}.
We have
P({X > 3} ∩ {X > 1}) P{X > 3}
P{X > 3|X > 1} = =
P{X > 1} P{X > 1}
n
1 − 21n (1 + n + 2 ) 2n − 1 − n − n2



= = ,
1 − 21n 2n − 1
where
we used that the probability to get k heads out of n tosses is given
by n 1
k 2n .
Exercise 8.3 Let F denote the event the a fair coin is used and H the event
that the first n outcome of the coin are heads. We want to find P(F|H). We
know that
1
P(F) = P{outcome of the die is odd} =
2
and that
P(H|F) = 2−n P(H|Fc ) = pn .
We can use Bayes’ theorem to obtain
P(H|F)P(F) 2−n · 12 2−n
P(F|H) = = = .
P(H|F)P(F) + P(H|Fc )P(Fc ) 2−n · 12 + pn · 1
2
2−n + pn

Exercise 8.4 We will use the Law of Total Probability with an infinite num-
ber of events. Indeed, for every n > 1, the events {I = n} are disjoint (we
can’t choose two different integers) and their union is Ω (one integer is
necessarily chosen). Hence, letting H denote the event that the outcome is
heads, we have
P(H|I = n) = e−n .
Then, by the Law of Total Probability, we have
∞
X ∞
X ∞
X 1 1
P(H) = P(H|I = n)P{I = n} = e−n 2−n = (2e)−n = 1
−1 = ,
1 − 2e 2e −1
n=1 n=1 n=1
P∞
because n=0 xn = 1
1−x for |x| < 1.
Exercise 8.5 See Ash’s exercise 1.6.5.
Exercise 8.6 See Ash’s exercise 1.6.6.
Exercise 8.7 Let D denote the event that a random person has the disease,
P the event that the test is positive and R the event that the person has the
rash. We want to find P(D|R). We know that
P(D) = 0.2 P(P|D) = 0.9 P(P|Dc ) = 0.3 and P(R|P) = 0.25.
B. Solutions 227

First of all, let’s notice that we have P(R|D) = P(R|P∩D)P(P|D) = 0.25·0.9 =

0.225 and P(R|Dc ) = P(R|P ∩ Dc )P(P|Dc ) = 0.25 · 0.3 = 0.075. Now, by
Bayes’ theorem, we have
P(R|D)P(D) 0.225 · 0.2 0.045 3
P(D|R) = = = = .
P(R|D)P(D) + P(R|Dc )P(Dc ) 0.225 · 0.2 + 0.075 · 0.8 0.105 7

Exercise 8.8 Let A denote the event “the customer has an accident within
one year” and let R denote the event “the customer is likely to have acci-
dents”.
(a) We want to find P(A). By the Law of Total Probability, we have
P(A) = P(A | R)P(R) + P(A | Rc )P(Rc ) = (0.4 × 0.3) + (0.2 × 0.7) = 0.26.
(b) We want to compute P(R | A). The defnition of conditional proaba-
bility leads to
P(A | R)P(R) 0.4 × 0.3
P(R | A) = = = 0.46,
P(A) 0.26
where we used the result in (a).
Exercise 8.9 Let Ri denote the event “the receiver gets an i” and Ei the
event “the transmitter sends an i” (i ∈ {0, 1}).
(a) We want to find P(R0 ). By the Law of Total Probability,
P(R0 ) = P(R0 | E0 )P(E0 )+P(R0 | E1 )P(E1 ) = (0.8×0.45)+(0.1×0.55) = 0.415,
as E0 = Ec1 .
(b) We want to compute P(E0 | R0 ). The definition of conditional
probability leads to
P(R0 | E0 )P(E0 ) 0.8 × 0.45
P(E0 | R0 ) = = = 0.867,
P(R0 ) 0.415
where we used the result in (a).
Exercise 8.10 Let I,L and C be the events “the voter is independent, demo-
crat or republican”, respectively. Let V be the event “he actually voted in
the election”.
(a) By the Law of Total Probability, we have
P(V) = P(V|I)P(I) + P(V|L)P(L) + P(V|C)P(C) = 0.4862.
(b) We first compute P(I|V). By Bayes’ theorem, we have
P(V|I)P(I) 0.35 · 0.46
P(I|V) = = = 0.331.
P(V) 0.4862
228 B. Solutions

Similarly, we have
P(V|L)P(L) 0.62 · 0.30
P(L|V) = = = 0.383,
P(V) 0.4862
and
P(V|C)P(C) 0.58 · 0.24
P(C|V) = = = 0.286.
P(V) 0.4862
Exercise 8.11 Let An be the event “John drives on the n-th day” and Rn be
the event “he is late on the n-th day”.
(a) Let’s compute P(An ), we have
P(An ) = P(An |An−1 )P(An−1 ) + P(An |Acn−1 )P(Acn−1 )
1 1
= P(An−1 ) + (1 − P(An−1 ))
2 4
1 1
= P(An−1 ) + ,
4 4
c
where the event An stands for “John takes the train on the n-th
day.” Iterating this formula n − 1 times, we obtain
X  1 i  1 n−1
n−1
 n−1 !
1 1 1 − ( 14 )n−1
P(An ) = P(A1 ) + = p+
4
i=1
4 4 4 1 − 41
 n−1  n−1 !
1 1 1
= p+ 1− .
4 3 4
(b) By the Law of Total Probability, we have
P(Rn ) = P(Rn |An )P(An ) + P(Rn |Acn )P(Acn )
1 1
= P(An ) + (1 − P(An ))
2 4
1 1
= P(An ) + = P(An+1 ).
4 4
By (a), we then have
 n   n 
1 1 1
P(Rn ) = p+ 1− .
4 3 4
(c) Let’s compute limn→∞ P(An ). We know that limn→∞ ( 14 )n−1 = 0.
Hence, limn→∞ P(An ) = 13 . Similarly, we have limn→∞ P(Rn ) =
limn→∞ P(An+1 ) = 31 .
B. Solutions 229

Exercise 9.1

(a) The events ”‘getting a spade”’ and ”‘getting a heart”’ are disjoint
but not independent.
(b) The events ”‘getting a spade”’ and ”‘getting a king”’ are indepen-
dent (check the definition) and not disjoint: you can get the king
of spades.
(c) The events ”‘getting a king”’ and ”‘getting a queen and a jack”’
are disjoint (obvious) and independent. As the probability of the
second event is zero, this is easy to check.
(d) The events ”‘getting a heart”’ and ”‘getting a red king”’ are not
disjoint and not independent.

Exercise 9.2 The number of ones (resp. twos) is comprised between 0 and
6. Hence, we have the following possibilities : three ones and no two,
four ones and one two. (Other possibilities are not compatible with the
experiment.) Hence, noting A the event of which we want the probability,
we have

P(A) = P{three 1’s, no 2 (and three others)} + P{four one’s, one two (and one other)}
6 6

3
2

3 ·4 4 · 1 ·4
= + .
66 66
Indeed, we have to choose 3 positions among 6 for the ones and four
choices for each of the other values for the first probability and we have
to choose 4 positions among 6 for the ones, one position among the two
remaining for the two and we have 4 choices for the last value for the
second probability. The total number of results is 66 (six possible values
for each roll of a die).
Exercise 9.3

(a) If A is independent of itself, then P(A) = P(A ∩ A) = P(A)P(A) =

P(A)2 . The only possible solutions to this equation are P(A) = 0
or P(A) = 1.
(b) Let B be any event. If P(A) = 0, then A ∩ B ⊂ A, hence 0 6
P(A∩B) 6 P(A) = 0. As a consequence, P(A∩B) = 0 = P(A)P(B).
On another hand, if P(A) = 1, then P(Ac ) = 0. Hence, by the first
part, Ac is independent of any event B. This implies that A is
independent of any event B by the properties of independence.

Exercise 9.4 The sample space for this experiment is

Ω = {(P, P, P), (P, P, F), (P, F, P), (P, F, F), (F, P, P), (F, P, F), (F, F, P), (F, F, F)}.
230 B. Solutions

All outcomes are equally likely and then, P{ω} = 18 , for all ω ∈ Ω. More-
over, counting the favorable cases for each event, we see that
4 1
P(G1 ) = = = P(G2 ) = P(G3 )
8 2
2 1
P(G1 ∩ G2 ) = = = P(G1 )P(G2 ).
8 4
Similarly, we find that P(G1 ∩ G3 ) = P(G1 )P(G3 ) and that P(G2 ∩ G3 ) =
P(G2 )P(G3 ). The events G1 , G2 and G3 are pairwise independent.
However,
2 1 1
P(G1 ∩ G2 ∩ G3 ) = = 6= P(G1 ) · P(G2 ) · P(G3 ) = ,
8 4 8
hence G1 , G2 and G3 are not independent. Actually, it is to see that if G1
and G2 occur, then G3 occurs as well, which explains the dependence.
Exercise 9.5 We consider that having 4 children is the result of 4 indepen-
dent trials, each one being a success (girl) with probability 0.48 or a failure
(boy) with probability 0.52. Let Ei be the event “the i-th child is a girl”.
(a) Having children with all the same gender corresponds to the
event { 4 successes or 0 success}. Hence, P(“all children have the
same gender”) = P(“4 successes”) + P(“0 success”) = (0.48)4 +
(0.52)4 .
(b) The fact that the three oldest children are boys and the youngest
is a girl corresponds to the event Ec1 ∩ Ec2 ∩ Ec3 ∩ E4 . Hence P(“three
oldest are boys and the youngest is a girl”) = (0.52)3 (0.48).
(c) Having three boys comes back to having 1 success among the 4
trials. Hence, P(“exactly three boys”) = 43 (0.52)3 (0.48).

(d) The two oldest are boys, the other do not matter. This comes back
to having two failures among the first two trials. Hence, P(“the
two oldest are boys”) = (0.52)2 .
(e) Let’s first compute the probability that there is no girl. This equals
the probability of no sucess, that is (0.52)4 . Hence, P(“at least one
girl”) = 1 − P(“no girl”) = 1 − (0.52)4 .
B. Solutions 231

Exercise 10.1 The sample space is Ω = {H, T }3 := {HHH, HHT , HT H, HT T , T HH, T HT , T T H, T T T }.

These eight outcomes are equally likely, hence the probability measure is
given by P{ω} = 18 for all ω ∈ Ω. The random variable X can be defined
by
X
3
X(ω) = 1{H} (ωi ) when ω = (ω1 ω2 ω3 ).
i=1
Otherwise, one can define X this way :


 0 if ω = HHH,

1 if ω = HHT , HT H, T HH,
X(ω) =

 2 if ω = HT T , T HT , T T H,

3 if ω = T T T .

Exercise 10.2 The sample space is Ω = {1, 2, 3, 4, 5, 6}3 := {(ω1 , ω2 , ω3 ) :

ω1 , ω2 , ω3 ∈ {1, 2, 3, 4, 5, 6}}. There are 216 equally likely outcomes, hence
1
the probability measure is given by P{ω} = 216 for all ω ∈ Ω. The random
variable X can be defined by
X(ω) = ω1 · ω2 · ω3 when ω = (ω1 , ω2 , ω3 ).
232 B. Solutions

Exercise 11.1 See Ash’s exercise 2.3.2.

k 5
3−k
Exercise 11.2 P(X = k) = k3 16 6 , which gives
x 0 1 2 3
125 75 15 1
f(x) 216 216 216 216
f(x) = 0 for all x 6= 0, 1, 2, 3.
Exercise 11.3
(a)
 1

 if x = 1, 9, 16, 25 or 36,


36

 1
if x = 2, 3, 5, 8, 10, 15, 18, 20, 24 or 30,
 18
f(x) = 1
if x = 4,


12

 1
if x = 6 or 12,

 9

0 otherwise.
(b)
x 1 2 3 4 5 6
1 1 5 7 1 11
f(x) 36 12 36 36 4 36
f(x) = 0 for all x 6= 0, . . . , 6.
Exercise 11.4 The random variable X counts the number of even outcomes,
when we roll a fair die twice. Its probability mass function is
x 0 1 2
1 1 1
f(x) 4 2 4
f(x) = 0 for all x 6= 0, 1, 2.
Exercise 11.5
(a) There are 53 = 10 ways of picking the balls. The maximum num-

ber can only be 3, 4 or 5.

x 1 2 3 4 5
1 3 6
f(x) 0 0 10 10 10
f(x) = 0 for all x 6= 1, . . . , 5.
(b) The minimum number can only be 1, 2 or 3.
x 1 2 3 4 5
6 3 1
f(x) 10 10 10 0 0
f(x) = 0 for all x 6= 1, . . . , 5.
Exercise 11.6 See Ash’s exercise 1.5.4.
B. Solutions 233

Exercise 12.1
(a) The random variable X has a geometric distribution with param-
eter p, hence P{X = n} = p(1 − p)n−1 . Then,
∞
X X∞ X∞ ∞
X
n−1 n−1 1
P{X = n} = p(1 − p) =p (1 − p) =p (1 − p)n = p = 1.
1 − (1 − p)
n=1 n=1 n=1 n=0
by the standard formula for geometric series.
(b) The random variable Y has a Poisson distribution with parameter
n
λ, hence P{Y = n} = e−λ λn! . Then,
∞
X ∞
X ∞
X λn
λn
P{Y = n} = e−λ = e−λ = e−λ eλ = 1,
n! n!
n=0 n=0 n=0
by the standard series expansion for exponentials.
Exercise 12.2
(a) Let X be the r.v. counting the number of cars having an accident
this day. The r.v. X has a binomial distribution with parameters
n = 10, 000 and p = 0.002. As p is small, n is large and np is
not too large, nor too small, we can approximate X by a Poisson
random variable with parameter λ = np = 20. Then, we have
λ15 2015
P{X = 15} ' e−λ = e−20 = 5.16%.
15! 15!
We notice that the exact value of P{X = 15} would be precisely
5.16%.
(b) As above, let Y be the r.v. counting the number of gray cars having
an accident this day. By a similar argument as in (a) and as one car
out of 5 is gray, the r.v. Y follows a binomial random variable with
parameters n = 2, 000 and p = 0.002. We can again approximate
by a Poisson distribution of parameter λ = np = 4. Then, we have
λ3 43
P{Y = 3} ' e−λ = e−4 = 19.54%.
3! 3!
The exact value would be P{Y = 3} = 19.55%.
234 B. Solutions

Exercise 13.1
(a) We can easilly check that F(x) is a non-decreasing function, that
limx→∞ F(x) = 1, that limx→−∞ F(x) = 0 and that F is right-
continuous. (A plot can help.) Hence, F is a cumulative distri-
bution function.
(b) We will use the properties of CDFs to compute the probabilities.
Namely, we have
 
1 1 1 1
P{X = 2} = F(2) − F(2−) = ·2+ − = .
6 3 3 3
1
(c) P{X < 2} = F(2−) = limx↑2 3 = 13 .
(d) As the two events are disjoint, we have
1 3 1 3
P X = 2 or 6X< = P{X = 2} + P 6X6
2 2 2 2
= P{X = 2} + (F (3/2−) − F (1/2−))
 
1 1 1 7
= + − =
3 3 12 12
(e) As 2 is included in [ 21 ; 3], we have
1 1
P X = 2 or 6X63 = P 6X63
2 2
= F(3) − F (1/2−)
5 1 3
= − = .
6 12 4
B. Solutions 235

Exercise 14.1 The random variable X is neither discrete, nor continuous.

Indeed, F has jumps, which prevents it from being continuous and the
portions between jumps are not always constant.
Exercise 14.2
R+∞
(a) We need to find c such that −∞ f(x)dx = 1. In order for f to be a
pdf, we need c > 0. Moreover,
Z +∞ Z2
x3 2
 
2
f(x)dx = c (4 − x )dx = c 4x −
−∞ −2 3 −2
   
8 8 32
= c 8− − −8 + = c.
3 3 3
3
Hence, c = 32 .
Rx
(b) The cdf F of X is given by F(x) = −∞ f(u)du. As x 6 −2, the
Rx
integral vanishes. Moreover, as x > 2, we have −∞ f(u)du =
R+∞
−∞ f(u)du = 1. Then, for −2 < x < 2,
Zx Z
3 x u3 x
 
2 3
f(u)du = (4 − u )du = 4u −
−∞ 32 −2 32 3 −2
3 x3 16
     
3 x 8 3
= 4x − − −8 + = 4x − +
32 3 3 32 3 3
1
16 + 12x − x3 .


=
32
Finally,


 0 if x 6 −2,
1 3
F(x) = 32 (16 + 12x − x ) if − 2 < x < 2,

 1 if x > 2.

Exercise 14.3
R+∞
(a) We need to find c such that −∞ f(x)dx = 1. In order for f to be a
pdf, we need c > 0. Moreover,
Z +∞ Zπ Zπ
2 2 1 + cos(2x)
2
f(x)dx = c cos (x) dx = c dx
−∞ 0 0 2
  π2
x sin(2x) πc
= c + = .
2 4 0 4
4
Hence, c = π.
Rx
(b) The cdf F of X is given by F(x) = −∞ f(u)du. As x 6 0, the
Rx
integral vanishes. Moreover, as x > π2 , we have −∞ f(u)du =
236 B. Solutions

R+∞
−∞ f(u)du = 1. Then. for 0 < x < π2 ,
Zx Z x
4 x
 
2 4 u sin(2u)
f(u)du = cos (u) du = +
−∞ π 0 π 2 4 0
 
2 sin(2x)
= x+ .
π 2
Finally,


 0 if x 6 0,
  
2 sin(2x) π
F(x) = π x+ if 0 < x < 2,


2
 1 if x > π
2.

Exercise 14.4 For each question, we need to find the right set (union of
Rb
intervals) and integrate f over it. The fact that for a, b > 0, a f(x) dx =
1 −a
2 (e − e−b ) is used throughout.
(a) By symmetry around 0, we have
1
P{|X| 6 2} = 2 · P{0 6 X 6 2} = 2 · (1 − e−2 ) = 1 − e−2 .
2
(b) We have {|X| 6 2 or X > 0} ⇔ {X > −2}. Hence,
Z∞ Z Z
1 0 x 1 ∞
P{|X| 6 2 or X > 0} = f(x) dx = e dx + f(x) dx
−2 2 −2 2 0
1 1 1
= (1 − e−2 ) + = 1 − e−2 .
2 2 2
(c) We have {|X| 6 2 or X 6 −1} ⇔ {X 6 2}. Moreover, by symmetry,
P{X 6 2} = P{X > −2} = 1 − 12 e−2 , by the result in (b).
(d) The condition |X| + |X − 3| 6 3 corresponds to 0 6 X 6 3. Hence,
1
P{|X| + |X − 3| 6 3} = P{0 6 X 6 3} = (1 − e−3 ).
2
(e) We have X3 −X2 −X+2 = (X−2)(X2 +X+1). Hence, X3 −X2 −X+2 >
0 if and only if X > 2. Then, using the result in (c)
1
P{X3 − X2 − X + 2 > 0} = P{X > 2} = e−2 .
2
(f) We have
esin(πX) > 1 ⇔ sin(πX) > 0
⇔ X ∈ [2k, 2k + 1] for some k ∈ Z.
Now by symmetry, P{−2k 6 X − 2k + 1} = P{2k − 1 6 X 6 2k}.
Hence, P{X ∈ [2k, 2k + 1] for some k ∈ Z} = P{X > 0} = 21 .
B. Solutions 237

(g) As X is a continuous random variable,

∞
X
P{X ∈ N} = P{X = n} = 0,
n=0

as P{X = x} = 0 for every x for a continuous random variable.

Exercise 14.5
R+∞
(a) In order for f to be a pdf, we need c > 0. Let’s compute −∞ f(x)dx:
Z +∞ Z +∞
1 √ +∞
f(x)dx = c √ dx = c(2 x) = +∞,
−∞ 1 x 1

for all c > 0. Hence, there is no value of c for which f is a pdf.

(b) We need to check the properties of a cdf. First, F is non-decreasing.
1 1
Indeed, if 0 < x 6 y, then 0 6 e− x 6 e− y .
The function F is right-continuous. For x 6= 0, this is obvious.
1
At x = 0, we have limx↓0 e− x = limy↑∞ e−y = 0.
Finally, limx→−∞ F(x) = 0 by definition and
1
lim F(x) = lim e− x = lim e−z = 1.
x→+∞ x→+∞ z→0

The function F is a cdf. The density function is given by f(x) =

F 0 (x).Hence, f(x) = 0 pour x < 0. For x > 0, we have
d −1 1 1
f(x) = e x = 2 e− x .
dx x
0 (0) = 0 and
At x = 0, we have F−
1
0 F(0 + h) − F(0) e− h
F+ (0) = lim = lim = lim xe−x = 0.
h↓0 h h↓0 h x↑+∞

Hence, F 0 (0) = 0 and we finally have


0 if x 6 0,
f(x) = 1 − x1
x2
e if x > 0.

Exercise 14.6
R+∞
(a) In order for f to be a pdf, we need c > 0. Let’s compute −∞ f(x)dx:
Z +∞ Z +∞
1 +∞ πc
f(x)dx = c 2
dx = c(arctan(x)) 0 = .
−∞ 0 1+x 2
2
Then, taking c = π, f is a pdf. This is a Cauchy distribution.
238 B. Solutions

(b) We need to check the properties of a cdf. First check that F is

non-decreasing. It is enough to check that g(x) = √ x 2 is non-
1+x
decreasing. For x > 0, we have
s r
x2 1
g(x) = 2
= 1− ,
1+x 1 + x2
which is non-decreasing on [0, +∞). The function g being an odd
function, this is also true for x < y 6 0. Finally, for x < 0 < y, we
have g(x) < 0 < g(y). Hence, F is non-decreasing.
It is obvious to see that F is right-continuous.
Finally,
 s 
x2 
 
1 x 1
lim F(x) = 1 + lim √ = 1+ lim = 1.
x→+∞ 2 x→+∞ 1 + x2 2 x→+∞ 1 + x2

and
   
1 x 1 x
lim F(x) = 1 + lim √ = 1 − lim √ = 0.
x→−∞ 2 x→−∞ 1 + x2 2 x→+∞ 1 + x2
The function F is a cdf. The density function is given by f(x) =
0
F (x). We have
 √1 + x2 − x √2x
1 + x2 − x2

d 1 x 2 1+x2 1
f(x) = 1+ √ = 2
= 3 = 3 ,
dx 2 1 + x2 2(1 + x ) 2(1 + x2 ) 2 2(1 + x2 ) 2
for all x ∈ R.
B. Solutions 239

Exercise 17.1 We use the formula developped in class. We have n = 10, 000,
p= 7940, b = 8080, p = 0.8. Hence, np = 8, 000, np(1 − p) = 1, 600 and
a
np(1 − p) = 40. Now,
   
8, 080 − 8, 000 7, 940 − 8, 000
P{7940 6 X 6 8080} = Φ −Φ = Φ(2) − Φ(−1.5)
40 40
= Φ(2) − 1 + Φ(1.5) = 0.9772 + 0.9332 − 1 = 0.9104.
Hence, there is 91.04% probability to find between 7,940 and 8,080 suc-
cesses.
240 B. Solutions

Exercise 19.1 Let g : (−1, 1) → ( e1 , e) defined by g(x) = e−x . Then, g is

one-to-one. Its inverse is the solution of y = e−x which is x = − log y. As
1
2 if x ∈ (−1, 1),
fX (x) =
0 otherwise,
we have
1
dx 1 if y ∈ ( e1 , e),
fY (y) = fX (x)| | = fX (− log y) = 2y
dy y 0 otherwise.

Exercise 19.2 Let g : (0, ∞) → (0, ∞) defined by g(x) = x2 . Then, g is

√ √
one-to-one. Its inverse is the solution of y = x2 which is x = y (not − y
since x > 0). As
λ e−λx if x > 0,
fX (x) =
0 if x 6 0,
we have
 √
dx √ 1 λe−λ y
√ if y > 0,
fY (y) = fX (x)| | = fX ( y) √ = 2 y
dy 2 y 0 if y 6 0.
Exercise 19.3
(a) We have Y = g(X), with g : R → (0, ∞) given by g(x) = ex . Then,
g is one-to-one. Its inverse is the solution to y = ex which is
x = log y. As X is a standard normal random variable, fX (x) =
x 2
√1 e− 2 . Then, for y > 0,
2π
1 (log y)2 1 (log y)2
fY (y) = √ e− 2 |(log y) 0 | = √ e− 2 ,
2π 2π y
and fY (y) = 0 if y 6 0.

(b) We have h(Z) = X, with h : R → R given by h(z) = z3 +z+1.

This function is one-to-one. Indeed, h 0 (z) = 3z2 + 1 > 0 for all z ∈
R, hence h is strictly increasing. Moreover, limz→+∞ h(z) = +∞
and limz→−∞ h(z) = −∞, which ensures the bijectivity of h. The
r.v. Z is then given by Z = h−1 (X) which (conveniently!) has
solution X = h(Z).
The probability density function fZ is then fZ (z) = fX (h(z))|h 0 (z)|.
As X is a standard normal random variable,
1 (z3 +z+1)2
fZ (z) = √ (3z2 + 1)e− 2 ,
2π
for all z ∈ R.
B. Solutions 241

Exercise 19.4
(a) We have Y = g(X), with g : (0, ∞) → R given by g(x) = log x. The
function g is one-to-one. Its inverse is the solution to y = log x
which is x = ey . Hence :
fY (y) = fX (ey )|(ey ) 0 |.
As X is an exponential r.v. with parameter λ, fX (x) = λe−λx , for
x > 0 and fX (x) = 0 otherwise. Hence, for y ∈ R,
y y
fY (y) = λe−λe |ey | = λey−λe .
(b) We have h(Z) = X, with h : − π2 , π2 → R given by h(z) = z +

tan(z). This function is one-to-one. Indeed, h 0 (z) = 1 + cos21(z) >

0 for all z ∈ − π2 , π2 , hence g is strictly increasing. Moreover,

limz→ π2 h(z) = +∞ and limz→− π2 h(z) = −∞, which ensures the

bijectivity of h. The r.v. Z is then given by Z = h−1 (X), the inverse
of which is (conveniently) given by X = h(Z).
The probability density function fZ is then fZ (z) = fX (h(z))|h 0 (z)|.
As X is a standard normal random variable,
 
1 1 −
(z+tan(z))2
fZ (z) = √ 1+ e 2 ,
2π cos2 (z)
for all z ∈ − π2 , π2 .

Exercise 19.5 Let g : [1, ∞) → [1, ∞) defined by

2x if x > 2,
g(x) =
x2 if x < 2,
Then, as it is increasing, the function g is one-to-one. As
1
x2
if x > 1,
fX (x) =
0 otherwise,
we have
fY (y) = fX (h(y))|h 0 (y)|.
We have to consider two cases. First, for 1 6 y < 4, we have
√ 1
h(y) = g−1 (y) = y, h0 (x) = √ .
2 y
Then, for y > 4, we have
y 1
h(y) = g−1 (y) = , h0 (x) = .
2 2
242 B. Solutions

Hence,



1
2y3/2
if y ∈ [1, 4),
fY (y) = 2
if y > 4,


y2
0 otherwise.

Exercise 19.6
(a) We have Y = g(X), with g : R → R given by g(x) = x2 . The
function g is not one-to-one from R into R. First find the cumu-
lative distribution function FY from the definition. For y 6 0,
FY (y) = P{Y 6 y} = P{X2 6 y} = 0. For y > 0,
√ √ √ √
FY (y) = P{Y 6 y} = P{X2 6 y} = P{− y 6 X 6 y} = F( y) − F(− y),

where F is the CDF of X. We then find the probability density

function fY by taking the derivative. For y < 0, fY (y) = 0. For
y > 0,
1 √ 1 √ 1 √ √
fY (y) = FY0 (y) = √ f( y) + √ f(− y) = √ (f( y) + f(− y)).
2 y 2 y 2 y
Finally,
 √ √
1
2 y (f( y)
√ + f(− y)) if y > 0,
fY (y) =
0 if y < 0.

Alternatively, we could have applied the formula for fY once to

each solution of x2 = y and then added the two. So
√ √ √ √
fY (y) = f( y)|( y) 0 | + f(− y)|(− y) 0 |

which gives the same answer as above.

(b) We will apply the formula obtained in (a) when X is a standard
x 2
normal r.v., namely with f(x) = √1 e− 2 . In this case, for y > 0,
2π

1 1 y y 1 1 y
fY (y) = √ √ (e− 2 + e− 2 ) = √ y− 2 e− 2 ,
2 y 2π 2π
which corresponds to a Gamma density function with parameters
1 y
α = 12 and λ = 12 . Indeed, y− 2 e− 2 appears in the Gamma density
and the constant is necessarily the right one, determined by the
R+∞ √
property 0 fY (y)dy = 1. In particular, Γ ( 12 ) = π.
v2 v2
Exercise 19.7 Let g : [0, π2 ] → [0, g0 ] be defined by g(x) = g0 sin(2θ). The
function g is not one-to-one as every possible sine value can be obtained
B. Solutions 243

in two ways, namely as 1

2 arcsin( gR
v2
) and π
2 − 1
2 arcsin( gR
v2
). Hence, for
0 0
v20
06r6 g,
     
1 gr π 1 gr π
FR (r) = P(R 6 r) = P 0 6 θ 6 arcsin or − arcsin 6θ6
2 v20 2 2 v20 2
      
2 1 gr 2 π π 1 gr
= arcsin 2
+ − − arcsin
π 2 v π 2 2 2 v20
 0
2 gr
= arcsin .
π v20
v20
Now, we find that for 0 6 r 6 g,
 
2 g gr
fR (r) = FR0 (r)
= arcsin 0
π v20 v20
2g 1
= 2
r
πv0 2 2
1 − gvr4
0

2g 1
= q
π v4 − g2 r2
0

Finally,
 v20
2g √ 1
π if 0 6 r 6 g,
fR (r) = v40 −g2 r2
0 otherwise.
244 B. Solutions

Exercise 21.1
(a) No, X and Y are not independent. For instance, we have fX (1) =
0.4 + 0.3 = 0.7, fY (2) = 0.3 + 0.1 = 0.4. Hence, fX (1)fY (2) =
0.7 · 0.4 = 0.28 6= 0.3 = f(1, 2).
(b) We have
P(XY 6 2) = 1 − P(XY > 2) = 1 − P(X = 2, Y = 2) = 1 − 0.1 = 0.9.
Exercise 21.2
(a) The set of possible values for X1 and X2 is {1, . . . , 6}. By definition,
we always have X1 6 X2 . We have to ccompute f(x1 , x2 ) = P{X1 =
x1 , X2 = x2 }. If x1 = x2 , both outcomes have to be the same, equal
to x1 . There is only one possible roll for this, namely (x1 , x1 ) and
1
f(x1 , x2 ) = 36 . If x1 < x2 , one dice has to be x1 , the other one x2 .
There are two possible rolls for this to happen, namely (x1 , x2 ) and
1
(x2 , x1 ). We obtain f(x1 , x2 ) = 18 . Then, for x1 , x2 ∈ {1, 2, 3, 4, 5, 6},
 1

 36 if x1 = x2 ,
f(x1 , x2 ) = 1
if x1 < x2 ,


18
0 otherwise.
(b) In order to find the density of X1 , we have to add all the probabili-
P
ties for which X1 takes a precise value (i.e. fX1 (x1 ) = 6i=1 f(x1 , i)).
The following table sums up the results (as in the example in
class).
x1 |x2 1 2 3 4 5 6 fX1 (x1 )
1 1 1 1 1 1 11
1 36 18 18 18 18 18 36
1 1 1 1 1 9
2 0 36 18 18 18 18 36
1 1 1 1 7
3 0 0 36 18 18 18 36
1 1 1 5
4 0 0 0 36 18 18 36
1 1 3
5 0 0 0 0 36 18 36
1 1
6 0 0 0 0 0 36 36

1 3 5 7 9 11
fX2 (x2 ) 36 36 36 36 36 36
(c) They are not independent. Namely, f(x1 , x2 ) 6= fX1 (x1 )fX2 (x2 ). For
instance, f(6, 1) = 0 6= 3612 = fX1 (6)fX2 (1).
Exercise 21.3
(a) The set of possible values for X1 is {4, 5, 6, 7, 8} and the set of pos-
sible values for X2 is {4, 6, 8, 9, 12, 16}. We can see that the values
B. Solutions 245

of X1 and X2 only correspond to one exact possible draw (up to

the symmetry). Hence, possible values (4, 4), (6, 9) and (8, 16) for
(X1 , X2 ) respectivley correspond to the draws of (2, 2), (3, 3) and
(4, 4). Their probability is 19 . Possible values (5, 6), (6, 8) and (7, 12)
for (X1 , X2 ) respectivley correspond to the draws of (2, 3), (2, 4)
and (3, 4) (and their symmetric draws). Their probability is 29 .
Other pairs are not possible and have probability 0.
(b) In order to find the density of X1 , we have to add all the probabili-
P
ties for which X1 takes a precise value (i.e. fX1 (x1 ) = 6i=1 f(x1 , i)).
The following table sums up the results (as in the example in
class).
x1 |x2 4 6 8 9 12 16 fX1 (x1 )
1 1
4 9 0 0 0 0 0 9
2 2
5 0 9 0 0 0 0 9
2 1 3
6 0 0 9 9 0 0 9
2 2
7 0 0 0 0 9 0 9
1 1
8 0 0 0 0 0 9 9

1 2 2 1 2 1
fX2 (x2 ) 9 9 9 9 9 9
(c) They are not independent. Namely, f(x1 , x2 ) 6= fX1 (x1 )fX2 (x2 ). For
2
instance, f(5, 4) = 0 6= 81 = fX1 (5)fX2 (4).
Exercise 21.4 By Theorem 21.1 the transformation is


 0
√ if u < 0,



 3u if 0 6 u 6 1/3,
G(u) = 2 if 1/3 < u 6 2/3,



 6u − 2 if 2/3 < u < 1,


4 if u > 1.
Note that technically Theorem 21.1 asks for the CDF to be strictly in-
creasing on (−∞, ∞). Here, the CDF of U is constant on (−∞, 0) and
then on (1, ∞). However, since U never takes values in these intervals, we
should still be able to apply the theorem. Check the proof of the theorem
and make sure you understand why it still works.
246 B. Solutions

Exercise 23.1 We can see that the distribution of (X, Y) is uniform on the
square [−1, 1]2 . Hence, we can use a ratio of surfaces to compute the
probabilities. (In most cases a drawing of the domain can help.)
(a) We have P{X + Y 6 12 } = P{Y 6 1
− X} = 1 − P{Y > 12 − X}. Now the
2
surface corresponding to {Y > 1
− X} is a triangle and we have
2
1 3 2


1 · 23
P{X + Y 6 } = 1 − 2 2 = .
2 4 32
(b) The domain corresponding to {X − Y 6 2 } has exactly the same
1

shape as the one in (a). Hence, P{X − Y 6 12 } = 23

32 .
(c) We have XY > 14 ⇔ Y > 4X
1
if X > 0 and XY > 41 ⇔ Y < 4X1
if X < 0.
Now, we can write the surface of the domain corresponding to
XY > 14 as
Z1 Z1 Z1     1
1 ln(x) 3 − ln(4)
2 dx dy = 2 dx 1 − =2 x− = .
1
4 1/4x 1
4
4x 4 1
2
4

Hence, P{XY 6 41 } = 1 − P{XY > 14 } = 1 − 3−ln(4)

8 = 5+ln(4)
8 .
Y 1 X Y 1 X
(d) We have X 6 2 ⇔ Y 6 2 if X > 0 and X 6 2 ⇔ Y > 2 if X < 0.
Hence, the surface corresponding to { X Y
6 12 } is the union of two
trapezoids with surface 54 each. Hence, P{ XY
6 12 } = 2 · 5/4 5
4 = 8.
 Y Y2
= P Y 2 6 X4 = P − |X| |X|
2
(e) We have P X 6 12 = P X 1
2 6 4 2 6Y 6 2 .
We can easilly identify the surface as the union of two triangles of
surface 21 each and, hence,
Y 1 1/2 1
P 6 =2· = .
X 2 4 4
(f) We have P{|X| + |Y| 6 1} = P{|Y| 6 1 − |X|} = P{|X| − 1 6 Y 6 1 − |X|}.
The surface is then a square with
√ corners (0, 1), (−1, 0), (0, −1) and
(1, 0). The sides have length 2 and
√
( 2)2 1
P{|X| + |Y| 6 1} = = .
4 2
(g) We have P{|Y| 6 e } = P{−e 6 Y 6 eX }. This condition only
X X

matters when X < 0. Hence,

Z Z ex Z
1 0 1 1 1 0 1 1 1
P{|Y| 6 eX } = + dx dy = + dx ex = + (1−e−1 ) = 1− .
2 −1 −e x 4 2 2 −1 2 2 2e
B. Solutions 247

Exercise 24.1
(a) We must choose c such that
Z +∞ Z +∞
f(x, y)dxdy = 1.
−∞ −∞

But,
Z +∞ Z +∞ Z1 Z1 Z1  2 x=1
x
f(x, y) dxdy = c (x + y) dxdy = c + xy dy
−∞ −∞ 0 0 0 2 x=0
Z1  1
y y2
   
1 1 1
= c + y dy = c + =c + = c.
0 2 2 2 0 2 2
Hence, c = 1.
(b) Observe that
ZZ Z1 Zy
P{X < Y} = f(x, y) dxdy = (x + y) dxdy
0 0
{(x,y) : x<y}
Z1  x=y Z1  1
x2 3 2 3 y3 1
= + xy dy = y dy = = .
0 2 x=0 2 0 2 3 0 2
(c) For x 6∈ [0, 1], fX (x) = 0. For x ∈ [0, 1],
Z +∞ Z1 1
y2

1
fX (x) = f(x, y) dy = (x + y) dy = xy + = + x.
−∞ 0 2 0 2
By symmetry, fY (y) = fX (y) for all y ∈ R.
(d) We can write P{X = Y} as
ZZ Z∞ Zy
P{X = Y} = f(x, y) dxdy = f(x, y) dxdy = 0,
−∞ y
{(x,y) : x=y}

for all density function f. Hence, P{X = Y} = 0 for all jointly

continuous random variables.
Exercise 24.2
(a) First of all, observe that fX (x) = 0 if x 6∈ [0, 1]. Then, for 0 6 x 6 1,
Z Zx Z1
fX (x) = f(x, y)dy = 4xy dy + 6x2 dy = 2x3 + 6x2 (1 − x) = 6x2 − 4x3 .
R 0 x

Moreover, fY (y) = 0 for y 6∈ [0, 1]. Then, for 0 6 y 6 1,

Z Zy Z1
fY (y) = f(x, y)dx = 6x dx + 4xy dx = 2y3 + 2y(1 − y2 ) = 2y.
2
R 0 y
248 B. Solutions

(b) We have
P(A ∪ B) = P(A) + P(B) − P(A ∩ B)
= P{X 6 1/2} + P{Y 6 1/2} − P{X 6 1/2, Y 6 1/2}
Z 1/2 Z 1/2 Z 1/2 Zx Z 1/2 !
2 3 2
= (6x − 4x ) dx + 2y dy − dx 4xy dy + 6x dy
0 0 0 0 x

1/2 1/2
Z 1/2
= (2x3 − x4 ) + y2 − dx(3x2 − 4x3 )
0 0 0
 
1 1 1 1/2
= − + − (x3 − x4 )
4 16 4 0
 
7 1 1 6 3
= − − = = .
16 8 16 16 8
Exercise 24.3 First of all, fX (x) = 0 if x < 0. Now, for x > 0,
Z Zx x
fX (x) = f(x, y) dy = 2 e−(x+y) dy = 2e−x (−e−y ) = 2e−x (1 − e−x ).
R 0 0

We have fY (y) = 0 for y < 0. For y > 0,

Z Z∞ ∞
fY (y) = f(x, y) dx = 2 e−(x+y) dx = 2e−y (−e−x ) = 2e−2y .
R y y

Exercise 24.4 See Ash’s exercise 2.7.3.

Exercise 24.5 See Ash’s exercise 2.7.8.
B. Solutions 249

Exercise 25.1 There are two ways to proceed.

One way is to first compute the CDF. There are two cases to distin-
guish. If 0 < z 6 1, then the parabola y = zx2 intersects the square at the
right-most side. Then in that case,
Z1
P{Y/X 6 z} = P{Y 6 zX } = zx2 dx = z/3.
2 2
0

But when
√ z > 3 then the parabola intersects the square at its top side at
x = 1/ z and
Z 1/√z Z1
z 1 2
P{Y/X 6 z} = P{Y 6 zX } =
2 2 2
zx dx+ √ 1 dx = √ 3 +1− √ = 1− √ .
0 1/ z 3( z) z 3 z

Therefore, the pdf is 1/3 when z ∈ (0, 1] and 1/(3z3/2 ) when z > 1.
Alternatively, one can use the pdf method: let W = X and Z = Y/X2 .
Then X = W and Y = ZW 2 . The Jacobian matrix is
 
1 0
2zw w2
and its determinant is w2 . So
fZ,W (z, w) = 1 × w2 = w2 .
The crucial thing though is the domain! The above formula is valid if
0 < x < 1 and 0 < y < 1 which becomes 0 < w < 1 and 0 < zw2 < 1. The
pdf is 0 otherwise. So if 0 < z < 1 then 0 < w < 1 and while if z > 1 we
gave 0 < w < 1/sqrt(z). Finally, the pdf of Z is
Z1
1
fZ (z) = w2 dw = if 0 < z < 1
0 3
and
Z 1/√z
1
fZ (z) = w2 dw = if z > 1.
0 3z3/2
Exercise 25.2 First of all, by independence,
e−(x+y) if x > 0, y > 0,
fX,Y (x, y) = fX (x)fY (y) =
0 otherwise.
(a) We will use the transformation U = X, Z = X + Y. This transfor-
mation is bijective with inverse given by X = U, Y = Z − U. The
Jacobian of this transformation is given by
∂x ∂x
!  
∂u ∂z 1 0
J(u, z) = det = det = 1.
∂y ∂y −1 1
∂u ∂z
250 B. Solutions

Now,
fU,Z (u, z) = fX,Y (x(u, z), y(u, z))|J(u, z)| = e−(u+(z−u)) = e−z ,
for u > 0, z > 0 and u 6 z. The latter condition comes from y > 0.
Hence,
e−z if u > 0, z > 0 and u 6 z,
fU,Z (u, z) =
0 otherwise.
Finally, fZ (z) = 0 if z < 0 and, for z > 0,
Z Zz
fZ (z) = fU,Z (u, z) du = e−z du = ze−z .
R 0
Y
(b) Similarly as above, we will consider V = X, W = X . This trans-
formation is bijective with inverse X = V, Y = VW. The Jacobian
of this transformation is given by
∂x ∂x
!  
∂v ∂w 1 0
J(v, w) = det = det = v.
∂y ∂y w v
∂v ∂w
Now,
fV,W (v, w) = fX,Y (x(v, w), y(v, w))|J(v, w)| = e−(v+vw) · v = ve−(1+w)v ,
for v > 0, w > 0. Hence,
ve−(1+w)v if v > 0, w > 0,
fV,W (v, w) =
0 otherwise.
Finally, fW (w) = 0 if w < 0 and, for w > 0,
Z Z∞
1
fW (w) = fV,W (v, w) dv = ve−(1+w)v dv = ,
R 0 (1 + w)2
where we used the properties of Gamma integrals on p.73 of the
Lecture Notes.
Exercise 25.3 First of all, by independence,
1 − 1 (x2 +y2 )
fX,Y (x, y) = fX (x)fY (y) = e 2 .
2π
Y
We will consider U = X, Z = X . This transformation is bijective with
inverse X = U, Y = UZ. The Jacobian of this transformation is given by
∂x ∂x
!  
∂u ∂z 1 0
J(u, z) = det = det = u.
∂y ∂y z u
∂u ∂z
Now,
1 − 1 (u2 +u2 z2 ) 1 1 2 2
fU,Z (u, z) = fX,Y (x(u, z), y(u, z))|J(u, z)| = e 2 ·|u| = |u|e− 2 (1+z )u ,
2π 2π
B. Solutions 251

for all u, z ∈ R. Finally, for z ∈ R,

Z Z Z
1 ∞ − 12 (1+z2 )u2 1 ∞ − 1 (1+z2 )u2
fZ (z) = fU,Z (u, z) du = |u|e du = ue 2 du
R 2π −∞ π 0
1 1 2 2 ∞ 1
= − 2
e− 2 (1+z )u = .
π(1 + z ) 0 π(1 + z2 )
Exercise 25.4 See Ash’s exercise 2.8.5.
Exercise 25.5 See Ash’s exercise 2.8.6.
Exercise 25.6 See Ash’s exercise 2.8.8.
Exercise 25.7 First of all, by independence,
e−(x+y+z) if x > 0, y > 0, z > 0,
f(x, y, z) = fX (x)fY (y)fZ (z) =
0 otherwise.
Now, letting A = {(x, y, z) ∈ R3 : x > 2y > 3z}, we have
ZZZ
P{X > 2Y > 3Z} = f(x, y, z) dxdydz
Z∞ Z∞A
Z∞
= dz dy dx e−(x+y+z)
3
0 2z 2y
Z∞ Z∞ Z∞
= dz e−z dy e−y dx e−x
3
0 2z 2y
Z∞ Z∞
= dz e−z dy e−3y
3
0 2z
Z∞
1 11 2
= dz e− 2 z = .
3 0 33
Exercise 25.8 Let X (resp. Y) be the number of minutes after 10 am at
which the woman (resp. man) arrives. The random variables X and Y are
independent and both unifomrly distributed on the interval [0, 60]. Then,
the vector (X, Y) is uniforlmy distributed on the square [0, 60]2 . We want
to find the probability p that both people arrive within an interval of 10
minutes one from the other. In other words, we want p = P({X 6 Y 6
X + 10} ∪ {Y < X 6 Y + 10}). Both events are disjoint and, by symmetry,
have the same probability. Hence, p = 2P{X 6 Y 6 X+10} = 2P{(X, Y) ∈ A},
with A = {(x, y) ∈ [0, 60]2 : x 6 y 6 x + 10}. We know that
Area(A) Area(A)
p=2· 2
= .
Aire([0, 60] ) 1800
Let’s compute the area of A (a picture can help). The set A is a trpezoid
made of the triangle T1 = {(x, y) ∈ [0, 60]2 : x 6 y} minus the triangle
T2 = {(x, y) ∈ [0, 60]2 : y 6 x+10}. We have Area(A) = Area(T1 )− Area(T2 ).
252 B. Solutions

Triangle T1 (resp. T2 ) has size of length 60 (resp. 50). Then, we find

2 2
Area(A) = 602 − 502 = 1100 550 11
2 = 550. Finally, p = 1800 = 36 = 0.3056.
Exercise 25.9
(a) Let X (resp. Y) the number of minutes John will have to wait
for the bus (resp. the train). As John doesn’t know the exact
schedules, we assume that the random variables X and Y are in-
dependent and unifomrly distributed on the interval [0, 20] (resp.
[0, 10]). Hence, the random vector (X, Y) is uniformly distributed
on the rectangle [0, 20] × [0, 10]. We want to find the probability p
that the total travel time with public transportation is larger than
27. In other words, p = P{X + Y + 12 > 27} = P{X + Y > 15}. Hence,
p = P{(X, Y) ∈ A}, with A = {(x, y) ∈ [0, 20] × [0, 10] : x + y > 15}.
We know that
Area(A) Area(A)
p= = .
Area([0, 20] × [0, 10]) 200
Let’s compute Area(A). The set A is a trapezoid with small base
5, large base 15 and height 10. Hence, Area(A) = (5+15)·10
2 = 100.
100
Finally, p = 200 = 12 .
(b) If we know that the buses are systematically 2 minutes late, it
doesn’t change anything to the problem above. As John doesn’t
know the exact schedule, the uniform assumption remains un-
changed.
Exercise 25.10 First of all, by independence,
1 − 1 2 (x2 +y2 )
fX,Y (x, y) = fX (x)fY (y) = e 2σ .
2πσ2
We will consider the transformation X = R cos(Θ), Y = R sin(Θ). This
transformation is bijective, it is the polar change of coordinates. The Jaco-
bian of this transformation is given by
∂x ∂x
!  
∂r ∂θ cos(θ) −r sin(θ)
J(r, θ) = det = det = r.
∂y ∂y sin(θ) r cos(θ)
∂r ∂θ

Now,
1 − 1 2 (r2 cos2 (θ)+r2 sin2 (θ)) 1 r2
fR,Θ (r, θ) = fX,Y (x(r, θ), y(r, θ))|J(r, θ)| = 2
e 2σ ·|r| = 2
re− 2σ2 ,
2πσ 2πσ
for r > 0 and 0 6 θ < 2π. We can immediately conclude that R and Θ
are independent. Namely, it is easy to see that we can write fR,Θ (r, θ) =
g(r)h(θ) for suitable functions g and h. Finally, a direct integration shows
B. Solutions 253

that 
 r2
r
σ2
e− 2σ2 if r > 0,
fR (r) =
 0 otherwise.
and  1
2π if 0 6 θ < 2π,
fΘ (θ) =
0 otherwise.
Exercise 25.11 See Ash’s exercise 2.8.16.
254 B. Solutions

Exercise 27.1 We have

18 20
P{X = +1} = , P{X = −1} = .
38 38
Hence,
18 20 2
E[X] = (+1) × + (−1) × = − ' −0.0526.
38 38 38
This means that on average you lose 5.26 cents per bet.
Exercise 27.2 Let the sample space Ω be {1, 2, 5, 10, 20,“0”, “00”}. Denote
by ω the outcome of the wheel. The probability measure P is given by
ω 1 2 5 10 20 0 00
22 15 7 4 2 1 1
P{ω} 52 52 52 52 52 52 52
(a) Let H be the random variable given the profit of the player when
he bets $1 on each of the possible numbers or symbols. The pos-
sible values for H are
ω 1 2 5 10 20 0 00
H(ω) −5 −4 −1 4 14 34 34
(Remember that the player gets the $1 back if he wins.)
The probability mass function of H is
x −5 −4 −1 4 14 34
22 15 7 4 2 2
P{H = x} 52 52 52 52 52 52
Hence, the expectation is
22 15 7 4 2 2 65
E[H] = (−5)· +(−4)· +(−1)· +4· +14· +34· = − = −1.25.
52 52 52 52 52 52 52
(b) For m ∈ {1, 2, 5, 10, 20,“0”, “00”}, let Hm be the profit of the player
when he bets $1 on the number or symbol m. Then, Hm can only
take two values and its mass function is
x −1 m
, if m ∈ {1, 2, 5, 10, 20},
P{Hm = x} 1 − pm pm

x −1 40
, if m ∈ {0, 00},
51 1
P{Hm = x} 52 52
where pm = P{ω = m}. Hence, E[Hm ] = mpm + (−1)(1 − pm ).
The numerical results are presented in the following table:
m 1 2 5 10 20 0 00
8 7
E[Hm ] − 52 − 52 − 10
52
8
− 52 − 10
52 − 11
52 − 11
52
B. Solutions 255

Hence, betting on “0” or “00” gives the worst expectation and bet
on “2” gives the best. We notice that the expected values are all
negative and, hence, this game is always in favor of the organiser.
Exercise 27.3 We know that for a Geometric random variable, f(k) = P{X =
k} = p(1 − p)k−1 for k > 1. Hence, we have
X∞ X∞
E[X] = kp(1 − p)k−1 = p kqk−1 ,
k=1 k=1

with q := 1 − p. The trick to compute this sum is to remark that kqk−1 is

the derivative with repsect to q of qk . Hence, we can write
X∞ X ∞ ∞
d X k
!  
k−1 d k d 1
E[X] = p kq =p (q ) = p q =p
dq dq dq 1 − q
k=1 k=1 k=1
p 1
= 2
= .
(1 − q) p
256 B. Solutions

Exercise 29.1 For an exponential random variable, we have f(x) = λe−λx

for x > 0, f(x) = 0 otherwise. Hence, by an integration by parts, we have

Z∞
E[X] = xλe−λx dx
0
Z∞
∞
−xe−λx 0 e−λx dx

= +
0
∞
e−λx

= −
λ 0
1
= .
λ

Then, again using integration by parts and the results above, we have

Z∞
2
E[X ] = x2 λe−λx dx
0
Z∞
∞
−x2 e−λx 0 xe−λx dx

= +2
0
2
= .
λ2

Exercise 29.2 First of all, if n is odd, we have

Z∞
n 1 x2
E[X ] = √ xn e− 2 dx = 0,
2π −∞

x2
by the symmetry of the function x 7→ xn e− 2 (the function is odd). More-
over, if n = 2, we have seen in class that E[X2 ] = 1. Let’s prove the result
by induction. Assume the result is true for all even numbers up to n − 2
and let’s compute E[Xn ]. Using an integration by parts, we have

Z∞ Z∞
1 x2 1 x2
E[Xn ] = √ xn e− 2 dx = √ xn−1 xe− 2 dx
2π −∞ 2π −∞
∞ Z
(n − 1) ∞ n−2 − x2

1 n−1 − x2
2
= √ −x e + √ x e 2 dx
2π −∞ 2π −∞
= (n − 1)E[Xn−2 ] = (n − 1) · (n − 3)(n − 5) · · · 1.

Hence, by induction the result is true for every even integer n.

B. Solutions 257

Exercise 29.3 By the definition of expectation and using integration by

parts, we have
Z∞ Z3 Z∞
−x
E[c(X)] = c(x)f(x) dx = 2 xe dx + (2 + 6(x − 3))xe−x dx
−∞ 0 3
Z3 Z∞
−x 3 −x −x ∞
(12x − 16)e−x dx



= 2 −xe 0
+2 e dx + −(2 + 6(x − 3))xe 3
+
0 3
Z∞
3 ∞
= −6e−3 + 2 (−e−x ) 0 + 6e−3 + −(12x − 16)e−x 12e−x dx


3
+
3

∞
= 2 − 2e−3 + 20e−3 + −12e−x 3
= 2 + 18e−3 + 12e−3 = 2 + 30e−3 .
258 B. Solutions

Exercise 30.1 We remind that if X is exponentially distributed with param-

eter 1, then E[X] = 1.
(a) We have E[XY] = E[X]E[Y] = 1 · 1 = 1, since X and Y are indepen-
dent.
(b) We have E[X − Y] = E[X] − E[Y] = 1 − 1 = 0.
(c) This is Example 30.2 on page 147 in the Lecture Notes.
Exercise 30.2 The random variables X and Y have joint density function
f(x, y) = 14 if −1 6 x 6 1 and −1 6 y 6 1, f(x, y) = 0 otherwise. Hence,
Z Z1 Z Zx Z1 !
1 1 1 1
E[max(X, Y)] = dx dy max(x, y) = dx dy x + dy y
4 −1 −1 4 −1 −1 x
Z1 Z 1
1 − x2 1 1 1 (x + 1)3
 
1 1
= dx x(x + 1) + = dx (x + 1)2 = =
4 −1 2 8 −1 8 3 3
−1

Exercise 30.3 This corresponds to Examples 29.7 on page 144 and 30.8 on
page 149 in the Lecture Notes.
Exercise 30.4 This corresponds to Example 31.1 on page 153 in the Lecture
Notes.
Exercise 30.5 First of all, let’s notice that Y 2 + Z2 = cos2 (X) + sin2 (X) = 1
and that YZ = cos(X) sin(X) = sin(2X)2 . Hence, we have
Z
1 1 2π
E[YZ] = E[sin(2X)] = sin(2x) dx = 0
2 4π 0
Moreover,
Z
1 2π
E[Y] = E[cos(X)] = cos(x) dx = 0.
2π 0
Similarly, E[Z] = 0 and E[YZ] = E[Y]E[Z]. Then, as E[Y] = 0,
Z 2π 2π
2 1
2 2 1 sin(2x) 1
Var(Y) = E[Y ] = E[cos (X)] = cos (x) dx = (x − ) = .
2π 0 4π 2 2
0

Simlarly, we can show that Var(Z) = 21 . Moreover, as E[Y + Z] = 0,

Var(Y + Z) = E[(Y + Z)2 ] = E[Y 2 + Z2 + 2YZ] = 1 + 2E[YZ] = 1.
Hence, Var(Y + Z) = Var(Y) + Var(Z). Nevertheless, we have,
1
P(Y > 1/2) = P(cos(X) > 1/2) = P(−π/3 < X < π/3) = ,
3
1
P(Z > 1/2) = P(sin(X) > 1/2) = P(π/6 < X < 5π/6) = ,
3
B. Solutions 259

and
1 1
P(Y > 1/2, Z > 1/2) = P(π/6 < X < π/3) = 6= ,
12 9
which proves that Y and Z are not independent.
Exercise 30.6 See Ash’s exercise 3.2.8.
Exercise 30.7 This corresponds to Theorem 30.3 on page 148 in the Lecture
Notes.
260 B. Solutions

Exercise 32.1 See Ash’s exercise 3.4.1.

Exercise 32.2 See Ash’s exercise 3.4.3. (Go to the link on Ash’s website
saying Solutions to Problems Not Solved in the Text.)
Exercise 32.3 See Ash’s exercise 3.4.4.
Exercise 32.4 Recall that a direct computation shows that
Var(X + Y) = Var(X) + Var(Y) + 2Cov(X, Y).
This proves the result when n = 2. Now, we proceed by induction. Let
us assume the result is true for n and prove it for n + 1. Noting Sn =
X1 + · · · + Xn , we have
Var(X1 + · · · + Xn+1 ) = Var(Sn + Xn+1 )
= Var(Sn ) + Var(Xn+1 ) + 2Cov(Sn , Xn+1 )
X
n+1 X
n X
i−1
= Var(Xi ) + 2 Cov(Xi , Xj ) + 2Cov(Xn+1 , X1 + · · · + Xn )
i=1 i=1 j=1

X
n+1 X
n X
i−1 X
n
= Var(Xi ) + 2 Cov(Xi , Xj ) + 2 Cov(Xn+1 , Xj )
i=1 i=1 j=1 j=1

X
n+1 XX
n+1 i−1
= Var(Xi ) + 2 Cov(Xi , Xj ).
i=1 i=1 j=1
B. Solutions 261

Exercise 33.1 See Ash’s exercise 3.5.2. (Go to the link on Ash’s website
saying Solutions to Problems Not Solved in the Text.)
Exercise 33.2 Consider an element x. If it does not belong to any Ai ,
then all of the indicator functions in the formula take the value 0 and the
formula says 0 = 0, which is true.
If x does belong to at least one Ai , then consider all sets Ai to which
x belongs. There is no loss in generality when assuming these sets are
A1 , · · · , Ar for some r > 1. (Otherwise, simply rename the sets!) The left-
hand side of the formula is 1. So we need to show that the right-hand side
is also 1.
The indicator functions on the right-hand side take the value 0 unless
all the indices j1 , . . . , ji are among {1, . . . , r}, in which case the indicator
function takes the value 1. Moreover, for a given i 6 r the number of
possible choices of distinct integers j1 , . . . , ji from {1, . . . , r} is ri . Hence,

the right-hand side in fact equals

X
r  
i−1 r
(−1) .
i
i=1

Now since (1 − 1)r = 0 we can use the binomial formula to write

Xr  
i r
(−1) = 0.
i
i=0
But then
X
r   X
r  
i−1 r i r
(−1) =1− (−1) = 1.
i i
i=1 i=0
Exercise 33.3 See Ash’s exercise 3.5.6.
Exercise 33.4 See Ash’s exercise 3.7.1.
Exercise 33.5 See Ash’s exercise 3.7.3. (Go to the link on Ash’s website
saying Solutions to Problems Not Solved in the Text.)
262 B. Solutions

Exercise 34.1 See Ash’s exercise 4.4.4.

Exercise 34.2 See Ash’s exercise 4.4.7.
B. Solutions 263

Exercise 35.1 See Ash’s exercise 4.2.5.

Exercise 35.2 See Ash’s exercise 4.3.1.
Exercise 35.3 See Ash’s exercise 4.3.2.
Exercise 35.4 See Ash’s exercise 4.3.4.
Exercise 35.5 See Ash’s exercise 4.4.1.
Exercise 35.6 See Ash’s exercise 4.4.3.
Exercise 35.7 See Ash’s exercise 4.4.5.
Exercise 35.8 See Ash’s exercise 4.4.9.
Exercise 35.9 See Ash’s exercise 4.4.10.
Exercise 35.10 See Ash’s exercise 4.4.11.
Exercise 35.11 See Ash’s exercise 4.4.16.
Exercise 35.12 See Ash’s exercise 4.4.17.

Note: Some of the solutions are not in Ash’s book itself, but in a pdf
file on his site, under a link that says Solutions to Problems Not Solved in
the Text.
264 B. Solutions

Exercise 36.1 By Example 36.13,

 α  β
λ λ
MX (t) = and MY (t) =
λ−t λ−t
Then,
α+β

λ
MX+Y (t) = MX (t)MY (t) =
λ−t
and X + Y follows a Gamma distribution with parameters α + β and λ.
σ2i t2
Exercise 36.2 For i = 1, . . . , n, we have MXi (t) = exp(µi t + 2 ). Then,
MX1 +···+Xn (t) = MX1 (t) · · · · · MXn (t)
t2
= exp((µ1 + · · · + µn )t + (σ21 + · · · + σ2n )
).
2
Identifying the moment generating function, this proves the result.
Exercise 36.3
(a) It’s not an mgf, it can take negative values.
(b) It’s not an mgf, M(0) 6= 1.
(c) It is the mgf of an exponential random variable with parameter
λ = 1. (See Example 36.13)
(d) It is the mgf of a discrete random variable taking values −2, 0, 2, 13
1 1 1 1
with respective probabilities 12 , 3 , 2 , 12 .
Exercise 36.4
MY (t) = E[etY ] = E[et(aX+b) ] = E[ebt etaX ] = ebt E[etaX ] = ebt MX (at).
Exercise 36.5 To be added in the future.
B. Solutions 265

Exercise 37.1 For i = 1, . . . , n, we have MXi (t) = exp(λ(et − 1)). Hence,

MX1 +···+Xn (t) = MX1 (t) · · · · · MXn (t) = exp(nλ(et − 1)).
As a consequence, X1 + . . . + Xn has a Poisson distribution with parameter
nλ.
X1 +···+X
√ n −nλ , we have
Then, setting Zn = nλ
√
√t (X1 +···+Xn )
MZn (t) = E[etZn ] = e −t nλ
E[e nλ ]
√ t √
−t nλ −t nλ √t
=e MX1 +···+Xn ( √ ) = e exp(nλ(e nλ − 1)).
nλ
Using de l’Hospital’s rule, we can prove that, as n → ∞, this function
2
converges to exp( t2 ), hence to a standard normal distribution.
Exercise 37.2 For t < 1, the mgf is given by
Z∞ Z∞ Z∞
tX tx tx −(x+2)
MX (t) = E[e ] = e f(x) dx = e e dx = e(t−1)x−2 dx
−∞ −2 −2
∞
1 1
= e(t−1)x−2 = e−2t .
t−1 −2 1−t
Then,
0 (2t − 1)e−2t 0
MX (t) = and E[X] = MX (0) = −1,
(1 − t)2
and
00 2(2t2 − 2t + 1)e−2t 00
MX (t) = and E[X2 ] = MX (0) = 2.
(1 − t)3
λ t
ne
Exercise 37.3 We have MXn (t) = λ
1−(1− n )et
. We have
λ n t
t ne
MXn /n (t) = MXn ( ) = λ t .
n 1 − (1 − n )e n
Then,
λheht λ
lim MXn /n (t) = lim = .
n→∞ h→0 1 − (1 − λh)eht λ−t
Identifying the mgf, we see that Xnn converges in distribution to an expo-
nential random variable of parameter λ.
Exercise 37.4 To be added in the future.
266 B. Solutions

Exercise 38.1 The probability is approximately 0.8512.

Exercise 38.2 (a) The probability is approximately 0.9393; (b) a = 11.65.
Exercise 38.3 The probability is approximately 0.4359.
Exercise 38.4 The random variable X has binomial distribution with pa-
rameters n = 10, 000 and p = 0.8. Hence, E[X] = np = 8000 and Var(X) =
np(1 − p) = 1600. Now, by the Central Limit Theorem, we know that
Z = (X − np)/ np(1 − p) = X−8000
p
√
1600
follows approximately a N(0, 1) dis-
tribution. Hence,
 
7940 − 8000 X − 8000 8080 − 8000
P(7940 6 X 6 8080) = P 6 6
40 40 40
   
3 3
= P − 6 Z 6 2 ' Φ(2) − Φ − = 0.977 − 0.067 = 0.910.
2 2
The probability that the player scores between 7940 and 8080 baskets is
approximately 91%.
 Appendix C

Figure C.1. c 1991 Introduction to Probability and Mathematical Statis-

tics, 2nd Edition, by Bain & Engelhardt

267