Lecture Notes For Introductory Probability

Lecture Notes for Introductory Probability
Janko Gravner
Mathematics Department
University of California
Davis, CA 95616
[email protected]
June 9, 2011
These notes were started in January 2009 with help from Christopher Ng, a student in
Math 135A and 135B classes at UC Davis, who typeset the notes he took during my lectures.
This text is not a treatise in elementary probability and has no lofty goals; instead, its aim is
to help a student achieve the prociency in the subject required for a typical exam and basic
real-life applications. Therefore, its emphasis is on examples, which are chosen without much
redundancy. A reader should strive to understand every example given and be able to design
and solve a similar one. Problems at the end of chapters and on sample exams (the solutions
to all of which are provided) have been selected from actual exams, hence should be used as a
test for preparedness.
I have only one tip for studying probability: you cannot do it half-heartedly. You have to
devote to this class several hours per week of concentrated attention to understand the subject
enough so that standard problems become routine. If you think that coming to class and reading
the examples while also doing something else is enough, youre in for an unpleasant surprise on
the exams.
This text will always be available free of charge to UC Davis students. Please contact me if
you spot any mistake. I am thankful to Marisano James for numerous corrections and helpful
suggestions.
Copyright 2010, Janko Gravner
1 INTRODUCTION 1
1 Introduction
The theory of probability has always been associated with gambling and many most accessible
examples still come from that activity. You should be familiar with the basic tools of the
gambling trade: a coin, a (six-sided) die, and a full deck of 52 cards. A fair coin gives you Heads
(H) or Tails (T) with equal probability, a fair die will give you 1, 2, 3, 4, 5, or 6 with equal
probability, and a shued deck of cards means that any ordering of cards is equally likely.
Example 1.1. Here are typical questions that we will be asking and that you will learn how to
answer. This example serves as an illustration and you should not expect to understand how to
get the answer yet.
Start with a shued deck of cards and distribute all 52 cards to 4 players, 13 cards to each.
What is the probability that each player gets an Ace? Next, assume that you are a player and
you get a single Ace. What is the probability now that each player gets an Ace?
Answers. If any ordering of cards is equally likely, then any position of the four Aces in the
deck is also equally likely. There are
_
52
4
_
possibilities for the positions (slots) for the 4 aces. Out
of these, the number of positions that give each player an Ace is 13
4
: pick the rst slot among
the cards that the rst player gets, then the second slot among the second players cards, then
the third and the fourth slot. Therefore, the answer is
13
4
(
52
4
)
0.1055.
After you see that you have a single Ace, the probability goes up: the previous answer needs
to be divided by the probability that you get a single Ace, which is
13(
39
3
)
(
52
4
)
0.4388. The answer
then becomes
13
4
13(
39
3
)
0.2404.
Here is how you can quickly estimate the second probability during a card game: give the
second ace to a player, the third to a dierent player (probability about 2/3) and then the last
to the third player (probability about 1/3) for the approximate answer 2/9 0.22.
History of probability
Although gambling dates back thousands of years, the birth of modern probability is considered
to be a 1654 letter from the Flemish aristocrat and notorious gambler Chevalier de Mere to the
mathematician and philosopher Blaise Pascal. In essence the letter said:
I used to bet even money that I would get at least one 6 in four rolls of a fair die.
The probability of this is 4 times the probability of getting a 6 in a single die, i.e.,
4/6 = 2/3; clearly I had an advantage and indeed I was making money. Now I bet
even money that within 24 rolls of two dice I get at least one double 6. This has the
same advantage (24/6
2
= 2/3), but now I am losing money. Why?
As Pascal discussed in his correspondence with Pierre de Fermat, de Meres reasoning was faulty;
after all, if the number of rolls were 7 in the rst game, the logic would give the nonsensical
probability 7/6. Well come back to this later.
1 INTRODUCTION 2
Example 1.2. In a family with 4 children, what is the probability of a 2:2 boy-girl split?
One common wrong answer:
1
5
, as the 5 possibilities for the number of boys are not equally
likely.
Another common guess: close to 1, as this is the most balanced possibility. This repre-
sents the mistaken belief that symmetry in probabilities should very likely result in symmetry in
the outcome. A related confusion supposes that events that are probable (say, have probability
around 0.75) occur nearly certainly.
Equally likely outcomes
Suppose an experiment is performed, with n possible outcomes comprising a set S. Assume
also that all outcomes are equally likely. (Whether this assumption is realistic depends on the
context. The above Example 1.2 gives an instance where this is not a reasonable assumption.)
An event E is a set of outcomes, i.e., E S. If an event E consists of m dierent outcomes
(often called good outcomes for E), then the probability of E is given by:
(1.1) P(E) =
m
n
.
Example 1.3. A fair die has 6 outcomes; take E = 2, 4, 6. Then P(E) =
1
2
.
What does the answer in Example 1.3 mean? Every student of probability should spend
some time thinking about this. The fact is that it is very dicult to attach a meaning to P(E)
if we roll a die a single time or a few times. The most straightforward interpretation is that for
a very large number of rolls about half of the outcomes will be even. Note that this requires
at least the concept of a limit! This relative frequency interpretation of probability will be
explained in detail much later. For now, take formula (1.1) as the denition of probability.
2 COMBINATORICS 3
2 Combinatorics
Example 2.1. Toss three fair coins. What is the probability of exactly one Heads (H)?
There are 8 equally likely outcomes: HHH, HHT, HTH, HTT, THH, THT, TTH, TTT. Out
of these, 3 have exactly one H. That is, E = HTT, THT, TTH, and P(E) = 3/8.
Example 2.2. Let us now compute the probability of a 2:2 boy-girl split in a four-children
family. We have 16 outcomes, which we will assume are equally likely, although this is not quite
true in reality. We list the outcomes below, although we will soon see that there is a better way.
BBBB BBBG BBGB BBGG
BGBB BGBG BGGB BGGG
GBBB GBBG GBGB GBGG
GGBB GGBG GGGB GGGG
We conclude that
P(2:2 split) =
6
16
=
3
8
,
P(1:3 split or 3:1 split) =
8
16
=
1
2
,
P(4:0 split or 0:4 split) =
2
16
=
1
8
.
Example 2.3. Roll two dice. What is the most likely sum?
Outcomes are ordered pairs (i, j), 1 i 6, 1 j 6.
sum no. of outcomes
2 1
3 2
4 3
5 4
6 5
7 6
8 5
9 4
10 3
11 2
12 1
Our answer is 7, and P(sum = 7) =
6
36
=
1
6
.
2 COMBINATORICS 4
How to count?
Listing all outcomes is very inecient, especially if their number is large. We will, therefore,
learn a few counting techniques, starting with a trivial, but conceptually important fact.
Basic principle of counting. If an experiment consists of two stages and the rst stage has
m outcomes, while the second stage has n outcomes regardless of the outcome at the rst
stage, then the experiment as a whole has mn outcomes.
Example 2.4. Roll a die 4 times. What is the probability that you get dierent numbers?
At least at the beginning, you should divide every solution into the following three steps:
Step 1: Identify the set of equally likely outcomes. In this case, this is the set of all ordered
4-tuples of numbers 1, . . . , 6. That is, (a, b, c, d) : a, b, c, d 1, . . . , 6.
Step 2: Compute the number of outcomes. In this case, it is therefore 6
4
.
Step 3: Compute the number of good outcomes. In this case it is 6 5 4 3. Why? We
have 6 options for the rst roll, 5 options for the second roll since its number must dier
from the number on the rst roll; 4 options for the third roll since its number must not
appear on the rst two rolls, etc. Note that the set of possible outcomes changes from
stage to stage (roll to roll in this case), but their number does not!
The answer then is
6543
6
4
=
5
18
0.2778.
Example 2.5. Let us now compute probabilities for de Meres games.
In Game 1, there are 4 rolls and he wins with at least one 6. The number of good events is
6
4
5
4
, as the number of bad events is 5
4
. Therefore
P(win) = 1
_
5
6
_
4
0.5177.
In Game 2, there are 24 rolls of two dice and he wins by at least one pair of 6s rolled. The
number of outcomes is 36
24
, the number of bad ones is 35
24
, thus the number of good outcomes
equals 36
24
35
24
. Therefore
P(win) = 1
_
35
36
_
24
0.4914.
Chevalier de Mere overcounted the good outcomes in both cases. His count 4 6
3
in Game
1 selects a die with a 6 and arbitrary numbers for other dice. However, many outcomes have
more than one six and are hence counted more than once.
One should also note that both probabilities are barely dierent from 1/2, so de Mere was
gambling a lot to be able to notice the dierence.
2 COMBINATORICS 5
Permutations
Assume you have n objects. The number of ways to ll n ordered slots with them is
n (n 1) . . . 2 1 = n!,
while the number of ways to ll k n ordered slots is
n(n 1) . . . (n k + 1) =
n!
(n k)!
.
Example 2.6. Shue a deck of cards.
P(top card is an Ace) =
1
13
=
451!
52!
.
P(all cards of the same suit end up next to each other) =
4!(13!)
4
52!
4.510
28
. This event
never happens in practice.
P(hearts are together) =
40!13!
52!
= 6 10
11
.
To compute the last probability, for example, collect all hearts into a block; a good event is
specied by ordering 40 items (the block of hearts plus 39 other cards) and ordering the hearts
within their block.
Before we go on to further examples, let us agree that when the text says without further
elaboration, that a random choice is made, this means that all available choices are equally
likely. Also, in the next problem (and in statistics in general) sampling with replacement refers
to choosing, at random, an object from a population, noting its properties, putting the object
back into the population, and then repeating. Sampling without replacement omits the putting
back part.
Example 2.7. A bag has 6 pieces of paper, each with one of the letters, E, E, P, P, P, and
R, on it. Pull 6 pieces at random out of the bag (1) without, and (2) with replacement. What
is the probability that these pieces, in order, spell PEPPER?
There are two problems to solve. For sampling without replacement:
1. An outcome is an ordering of the pieces of paper E
1
E
2
P
1
P
2
P
3
R.
2. The number of outcomes thus is 6!.
3. The number of good outcomes is 3!2!.
The probability is
3!2!
6!
=
1
60
.
For sampling with replacement, the answer is
3
3
2
2
6
6
=
1
26
3
, quite a lot smaller.
2 COMBINATORICS 6
Example 2.8. Sit 3 men and 3 women at random (1) in a row of chairs and (2) around a table.
Compute P(all women sit together). In case (2), also compute P(men and women alternate).
In case (1), the answer is
4!3!
6!
=
1
5
.
For case (2), pick a man, say John Smith, and sit him rst. Then, we reduce to a row
problem with 5! outcomes; the number of good outcomes is 3! 3!. The answer is
3
10
. For the
last question, the seats for the men and women are xed after John Smith takes his seat and so
the answer is
3!2!
5!
=
1
10
.
Example 2.9. A group consists of 3 Norwegians, 4 Swedes, and 5 Finns, and they sit at random
around a table. What is the probability that all groups end up sitting together?
The answer is
3!4!5!2!
11!
. Pick, say, a Norwegian (Arne) and sit him down. Here is how you
count the good events. There are 3! choices for ordering the group of Norwegians (and then sit
them down to one of both sides of Arne, depending on the ordering). Then, there are 4! choices
for arranging the Swedes and 5! choices for arranging the Finns. Finally, there are 2! choices to
order the two blocks of Swedes and Finns.
Combinations
Let
_
n
k
_
be the number of dierent subsets with k elements of a set with n elements. Then,
_
n
k
_
=
n(n 1) . . . (n k + 1)
k!
=
n!
k!(n k)!
To understand why the above is true, rst choose a subset, then order its elements in a row
to ll k ordered slots with elements from the set with n objects. Then, distinct choices of a
subset and its ordering will end up as distinct orderings. Therefore,
_
n
k
_
k! = n(n 1) . . . (n k + 1).
We call
_
n
k
_
= n choose k or a binomial coecient (as it appears in the binomial theorem:
(x +y)
n
=

n
k=0
_
n
k
_
x
k
y
nk
). Also, note that
_
n
0
_
=
_
n
n
_
= 1 and
_
n
k
_
=
_
n
n k
_
.
The multinomial coecients are more general and are dened next.
2 COMBINATORICS 7
The number of ways to divide a set of n elements into r (distinguishable) subsets of
n
1
, n
2
, . . . , n
r
elements, where n
1
+. . . +n
r
= n, is denoted by
_
n
n
1
...n
r
_
and
_
n
n
1
. . . n
r
_
=
_
n
n
1
__
n n
1
n
2
__
n n
1
n
2
n
3
_
. . .
_
n n
1
. . . n
r1
n
r
_
=
n!
n
1
!n
2
! . . . n
r
!
To understand the slightly confusing word distinguishable, just think of painting n
1
elements
red, then n
2
dierent elements blue, etc. These colors distinguish among the dierent subsets.
Example 2.10. A fair coin is tossed 10 times. What is the probability that we get exactly 5
Heads?
P(exactly 5 Heads) =
_
10
5
_
2
10
0.2461,
as one needs to choose the position of the ve heads among 10 slots to x a good outcome.
Example 2.11. We have a bag that contains 100 balls, 50 of them red and 50 blue. Select 5
balls at random. What is the probability that 3 are blue and 2 are red?
The number of outcomes is
_
100
5
_
and all of them are equally likely, which is a reasonable
interpretation of select 5 balls at random. The answer is
P(3 are blue and 2 are red) =
_
50
3
__
50
2
_
_
100
5
_ 0.3189
Why should this probability be less than a half? The probability that 3 are blue and 2 are
red is equal to the probability of 3 are red and 2 are blue and they cannot both exceed
1
2
, as
their sum cannot be more than 1. It cannot be exactly
1
2
either, because other possibilities (such
as all 5 chosen balls red) have probability greater than 0.
Example 2.12. Here we return to Example 1.1 and solve it more slowly. Shue a standard
deck of 52 cards and deal 13 cards to each of the 4 players.
What is the probability that each player gets an Ace? We will solve this problem in two
ways to emphasize that you often have a choice in your set of equally likely outcomes.
The rst way uses an outcome to be an ordering of 52 cards:
1. There are 52! equally likely outcomes.
2. Let the rst 13 cards go to the rst player, the second 13 cards to the second player,
etc. Pick a slot within each of the four segments of 13 slots for an Ace. There are 13
4
possibilities to choose these four slots for the Aces.
3. The number of choices to ll these four positions with (four dierent) Aces is 4!.
4. Order the rest of the cards in 48! ways.
2 COMBINATORICS 8
The probability, then, is
13
4
4!48!
52!
.
The second way, via a small leap of faith, assumes that each set of the four positions of the
four Aces among the 52 shued cards is equally likely. You may choose to believe this intuitive
fact or try to write down a formal proof: the number of permutations that result in a given set
F of four positions is independent of F. Then:
1. The outcomes are the positions of the 4 Aces among the 52 slots for the shued cards of
the deck.
2. The number of outcomes is
_
52
4
_
.
3. The number of good outcomes is 13
4
, as we need to choose one slot among 13 cards that
go to the rst player, etc.
13
4
(
52
4
)
, which agrees with the number we obtained the rst way.
Let us also compute P(one person has all four Aces). Doing the problem the second way,
we get
1. The number of outcomes is
_
52
4
_
.
2. To x a good outcome, pick one player (
_
4
1
_
choices) and pick four slots for the Aces for
that player (
_
13
4
_
choices).
The answer, then, is
(
4
1
)(
13
4
)
(
52
4
)
= 0.0106, a lot smaller than the probability of each player getting
an Ace.
Example 2.13. Roll a die 12 times. P(each number appears exactly twice)?
1. An outcome consists of lling each of the 12 slots (for the 12 rolls) with an integer between
1 and 6 (the outcome of the roll).
2. The number of outcomes, therefore, is 6
12
.
3. To x a good outcome, pick two slots for 1, then pick two slots for 2, etc., with
_
12
2
__
10
2
_
. . .
_
2
2
_
choices.
(
12
2
)(
10
2
)...(
2
2
)
6
12
.
What is P(1 appears exactly 3 times, 2 appears exactly 2 times)?
To x a good outcome now, pick three slots for 1, two slots for 2, and ll the remaining 7
slots by numbers 3, . . . , 6. The number of choices is
_
12
3
__
9
2
_
4
7
and the answer is
(
12
3
)(
9
2
)4
7
6
12
.
Example 2.14. We have 14 rooms and 4 colors, white, blue, green, and yellow. Each room is
painted at random with one of the four colors. There are 4
14
equally likely outcomes, so, for
2 COMBINATORICS 9
example,
P(5 white, 4 blue, 3 green, 2 yellow rooms) =
_
14
5
__
9
4
__
5
3
__
2
2
_
4
14
.
Example 2.15. A middle row on a plane seats 7 people. Three of them order chicken (C) and
the remaining four pasta (P). The ight attendant returns with the meals, but has forgotten
who ordered what and discovers that they are all asleep, so she puts the meals in front of them
at random. What is the probability that they all receive correct meals?
A reformulation makes the problem clearer: we are interested in P(3 people who ordered C
get C). Let us label the people 1, . . . , 7 and assume that 1, 2, and 3 ordered C. The outcome
is a selection of 3 people from the 7 who receive C, the number of them is
_
7
3
_
, and there is a
single good outcome. So, the answer is
1
(
7
3
)
=
1
35
. Similarly,
P(no one who ordered C gets C) =
_
4
3
_
_
7
3
_ =
4
35
,
P(a single person who ordered C gets C) =
3
_
4
2
_
_
7
3
_ =
18
35
,
P(two persons who ordered C get C) =
_
3
2
_
4
_
7
3
_ =
12
35
.
Problems
1. A California licence plate consists of a sequence of seven symbols: number, letter, letter,
letter, number, number, number, where a letter is any one of 26 letters and a number is one
among 0, 1, . . . , 9. Assume that all licence plates are equally likely. (a) What is the probability
that all symbols are dierent? (b) What is the probability that all symbols are dierent and the
rst number is the largest among the numbers?
2. A tennis tournament has 2n participants, n Swedes and n Norwegians. First, n people are
chosen at random from the 2n (with no regard to nationality) and then paired randomly with
the other n people. Each pair proceeds to play one match. An outcome is a set of n (ordered)
pairs, giving the winner and the loser in each of the n matches. (a) Determine the number of
outcomes. (b) What do you need to assume to conclude that all outcomes are equally likely?
(c) Under this assumption, compute the probability that all Swedes are the winners.
3. A group of 18 Scandinavians consists of 5 Norwegians, 6 Swedes, and 7 Finns. They are seated
at random around a table. Compute the following probabilities: (a) that all the Norwegians
sit together, (b) that all the Norwegians and all the Swedes sit together, and (c) that all the
Norwegians, all the Swedes, and all the Finns sit together.
2 COMBINATORICS 10
4. A bag contains 80 balls numbered 1, . . . , 80. Before the game starts, you choose 10 dierent
numbers from amongst 1, . . . , 80 and write them on a piece of paper. Then 20 balls are selected
(without replacement) out of the bag at random. (a) What is the probability that all your
numbers are selected? (b) What is the probability that none of your numbers is selected? (c)
What is the probability that exactly 4 of your numbers are selected?
5. A full deck of 52 cards contains 13 hearts. Pick 8 cards from the deck at random (a) without
replacement and (b) with replacement. In each case compute the probability that you get no
hearts.
Solutions to problems
1. (a)
10987262524
10
4
26
3
, (b) the answer in (a) times
1
4
.
2. (a) Divide into two groups (winners and losers), then pair them:
_
2n
n
_
n!. Alternatively, pair
the rst player, then the next available player, etc., and, then, at the end choose the winners
and the losers: (2n 1)(2n 3) 3 1 2
n
. (Of course, these two expressions are the same.)
(b) All players are of equal strength, equally likely to win or lose any match against any other
player. (c) The number of good events is n!, the choice of a Norwegian paired with each Swede.
3. (a)
13!5!
17!
, (b)
8!5!6!
17!
, (c)
2!7!6!5!
17!
.
4. (a)
(
70
10
)
(
80
20
)
, (b)
(
70
20
)
(
80
20
)
, (c)
(
10
4
)(
70
16
)
(
80
20
)
.
5. (a)
(
39
8
)
(
52
8
)
, (b)
_
3
4
_
8
.
3 AXIOMS OF PROBABILITY 11
3 Axioms of Probability
The question here is: how can we mathematically dene a random experiment? What we have
are outcomes (which tell you exactly what happens), events (sets containing certain outcomes),
and probability (which attaches to every event the likelihood that it happens). We need to
agree on which properties these objects must have in order to compute with them and develop
a theory.
When we have nitely many equally likely outcomes all is clear and we have already seen
many examples. However, as is common in mathematics, innite sets are much harder to deal
with. For example, we will soon see what it means to choose a random point within a unit circle.
On the other hand, we will also see that there is no way to choose at random a positive integer
remember that at random means all choices are equally likely, unless otherwise specied.
Finally, a probability space is a triple (, T, P). The rst object is an arbitrary set,
repesenting the set of outcomes, sometimes called the sample space.
The second object T is a collection of events, that is, a set of subsets of . Therefore, an
event A T is necessarily a subset of . Can we just say that each A is an event? In this
course you can assume so without worry, although there are good reasons for not assuming so
in general ! I will give the denition of what properties T needs to satisfy, but this is only for
illustration and you should take a course in measure theory to understand what is really going
on. Namely, T needs to be a -algebra, which means (1) T, (2) A T = A
c
T, and
(3) A
1
, A
2
, T =
i=1
A
i
T.
What is important is that you can take the complement A
c
of an event A (i.e., A
c
happens
when A does not happen), unions of two or more events (i.e., A
1
A
2
happens when either A
1
or A
2
happens), and intersections of two or more events (i.e., A
1
A
2
happens when both A
1
and A
2
happen). We call events A
1
, A
2
, . . . pairwise disjoint if A
i
A
j
= if i ,= j that is,
at most one of these events can occur.
Finally, the probability P is a number attached to every event A and satises the following
three axioms:
Axiom 1. For every event A, P(A) 0.
Axiom 2. P() = 1.
Axiom 3. If A
1
, A
2
, . . . is a sequence of pairwise disjoint events, then
P(
_
i=1
A
i
) =
i=1
P(A
i
).
Whenever we have an abstract denition such as this one, the rst thing to do is to look for
examples. Here are some.
Example 3.1. = 1, 2, 3, 4, 5, 6,
P(A) =
(number of elements in A)
6
.
The random experiment here is rolling a fair die. Clearly, this can be generalized to any nite
set with equally likely outcomes.
Example 3.2. = 1, 2, . . . and you have numbers p
1
, p
2
, . . . 0 with p
1
+p
2
+. . . = 1. For
any A ,
P(A) =
iA
p
i
.
For example, toss a fair coin repeatedly until the rst Heads. Your outcome is the number of
tosses. Here, p
i
=
1
2
i
.
Note that p
i
cannot be chosen to be equal, as you cannot make the sum of innitely many
equal numbers to be 1.
Example 3.3. Pick a point from inside the unit circle centered at the origin. Here, = (x, y) :
x
2
+y
2
< 1 and
P(A) =
(area of A)
.
It is important to observe that if A is a singleton (a set whose element is a single point), then
P(A) = 0. This means that we cannot attach the probability to outcomes you hit a single
point (or even a line) with probability 0, but a fatter set with positive area you hit with
positive probability.
Another important theoretical remark: this is a case where A cannot be an arbitrary subset
of the circle for some sets area cannot be dened!
Consequences of the axioms
(C0) P() = 0.
Proof . In Axiom 3, take all sets to be .
(C1) If A
1
A
2
= , then P(A
1
A
2
) = P(A
1
) +P(A
2
).
Proof . In Axiom 3, take all sets other than rst two to be .
(C2)
P(A
c
) = 1 P(A).
Proof . Apply (C1) to A
1
= A, A
2
= A
c
.
(C3) 0 P(A) 1.
Proof . Use that P(A
c
) 0 in (C2).
(C4) If A B, P(B) = P(A) +P(B A) P(A).
Proof . Use (C1) for A
1
= A and A
2
= B A.
(C5) P(A B) = P(A) +P(B) P(A B).
Proof . Let P(A B) = p
1
, P(AB) = p
12
and P(B A) = p
2
, and note that A B, AB, and
BA are pairwise disjoint. Then P(A) = p
1
+p
12
, P(B) = p
2
+p
12
, and P(AB) = p
1
+p
2
+p
12
.
(C6)
P(A
1
A
2
A
3
) = P(A
1
) +P(A
2
) +P(A
3
)
P(A
1
A
2
) P(A
1
A
3
) P(A
2
A
2
)
+P(A
1
A
2
A
3
)
and more generally
P(A
1
A
n
) =
n
i=1
P(A
i
)
1i<jn
P(A
i
A
j
) +
1i<j<kn
P(A
i
A
j
A
k
) +. . .
+ (1)
n1
P(A
1
A
n
).
This is called the inclusion-exclusion formula and is commonly used when it is easier to compute
probabilities of intersections than of unions.
Proof . We prove this only for n = 3. Let p
1
= P(A
1
A
c
2
A
c
3
), p
2
= P(A
c
1
A
2
A
c
3
),
p
3
= P(A
c
1
A
c
2
A
3
), p
12
= P(A
1
A
2
A
c
3
), p
13
= P(A
1
A
c
2
A
3
), p
23
= P(A
c
1
A
2
A
3
),
and p
123
= P(A
1
A
2
A
3
). Again, note that all sets are pairwise disjoint and that the right
hand side of (6) is
(p
1
+p
12
+p
13
+p
123
) + (p
2
+p
12
+p
23
+p
123
) + (p
3
+p
13
+p
23
+p
123
)
(p
12
+p
123
) (p
13
+p
123
) (p
23
+p
123
)
+p
123
= p
1
+p
2
+p
3
+p
12
+p
13
+p
23
+p
123
= P(A
1
A
2
A
3
).
Example 3.4. Pick an integer in [1, 1000] at random. Compute the probability that it is
divisible neither by 12 nor by 15.
The sample space consists of the 1000 integers between 1 and 1000 and let A
r
be the subset
consisting of integers divisible by r. The cardinality of A
r
is 1000/r|. Another simple fact
is that A
r
A
s
= A
lcm(r,s)
, where lcm stands for the least common multiple. Our probability
equals
1 P(A
12
A
15
) = 1 P(A
12
) P(A
15
) +P(A
12
A
15
)
= 1 P(A
12
) P(A
15
) +P(A
60
)
= 1
83
1000

66
1000
+
16
1000
= 0.867.
Example 3.5. Sit 3 men and 4 women at random in a row. What is the probability that either
all the men or all the women end up sitting together?
Here, A
1
= all women sit together, A
2
= all men sit together, A
1
A
2
= both women
and men sit together, and so the answer is
P(A
1
A
2
) = P(A
1
) +P(A
2
) P(A
1
A
2
) =
4! 4!
7!
+
5! 3!
7!

2! 3! 4!
7!
.
Example 3.6. A group of 3 Norwegians, 4 Swedes, and 5 Finns is seated at random around a
table. Compute the probability that at least one of the three groups ends up sitting together.
Dene A
N
= Norwegians sit together and similarly A
S
, A
F
. We have
P(A
N
) =
3! 9!
11!
, P(A
S
) =
4! 8!
11!
, P(A
N
) =
5! 7!
11!
,
P(A
N
A
S
) =
3! 4! 6!
11!
, P(A
N
A
F
) =
3! 5! 5!
11!
, P(A
S
A
F
) =
4! 5! 4!
11!
,
P(A
N
A
S
A
F
) =
3! 4! 5! 2!
11!
.
Therefore,
P(A
N
A
S
A
F
) =
3! 9! + 4! 8! + 5! 7! 3! 4! 6! 3! 5! 5! 4! 5! 4! + 3! 4! 5! 2!
11!
.
Example 3.7. Matching problem. A large company with n employees has a scheme according
to which each employee buys a Christmas gift and the gifts are then distributed at random to
the employees. What is the probability that someone gets his or her own gift?
Note that this is dierent from asking, assuming that you are one of the employees, for the
probability that you get your own gift, which is
1
n
.
Let A
i
= ith employee gets his or her own gift. Then, what we are looking for is
P(
n
_
i=1
A
i
).
We have
P(A
i
) =
1
n
(for all i),
P(A
i
A
j
) =
(n 2)!
n!
=
1
n(n 1)
(for all i < j),
P(A
i
A
j
A
k
) =
(n 3)!
n!
=
1
n(n 1)(n 2)
(for all i < j < k),
. . .
P(A
1
A
n
) =
1
n!
.
Therefore, our answer is
n
1
n

_
n
2
_
1
n(n 1)
+
_
n
3
_
1
n(n 1)(n 2)
. . . + (1)
n1
1
n!
= 1
1
2!
+
1
3!
+. . . + (1)
n1
1
n!
1
1
e
0.6321 (as n ).
Example 3.8. Birthday Problem. Assume that there are k people in the room. What is
the probability that there are two who share a birthday? We will ignore leap years, assume
all birthdays are equally likely, and generalize the problem a little: from n possible birthdays,
sample k times with replacement.
P(a shared birthday) = 1 P(no shared birthdays) = 1
n (n 1) (n k + 1)
n
k
.
When n = 365, the lowest k for which the above exceeds 0.5 is, famously, k = 23. Some values
are given in the following table.
k prob. for n = 365
10 0.1169
23 0.5073
41 0.9032
57 0.9901
70 0.9992
Occurences of this problem are quite common in various contexts, so we give another example.
Each day, the Massachusetts lottery chooses a four digit number at random, with leading 0s
allowed. Thus, this is sampling with replacement from among n = 10, 000 choices each day. On
February 6, 1978, the Boston Evening Globe reported that
During [the lotterys] 22 months existence [...], no winning number has ever been
repeated. [David] Hughes, the expert [and a lottery ocial] doesnt expect to see
duplicate winners until about half of the 10, 000 possibilities have been exhausted.
What would an informed reader make of this? Assuming k = 660 days, the probability of no
repetition works out to be about 2.19 10
10
, making it a remarkably improbable event. What
happened was that Mr. Hughes, not understanding the Birthday Problem, did not check for
repetitions, condent that there would not be any. Apologetic lottery ocials announced later
that there indeed were repetitions.
Example 3.9. Coupon Collector Problem. Within the context of the previous problem, assume
that k n and compute P(all n birthdays are represented).
More often, this is described in terms of cereal boxes, each of which contains one of n dierent
cards (coupons), chosen at random. If you buy k boxes, what is the probability that you have
a complete collection?
When k = n, our probability is
n!
n
n
. More generally, let
A
i
= ith birthday is missing.
Then, we need to compute
1 P(
n
_
i=1
A
i
).
Now,
P(A
i
) =
(n 1)
k
n
k
(for all i)
P(A
i
A
j
) =
(n 2)
k
n
k
(for all i < j)
. . .
P(A
1
A
n
) = 0
and our answer is
1 n
_
n 1
n
_
k
+
_
n
2
__
n 2
n
_
k
. . . + (1)
n1
_
n
n 1
__
1
n
_
k
=
n
i=0
_
n
i
_
(1)
i
_
1
i
n
_
k
.
This must be
n!
n
n
for k = n, and 0 when k < n, neither of which is obvious from the formula.
Neither will you, for large n, get anything close to the correct numbers when k n if you try to
compute the probabilities by computer, due to the very large size of summands with alternating
signs and the resulting rounding errors. We will return to this problem later for a much more
ecient computational method, but some numbers are in the two tables below. Another remark
for those who know a lot of combinatorics: you will perhaps notice that the above probability
is
n!
n
k
S
k,n
, where S
k,n
is the Stirling number of the second kind.
k prob. for n = 6
13 0.5139
23 0.9108
36 0.9915
k prob. for n = 365
1607 0.0101
1854 0.1003
2287 0.5004
2972 0.9002
3828 0.9900
4669 0.9990
More examples with combinatorial avor
We will now do more problems which would rather belong to the previous chapter, but are a
little harder, so we do them here instead.
Example 3.10. Roll a die 12 times. Compute the probability that a number occurs 6 times
and two other numbers occur three times each.
The number of outcomes is 6
12
. To count the number of good outcomes:
1. Pick the number that occurs 6 times:
_
6
1
_
= 6 choices.
2. Pick the two numbers that occur 3 times each:
_
5
2
_
choices.
3. Pick slots (rolls) for the number that occurs 6 times:
_
12
6
_
choices.
4. Pick slots for one of the numbers that occur 3 times:
_
6
3
_
choices.
Therefore, our probability is
6(
5
2
)(
12
6
)(
6
3
)
6
12
.
Example 3.11. You have 10 pairs of socks in the closet. Pick 8 socks at random. For every i,
compute the probability that you get i complete pairs of socks.
There are
_
20
8
_
outcomes. To count the number of good outcomes:
1. Pick i pairs of socks from the 10:
_
10
i
_
choices.
2. Pick pairs which are represented by a single sock:
_
10i
82i
_
choices.
3. Pick a sock from each of the latter pairs: 2
82i
choices.
Therefore, our probability is
2
82i
(
10i
82i
)(
10
i
)
(
20
8
)
.
Example 3.12. Poker Hands. In the denitions, the word value refers to A, K, Q, J, 10, 9,
8, 7, 6, 5, 4, 3, 2. This sequence orders the cards in descending consecutive values, with one
exception: an Ace may be regarded as 1 for the purposes of making a straight (but note that,
for example, K, A, 1, 2, 3 is not a valid straight sequence A can only begin or end a straight).
From the lowest to the highest, here are the hands:
(a) one pair: two cards of the same value plus 3 cards with dierent values
J J 9 Q 4
(b) two pairs: two pairs plus another card of dierent value
J J 9 9 3
(c) three of a kind: three cards of the same value plus two with dierent values
Q Q Q 9 3
(d) straight: ve cards with consecutive values
5 4 3 2 A
(e) ush: ve cards of the same suit
K 9 7 6 3
(f) full house: a three of a kind and a pair
J J J 3 3
(g) four of a kind: four cards of the same value
K K K K 10
(e) straight ush: ve cards of the same suit with consecutive values
A K Q J 10
Here are the probabilities:
hand no. combinations approx. prob.
one pair 13
_
12
3
_
_
4
2
_
4
3
0.422569
two pairs
_
13
2
_
11
_
4
2
_
_
4
2
_
4 0.047539
three of a kind 13
_
12
2
_
_
4
3
_
4
2
0.021128
straight 10 4
5
0.003940
ush 4
_
13
5
_
0.001981
full house 13 12
_
4
3
_
_
4
2
_
0.001441
four of a kind 13 12 4 0.000240
straight ush 10 4 0.000015
Note that the probabilities of a straight and a ush above include the possibility of a straight
ush.
Let us see how some of these are computed. First, the number of all outcomes is
_
52
5
_
=
2, 598, 960. Then, for example, for the three of a kind, the number of good outcomes may be
obtained by listing the number of choices:
1. Choose a value for the triple: 13.
2. Choose the values of other two cards:
_
12
2
_
.
3. Pick three cards from the four of the same chosen value:
_
4
3
_
.
4. Pick a card (i.e., the suit) from each of the two remaining values: 4
2
.
One could do the same for one pair:
1. Pick a number for the pair: 13.
2. Pick the other three numbers:
_
12
3
_
3. Pick two cards from the value of the pair:
_
4
2
_
.
4. Pick the suits of the other three values: 4
3
And for the ush:
1. Pick a suit: 4.
2. Pick ve numbers:
_
13
5
_
Our nal worked out case is straight ush:
1. Pick a suit: 4.
2. Pick the beginning number: 10.
We end this example by computing the probability of not getting any hand listed above,
that is,
P(nothing) = P(all cards with dierent values) P(straight or ush)
=
_
13
5
_
4
5
_
52
5
_ (P(straight) +P(ush) P(straight ush))
=
_
13
5
_
4
5
10 4
5
4
_
13
5
_
+ 40
_
52
5
_
0.5012.
Example 3.13. Assume that 20 Scandinavians, 10 Finns, and 10 Danes, are to be distributed
at random into 10 rooms, 2 per room. What is the probability that exactly 2i rooms are mixed,
i = 0, . . . 5?
This is an example when careful thinking about what the outcomes should be really pays
o. Consider the following model for distributing the Scandinavians into rooms. First arrange
them at random into a row of 20 slots S1, S2, . . . , S20. Assume that room 1 takes people in
slots S1, S2, so let us call these two slots R1. Similarly, room 2 takes people in slots S3, S4, so
let us call these two slots R2, etc.
Now, it is clear that we only need to keep track of the distribution of 10 Ds into the 20 slots,
corresponding to the positions of the 10 Danes. Any such distribution constitutes an outcome
and they are equally likely. Their number is
_
20
10
_
.
To get 2i (for example, 4) mixed rooms, start by choosing 2i (ex., 4) out of the 10 rooms
which are going to be mixed; there are
_
10
2i
_
choices. You also need to decide into which slot in
each of the 2i chosen mixed rooms the D goes, for 2
2i
choices.
Once these two choices are made, you still have 10 2i (ex., 6) Ds to distribute into 5 i
(ex., 3) rooms, as there are two Ds in each of these rooms. For this, you need to choose 5 i
(ex., 3) rooms from the remaining 10 2i (ex., 6), for
_
102i
5i
_
choices, and this choice xes a
good outcome.
The nal answer, therefore, is
_
10
2i
_
2
2i
_
102i
5i
_
_
20
10
_ .
Problems
1. Roll a single die 10 times. Compute the following probabilities: (a) that you get at least one
6; (b) that you get at least one 6 and at least one 5; (c) that you get three 1s, two 2s, and ve
3s.
2. Three married couples take seats around a table at random. Compute P(no wife sits next to
her husband).
3. A group of 20 Scandinavians consists of 7 Swedes, 3 Finns, and 10 Norwegians. A committee
of ve people is chosen at random from this group. What is the probability that at least one of
the three nations is not represented on the committee?
4. Choose each digit of a 5 digit number at random from digits 1, . . . 9. Compute the probability
that no digit appears more than twice.
5. Roll a fair die 10 times. (a) Compute the probability that at least one number occurs exactly
6 times. (b) Compute the probability that at least one number occurs exactly once.
6. A lottery ticket consists of two rows, each containing 3 numbers from 1, 2, . . . , 50. The
drawing consists of choosing 5 dierent numbers from 1, 2, . . . , 50 at random. A ticket wins if
its rst row contains at least two of the numbers drawn and its second row contains at least two
of the numbers drawn. The four examples below represent the four types of tickets:
Ticket 1 Ticket 2 Ticket 3 Ticket 4
1 2 3
4 5 6
1 2 3
1 2 3
1 2 3
2 3 4
1 2 3
3 4 5
For example, if the numbers 1, 3, 5, 6, 17 are drawn, then Ticket 1, Ticket 2, and Ticket 4 all
win, while Ticket 3 loses. Compute the winning probabilities for each of the four tickets.
1. (a) 1 (5/6)
10
. (b) 1 2 (5/6)
10
+ (2/3)
10
. (c)
_
10
3
__
7
2
_
6
10
.
2. The complement is the union of the three events A
i
= couple i sits together, i = 1, 2, 3.
Moreover,
P(A
1
) =
2
5
= P(A
2
) = P(A
3
),
P(A
1
A
2
) = P(A
1
A
3
) = P(A
2
A
3
) =
3! 2! 2!
5!
=
1
5
,
P(A
1
A
2
A
3
) =
2! 2! 2! 2!
5!
=
2
15
.
For P(A
1
A
2
), for example, pick a seat for husband
3
. In the remaining row of 5 seats, pick the
ordering for couple
1
, couple
2
, and wife
3
, then the ordering of seats within each of couple
1
and
couple
2
. Now, by inclusion-exclusion,
P(A
1
A
2
A
3
) = 3
2
5
3
1
5
+
2
15
=
11
15
,
and our answer is
4
15
.
3. Let A
1
= the event that Swedes are not represented, A
2
= the event that Finns are not
represented, and A
3
= the event that Norwegians are not represented.
P(A
1
A
2
A
3
) = P(A
1
) +P(A
2
) +P(A
3
) P(A
1
A
2
) P(A
1
A
3
) P(A
2
A
3
)
+P(A
1
A
2
A
3
)
=
1
_
20
5
_
__
13
5
_
+
_
17
5
_
+
_
10
5
_
_
10
5
_
0
_
7
5
_
+ 0
_
4. The number of bad events is 9
_
5
3
_
8
2
+ 9
_
5
4
_
8 + 9. The rst term is the number of
numbers in which a digit appears 3 times, but no digit appears 4 times: choose a digit, choose 3
positions lled by it, then ll the remaining position. The second term is the number of numbers
in which a digit appears 4 times, but no digit appears 5 times, and the last term is the number
of numbers in which a digit appears 5 times. The answer then is
1
9
_
5
3
_
8
2
+ 9
_
5
4
_
8 + 9
9
5
.
5. (a) Let A
i
be the event that the number i appears exactly 6 times. As A
i
are pairwise
disjoint,
P(A
1
A
2
A
3
A
4
A
5
A
6
) = 6
_
10
6
_
5
4
6
10
.
(b) (a) Now, A
i
is the event that the number i appears exactly once. By inclusion-exclusion,
P(A
1
A
2
A
3
A
4
A
5
A
6
)
= 6P(A
1
)
_
6
2
_
P(A
1
A
2
)
+
_
6
3
_
P(A
1
A
2
A
3
)
_
6
4
_
P(A
1
A
2
A
3
A
4
)
+
_
6
5
_
P(A
1
A
2
A
3
A
4
A
5
)
_
6
6
_
P(A
1
A
2
A
3
A
4
A
5
A
6
)
= 6 10
5
9
6
10
_
6
2
_
10 9
4
8
6
10
+
_
6
3
_
10 9 8
3
7
6
10
_
6
4
_
10 9 8 7
2
6
6
10
+
_
6
5
_
10 9 8 7 6
1
6
10
0.
6. Below, a hit is shorthand for a chosen number.
P(ticket 1 wins) = P(two hits on each line) +P(two hits on one line, three on the other)
=
3 3 44 + 2 3
_
50
5
_ =
402
_
50
5
_,
and
P(ticket 2 wins) = P(two hits among 1, 2, 3) +P(three hits among 1, 2, 3)
=
3
_
47
3
_
+
_
47
2
_
_
50
5
_ =
49726
_
50
5
_ ,
and
P(ticket 3 wins) = P(2, 3 both hit) +P(1, 4 both hit and one of 2, 3 hit)
=
_
48
3
_
+ 2
_
46
2
_
_
50
5
_ =
19366
_
50
5
_ ,
and, nally,
P(ticket 4 wins) = P(3 hit, at least one additional hit on each line) +P(1, 2, 4, 5 all hit)
=
(4
_
45
2
_
+ 4 45 + 1) + 45
_
50
5
_ =
4186
_
50
5
_
4 CONDITIONAL PROBABILITY AND INDEPENDENCE 24
4 Conditional Probability and Independence
Example 4.1. Assume that you have a bag with 11 cubes, 7 of which have a fuzzy surface and
4 are smooth. Out of the 7 fuzzy ones, 3 are red and 4 are blue; out of 4 smooth ones, 2 are red
and 2 are blue. So, there are 5 red and 6 blue cubes. Other than color and fuzziness, the cubes
have no other distinguishing characteristics.
You plan to pick a cube out of the bag at random, but forget to wear gloves. Before you
start your experiment, the probability that the selected cube is red is 5/11. Now, you reach into
the bag, grab a cube, and notice it is fuzzy (but you do not take it out or note its color in any
other way). Clearly, the probability should now change to 3/7!
Your experiment clearly has 11 outcomes. Consider the events R, B, F, S that the selected
cube is red, blue, fuzzy, and smooth, respectively. We observed that P(R) = 5/11. For the
probability of a red cube, conditioned on it being fuzzy, we do not have notation, so we introduce
it here: P(R[F) = 3/7. Note that this also equals
P(R F)
P(F)
=
P(the selected ball is red and fuzzy)
P(the selected ball is fuzzy)
.
This conveys the idea that with additional information the probability must be adjusted.
This is common in real life. Say bookies estimate your basketball teams chances of winning a
particular game to be 0.6, 24 hours before the game starts. Two hours before the game starts,
however, it becomes known that your teams star player is out with a sprained ankle. You
cannot expect that the bookies odds will remain the same and they change, say, to 0.3. Then,
the game starts and at half-time your team leads by 25 points. Again, the odds will change, say
to 0.8. Finally, when complete information (that is, the outcome of your experiment, the game
in this case) is known, all probabilities are trivial, 0 or 1.
For the general denition, take events A and B, and assume that P(B) > 0. The conditional
probability of A given B equals
P(A[B) =
P(A B)
P(B)
.
Example 4.2. Here is a question asked on Wall Street job interviews. (This is the original
formulation; the macabre tone is not unusual for such interviews.)
Lets play a game of Russian roulette. You are tied to your chair. Heres a gun, a revolver.
Heres the barrel of the gun, six chambers, all empty. Now watch me as I put two bullets into
the barrel, into two adjacent chambers. I close the barrel and spin it. I put a gun to your head
and pull the trigger. Click. Lucky you! Now Im going to pull the trigger one more time. Which
would you prefer: that I spin the barrel rst or that I just pull the trigger?
Assume that the barrel rotates clockwise after the hammer hits and is pulled back. You are
given the choice between an unconditional and a conditional probability of death. The former,
if the barrel is spun again, remains 1/3. The latter, if the trigger is pulled without the extra
spin, equals the probability that the hammer clicked on an empty slot, which is next to a bullet
in the counterclockwise direction, and equals 1/4.
For a xed condition B, and acting on events A, the conditional probability Q(A) = P(A[B)
satises the three axioms in Chapter 3. (This is routine to check and the reader who is more the-
oretically inclined might view it as a good exercise.) Thus, Q is another probability assignment
and all consequences of the axioms are valid for it.
Example 4.3. Toss two fair coins, blindfolded. Somebody tells you that you tossed at least
one Heads. What is the probability that both tosses are Heads?
Here A = both H, B = at least one H, and
P(A[B) =
P(A B)
P(B)
=
P(both H)
P(at least one H)
=
1
4
3
4
=
1
3
.
Example 4.4. Toss a coin 10 times. If you know (a) that exactly 7 Heads are tossed, (b) that
at least 7 Heads are tossed, what is the probability that your rst toss is Heads?
For (a),
P(rst toss H[exactly 7 Hs) =
_
9
6
_
1
2
10
_
10
7
_
1
2
10
=
7
10
.
Why is this not surprising? Conditioned on 7 Heads, they are equally likely to occur on any
given 7 tosses. If you choose 7 tosses out of 10 at random, the rst toss is included in your
choice with probability
7
10
.
For (b), the answer is, after canceling
1
2
10
,
_
9
6
_
+
_
9
7
_
+
_
9
8
_
+
_
9
9
_
_
10
7
_
+
_
10
8
_
+
_
10
9
_
+
_
10
10
_ =
65
88
0.7386.
Clearly, the answer should be a little larger than before, because this condition is more advan-
tageous for Heads.
Conditional probabilities are sometimes given, or can be easily determined, especially in
sequential random experiments. Then, we can use
P(A
1
A
2
) = P(A
1
)P(A
2
[A
1
),
P(A
1
A
2
A
3
) = P(A
1
)P(A
2
[A
1
)P(A
3
[A
1
A
2
),
etc.
Example 4.5. An urn contains 10 black and 10 white balls. Draw 3 (a) without replacement,
and (b) with replacement. What is the probability that all three are white?
We already know how to do part (a):
1. Number of outcomes:
_
20
3
_
.
2. Number of ways to select 3 balls out of 10 white ones:
_
10
3
_
.
Our probability is then
(
10
3
)
(
20
3
)
.
To do this problem another way, imagine drawing the balls sequentially. Then, we are
computing the probability of the intersection of the three events: P(1st ball is white, 2nd ball
is white, and 3rd ball is white). The relevant probabilities are:
1. P(1st ball is white) =
1
2
.
2. P(2nd ball is white[1st ball is white) =
9
19
.
3. P(3rd ball is white[1st two picked are white) =
8
18
.
Our probability is, then, the product
1
2

9
19

8
18
, which equals, as it must, what we obtained
before.
This approach is particularly easy in case (b), where the previous colors of the selected balls
do not aect the probabilities at subsequent stages. The answer, therefore, is
_
1
2
_
3
.
Theorem 4.1. First Bayes formula. Assume that F
1
, . . . , F
n
are pairwise disjoint and that
F
1
. . . F
n
= , that is, exactly one of them always happens. Then, for an event A,
P(A) = P(F
1
)P(A[F
1
)+P(F
2
)P(A[F
2
)+. . .+P(F
n
)P(A[F
n
) .
Proof.
P(F
1
)P(A[F
1
) +P(F
2
)P(A[F
2
) +. . . +P(F
n
)P(A[F
n
) = P(A F
1
) +. . . +P(A F
n
)
= P((A F
1
) . . . (A F
n
))
= P(A (F
1
. . . F
n
))
= P(A ) = P(A)
We call an instance of using this formula computing the probability by conditioning on
which of the events F
i
happens. The formula is useful in sequential experiments, when you
face dierent experimental conditions at the second stage depending on what happens at the
rst stage. Quite often, there are just two events F
i
, that is, an event F and its complement
F
c
, and we are thus conditioning on whether F happens or not.
Example 4.6. Flip a fair coin. If you toss Heads, roll 1 die. If you toss Tails, roll 2 dice.
Compute the probability that you roll exactly one 6.
Here you condition on the outcome of the coin toss, which could be Heads (event F) or Tails
(event F
c
). If A = exactly one 6, then P(A[F) =
1
6
, P(A[F
c
) =
25
36
, P(F) = P(F
c
) =
1
2
and
so
P(A) = P(F)P(A[F) +P(F
c
)P(A[F
c
) =
2
9
.
Example 4.7. Roll a die, then select at random, without replacement, as many cards from the
deck as the number shown on the die. What is the probability that you get at least one Ace?
Here F
i
= number shown on the die is i, for i = 1, . . . , 6. Clearly, P(F
i
) =
1
6
. If A is the
event that you get at least one Ace,
1. P(A[F
1
) =
1
13
,
2. In general, for i 1, P(A[F
i
) = 1
(
48
i
)
(
52
i
)
.
Therefore, by Bayes formula,
P(A) =
1
6
_
1
13
+ 1
_
48
2
_
_
52
2
_ + 1
_
48
3
_
_
52
3
_ + 1
_
48
4
_
_
52
4
_ + 1
_
48
5
_
_
52
5
_ + 1
_
48
6
_
_
52
6
_
_
.
Example 4.8. Coupon collector problem, revisited. As promised, we will develop a compu-
tationally much better formula than the one in Example 3.9. This will be another example of
conditioning, whereby you (1) reinterpret the problem as a sequential experiment and (2) use
Bayes formula with conditions F
i
being relevant events at the rst stage of the experiment.
Here is how it works in this example. Let p
k,r
be the probability that exactly r (out of a
total of n) birthdays are represented among k people; we call the event A. We will x n and let
k and r be variable. Note that p
k,n
is what we computed by the inclusion-exclusion formula.
At the rst stage you have k 1 people; then the kth person arrives on the scene. Let F
1
be the event that there are r birthdays represented among the k 1 people and let F
2
be the
event that there are r 1 birthdays represented among the k 1 people. Let F
3
be the event
that any other number of birthdays occurs with k 1 people. Clearly, P(A[F
3
) = 0, as the
newcomer contributes either 0 or 1 new birthdays. Moreover, P(A[F
1
) =
r
n
, the probability that
the newcomer duplicates one of the existing r birthdays, and P(A[F
2
) =
nr+1
n
, the probability
that the newcomer does not duplicate any of the existing r 1 birthdays. Therefore,
p
k,r
= P(A) = P(A[F
1
)P(F
1
) +P(A[F
2
)P(F
2
) =
r
n
p
k1,r
+
n r + 1
n
p
k1,r1
,
for k, r 1, and this, together with the boundary conditions
p
0,0
= 1,
p
k,r
= 0, for 0 k < r,
p
k,0
= 0, for k > 0,
makes the computation fast and precise.
Theorem 4.2. Second Bayes formula. Let F
1
, . . . , F
n
and A be as in Theorem 4.1. Then
P(F
j
[A) =
P(F
j
A)
P(A)
=
P(A[F
j
)P(F
j
)
P(A[F
1
)P(F
1
) +. . . +P(A[F
n
)P(F
n
)
.
An event F
j
is often called a hypothesis, P(F
j
) its prior probability, and P(F
j
[A) its posterior
probability.
Example 4.9. We have a fair coin and an unfair coin, which always comes out Heads. Choose
one at random, toss it twice. It comes out Heads both times. What is the probability that the
coin is fair?
The relevant events are F = fair coin, U = unfair coin, and B = both tosses H. Then
P(F) = P(U) =
1
2
(as each coin is chosen with equal probability). Moreover, P(B[F) =
1
4
, and
P(B[U) = 1. Our probability then is
1
2

1
4
1
2

1
4
+
1
2
1
=
1
5
.
Example 4.10. A factory has three machines, M
1
, M
2
and M
3
, that produce items (say,
lightbulbs). It is impossible to tell which machine produced a particular item, but some are
defective. Here are the known numbers:
machine proportion of items made prob. any made item is defective
M
1
0.2 0.001
M
2
0.3 0.002
M
3
0.5 0.003
You pick an item, test it, and nd it is defective. What is the probability that it was made
by machine M
2
?
The best way to think about this random experiment is as a two-stage procedure. First you
choose a machine with the probabilities given by the proportion. Then, that machine produces
an item, which you then proceed to test. (Indeed, this is the same as choosing the item from a
large number of them and testing it.)
Let D be the event that an item is defective and let M
i
also denote the event that the
item was made by machine i. Then, P(D[M
1
) = 0.001, P(D[M
2
) = 0.002, P(D[M
3
) = 0.003,
P(M
1
) = 0.2, P(M
2
) = 0.3, P(M
3
) = 0.5, and so
P(M
2
[D) =
0.002 0.3
0.001 0.2 + 0.002 0.3 + 0.003 0.5
0.26.
Example 4.11. Assume 10% of people have a certain disease. A test gives the correct diagnosis
with probability of 0.8; that is, if the person is sick, the test will be positive with probability 0.8,
but if the person is not sick, the test will be positive with probability 0.2. A random person from
the population has tested positive for the disease. What is the probability that he is actually
sick? (No, it is not 0.8!)
Let us dene the three relevant events: S = sick, H = healthy, T = tested positive.
Now, P(H) = 0.9, P(S) = 0.1, P(T[H) = 0.2 and P(T[S) = 0.8. We are interested in
P(S[T) =
P(T[S)P(S)
P(T[S)P(S) +P(T[H)P(H)
=
8
26
31%.
Note that the prior probability P(S) is very important! Without a very good idea about what
it is, a positive test result is dicult to evaluate: a positive test for HIV would mean something
very dierent for a random person as opposed to somebody who gets tested because of risky
behavior.
Example 4.12. O. J. Simpsons rst trial , 1995. The famous sports star and media personality
O. J. Simpson was on trial in Los Angeles for murder of his wife and her boyfriend. One of the
many issues was whether Simpsons history of spousal abuse could be presented by prosecution
at the trial; that is, whether this history was probative, i.e., it had some evidentiary value,
or whether it was merely prejudicial and should be excluded. Alan Dershowitz, a famous
professor of law at Harvard and a consultant for the defense, was claiming the latter, citing the
statistics that < 0.1% of men who abuse their wives end up killing them. As J. F. Merz and
J. C. Caulkins pointed out in the journal Chance (Vol. 8, 1995, pg. 14), this was the wrong
probability to look at!
We need to start with the fact that a woman is murdered. These numbered 4, 936 in 1992,
out of which 1, 430 were killed by partners. In other words, if we let
A = the (murdered) woman was abused by the partner,
M = the woman was murdered by the partner,
then we estimate the prior probabilities P(M) = 0.29, P(M
c
) = 0.71, and what we are interested
in is the posterior probability P(M[A). It was also commonly estimated at the time that
about 5% of the women had been physically abused by their husbands. Thus, we can say that
P(A[M
c
) = 0.05, as there is no reason to assume that a woman murdered by somebody else
was more or less likely to be abused by her partner. The nal number we need is P(A[M).
Dershowitz states that a considerable number of wife murderers had previously assaulted
them, although some did not. So, we will (conservatively) say that P(A[M) = 0.5. (The
two-stage experiment then is: choose a murdered woman at random; at the rst stage, she is
murdered by her partner, or not, with stated probabilities; at the second stage, she is among
the abused women, or not, with probabilities depending on the outcome of the rst stage.) By
Bayes formula,
P(M[A) =
P(M)P(A[M)
P(M)P(A[M) +P(M
c
)P(A[M
c
)
=
29
36.1
0.8.
The law literature studiously avoids quantifying concepts such as probative value and reasonable
doubt. Nevertheless, we can probably say that 80% is considerably too high, compared to the
prior probability of 29%, to use as a sole argument that the evidence is not probative.
Independence
Events A and B are independent if P(A B) = P(A)P(B) and dependent (or correlated)
otherwise.
Assuming that P(B) > 0, one could rewrite the condition for independence,
P(A[B) = P(A),
so the probability of A is unaected by knowledge that B occurred. Also, if A and B are
independent,
P(A B
c
) = P(A) P(A B) = P(A) P(A)P(B) = P(A)(1 P(B)) = P(A)P(B
c
),
so A and B
c
are also independent knowing that B has not occurred also has no inuence on
the probability of A. Another fact to notice immediately is that disjoint events with nonzero
probability cannot be independent: given that one of them happens, the other cannot happen
and thus its probability drops to zero.
Quite often, independence is an assumption and it is the most important concept in proba-
bility.
Example 4.13. Pick a random card from a full deck. Let A = card is an Ace and R =
card is red. Are A and R independent?
We have P(A) =
1
13
, P(R) =
1
2
, and, as there are two red Aces, P(A R) =
2
52
=
1
26
.
The two events are independent the proportion of aces among red cards is the same as the
proportion among all cards.
Now, pick two cards out of the deck sequentially without replacement. Are F = rst card
is an Ace and S = second card is an Ace independent?
Now P(F) = P(S) =
1
13
and P(S[F) =
3
51
, so they are not independent.
Example 4.14. Toss 2 fair coins and let F = Heads on 1st toss, S = Heads on 2nd toss.
These are independent. You will notice that here the independence is in fact an assumption.
How do we dene independence of more than two events? We say that events A
1
, A
2
, . . . , A
n
are independent if
P(A
i
1
. . . A
i
k
) = P(A
i
1
)P(A
i
2
) P(A
i
k
),
where 1 i
1
< i
2
< . . . < i
k
n are arbitrary indices. The occurrence of any combination
of events does not inuence the probability of others. Again, it can be shown that, in such a
collection of independent events, we can replace an A
i
by A
c
i
and the events remain independent.
Example 4.15. Roll a four sided fair die, that is, choose one of the numbers 1, 2, 3, 4 at
random. Let A = 1, 2, B = 1, 3, C = 1, 4. Check that these are pairwise independent
(each pair is independent), but not independent.
Indeed, P(A) = P(B) = P(C) =
1
2
and P(AB) = P(AC) = P(BC) =
1
4
and pairwise
independence follows. However,
P(A B C) =
1
4
,=
1
8
.
The simple reason for lack of independence is
A B C,
so we have complete information on the occurrence of C as soon as we know that A and B both
happen.
Example 4.16. You roll a die, your friend tosses a coin.
If you roll 6, you win outright.
If you do not roll 6 and your friend tosses Heads, you lose outright.
If neither, the game is repeated until decided.
What is the probability that you win?
One way to solve this problem certainly is this:
P(win) = P(win on 1st round) +P(win on 2nd round) +P(win on 3rd round) +. . .
=
1
6
+
_
5
6

1
2
_
1
6
+
_
5
6

1
2
_
2
1
6
+. . . ,
and then we sum the geometric series. Important note: we have implicitly assumed independence
between the coin and the die, as well as between dierent tosses and rolls. This is very common
in problems such as this!
You can avoid the nuisance, however, by the following trick. Let
D = game is decided on 1st round,
W = you win.
The events D and W are independent, which one can certainly check by computation, but, in
fact, there is a very good reason to conclude so immediately. The crucial observation is that,
provided that the game is not decided in the 1st round, you are thereafter facing the same game
with the same winning probability; thus
P(W[D
c
) = P(W).
In other words, D
c
and W are independent and then so are D and W, and therefore
P(W) = P(W[D).
This means that one can solve this problem by computing the relevant probabilities for the 1st
round:
P(W[D) =
P(W D)
P(D)
=
1
6
1
6
+
5
6
1
2
=
2
7
,
which is our answer.
Example 4.17. Craps. Many casinos allow you to bet even money on the following game. Two
dice are rolled and the sum S is observed.
If S 7, 11, you win immediately.
If S 2, 3, 12, you lose immediately.
If S 4, 5, 6, 8, 9, 10, the pair of dice is rolled repeatedly until one of the following
happens:
S repeats, in which case you win.
7 appears, in which case you lose.
What is the winning probability?
Let us look at all possible ways to win:
1. You win on the rst roll with probability
8
36
.
2. Otherwise,
you roll a 4 (probability
3
36
), then win with probability
3
36
3
36
+
6
36
=
3
3+6
=
1
3
;
4
36
4
4+6
=
2
5
;
5
36
5
5+6
=
5
11
;
5
36
5
5+6
=
5
11
;
4
36
4
4+6
=
2
5
;
3
36
3
3+6
=
1
3
.
Using Bayes formula,
P(win) =
8
36
+ 2
_
3
36

1
3
+
4
36

2
5
+
5
36

5
11
_
0.4929,
a decent game by casino standards.
Bernoulli trials
Assume n independent experiments, each of which is a success with probability p and, thus,
failure with probability 1 p.
In a sequence of n Bernoulli trials, P(exactly k successes) =
_
n
k
_
p
k
(1 p)
nk
.
This is because the successes can occur on any subset S of k trials out of n, i.e., on any
S 1, . . . , n with cardinality k. These possibilities are disjoint, as exactly k successes cannot
occur on two dierent such sets. There are
_
n
k
_
such subsets; if we x such an S, then successes
must occur on k trials in S and failures on all n k trials not in S; the probability that this
happens, by the assumed independence, is p
k
(1 p)
nk
.
Example 4.18. A machine produces items which are independently defective with probability
p. Let us compute a few probabilities:
1. P(exactly two items among the rst 6 are defective) =
_
6
2
_
p
2
(1 p)
4
.
2. P(at least one item among the rst 6 is defective) = 1 P(no defects) = 1 (1 p)
6
3. P(at least 2 items among the rst 6 are defective) = 1 (1 p)
6
6p(1 p)
5
4. P(exactly 100 items are made before 6 defective are found) equals
P(100th item defective, exactly 5 items among 1st 99 defective) = p
_
99
5
_
p
5
(1 p)
94
.
Example 4.19. Problem of Points. This involves nding the probability of n successes before
m failures in a sequence of Bernoulli trials. Let us call this probability p
n,m
.
p
n,m
= P(in the rst m+n 1 trials, the number of successes is n)
=
n+m1
k=n
_
n +m1
k
_
p
k
(1 p)
n+m1k
.
The problem is solved, but it needs to be pointed out that computationally this is not the best
formula. It is much more ecient to use the recursive formula obtained by conditioning on the
outcome of the rst trial. Assume m, n 1. Then,
p
n,m
= P(rst trial is a success) P(n 1 successes before m failures)
+P(rst trial is a failure) P(n successes before m1 failures)
= p p
n1,m
+ (1 p) p
n,m1
.
Together with boundary conditions, valid for m, n 1,
p
n,0
= 0, p
0,m
= 1,
which allows for very speedy and precise computations for large m and n.
Example 4.20. Best of 7. Assume that two equally matched teams, A and B, play a series
of games and that the rst team that wins four games is the overall winner of the series. As
it happens, team A lost the rst game. What is the probability it will win the series? Assume
that the games are Bernoulli trials with success probability
1
2
.
We have
P(A wins the series) = P(4 successes before 3 failures)
=
6
k=4
_
6
k
__
1
2
_
6
=
15 + 6 + 1
2
6
0.3438.
Example 4.21. Banach Matchbox Problem. A mathematician carries two matchboxes, each
originally containing n matches. Each time he needs a match, he is equally likely to take it from
either box. What is the probability that, upon reaching for a box and nding it empty, there
are exactly k matches still in the other box? Here, 0 k n.
Let A
1
be the event that matchbox 1 is the one discovered empty and that, at that instant,
matchbox 2 contains k matches. Before this point, he has used n + n k matches, n from
matchbox 1 and n k from matchbox 2. This means that he has reached for box 1 exactly n
times in (n + n k) trials and for the last time at the (n + 1 + n k)th trial. Therefore, our
probability is
2 P(A
1
) = 2
1
2
_
2n k
n
_
1
2
2nk
=
_
2n k
n
_
1
2
2nk
.
Example 4.22. Each day, you independently decide, with probability p, to ip a fair coin.
Otherwise, you do nothing. (a) What is the probability of getting exactly 10 Heads in the rst
20 days? (b) What is the probability of getting 10 Heads before 5 Tails?
For (a), the probability of getting Heads is p/2 independently each day, so the answer is
_
20
10
_
_
p
2
_
10
_
1
p
2
_
10
.
For (b), you can disregard days at which you do not ip to get
14
k=10
_
14
k
_
1
2
14
.
Example 4.23. You roll a die and your score is the number shown on the die. Your friend rolls
ve dice and his score is the number of 6s shown. Compute (a) the probability of event A that
the two scores are equal and (b) the probability of event B that your friends score is strictly
larger than yours.
In both cases we will condition on your friends score this works a little better in case (b)
than conditioning on your score. Let F
i
, i = 0, . . . , 5, be the event that your friends score is i.
Then, P(A[F
i
) =
1
6
if i 1 and P(A[F
0
) = 0. Then, by the rst Bayes formula, we get
P(A) =
5
i=1
P(F
i
)
1
6
=
1
6
(1 P(F
0
)) =
1
6

5
5
6
6
0.0997.
Moreover, P(B[F
i
) =
i1
6
if i 2 and 0 otherwise, and so
P(B) =
5
i=1
P(F
i
)
i 1
6
=
1
6
5
i=1
i P(F
i
)
1
6
5
i=1
P(F
i
)
=
1
6
5
i=1
i P(F
i
)
1
6
+
5
5
6
6
=
1
6
5
i=1
i
_
5
i
__
1
6
_
i
_
5
6
_
5i
1
6
+
5
5
6
6
=
1
6

5
6

1
6
+
5
5
6
6
0.0392.
The last equality can be obtained by computation, but we will soon learn why the sum has to
equal
5
6
.
Problems
1. Consider the following game. Pick one card at random from a full deck of 52 cards. If you
pull an Ace, you win outright. If not, then you look at the value of the card (K, Q, and J count
as 10). If the number is 7 or less, you lose outright. Otherwise, you select (at random, without
replacement) that number of additional cards from the deck. (For example, if you picked a 9
the rst time, you select 9 more cards.) If you get at least one Ace, you win. What are your
chances of winning this game?
2. An item is defective (independently of other items) with probability 0.3. You have a method
of testing whether the item is defective, but it does not always give you correct answer. If
the tested item is defective, the method detects the defect with probability 0.9 (and says it is
good with probability 0.1). If the tested item is good, then the method says it is defective with
probability 0.2 (and gives the right answer with probability 0.8).
A box contains 3 items. You have tested all of them and the tests detect no defects. What
is the probability that none of the 3 items is defective?
3. A chocolate egg either contains a toy or is empty. Assume that each egg contains a toy with
probability p, independently of other eggs. You have 5 eggs; open the rst one and see if it has
a toy inside, then do the same for the second one, etc. Let E
1
be the event that you get at least
4 toys and let E
2
be the event that you get at least two toys in succession. Compute P(E
1
) and
P(E
2
). Are E
1
and E
2
independent?
4. You have 16 balls, 3 blue, 4 green, and 9 red. You also have 3 urns. For each of the 16 balls,
you select an urn at random and put the ball into it. (Urns are large enough to accommodate any
number of balls.) (a) What is the probability that no urn is empty? (b) What is the probability
that each urn contains 3 red balls? (c) What is the probability that each urn contains all three
colors?
5. Assume that you have an nelement set U and that you select r independent random subsets
A
1
, . . . , A
r
U. All A
i
are chosen so that all 2
n
choices are equally likely. Compute (in a simple
closed form) the probability that the A
i
are pairwise disjoint.
1. Let
F
1
= Ace rst time,
F
8
= 8 rst time,
F
9
= 9 rst time,
F
10
= 10, J, Q, or K rst time.
Also, let W be the event that you win. Then
P(W[F
1
) = 1,
P(W[F
8
) = 1
_
47
8
_
_
51
8
_,
P(W[F
9
) = 1
_
47
9
_
_
51
9
_,
P(W[F
10
) = 1
_
47
10
_
_
51
10
_,
and so,
P(W) =
4
52
+
4
52
_
1
_
47
8
_
_
51
8
_
_
+
4
52
_
1
_
47
9
_
_
51
9
_
_
+
16
52
_
1
_
47
10
_
_
51
10
_
_
.
2. Let F = none is defective and A = test indicates that none is defective. By the second
Bayes formula,
P(F[A) =
P(A F)
P(A)
=
(0.7 0.8)
3
(0.7 0.8 + 0.3 0.1)
3
=
_
56
59
_
3
.
3. P(E
1
) = 5p
4
(1 p) +p
5
= 5p
4
4p
5
and P(E
2
) = 1 (1 p)
5
5p(1 p)
4
_
4
2
_
p
2
(1 p)
3
p
3
(1 p)
2
. As E
1
E
2
, E
1
and E
2
are not independent.
4. (a) Let A
i
= the event that the i-th urn is empty.
P(A
1
) = P(A
2
) = P(A
3
) =
_
2
3
_
16
,
P(A
1
A
2
) = P(A
1
A
3
) = P(A
2
A
3
) =
_
1
3
_
16
,
P(A
1
A
2
A
3
) = 0.
Hence, by inclusion-exclusion,
P(A
1
A
2
A
3
) =
2
16
1
3
15
,
and
P(no urns are empty) = 1 P(A
1
A
2
A
3
)
= 1
2
16
1
3
15
.
(b) We can ignore other balls since only the red balls matter here.
Hence, the result is:
9!
3!3!3!
3
9
=
9!
8 3
12
.
(c) As
P(at least one urn lacks blue) = 3
_
2
3
_
3
3
_
1
3
_
3
,
P(at least one urn lacks green) = 3
_
2
3
_
4
3
_
1
3
_
4
,
P(at least one urn lacks red) = 3
_
2
3
_
9
3
_
1
3
_
9
,
we have, by independence,
P(each urn contains all 3 colors) =
_
1
_
3
_
2
3
_
3
3
_
1
3
_
3
__
_
1
_
3
_
2
3
_
4
3
_
1
3
_
4
__
_
1
_
3
_
2
3
_
9
3
_
1
3
_
9
__
.
5. This is the same as choosing an r n matrix in which every entry is independently 0 or
1 with probability 1/2 and ending up with at most one 1 in every column. Since columns are
independent, this gives ((1 +r)2
r
)
n
.
Interlude: Practice Midterm 1
This practice exam covers the material from the rst four chapters. Give yourself 50 minutes to
solve the four problems, which you may assume have equal point score.
1. Ten fair dice are rolled. What is the probability that:
(a) At least one 1 appears.
(b) Each of the numbers 1, 2, 3 appears exactly twice, while the number 4 appears four times.
(c) Each of the numbers 1, 2, 3 appears at least once.
2. Five married couples are seated at random around a round table.
(a) Compute the probability that all couples sit together (i.e., every husband-wife pair occupies
adjacent seats).
(b) Compute the probability that at most one wife does not sit next to her husband.
3. Consider the following game. A player rolls a fair die. If he rolls 3 or less, he loses immediately.
Otherwise he selects, at random, as many cards from a full deck as the number that came up
on the die. The player wins if all four Aces are among the selected cards.
(a) Compute the winning probability for this game.
(b) Smith tells you that he recently played this game once and won. What is the probability
that he rolled a 6 on the die?
probability p (0, 1), independently of other eggs. Each toy is, with equal probability, red,
white, or blue (again, independently of other toys). You buy 5 eggs. Let E
1
be the event that
you get at most 2 toys and let E
2
be the event that you get you get at least one red and at least
one white and at least one blue toy (so that you have a complete collection).
(a) Compute P(E
1
). Why is this probability very easy to compute when p = 1/2?
(b) Compute P(E
2
).
(c) Are E
1
and E
2
independent? Explain.
Solutions to Practice Midterm 1
1. Ten fair dice are rolled. What is the probability that:
(a) At least one 1 appears.
Solution:
1 P(no 1 appears) = 1
_
5
6
_
10
.
(b) Each of the numbers 1, 2, 3 appears exactly twice, while the number 4 appears four
times.
Solution:
_
10
2
__
8
2
__
6
2
_
6
10
=
10!
2
3
4! 6
10
.
(c) Each of the numbers 1, 2, 3 appears at least once.
Solution:
Let A
i
be the event that the number i does not appear. We know the following:
P(A
1
) = P(A
2
) = P(A
3
) =
_
5
6
_
10
,
P(A
1
A
2
) = P(A
1
A
3
) = P(A
2
A
3
) =
_
4
6
_
10
,
P(A
1
A
2
A
3
) =
_
3
6
_
10
.
Then,
P(1, 2, and 3 each appear at least once)
= P((A
1
A
2
A
3
)
c
)
= 1 P(A
1
) P(A
2
) P(A
3
)
+P(A
1
A
2
) +P(A
2
A
3
) +P(A
1
A
3
)
P(A
1
A
2
A
3
)
= 1 3
_
5
6
_
10
+ 3
_
4
6
_
10
_
3
6
_
10
.
2. Five married couples are seated at random around a round table.
(a) Compute the probability that all couples sit together (i.e., every husband-wife pair
occupies adjacent seats).
Solution:
Let i be an integer in the set 1, 2, 3, 4, 5. Denote each husband and wife as h
i
and w
i
, respectively.
i. Fix h
1
onto one of the seats.
ii. There are 9! ways to order the remaining 9 people in the remaining 9 seats. This
is our sample space.
iii. There are 2 ways to order w
1
.
iv. Treat each couple as a block and the remaining 8 seats as 4 pairs (where each
pair is two adjacent seats). There are 4! ways to seat the remaining 4 couples
into 4 pairs of seats.
v. There are 2
4
ways to order each h
i
and w
i
within its pair of seats.
Therefore, our solution is
2 4! 2
4
9!
.
(b) Compute the probability that at most one wife does not sit next to her husband.
Solution:
Let A be the event that all wives sit next to their husbands and let B be the event
that exactly one wife does not sit next to her husband. We know that P(A) =
2
5
4!
9!
from part (a). Moreover, B = B
1
B
2
B
3
B
4
B
5
, where B
i
is the event that w
i
does not sit next to h
i
and the remaining couples sit together. Then, B
i
are disjoint
and their probabilities are all the same. So, we need to determine P(B
1
).
i. Fix h
1
onto one of the seats.
ii. There are 9! ways to order the remaining 9 people in the remaining 9 seats.
iii. Consider each of remaining 4 couples and w
1
as 5 blocks.
iv. As w
1
cannot be next to her husband, we have 3 positions for w
1
in the ordering
of the 5 blocks.
v. There are 4! ways to order the remaining 4 couples.
vi. There are 2
4
ways to order the couples within their blocks.
Therefore,
P(B
1
) =
3 4! 2
4
9!
.
Our answer, then, is
5
3 4! 2
4
9!
+
2
5
4!
9!
.
3. Consider the following game. The player rolls a fair die. If he rolls 3 or less, he loses
immediately. Otherwise he selects, at random, as many cards from a full deck as the
number that came up on the die. The player wins if all four Aces are among the selected
cards.
(a) Compute the winning probability for this game.
Solution:
Let W be the event that the player wins. Let F
i
be the event that he rolls i, where
i = 1, . . . , 6; P(F
i
) =
1
6
.
Since we lose if we roll a 1, 2, or 3, P(W[F
1
) = P(W[F
2
) = P(W[F
3
) = 0. Moreover,
P(W[F
4
) =
1
_
52
4
_,
P(W[F
5
) =
_
5
4
_
_
52
4
_,
P(W[F
6
) =
_
6
4
_
_
52
4
_.
Therefore,
P(W) =
1
6

1
_
52
4
_
_
1 +
_
5
4
_
+
_
6
4
__
.
(b) Smith tells you that he recently played this game once and won. What is the proba-
bility that he rolled a 6 on the die?
Solution:
P(F
6
[W) =
1
6

1
(
52
4
)

_
6
4
_
P(W)
=
_
6
4
_
1 +
_
5
4
_
+
_
6
4
_
=
15
21
=
5
7
.
probability p (0, 1), independently of other eggs. Each toy is, with equal probability,
red, white, or blue (again, independently of other toys). You buy 5 eggs. Let E
1
be the
event that you get at most 2 toys and let E
2
be the event that you get you get at least
one red and at least one white and at least one blue toy (so that you have a complete
collection).
(a) Compute P(E
1
). Why is this probability very easy to compute when p = 1/2?
Solution:
P(E
1
) = P(0 toys) +P(1 toy) +P(2 toys)
= (1 p)
5
+ 5p(1 p)
4
+
_
5
2
_
p
2
(1 p)
3
.
When p =
1
2
,
P(at most 2 toys) = P(at least 3 toys)
= P(at most 2 eggs are empty)
Therefore, P(E
1
) = P(E
c
1
) and so P(E
1
) =
1
2
.
(b) Compute P(E
2
).
Solution:
Let A
1
be the event that red is missing, A
2
the event that white is missing, and
A
3
the event that blue is missing.
P(E
2
) = P((A
1
A
2
A
3
)
c
)
= 1 3
_
1
p
3
_
5
+ 3
_
1
2p
3
_
5
(1 p)
5
.
(c) Are E
1
and E
2
independent? Explain.
Solution:
No: E
1
E
2
= .
5 DISCRETE RANDOM VARIABLES 45
5 Discrete Random Variables
A random variable is a number whose value depends upon the outcome of a random experiment.
Mathematically, a random variable X is a real-valued function on , the space of outcomes:
X : R.
Sometimes, when convenient, we also allow X to have the value or, more rarely, , but
this will not occur in this chapter. The crucial theoretical property that X should have is that,
for each interval B, the set of outcomes for which X B is an event, so we are able to talk
about its probability, P(X B). Random variables are traditionally denoted by capital letters
to distinguish them from deterministic quantities.
Example 5.1. Here are some examples of random variables.
1. Toss a coin 10 times and let X be the number of Heads.
2. Choose a random point in the unit square (x, y) : 0 x, y 1 and let X be its distance
from the origin.
3. Choose a random person in a class and let X be the height of the person, in inches.
4. Let X be value of the NASDAQ stock index at the closing of the next business day.
A discrete random variable X has nitely or countably many values x
i
, i = 1, 2, . . ., and
p(x
i
) = P(X = x
i
) with i = 1, 2, . . . is called the probability mass function of X. Sometimes X
is added as the subscript of its p. m. f., p = p
X
.
A probability mass function p has the following properties:
1. For all i, p(x
i
) > 0 (we do not list values of X which occur with probability 0).
2. For any interval B, P(X B) =

x
i
B
p(x
i
).
3. As X must have some value,

i
p(x
i
) = 1.
Example 5.2. Let X be the number of Heads in 2 fair coin tosses. Determine its p. m. f.
Possible values of X are 0, 1, and 2. Their probabilities are: P(X = 0) =
1
4
, P(X = 1) =
1
2
,
and P(X = 2) =
1
4
.
You should note that the random variable Y , which counts the number of Tails in the 2
tosses, has the same p. m. f., that is, p
X
= p
Y
, but X and Y are far from being the same
random variable! In general, random variables may have the same p. m. f., but may not even
be dened on the same set of outcomes.
Example 5.3. An urn contains 20 balls numbered 1, . . . , 20. Select 5 balls at random, without
replacement. Let X be the largest number among selected balls. Determine its p. m. f. and the
probability that at least one of the selected numbers is 15 or more.
The possible values are 5, . . . , 20. To determine the p. m. f., note that we have
_
20
5
_
outcomes,
and, then,
P(X = i) =
_
i1
4
_
_
20
5
_ .
Finally,
P(at least one number 15 or more) = P(X 15) =
20
i=15
P(X = i).
Example 5.4. An urn contains 11 balls, 3 white, 3 red, and 5 blue balls. Take out 3 balls at
random, without replacement. You win $1 for each red ball you select and lose a $1 for each
white ball you select. Determine the p. m. f. of X, the amount you win.
The number of outcomes is
_
11
3
_
. X can have values 3, 2, 1, 0, 1, 2, and 3. Let us start
with 0. This can occur with one ball of each color or with 3 blue balls:
P(X = 0) =
3 3 5 +
_
5
3
_
_
11
3
_ =
55
165
.
To get X = 1, we can have 2 red and 1 white, or 1 red and 2 blue:
P(X = 1) = P(X = 1) =
_
3
2
__
3
1
_
+
_
3
1
__
5
2
_
_
11
3
_ =
39
165
.
The probability that X = 1 is the same because of symmetry between the roles that the red
and the white balls play. Next, to get X = 2 we must have 2 red balls and 1 blue:
P(X = 2) = P(X = 2) =
_
3
2
__
5
1
_
_
11
3
_ =
15
165
.
Finally, a single outcome (3 red balls) produces X = 3:
P(X = 3) = P(X = 3) =
1
_
11
3
_ =
1
165
.
All the seven probabilities should add to 1, which can be used either to check the computations
or to compute the seventh probability (say, P(X = 0)) from the other six.
Assume that X is a discrete random variable with possible values x
i
, i = 1, 2 . . .. Then, the
expected value, also called expectation, average, or mean, of X is
EX =
i
x
i
P(X = x
i
) =
i
x
i
p(x
i
).
For any function, g : R R,
Eg(X) =
i
g(x
i
)P(X = x
i
).
Example 5.5. Let X be a random variable with P(X = 1) = 0.2, P(X = 2) = 0.3, and
P(X = 3) = 0.5. What is the expected value of X?
We can, of course, just use the formula, but let us instead proceed intuitively and see that
the denition makes sense. What, then, should the average of X be?
Imagine a large number n of repetitions of the experiment and measure the realization of X
in each. By the frequency interpretation of probability, about 0.2n realizations will have X = 1,
about 0.3n will have X = 2, and about 0.5n will have X = 3. The average value of X then
should be
1 0.2n + 2 0.3n + 3 0.5n
n
= 1 0.2 + 2 0.3 + 3 0.5 = 2.3,
which of course is the same as the formula gives.
Take a discrete random variable X and let = EX. How should we measure the deviation
of X from , i.e., how spread-out is the p. m. f. of X?
The most natural way would certainly be E[X [. The only problem with this is that
absolute values are annoying. Instead, we dene the variance of X
Var(X) = E(x )
2
.
The quantity that has the correct units is the standard deviation
(X) =
_
Var(X) =
_
E(X )
2
.
We will give another, more convenient, formula for variance that will use the following property
of expectation, called linearity:
E(
1
X
1
+
2
X
2
) =
1
EX
1
+
2
EX
2
,
valid for any random variables X
1
and X
2
and nonrandom constants
1
and
2
. This property
will be explained and discussed in more detail later. Then
Var(X) = E[(X )
2
]
= E[X
2
2X +
2
]
= E(X
2
) 2E(X) +
2
= E(X
2
)
2
= E(X
2
) (EX)
2
In computations, bear in mind that variance cannot be negative! Furthermore, the only way
that a random variable has 0 variance is when it is equal to its expectation with probability
1 (so it is not really random at all): P(X = ) = 1. Here is the summary:
The variance of a random variable X is Var(X) = E(XEX)
2
= E(X
2
)(EX)
2
.
Example 5.6. Previous example, continued. Compute Var(X).
E(X
2
) = 1
2
0.2 + 2
2
0.3 + 3
2
0.5 = 5.9,
(EX)
2
= (2.3)
2
= 5.29, and so Var(X) = 5.9 5.29 = 0.61 and (X) =
_
Var(X) 0.7810.
We will now look at some famous probability mass functions.
5.1 Uniform discrete random variable
This is a random variable with values x
1
, . . . , x
n
, each with equal probability 1/n. Such a random
variable is simply the random choice of one among n numbers.
Properties:
1. EX =
x
1
+...+x
n
n
.
2. VarX =
x
2
1
+...+x
2
n
n

_
x
1
+...+xn
n
_
2
.
Example 5.7. Let X be the number shown on a rolled fair die. Compute EX, E(X
2
), and
Var(X).
This is a standard example of a discrete uniform random variable and
EX =
1 + 2 +. . . + 6
6
=
7
2
,
EX
2
=
1 + 2
2
+. . . + 6
2
6
=
91
6
,
Var(X) =
91
6

_
7
2
_
2
=
35
12
.
5.2 Bernoulli random variable
This is also called an indicator random variable. Assume that A is an event with probability p.
Then, I
A
, the indicator of A, is given by
I
A
=
_
1 if A happens,
0 otherwise.
Other notations for I
A
include 1
A
and
A
. Although simple, such random variables are very
important as building blocks for more complicated random variables.
Properties:
1. EI
A
= p.
2. Var(I
A
) = p(1 p).
For the variance, note that I
2
A
= I
A
, so that E(I
2
A
) = EI
A
= p.
5.3 Binomial random variable
A Binomial(n,p) random variable counts the number of successes in n independent trials, each
of which is a success with probability p.
Properties:
1. Probability mass function: P(X = i) =
_
n
i
_
p
i
(1 p)
ni
, i = 0, . . . , n.
2. EX = np.
3. Var(X) = np(1 p).
The expectation and variance formulas will be proved in Chapter 8. For now, take them on
faith.
Example 5.8. Let X be the number of Heads in 50 tosses of a fair coin. Determine EX,
Var(X) and P(X 10)? As X is Binomial(50,
1
2
), so EX = 25, Var(X) = 12.5, and
P(X 10) =
10
i=0
_
50
i
_
1
2
50
.
Example 5.9. Denote by d the dominant gene and by r the recessive gene at a single locus.
Then dd is called the pure dominant genotype, dr is called the hybrid, and rr the pure recessive
genotype. The two genotypes with at least one dominant gene, dd and dr, result in the phenotype
of the dominant gene, while rr results in a recessive phenotype.
Assuming that both parents are hybrid and have n children, what is the probability that at
least two will have the recessive phenotype? Each child, independently, gets one of the genes at
random from each parent.
For each child, independently, the probability of the rr genotype is
1
4
. If X is the number of
rr children, then X is Binomial(n,
1
4
). Therefore,
P(X 2) = 1 P(X = 0) P(X = 1) = 1
_
3
4
_
n
n
1
4
_
3
4
_
n1
.
5.4 Poisson random variable
A random variable is Poisson(), with parameter > 0, if it has the probability mass function
given below.
Properties:
1. P(X = i) =

i
i!
e
, for i = 0, 1, 2, . . ..
2. EX = .
3. Var(X) = .
Here is how we compute the expectation:
EX =
i=1
i e
i
i!
= e
i=1
i1
(i 1)!
= e
= ,
and the variance computation is similar (and a good exercise!).
The Poisson random variable is useful as an approximation to a Binomial random variable
when the number of trials is large and the probability of success is small. In this context it is
often called the law of rare events, rst formulated by L. J. Bortkiewicz (in 1898), who studied
deaths by horse kicks in the Prussian cavalry.
Theorem 5.1. Poisson approximation to Binomial . When n is large, p is small, and = np
is of moderate size, Binomial(n,p) is approximately Poisson():
If X is Binomial(n, p), with p =

n
, then, as n ,
P(X = i) e
i
i!
,
for each xed i = 0, 1, 2, . . ..
Proof.
P(X = i) =
_
n
i
__
n
_
i
_
1

n
_
ni
=
n(n 1) . . . (n i + 1)
i!

i
n
i
_
1

n
_
n
_
1

n
_
i
=

i
i!

_
1

n
_
n
n(n 1) . . . (n i + 1)
n
i

1
_
1

n
_
i

i
i!
e
1 1,
as n .
The Poisson approximation is quite good: one can prove that the error made by computing
a probability using the Poisson approximation instead of its exact Binomial expression (in the
context of the above theorem) is no more than
min(1, ) p.
Example 5.10. Suppose that the probability that a person is killed by lighting in a year is,
independently, 1/(500 million). Assume that the US population is 300 million.
1. Compute P(3 or more people will be killed by lightning next year) exactly.
If X is the number of people killed by lightning, then X is Binomial(n, p), where n = 300
million and p = 1/ (500 million), and the answer is
1 (1 p)
n
np(1 p)
n1
_
n
2
_
p
2
(1 p)
n2
0.02311530.
2. Approximate the above probability.
As np =
3
5
, X is approximately Poisson(
3
5
), and the answer is
1 e

2
2
e
0.02311529.
3. Approximate P(two or more people are killed by lightning within the rst 6 months of
next year).
This highlights the interpretation of as a rate. If lightning deaths occur at the rate of
3
5
a year, they should occur at half that rate in 6 months. Indeed, assuming that lightning
deaths occur as a result of independent factors in disjoint time intervals, we can imagine
that they operate on dierent people in disjoint time intervals. Thus, doubling the time
interval is the same as doubling the number n of people (while keeping p the same), and
then np also doubles. Consequently, halving the time interval has the same p, but half as
many trials, so np changes to
3
10
and so =
3
10
as well. The answer is
1 e
0.0369.
4. Approximate P(in exactly 3 of next 10 years exactly 3 people are killed by lightning).
In every year, the probability of exactly 3 deaths is approximately

3
3!
e
, where, again,
=
3
5
. Assuming year-to-year independence, the number of years with exactly 3 people
killed is approximately Binomial(10,

3
3!
e
). The answer, then, is

_
10
3
__
3
3!
e
_
3
_
1

3
3!
e
_
7
4.34 10
6
.
5. Compute the expected number of years, among the next 10, in which 2 or more people are
killed by lightning.
By the same logic as above and the formula for Binomal expectation, the answer is
10(1 e
) 0.3694.
Example 5.11. Poisson distribution and law. Assume a crime has been committed. It is
known that the perpetrator has certain characteristics, which occur with a small frequency p
(say, 10
8
) in a population of size n (say, 10
8
). A person who matches these characteristics has
been found at random (e.g., at a routine trac stop or by airport security) and, since p is so
small, charged with the crime. There is no other evidence. What should the defense be?
Let us start with a mathematical model of this situation. Assume that N is the number of
people with given characteristics. This is a Binomial random variable but, given the assumptions,
we can easily assume that it is Poisson with = np. Choose a person from among these N, label
that person by C, the criminal. Then, choose at random another person, A, who is arrested.
The question is whether C = A, that is, whether the arrested person is guilty. Mathematically,
we can formulate the problem as follows:
P(C = A[ N 1) =
P(C = A, N 1)
P(N 1)
.
We need to condition as the experiment cannot even be performed when N = 0. Now, by the
rst Bayes formula,
P(C = A, N 1) =
k=1
P(C = A, N 1 [ N = k) P(N = k)
=
k=1
P(C = A[ N = k) P(N = k)
and
P(C = A[ N = k) =
1
k
,
so
P(C = A, N 1) =
k=1
1
k

k
k!
e
.
The probability that the arrested person is guilty then is
P(C = A[N 1) =
e
1 e
k=1
k
k k!
.
There is no closed-form expression for the sum, but it can be easily computed numerically. The
defense may claim that the probability of innocence, 1(the above probability), is about 0.2330
when = 1, presumably enough for a reasonable doubt.
This model was in fact tested in court, in the famous People v. Collins case, a 1968 jury
trial in Los Angeles. In this instance, it was claimed by the prosecution (on imsy grounds)
that p = 1/12, 000, 000 and n would have been the number of adult couples in the LA area, say
n = 5, 000, 000. The jury convicted the couple charged for robbery on the basis of the prose-
cutors claim that, due to low p, the chances of there being another couple [with the specied
characteristics, in the LA area] must be one in a billion. The Supreme Court of California
reversed the conviction and gave two reasons. The rst reason was insucient foundation for
the estimate of p. The second reason was that the probability that another couple with matching
characteristics existed was, in fact,
P(N 2 [ N 1) =
1 e
1 e
,
much larger than the prosecutor claimed, namely, for =
5
12
it is about 0.1939. This is about
twice the (more relevant) probability of innocence, which, for this , is about 0.1015.
5.5 Geometric random variable
A Geometric(p) random variable X counts the number trials required for the rst success in
independent trials with success probability p.
Properties:
1. Probability mass function: P(X = n) = p(1 p)
n1
, where n = 1, 2, . . ..
2. EX =
1
p
.
3. Var(X) =
1p
p
2
.
4. P(X > n) =

k=n+1
p(1 p)
k1
= (1 p)
n
.
5. P(X > n +k[X > k) =
(1p)
n+k
(1p)
k
= P(X > n).
We omit the proofs of the second and third formulas, which reduce to manipulations with
geometric series.
Example 5.12. Let X be the number of tosses of a fair coin required for the rst Heads. What
are EX and Var(X)?
As X is Geometric(
1
2
), EX = 2 and Var(X) = 2.
Example 5.13. You roll a die, your opponent tosses a coin. If you roll 6 you win; if you do
not roll 6 and your opponent tosses Heads you lose; otherwise, this round ends and the game
repeats. On the average, how many rounds does the game last?
P(game decided on round 1) =
1
6
+
5
6

1
2
=
7
12
,
and so the number of rounds N is Geometric(
7
12
), and
EN =
12
7
.
Problems
1. Roll a fair die repeatedly. Let X be the number of 6s in the rst 10 rolls and let Y the
number of rolls needed to obtain a 3. (a) Write down the probability mass function of X. (b)
Write down the probability mass function of Y . (c) Find an expression for P(X 6). (d) Find
an expression for P(Y > 10).
2. A biologist needs at least 3 mature specimens of a certain plant. The plant needs a year
to reach maturity; once a seed is planted, any plant will survive for the year with probability
1/1000 (independently of other plants). The biologist plants 3000 seeds. A year is deemed a
success if three or more plants from these seeds reach maturity.
(a) Write down the exact expression for the probability that the biologist will indeed end up
with at least 3 mature plants.
(b) Write down a relevant approximate expression for the probability from (a). Justify briey
the approximation.
(c) The biologist plans to do this year after year. What is the approximate probability that he
has at least 2 successes in 10 years?
(d) Devise a method to determine the number of seeds the biologist should plant in order to get
at least 3 mature plants in a year with probability at least 0.999. (Your method will probably
require a lengthy calculation do not try to carry it out with pen and paper.)
3. You are dealt one card at random from a full deck and your opponent is dealt 2 cards
(without any replacement). If you get an Ace, he pays you $10, if you get a King, he pays you
$5 (regardless of his cards). If you have neither an Ace nor a King, but your card is red and
your opponent has no red cards, he pays you $1. In all other cases you pay him $1. Determine
your expected earnings. Are they positive?
4. You and your opponent both roll a fair die. If you both roll the same number, the game
is repeated, otherwise whoever rolls the larger number wins. Let N be the number of times
the two dice have to be rolled before the game is decided. (a) Determine the probability mass
function of N. (b) Compute EN. (c) Compute P(you win). (d) Assume that you get paid
$10 for winning in the rst round, $1 for winning in any other round, and nothing otherwise.
Compute your expected winnings.
5. Each of the 50 students in class belongs to exactly one of the four groups A, B, C, or D. The
membership numbers for the four groups are as follows: A: 5, B: 10, C: 15, D: 20. First, choose
one of the 50 students at random and let X be the size of that students group. Next, choose
one of the four groups at random and let Y be its size. (Recall: all random choices are with
equal probability, unless otherwise specied.) (a) Write down the probability mass functions for
X and Y . (b) Compute EX and EY . (c) Compute Var(X) and Var(Y ). (d) Assume you have
s students divided into n groups with membership numbers s
1
, . . . , s
n
, and again X is the size
of the group of a randomly chosen student, while Y is the size of the randomly chosen group.
Let EY = and Var(Y ) =
2
. Express EX with s, n, , and .
Solutions
1. (a) X is Binomial(10,
1
6
):
P(X = i) =
_
10
i
__
1
6
_
i
_
5
6
_
10i
,
where i = 0, 1, 2, . . . , 10.
(b) Y is Geometric(
1
6
):
P(Y = i) =
1
6
_
5
6
_
i1
,
where i = 1, 2, . . ..
(c)
P(X 6) =
10
i=6
_
10
i
__
1
6
_
i
_
5
6
_
10i
.
(d)
P(Y > 10) =
_
5
6
_
10
.
2. (a) The random variable X, the number of mature plants, is Binomial
_
3000,
1
1000
_
.
P(X 3) = 1 P(X 2)
= 1 (0.999)
3000
3000(0.999)
2999
(0.001)
_
3000
2
_
(0.999)
2998
(0.001)
2
.
(b) By the Poisson approximation with = 3000
1
1000
= 3,
P(X 3) 1 e
3
3e
3
9
2
e
3
.
(c) Denote the probability in (b) by s. Then, the number of years the biologists succeeds is
approximately Binomial(10, s) and the answer is
1 (1 s)
10
10s(1 s)
9
.
(d) Solve
e
+e
+

2
2
e
= 0.001
for and then let n = 1000. The equation above can be solved by rewriting
= log 1000 + log(1 + +

2
2
)
and then solved by iteration. The result is that the biologist should plant 11, 229 seeds.
3. Let X be your earnings.
P(X = 10) =
4
52
,
P(X = 5) =
4
52
,
P(X = 1) =
22
52

_
26
2
_
_
51
2
_ =
11
102
,
P(X = 1) = 1
2
13

11
102
,
and so
EX =
10
13
+
5
13
+
11
102
1 +
2
13
+
11
102
=
4
13
+
11
51
> 0
4. (a) N is Geometric(
5
6
):
P(N = n) =
_
1
6
_
n1
5
6
,
where n = 1, 2, 3, . . ..
(b) EN =
6
5
.
(c) By symmetry, P(you win) =
1
2
.
(d) You get paid $10 with probability
5
12
, $1 with probability
1
12
, and 0 otherwise, so your
expected winnings are
51
12
.
5. (a)
x P(X = x) P(Y = x)
5 0.1 0.25
10 0.2 0.25
15 0.3 0.25
20 0.4 0.25
(b) EX = 15, EY = 12.5.
(c) E(X
2
) = 250, so Var(X) = 25. E(Y
2
) = 187.5, so Var(Y ) = 31.25.
(d) Let s = s
1
+. . . +s
n
. Then,
E(X) =
n
i=1
s
i
s
i
s
=
n
s
n
i=1
s
2
i

1
n
=
n
s
EY
2
=
n
s
(Var(Y ) + (EY )
2
) =
n
s
(
2
+
2
).
6 CONTINUOUS RANDOM VARIABLES 58
6 Continuous Random Variables
A random variable X is continuous if there exists a nonnegative function f so that, for every
interval B,
P(X B) =
_
B
f(x) dx,
The function f = f
X
is called the density of X.
We will assume that a density function f is continuous, apart from nitely many (possibly
innite) jumps. Clearly, it must hold that
_

f(x) dx = 1.
Note also that
P(X [a, b]) = P(a X b) =
_
b
a
f(x) dx,
P(X = a) = 0,
P(X b) = P(X < b) =
_
b
f(x) dx.
The function F = F
X
given by
F(x) = P(X x) =
_
x
f(s) ds
is called the distribution function of X. On an open interval where f is continuous,
F
(x) = f(x).
Density has the same role as the probability mass function for discrete random variables: it
tells which values x are relatively more probable for X than others. Namely, if h is very small,
then
P(X [x, x +h]) = F(x +h) F(x) F
(x) h = f(x) h.
By analogy with discrete random variables, we dene,
EX =
_

x f(x) dx,
Eg(X) =
_

g(x) f(x) dx,

and variance is computed by the same formula: Var(X) = E(X
2
) (EX)
2
.
Example 6.1. Let
f(x) =
_
cx if 0 < x < 4,
0 otherwise.
(a) Determine c. (b) Compute P(1 X 2). (c) Determine EX and Var(X).
For (a), we use the fact that density integrates to 1, so we have
_
4
0
cxdx = 1 and c =
1
8
. For
(b), we compute
_
2
1
x
8
dx =
3
16
.
Finally, for (c) we get
EX =
_
4
0
x
2
8
dx =
8
3
and
E(X
2
) =
_
4
0
x
3
8
dx = 8.
So, Var(X) = 8
64
9
=
8
9
.
Example 6.2. Assume that X has density
f
X
(x) =
_
3x
2
if x [0, 1],
0 otherwise.
Compute the density f
Y
of Y = 1 X
4
.
In a problem such as this, compute rst the distribution function F
Y
of Y . Before starting,
note that the density f
Y
(y) will be nonzero only when y [0, 1], as the values of Y are restricted
to that interval. Now, for y (0, 1),
F
Y
(y) = P(Y y) = P(1 X
4
y) = P(1 y X
4
) = P((1 y)
1
4
X) =
_
1
(1y)
1
4
3x
2
dx.
It follows that
f
Y
(y) =
d
dy
F
Y
(y) = 3((1 y)
1
4
)
2
1
4
(1 y)
3
4
(1) =
3
4
1
(1 y)
1
4
,
for y (0, 1), and f
Y
(y) = 0 otherwise. Observe that it is immaterial how f(y) is dened at
y = 0 and y = 1, because those two values contribute nothing to any integral.
As with discrete random variables, we now look at some famous densities.
6.1 Uniform random variable
Such a random variable represents the choice of a random number in [, ]. For [, ] = [0, 1],
this is ideally the output of a computer random number generator.
Properties:
1. Density: f(x) =
_
1
if x [, ],
0 otherwise.
2. EX =
+
2
.
3. Var(X) =
()
2
12
.
Example 6.3. Assume that X is uniform on [0, 1]. What is P(X Q)? What is the probability
that the binary expansion of X starts with 0.010?
As Q is countable, it has an enumeration, say, Q = q
1
, q
2
, . . . . By Axiom 3 of Chapter 3:
P(X Q) = P(
i
X = q
i
) =
i
P(X = q
i
) = 0.
Note that you cannot do this for sets that are not countable or you would prove that P(X
R) = 0, while we, of course, know that P(X R) = P() = 1. As X is, with probability 1,
irrational, its binary expansion is uniquely dened, so there is no ambiguity about what the
second question means.
Divide [0, 1) into 2
n
intervals of equal length. If the binary expansion of a number x [0, 1)
is 0.x
1
x
2
. . ., the rst n binary digits determine which of the 2
n
subintervals x belongs to: if you
know that x belongs to an interval I based on the rst n1 digits, then nth digit 1 means that
x is in the right half of I and nth digit 0 means that x is in the left half of I. For example, if
the expansion starts with 0.010, the number is in [0,
1
2
], then in [
1
4
,
1
2
], and then nally in [
1
4
,
3
8
].
Our answer is
1
8
, but, in fact, we can make a more general conclusion. If X is uniform on
[0, 1], then any of the 2
n
possibilities for its rst n binary digits are equally likely. In other
words, the binary digits of X are the result of an innite sequence of independent fair coin
tosses. Choosing a uniform random number on [0, 1] is thus equivalent to tossing a fair coin
innitely many times.
Example 6.4. A uniform random number X divides [0, 1] into two segments. Let R be the
ratio of the smaller versus the larger segment. Compute the density of R.
As R has values in (0, 1), the density f
R
(r) is nonzero only for r (0, 1) and we will deal
only with such rs.
F
R
(r) = P(R r) = P
_
X
1
2
,
X
1 X
r
_
+P
_
X >
1
2
,
1 X
X
r
_
= P
_
X
1
2
, X
r
r + 1
_
+P
_
X >
1
2
, X
1
r + 1
_
= P
_
X
r
r + 1
_
+P
_
X
1
r + 1
_ _
since
r
r + 1

1
2
and
1
r + 1

1
2
_
=
r
r + 1
+ 1
1
r + 1
=
2r
r + 1
For r (0, 1), the density, thus, equals
f
R
(r) =
d
dr
F
R
(r) =
2
(r + 1)
2
.
6.2 Exponential random variable
A random variable is Exponential(), with parameter > 0, if it has the probability mass
function given below. This is a distribution for the waiting time for some random event, for
example, for a lightbulb to burn out or for the next earthquake of at least some given magnitude.
Properties:
1. Density: f(x) =
_
e
x
if x 0,
0 if x < 0.
2. EX =
1
.
3. Var(X) =
1
2
.
4. P(X x) = e
x
.
5. Memoryless property: P(X x +y[X y) = e
x
.
The last property means that, if the event has not occurred by some given time (no matter
how large), the distribution of the remaining waiting time is the same as it was at the beginning.
There is no aging.
Proofs of these properties are integration exercises and are omitted.
Example 6.5. Assume that a lightbulb lasts on average 100 hours. Assuming exponential
distribution, compute the probability that it lasts more than 200 hours and the probability that
it lasts less than 50 hours.
Let X be the waiting time for the bulb to burn out. Then, X is Exponential with =
1
100
and
P(X 200) = e
2
0.1353,
P(X 50) = 1 e
1
2
0.3935.
6.3 Normal random variable
A random variable is Normal with parameters R and
2
> 0 or, in short, X is N(,
2
),
if its density is the function given below. Such a random variable is (at least approximately)
very common. For example, measurement with random error, weight of a randomly caught
yellow-billed magpie, SAT (or some other) test score of a randomly chosen student at UC Davis,
etc.
Properties:
1. Density:
f(x) = f
X
(x) =
1
2
e
(x)
2
2
2
,
where x (, ).
2. EX = .
3. Var(X) =
2
.
To show that
_

f(x) dx = 1
is a tricky exercise in integration, as is the computation of the variance. Assuming that the
integral of f is 1, we can use symmetry to prove that EX must be :
EX =
_

xf(x) dx =
_

(x )f(x) dx +
_

f(x) dx
=
_

(x )
1
2
e
(x)
2
2
2
dx +
=
_

z
1
2
e
z
2
2
2
dz +
= ,
where the last integral was obtained by the change of variable z = x and is zero because the
function integrated is odd.
Example 6.6. Let X be a N(,
2
) random variable and let Y = X +, with > 0. How is
Y distributed?
If X is a measurement with error X + amounts to changing the units and so Y should
still be normal. Let us see if this is the case. We start by computing the distribution function
of Y ,
F
Y
(y) = P(Y y)
= P(X + y)
= P
_
X
y
_
=
_ y
f
X
(x) dx
and, then, the density
f
Y
(y) = f
X
_
y
=
1
2
e
(y)
2
2
2
2
.
Therefore, Y is normal with EY = + and Var(Y ) = ()
2
.
In particular,
Z =
X
has EZ = 0 and Var(Z) = 1. Such a N(0, 1) random variable is called standard Normal. Its
distribution function F
Z
(z) is denoted by (z). Note that
f
Z
(z) =
1
2
e
z
2
/2
(z) = F
Z
(z) =
1
2
_
z
e
x
2
/2
dx.
The integral for (z) cannot be computed as an elementary function, so approximate values
are given in tables. Nowadays, this is largely obsolete, as computers can easily compute (z)
very accurately for any given z. You should also note that it is enough to know these values for
z > 0, as in this case, by using the fact that f
Z
(x) is an even function,
(z) =
_
z
f
Z
(x) dx =
_

z
f
Z
(x) dx = 1
_
z
f
Z
(x) dx = 1 (z).
In particular, (0) =
1
2
. Another way to write this is P(Z z) = P(Z z), a form which is
also often useful.
Example 6.7. What is the probability that a Normal random variable diers from its mean
by more than ? More than 2? More than 3?
In symbols, if X is N(,
2
), we need to compute P([X [ ), P([X [ 2), and
P([X [ 3).
In this and all other examples of this type, the letter Z will stand for an N(0, 1) random
variable.
We have
P([X [ ) = P
_
1
_
= P([Z[ 1) = 2P(Z 1) = 2(1 (1)) 0.3173.
Similarly,
P([X [ 2) = 2(1 (2)) 0.0455,
P([X [ 3) = 2(1 (3)) 0.0027.
Example 6.8. Assume that X is Normal with mean = 2 and variance
2
= 25. Compute
the probability that X is between 1 and 4.
Here is the computation:
P(1 X 4) = P
_
1 2
5

X 2
5

4 2
5
_
= P(0.2 Z 0.4)
= P(Z 0.4) P(Z 0.2)
= (0.4) (1 (0.2))
0.2347 .
Let S
n
be a Binomial(n, p) random variable. Recall that its mean is np and its variance
np(1 p). If we pretend that S
n
is Normal, then
S
n
np
np(1p)
is standard Normal, i.e., N(0, 1).
The following theorem says that this is approximately true if p is xed (e.g., 0.5) and n is large
(e.g., n = 100).
Theorem 6.1. De Moivre-Laplace Central Limit Theorem.
Let S
n
be Binomial(n, p), where p is xed and n is large. Then,
S
n
np
np(1p)
N(0, 1); more
precisely,
P
_
S
n
np
_
np(1 p)
x
_
(x)
as n , for every real number x.
We should also note that the above theorem is an analytical statement; it says that
k:0knp+x
np(1p)
_
n
k
_
p
k
(1 p)
nk
2
_
x
s
2
2
ds
as n , for every x R. Indeed it can be, and originally was, proved this way, with a lot of
computational work.
An important issue is the quality of the Normal approximation to the Binomial. One can
prove that the dierence between the Binomial probability (in the above theorem) and its limit
is at most
0.5 (p
2
+ (1 p)
2
)
_
np (1 p)
.
A commonly cited rule of thumb is that this is a decent approximation when np(1 p) 10;
however, if we take p = 1/3 and n = 45, so that np(1p) = 10, the bound above is about 0.0878,
too large for many purposes. Various corrections have been developed to diminish the error,
but they are, in my opinion, obsolete by now. In the situation when the above upper bound
on the error is too high, we should simply compute directly with the Binomial distribution and
not use the Normal approximation. (We will assume that the approximation is adequate in the
examples below.) Remember that, when n is large and p is small, say n = 100 and p =
1
100
, the
Poisson approximation (with = np) is much better!
Example 6.9. A roulette wheel has 38 slots: 18 red, 18 black, and 2 green. The ball ends at
one of these at random. You are a player who plays a large number of games and makes an even
bet of $1 on red in every game. After n games, what is the probability that you are ahead?
Answer this for n = 100 and n = 1000.
Let S
n
be the number of times you win. This is a Binomial(n,
9
19
) random variable.
P(ahead) = P(win more than half of the games)
= P
_
S
n
>
n
2
_
= P
_
S
n
np
_
np(1 p)
>
1
2
n np
_
np(1 p)
_
P
_
Z >
(
1
2
p)
n
_
p(1 p)
_
For n = 100, we get
P
_
Z >
5
90
_
0.2990,
and for n = 1000, we get
P
_
Z >
5
3
_
0.0478.
For comparison, the true probabilities are 0.2650 and 0.0448, respectively.
Example 6.10. What would the answer to the previous example be if the game were fair, i.e.,
you bet even money on the outcome of a fair coin toss each time.
Then, p =
1
2
and
P(ahead) P(Z > 0) = 0.5,
as n .
Example 6.11. How many times do you need to toss a fair coin to get 100 heads with probability
90%?
Let n be the number of tosses that we are looking for. For S
n
, which is Binomial(n,
1
2
), we
need to nd n so that
P(S
n
100) 0.9.
We will use below that n > 200, as the probability would be approximately
1
2
for n = 200 (see
the previous example). Here is the computation:
P
_
S
n
1
2
n
1
2
100
1
2
n
1
2
n
_
P
_
Z
100
1
2
n
1
2
n
_
= P
_
Z
200 n
n
_
= P
_
Z
_
n 200
n
__
= P
_
Z
n 200
n
_
=
_
n 200
n
_
= 0.9
Now, according to the tables, (1.28) 0.9, thus we need to solve
n200
n
= 1.28, that is,
n 1.28
n 200 = 0.
This is a quadratic equation in

n, with the only positive solution
n =
1.28 +
1.28
2
+ 800
2
.
Rounding up the number n we get from above, we conclude that n = 219.
Problems
1. A random variable X has the density function
f(x) =
_
c(x +
x) x [0, 1],
0 otherwise.
(a) Determine c. (b) Compute E(1/X). (c) Determine the probability density function of
Y = X
2
.
2. The density function of a random variable X is given by
f(x) =
_
a +bx 0 x 2,
0 otherwise.
We also know that E(X) = 7/6. (a) Compute a and b. (b) Compute Var(X).
3. After your complaint about their service, a representative of an insurance company promised
to call you between 7 and 9 this evening. Assume that this means that the time T of the call
is uniformly distributed in the specied interval.
(a) Compute the probability that the call arrives between 8:00 and 8:20.
(b) At 8:30, the call still hasnt arrived. What is the probability that it arrives in the next 10
minutes?
(c) Assume that you know in advance that the call will last exactly 1 hour. From 9 to 9:30,
there is a game show on TV that you wanted to watch. Let M be the amount of time of the
show that you miss because of the call. Compute the expected value of M.
4. Toss a fair coin twice. You win $1 if at least one of the two tosses comes out heads.
(a) Assume that you play this game 300 times. What is, approximately, the probability that
you win at least $250?
(b) Approximately how many times do you need to play so that you win at least $250 with
probability at least 0.99?
5. Roll a die n times and let M be the number of times you roll 6. Assume that n is large.
(a) Compute the expectation EM.
(b) Write down an approximation, in terms on n and , of the probability that M diers from
its expectation by less than 10%.
(c) How large should n be so that the probability in (b) is larger than 0.99?
Solutions
1. (a) As
1 = c
_
1
0
(x +
x) dx = c
_
1
2
+
2
3
_
=
7
6
c,
it follows that c =
6
7
.
(b)
6
7
_
1
0
1
x
(x +
x) dx =
18
7
.
(c)
F
r
(y) = P(Y y)
= P(X
y)
=
6
7
_

y
0
(x +
x) dx,
and so
f
Y
(y) =
_
3
7
(1 +y
1
4
) if y (0, 1),
0 otherwise.
2. (a) From
_
1
0
f(x) dx = 1 we get 2a + 2b = 1 and from
_
1
0
xf(x) dx =
7
6
we get 2a +
8
3
b =
7
6
.
The two equations give a = b =
1
4
.
(b) E(X
2
) =
_
1
0
x
2
f(x) dx =
5
3
and so Var(X) =
5
3
(
7
6
)
2
=
11
36
.
3. (a)
1
6
.
(b) Let T be the time of the call, from 7pm, in minutes; T is uniform on [0, 120]. Thus,
P(T 100[T 90) =
1
3
.
(c) We have
M =
_
_
0 if 0 T 60,
T 60 if 60 T 90,
30 if 90 T.
Then,
EM =
1
120
_
90
60
(t 60) dx +
1
120
_
120
90
30 dx = 11.25.
4. (a) P(win a single game) =
3
4
. If you play n times, the number X of games you win is
Binomial(n,
3
4
). If Z is N(0, 1), then
P(X 250) P
_
_
Z
250 n
3
4
_
n
3
4

1
4
_
_
.
For (a), n = 300 and the above expression is P(Z
10
3
), which is approximately 1 (3.33)
0.0004.
For (b), you need to nd n so that the above expression is 0.99 or so that
_
_
250 n
3
4
_
n
3
4

1
4
_
_
= 0.01.
The argument must be negative, hence
250 n
3
4
_
n
3
4

1
4
= 2.33.
If x =
3n, this yields

x
2
4.66x 1000 = 0
and solving the quadratic equation gives x 34.04, n > (34.04)
2
/3, n 387.
5. (a) M is Binomial(n,
1
6
), so EM =
n
6
.
(b)
P
_
M
n
6
<
n
6
0.1
_
P
_
_
[Z[ <
n
6
0.1
_
n
1
6

5
6
_
_
= 2
_
0.1
5
_
1.
(c) The above must be 0.99 and so
_
0.1
5
_
= 0.995,
0.1
5
= 2.57, and, nally, n 3303.
7 JOINT DISTRIBUTIONS AND INDEPENDENCE 70
7 Joint Distributions and Independence
Discrete Case
Assume that you have a pair (X, Y ) of discrete random variables X and Y . Their joint probability
mass function is given by
p(x, y) = P(X = x, Y = y)
so that
P((X, Y ) A) =
(x,y)A
p(x, y).
The marginal probability mass functions are the p. m. f.s of X and Y , given by
P(X = x) =
y
P(X = x, Y = y) =
y
p(x, y)
P(Y = y) =
x
P(X = x, Y = y) =
x
p(x, y)
Example 7.1. An urn has 2 red, 5 white, and 3 green balls. Select 3 balls at random and let
X be the number of red balls and Y the number of white balls. Determine (a) joint p. m. f. of
(X, Y ), (b) marginal p. m. f.s, (c) P(X Y ), and (d) P(X = 2[X Y ).
The joint p. m. f. is given by P(X = x, Y = y) for all possible x and y. In our case, x can
be 0, 1, or 2 and y can be 0, 1, 2, or 3. The values are given in the table.
yx 0 1 2 P(Y = y)
0 1/120 2 3/120 3/120 10/120
1 5 3/120 2 5 3/120 5/120 50/120
2 10 3/120 10 2/120 0 50/120
3 10/120 0 0 10/120
P(X = x) 56/120 56/120 8/120 1
The last row and column entries are the respective column and row sums and, therefore,
determine the marginal p. m. f.s. To answer (c) we merely add the relevant probabilities,
P(X Y ) =
1 + 6 + 3 + 30 + 5
120
=
3
8
,
and, to answer (d), we compute
P(X = 2, X Y )
P(X Y )
=
8
120
3
8
=
8
45
.
Two random variables X and Y are independent if
P(X A, Y B) = P(X A)P(Y B)
for all intervals A and B. In the discrete case, X and Y are independent exactly when
P(X = x, Y = y) = P(X = x)P(Y = y)
for all possible values x and y of X and Y , that is, the joint p. m. f. is the product of the
marginal p. m. f.s.
Example 7.2. In the previous example, are X and Y independent?
No, the 0s in the table are dead giveaways. For example, P(X = 2, Y = 2) = 0, but neither
P(X = 2) nor P(Y = 2) is 0.
Example 7.3. Most often, independence is an assumption. For example, roll a die twice and
let X be the number on the rst roll and let Y be the number on the second roll. Then, X and
Y are independent: we are used to assuming that all 36 outcomes of the two rolls are equally
likely, which is the same as assuming that the two random variables are discrete uniform (on
1, 2, . . . , 6) and independent.
Continuous Case
We say that (X, Y ) is a jointly continuous pair of random variables if there exists a joint density
f(x, y) 0 so that
P((X, Y ) S) =
__
S
f(x, y) dxdy,
where S is some nice (say, open or closed) subset of R
2
Example 7.4. Let (X, Y ) be a random point in S, where S is a compact (that is, closed and
bounded) subset of R
2
. This means that
f(x, y) =
_
1
area(S)
if (x, y) S,
0 otherwise.
The simplest example is a square of side length 1 where
f(x, y) =
_
1 if 0 x 1 and 0 y 1,
0 otherwise.
Example 7.5. Let
f(x, y) =
_
c x
2
y if x
2
y 1,
0 otherwise.
Determine (a) the constant c, (b) P(X Y ), (c) P(X = Y ), and (d) P(X = 2Y ).
For (a),
_
1
1
dx
_
1
x
2
c x
2
y dy = 1
c
4
21
= 1
and so
c =
21
4
.
For (b), let S be the region between the graphs y = x
2
and y = x, for x (0, 1). Then,
P(X Y ) = P((X, Y ) S)
=
_
1
0
dx
_
x
x
2
21
4
x
2
y dy
=
3
20
Both probabilities in (c) and (d) are 0 because a two-dimensional integral over a line is 0.
If f is the joint density of (X, Y ), then the two marginal densities, which are the densities of X
and Y , are computed by integrating out the other variable:
f
X
(x) =
_

f(x, y) dy,
f
Y
(y) =
_

f(x, y) dx.
Indeed, for an interval A, X A means that (X, Y ) S, where S = AR, and, therefore,
P(X A) =
_
A
dx
_

f(x, y) dy.
The marginal densities formulas follow from the denition of density. With some advanced
calculus expertise, the following can be checked.
Two jointly continuous random variables X and Y are independent exactly when the joint
density is the product of the marginal ones:
f(x, y) = f
X
(x) f
Y
(y),
for all x and y.
Example 7.6. Previous example, continued. Compute marginal densities and determine
whether X and Y are independent.
We have
f
X
(x) =
_
1
x
2
21
4
x
2
y dy =
21
8
x
2
(1 x
4
),
for x [1, 1], and 0 otherwise. Moreover,
f
Y
(y) =
_

y
y
21
4
x
2
y dx =
7
2
y
5
2
,
where y [0, 1], and 0 otherwise. The two random variables X and Y are clearly not indepen-
dent, as f(x, y) ,= f
X
(x)f
Y
(y).
Example 7.7. Let (X, Y ) be a random point in a square of length 1 with the bottom left corner
at the origin. Are X and Y independent?
f(x, y) =
_
1 (x, y) [0, 1] [0, 1],
0 otherwise.
The marginal densities are
f
X
(x) = 1,
if x [0, 1], and
f
Y
(y) = 1,
if y [0, 1], and 0 otherwise. Therefore, X and Y are independent.
Example 7.8. Let (X, Y ) be a random point in the triangle (x, y) : 0 y x 1. Are X
and Y independent?
Now
f(x, y) =
_
2 0 y x 1
0, otherwise.
The marginal densities are
f
X
(x) = 2x,
if x [0, 1], and
f
Y
(y) = 2(1 y),
if y [0, 1], and 0 otherwise. So X and Y are no longer distributed uniformly and no longer
independent.
We can make a more general conclusion from the last two examples. Assume that (X, Y ) is
a jointly continuous pair of random variables, uniform on a compact set S R
2
. If they are to
be independent, their marginal densities have to be constant, thus uniform on some sets, say A
and B, and then S = A B. (If A and B are both intervals, then S = A B is a rectangle,
which is the most common example of independence.)
Example 7.9. Mr. and Mrs. Smith agree to meet at a specied location between 5 and 6
p.m. Assume that they both arrive there at a random time between 5 and 6 and that their
arrivals are independent. (a) Find the density for the time one of them will have to wait for
the other. (b) Mrs. Smith later tells you she had to wait; given this information, compute the
probability that Mr. Smith arrived before 5:30.
Let X be the time when Mr. Smith arrives and let Y be the time when Mrs. Smith arrives,
with the time unit 1 hour. The assumptions imply that (X, Y ) is uniform on [0, 1] [0, 1].
For (a), let T = [X Y [, which has possible values in [0, 1]. So, x t [0, 1] and compute
(drawing a picture will also help)
P(T t) = P([X Y [ t)
= P(t X Y t)
= P(X t Y X +t)
= 1 (1 t)
2
= 2t t
2
,
and so
f
T
(t) = 2 2t.
For (b), we need to compute
P(X 0.5[X > Y ) =
P(X 0.5, X > Y )
P(X > Y )
=
1
8
1
2
=
1
4
.
Example 7.10. Assume that X and Y are independent, that X is uniform on [0, 1], and that
Y has density f
Y
(y) = 2y, for y [0, 1], and 0 elsewhere. Compute P(X +Y 1).
The assumptions determine the joint density of (X, Y )
f(x, y) =
_
2y if (x, y) [0, 1] [0, 1],
0 otherwise.
To compute the probability in question we compute
_
1
0
dx
_
1x
0
2y dy
or
_
1
0
dy
_
1y
0
2y dx,
whichever double integral is easier. The answer is
1
3
.
Example 7.11. Assume that you are waiting for two phone calls, from Alice and from Bob.
The waiting time T
1
for Alices call has expectation 10 minutes and the waiting time T
2
for
Bobs call has expectation 40 minutes. Assume T
1
and T
2
are independent exponential random
variables. What is the probability that Alices call will come rst?
We need to compute P(T
1
< T
2
). Assuming our unit is 10 minutes, we have, for t
1
, t
2
> 0,
f
T
1
(t
1
) = e
t
1
and
f
T
2
(t
2
) =
1
4
e
t
2
/4
,
so that the joint density is
f(t
1
, t
2
) =
1
4
e
t
1
t
2
/4
,
for t
1
, t
2
0. Therefore,
P(T
1
< T
2
) =
_

0
dt
1
_

t
1
1
4
e
t
1
t
2
/4
dt
2
=
_

0
e
t
1
dt
1
e
t
1
/4
=
_

0
e
5t
1
4
dt
1
=
4
5
.
Example 7.12. Buon needle problem. Parallel lines at a distance 1 are drawn on a large sheet
of paper. Drop a needle of length onto the sheet. Compute the probability that it intersects
one of the lines.
Let D be the distance from the center of the needle to the nearest line and let be the
acute angle relative to the lines. We will, reasonably, assume that D and are independent
and uniform on their respective intervals 0 D
1
2
and 0

2
. Then,
P(the needle intersects a line) = P
_
D
sin
<

2
_
= P
_
D <

2
sin
_
.
Case 1: 1. Then, the probability equals
_
/2
0
2
sin d
/4
=
/2
/4
=
2
.
When = 1, you famously get
2
, which can be used to get (very poor) approximations for .

Case 2: > 1. Now, the curve d =

2
sin intersects d =
1
2
at = arcsin
1
. The probability
equals
4
2
_
arcsin
1
0
sin d +
_
2
arcsin
1
1
2
_
=
4
2

1
2
_
2
1 +

4

1
2
arcsin
1
_
.
A similar approach works for cases with more than two random variables. Let us do an
example for illustration.
Example 7.13. Assume X
1
, X
2
, X
3
are uniform on [0, 1] and independent. What is P(X
1
+
X
2
+X
3
1)?
The joint density is
f
X
1
,X
2
,X
3
(x
1
, x
2
, x
3
) = f
X
1
(x
1
)f
X
2
(x
2
)f
X
3
(x
3
) =
_
1 if (x
1
, x
2
, x
3
) [0, 1]
3
,
0 otherwise.
Here is how we get the answer:
P(X
1
+X
2
+X
3
1) =
_
1
0
dx
1
_
1x
1
0
dx
2
_
1x
1
x
2
0
dx
3
=
1
6
.
In general, if X
1
, . . . , X
n
are independent and uniform on [0, 1],
P(X
1
+. . . +X
n
1) =
1
n!
Conditional distributions
The conditional p. m. f. of X given Y = y is, in the discrete case, given simply by
p
X
(x[Y = y) = P(X = x[Y = y) =
P(X = x, Y = y)
P(Y = y)
.
This is trickier in the continuous case, as we cannot divide by P(Y = y) = 0.
For a jointly continuous pair of random variables X and Y , we dene the conditional density of
X given Y = y as follows:
f
X
(x[Y = y) =
f(x, y)
f
Y
(y)
,
where f(x, y) is, of course, the joint density of (X, Y ).
Observe that when f
Y
(y) = 0,
_
f(x, y) dx = 0, and so f(x, y) = 0 for every x. So, we

have a
0
0
expression, which we dene to be 0.
Here is a physicists proof why this should be the conditional density formula:
P(X = x +dx[Y = y +dy) =
P(X = x +dx, Y = y +dy)
P(Y = y +dy)
=
f(x, y) dxdy
f
Y
(y) dy
=
f(x, y)
f
Y
(y)
dx
= f
X
(x[Y = y) dx .
Example 7.14. Let (X, Y ) be a random point in the triangle (x, y) : x, y 0, x + y 1.
Compute f
X
(x[Y = y).
The joint density f(x, y) equals 2 on the triangle. For a given y [0, 1], we know that, if
Y = y, X is between 0 and 1 y. Moreover,
f
Y
(y) =
_
1y
0
2 dx = 2(1 y).
Therefore,
f
X
(x[Y = y) =
_
1
1y
0 x 1 y,
0 otherwise.
In other words, given Y = y, X is distributed uniformly on [0, 1 y], which is hardly surprising.
Example 7.15. Suppose (X, Y ) has joint density
f(x, y) =
_
21
4
x
2
y x
2
y 1,
0 otherwise.
Compute f
X
(x[Y = y).
We compute rst
f
Y
(y) =
21
4
y
_

y
y
x
2
dx =
7
2
y
5
2
,
for y [0, 1]. Then,
f
X
(x[Y = y) =
21
4
x
2
y
7
2
y
5/2
=
3
2
x
2
y
3/2
,
where
y x
y.
Suppose we are asked to compute P(X Y [Y = y). This makes no literal sense because
the probability P(Y = y) of the condition is 0. We reinterpret this expression as
P(X y[Y = y) =
_

y
f
X
(x[Y = y) dx,
which equals
_

y
y
3
2
x
2
y
3/2
dx =
1
2
y
3/2
_
y
3/2
y
3
_
=
1
2
_
1 y
3/2
_
.
Problems
1. Let (X, Y ) be a random point in the square (x, y) : 1 x, y 1. Compute the conditional
probability P(X 0 [ Y 2X). (It may be a good idea to draw a picture and use elementary
geometry, rather than calculus.)
2. Roll a fair die 3 times. Let X be the number of 6s obtained and Y the number of 5s.
(a) Compute the joint probability mass function of X and Y .
(b) Are X and Y independent?
3. X and Y are independent random variables and they have the same density function
f(x) =
_
c(2 x) x (0, 1)
0 otherwise.
(a) Determine c. (b) Compute P(Y 2X) and P(Y < 2X).
4. Let X and Y be independent random variables, both uniformly distributed on [0, 1]. Let
Z = min(X, Y ) be the smaller value of the two.
(a) Compute the density function of Z.
(b) Compute P(X 0.5[Z 0.5).
(c) Are X and Z independent?
5. The joint density of (X, Y ) is given by
_
f(x, y) = 3x if 0 y x 1,
0 otherwise.
(a) Compute the conditional density of Y given X = x.
(b) Are X and Y independent?
1. After noting the relevant areas,
P(X 0[Y 2X) =
P(X 0, Y 2X)
P(Y 2X)
=
1
4
_
2
1
2

1
2
1
_
1
2
=
7
8
2. (a) The joint p. m. f. is given by the table
yx 0 1 2 3 P(Y = y)
0
4
3
216
34
2
216
34
216
1
216
125
216
1
34
2
216
324
216
3
216
0
75
216
2
34
216
3
216
0 0
15
216
3
1
216
0 0 0
1
216
P(X = x)
125
216
75
216
15
216
1
216
1
Alternatively, for x, y = 0, 1, 2, 3 and x +y 3,
P(X = x, Y = y) =
_
3
x
__
3 x
y
__
1
6
_
x+y
_
4
6
_
3xy
(b) No. P(X = 3, Y = 3) = 0 and P(X = 3)P(Y = 3) ,= 0.
3. (a) From
c
_
1
0
(2 x) dx = 1,
it follows that c =
2
3
.
(b) We have
P(Y 2X) = P(Y < 2X)
=
_
1
0
dy
_
1
y/2
4
9
(2 x)(2 y) dx
=
4
9
_
1
0
(2 y) dy
_
1
y/2
(2 x) dx
=
4
9
_
1
0
(2 y) dy
_
2
_
1
y
2
_
1
2
_
1
y
2
4
__
=
4
9
_
1
0
(2 y)
_
3
2
y +
y
2
8
_
dy
=
4
9
_
1
0
_
3
7
2
y +
5
4
y
2
y
3
8
_
dy
=
4
9
_
3
7
4
+
5
12

1
32
_
.
4. (a) For z [0, 1]
P(Z z) = 1 P(both X and Y are above z) = 1 (1 z)
2
= 2z z
2
,
so that
f
Z
(z) = 2(1 z),
for z [0, 1] and 0 otherwise.
(b) From (a), we conclude that P(Z 0.5) =
3
4
and P(X 0.5, Z 0.5) = P(X 0.5) =
1
2
,
so the answer is
2
3
.
(c) No: Z X.
5. (a) Assume that x [0, 1]. As
f
X
(x) =
_
x
0
3xdy = 3x
2
,
we have
f
Y
(y[X = x) =
f(x, y)
f
X
(x)
=
3x
3x
2
=
1
x
,
for 0 y x. In other words, Y is uniform on [0, x].
(b) As the answer in (a) depends on x, the two random variables are not independent.
This practice exam covers the material from chapters 5 through 7. Give yourself 50 minutes to
solve the four problems, which you may assume have equal point score.
1. A random variable X has density function
f(x) =
_
c(x +x
2
), x [0, 1],
0, otherwise.
(a) Determine c.
(b) Compute E(1/X).
(c) Determine the probability density function of Y = X
2
.
2. A certain country holds a presidential election, with two candidates running for oce. Not
satised with their choice, each voter casts a vote independently at random, based on the
outcome of a fair coin ip. At the end, there are 4,000,000 valid votes, as well as 20,000 invalid
votes.
(a) Using a relevant approximation, compute the probability that, in the nal count of valid
votes only, the numbers for the two candidates will dier by less than 1000 votes.
(b) Each invalid vote is double-checked independently with probability 1/5000. Using a relevant
approximation, compute the probability that at least 3 invalid votes are double-checked.
3. Toss a fair coin 5 times. Let X be the total number of Heads among the rst three tosses and
Y the total number of Heads among the last three tosses. (Note that, if the third toss comes
out Heads, it is counted both into X and into Y ).
(a) Write down the joint probability mass function of X and Y .
(b) Are X and Y independent? Explain.
(c) Compute the conditional probability P(X 2 [ X Y ).
4. Every working day, John comes to the bus stop exactly at 7am. He takes the rst bus
that arrives. The arrival of the rst bus is an exponential random variable with expectation 20
minutes.
Also, every working day, and independently, Mary comes to the same bus stop at a random
time, uniformly distributed between 7 and 7:30.
(a) What is the probability that tomorrow John will wait for more than 30 minutes?
(b) Assume day-to-day independence. Consider Mary late if she comes after 7:20. What is the
probability that Mary will be late on 2 or more working days among the next 10 working days?
(c) What is the probability that John and Mary will meet at the station tomorrow?
1. A random variable X has density function
f(x) =
_
c(x +x
2
), x [0, 1],
0, otherwise.
(a) Determine c.
Solution:
Since
1 = c
_
1
0
(x +x
2
) dx,
1 = c
_
1
2
+
1
3
_
,
1 =
5
6
c,
and so c =
6
5
.
(b) Compute E(1/X).
Solution:
E
_
1
X
_
=
6
5
_
1
0
1
x
(x +x
2
) dx
=
6
5
_
1
0
(1 +x) dx
=
6
5
_
1 +
1
2
_
=
9
5
.
(c) Determine the probability density function of Y = X
2
.
Solution:
The values of Y are in [0, 1], so we will assume that y [0, 1]. Then,
F(y) = P(Y y)
= P(X
2
y)
= P(X
y)
=
_

y
0
6
5
(x +x
2
) dx,
and so
f
Y
(y) =
d
dy
F
Y
(y)
=
6
5
(
y +y)
1
2
y
=
3
5
(1 +
y).
2. A certain country holds a presidential election, with two candidates running for oce. Not
satised with their choice, each voter casts a vote independently at random, based on the
outcome of a fair coin ip. At the end, there are 4,000,000 valid votes, as well as 20,000
invalid votes.
(a) Using a relevant approximation, compute the probability that, in the nal count of
valid votes only, the numbers for the two candidates will dier by less than 1000
votes.
Solution:
Let S
n
be the vote count for candidate 1. Thus, S
n
is Binomial(n, p), where n =
4, 000, 000 and p =
1
2
. Then, n S
n
is the vote count for candidate 2.
P([S
n
(n S
n
)[ 1000) = P(1000 2S
n
n 1000)
= P
_
_
500
_
n
1
2

1
2
S
n
n/2
_
n
1
2

1
2
500
_
n
1
2

1
2
_
_
P(0.5 Z 0.5)
= P(Z 0.5) P(Z 0.5)
= P(Z 0.5) (1 P(Z 0.5))
= 2P(Z 0.5) 1
= 2(0.5) 1
2 0.6915 1
= 0.383.
(b) Each invalid vote is double-checked independently with probability 1/5000. Using
a relevant approximation, compute the probability that at least 3 invalid votes are
double-checked.
Solution:
Now, let S
n
be the number of double-checked votes, which is Binomial
_
20000,
1
5000
_
and thus approximately Poisson(4). Then,
P(S
n
3) = 1 P(S
n
= 0) P(S
n
= 1) P(S
n
= 2)
1 e
4
4e
4
4
2
2
e
4
= 1 13e
4
.
3. Toss a fair coin 5 times. Let X be the total number of Heads among the rst three tosses
and Y the total number of Heads among the last three tosses. (Note that, if the third toss
comes out Heads, it is counted both into X and into Y ).
(a) Write down the joint probability mass function of X and Y .
Solution:
P(X = x, Y = y) is given by the table
xy 0 1 2 3
0 1/32 2/32 1/32 0
1 2/32 5/32 4/32 1/32
2 1/32 4/32 5/32 2/32
3 0 1/32 2/32 1/32
To compute these, observe that the number of outcomes is 2
5
= 32. Then,
P(X = 2, Y = 1) = P(X = 2, Y = 1, 3rd coin Heads)
+P(X = 2, Y = 1, 3rd coin Tails)
=
2
32
+
2
32
=
4
32
,
P(X = 2, Y = 2) =
2 2
32
+
1
32
=
5
32
,
P(X = 1, Y = 1) =
2 2
32
+
1
32
=
5
32
,
etc.
(b) Are X and Y independent? Explain.
Solution:
No,
P(X = 0, Y = 3) = 0 ,=
1
8

1
8
= P(X = 0)P(Y = 3).
(c) Compute the conditional probability P(X 2 [ X Y ).
Solution:
P(X 2[X Y ) =
P(X 2, X Y )
P(X Y )
=
1 + 4 + 5 + 1 + 2 + 1
1 + 2 + 1 + 5 + 4 + 1 + 5 + 2 + 1
=
14
22
=
7
11
.
4. Every working day, John comes to the bus stop exactly at 7am. He takes the rst bus that
arrives. The arrival of the rst bus is an exponential random variable with expectation 20
minutes.
Also, every working day, and independently, Mary comes to the same bus stop at a random
time, uniformly distributed between 7 and 7:30.
(a) What is the probability that tomorrow John will wait for more than 30 minutes?
Solution:
Assume that the time unit is 10 minutes. Let T be the arrival time of the bus.
It is Exponential with parameter =
1
2
. Then,
f
T
(t) =
1
2
e
t/2
,
for t 0, and
P(T 3) = e
3/2
.
(b) Assume day-to-day independence. Consider Mary late if she comes after 7:20. What
is the probability that Mary will be late on 2 or more working days among the next
10 working days?
Solution:
Let X be Marys arrival time. It is uniform on [0, 3]. Therefore,
P(X 2) =
1
3
.
The number of late days among 10 days is Binomial
_
10,
1
3
_
and, therefore,
P(2 or more late working days among 10 working days)
= 1 P(0 late) P(1 late)
= 1
_
2
3
_
10
10
1
3
_
2
3
_
9
.
(c) What is the probability that John and Mary will meet at the station tomorrow?
Solution:
We have
f
(T,X)
(t, x) =
1
6
e
t/2
,
for x [0, 3] and t 0. Therefore,
P(X T) =
1
3
_
3
0
dx
_

x
1
2
e
t/2
dt
=
1
3
_
3
0
e
x/2
dx
=
2
3
(1 e
3/2
).
8 MORE ON EXPECTATION AND LIMIT THEOREMS 87
8 More on Expectation and Limit Theorems
Given a pair of random variables (X, Y ) with joint density f and another function g of two
variables,
Eg(X, Y ) =
_ _
g(x, y)f(x, y) dxdy;
if instead (X, Y ) is a discrete pair with joint probability mass function p, then
Eg(X, Y ) =
x,y
g(x, y)p(x, y).
Example 8.1. Assume that two among the 5 items are defective. Put the items in a random
order and inspect them one by one. Let X be the number of inspections needed to nd the rst
defective item and Y the number of additional inspections needed to nd the second defective
item. Compute E[X Y [.
The joint p. m. f. of (X, Y ) is given by the following table, which lists P(X = i, Y = j),
together with [i j[ in parentheses, whenever the probability is nonzero:
ij 1 2 3 4
1 .1 (0) .1 (1) .1 (2) .1 (3)
2 .1 (1) .1 (0) .1 (1) 0
3 .1 (2) .1 (1) 0 0
4 .1 (3) 0 0 0
The answer is, therefore, E[X Y [ = 1.4.
Example 8.2. Assume that (X, Y ) is a random point in the right triangle (x, y) : x, y
0, x +y 1. Compute EX, EY , and EXY .
Note that the density is 2 on the triangle, and so,
EX =
_
1
0
dx
_
1x
0
x2 dy
=
_
1
0
2x(1 x) dx
= 2
_
1
2

1
3
_
=
1
3
,
and, therefore, by symmetry, EY = EX =
1
3
. Furthermore,
E(XY ) =
_
1
0
dx
_
1x
0
xy2 dy
=
_
1
0
2x
(1 x)
2
2
dx
=
_
1
0
(1 x)x
2
dx
=
1
3

1
4
=
1
12
.
Linearity and monotonicity of expectation
Theorem 8.1. Expectation is linear and monotone:
1. For constants a and b, E(aX +b) = aE(X) +b.
2. For arbitrary random variables X
1
, . . . , X
n
whose expected values exist,
E(X
1
+. . . +X
n
) = E(X
1
) +. . . +E(X
n
).
3. For two random variables X Y , we have EX EY .
Proof. We will check the second property for n = 2 and the continuous case, that is,
E(X +Y ) = EX +EY.
This is a consequence of the same property for two-dimensional integrals:
__
(x +y)f(x, y) dxdy =
__
xf(x, y) dxdy +
__
yf(x, y) dxdy.
To prove this property for arbitrary n (and the continuous case), one can simply proceed by
induction.
By the way we dened expectation, the third property is not immediately obvious. However,
it is clear that Z 0 implies EZ 0 and, applying this to Z = Y X, together with linearity,
establishes monotonicity.
We emphasize again that linearity holds for arbitrary random variables which do not need
to be independent! This is very useful. For example, we can often write a random variable X
as a sum of (possibly dependent) indicators, X = I
1
+ + I
n
. An instance of this method is
called the indicator trick.
Example 8.3. Assume that an urn contains 10 black, 7 red, and 5 white balls. Select 5 balls
(a) with and (b) without replacement and let X be the number of red balls selected. Compute
EX.
Let I
i
be the indicator of the event that the ith ball is red, that is,
I
i
= I
{ith ball is red}
=
_
1 if ith ball is red,
0 otherwise.
In both cases, X = I
1
+I
2
+I
3
+I
4
+I
5
.
In (a), X is Binomial(5,
7
22
), so we know that EX = 5
7
22
, but we will not use this knowledge.
Instead, it is clear that
EI
1
= 1 P(1st ball is red) =
7
22
= EI
2
= . . . = EI
5
.
Therefore, by additivity, EX = 5
7
22
.
For (b), one solution is to compute the p. m. f. of X,
P(X = i) =
_
7
i
__
15
5i
_
_
22
5
_ ,
where i = 0, 1, . . . , 5, and then
EX =
5
i=0
i
_
7
i
__
15
5i
_
_
22
5
_ .
However, the indicator trick works exactly as before (the fact that I
i
are now dependent does
not matter) and so the answer is also exactly the same, EX = 5
7
22
.
Example 8.4. Matching problem, revisited. Assume n people buy n gifts, which are then
assigned at random, and let X be the number of people who receive their own gift. What is
EX?
This is another problem very well suited for the indicator trick. Let
I
i
= I
{person i receives own gift}
.
Then,
X = I
1
+I
2
+. . . +I
n
.
Moreover,
EI
i
=
1
n
,
for all i, and so
EX = 1.
Example 8.5. Five married couples are seated around a table at random. Let X be the number
of wives who sit next to their husbands. What is EX?
Now, let
I
i
= I
{wife i sits next to her husband}
.
Then,
X = I
1
+. . . +I
5
,
and
EI
i
=
2
9
,
so that
EX =
10
9
.
Example 8.6. Coupon collector problem, revisited. Sample from n cards, with replacement,
indenitely. Let N be the number of cards you need to sample for a complete collection, i.e., to
get all dierent cards represented. What is EN?
Let N
i
be the number the number of additional cards you need to get the ith new card, after
you have received the (i 1)st new card.
Then, N
1
, the number of cards needed to receive the rst new card, is trivial, as the rst
card you buy is new: N
1
= 1. Afterward, N
2
, the number of additional cards needed to get
the second new card is Geometric with success probability
n1
n
. After that, N
3
, the number of
additional cards needed to get the third new card is Geometric with success probability
n2
n
. In
general, N
i
is geometric with success probability
ni+1
n
, i = 1, . . . , n, and
N = N
1
+. . . +N
n
,
so that
EN = n
_
1 +
1
2
+
1
3
+. . . +
1
n
_
.
Now, we have
n
i=2
1
i

_
n
1
1
x
dx
n1
i=1
1
i
,
by comparing the integral with the Riemman sum at the left and right endpoints in the division
of [1, n] into [1, 2], [2, 3], . . . , [n 1, n], and so
log n
n
i=1
1
i
log n + 1,
which establishes the limit
lim
n
EN
nlog n
= 1.
Example 8.7. Assume that an urn contains 10 black, 7 red, and 5 white balls. Select 5 balls
(a) with replacement and (b) without replacement, and let W be the number of white balls
selected, and Y the number of dierent colors. Compute EW and EY .
We already know that
EW = 5
5
22
in either case.
Let I
b
, I
r
, and I
w
be the indicators of the event that, respectively, black, red, and white balls
are represented. Clearly,
Y = I
b
+I
r
+I
w
,
and so, in the case with replacement
EY = 1
12
5
22
5
+ 1
15
5
22
5
+ 1
17
5
22
5
2.5289,
while in the case without replacement
EY = 1
_
12
5
_
_
22
5
_ + 1
_
15
5
_
_
22
5
_ + 1
_
17
5
_
_
22
5
_ 2.6209.
Expectation and independence
Theorem 8.2. Multiplicativity of expectation for independent factors.
The expectation of the product of independent random variables is the product of their expecta-
tions, i.e., if X and Y are independent,
E[g(X)h(Y )] = Eg(X) Eh(Y )
Proof. For the continuous case,
E[g(X)h(Y )] =
__
g(x)h(y)f(x, y) dxdy
=
__
g(x)h(y)f
X
(x)g
Y
(y) dxdy
=
_
g(x)f
X
(x) dx
_
h(y)f
Y
(y) dy
= Eg(X) Eh(Y )
Example 8.8. Let us return to a random point (X, Y ) in the triangle (x, y) : x, y 0, x+y
1. We computed that E(XY ) =
1
12
and that EX = EY =
1
3
. The two random variables have
E(XY ) ,= EX EY , thus, they cannot be independent. Of course, we already knew that they
were not independent.
If, instead, we pick a random point (X, Y ) in the square (x, y) : 0 x, y 1, X and Y
are independent and, therefore, E(XY ) = EX EY =
1
4
.
Finally, pick a random point (X, Y ) in the diamond of radius 1, that is, in the square with
corners at (0, 1), (1, 0), (0, 1), and (1, 0). Clearly, we have, by symmetry,
EX = EY = 0,
but also
E(XY ) =
1
2
_
1
1
dx
_
1|x|
1+|x|
xy dy
=
1
2
_
1
1
xdx
_
1|x|
1+|x|
y dy
=
1
2
_
1
1
xdx 0
= 0.
This is an example where E(XY ) = EX EY even though X and Y are not independent.
Computing expectation by conditioning
For a pair of random variables (X, Y ), we dene the conditional expectation of Y given
X = x by
E(Y [X = x) =
y
y P(Y = y[X = x) (discrete case),
=
_
y
y f
Y
(y[X = x) (continuous case).
Observe that E(Y [X = x) is a function of x; let us call it g(x) for a moment. We denote
g(X) by E(Y [X). This is the expectation of Y provided the value X is known; note again that
this is an expression dependent on X and so we can compute its expectation. Here is what we
get.
Theorem 8.3. Tower property.
The formula E(E(Y [X)) = EY holds; less mysteriously, in the discrete case
EY =
x
E(Y [X = x) P(X = x),
and, in the continuous case,
EY =
_
E(Y [X = x)f
X
(x) dx.
Proof. To verify this in the discrete case, we write out the expectation inside the sum:
y
yP(Y = y[X = x) P(X = x) =
y
y
P(X = x, Y = y)
P(X = x)
P(X = x)
=
y
yP(X = x, Y = y)
= EY
Example 8.9. Once again, consider a random point (X, Y ) in the triangle (x, y) : x, y
0, x +y 1. Given that X = x, Y is distributed uniformly on [0, 1 x] and so
E(Y [X = x) =
1
2
(1 x).
By denition, E(Y [X) =
1
2
(1 X), and the expectation of
1
2
(1 X) must, therefore, equal the
expectation of Y , and, indeed, it does as EX = EY =
1
3
, as we know.
Example 8.10. Roll a die and then toss as many coins as shown up on the die. Compute the
expected number of Heads.
Let X be the number on the die and let Y be the number of Heads. Fix an x 1, 2, . . . , 6.
Given that X = x, Y is Binomial(x,
1
2
). In particular,
E(Y [X = x) = x
1
2
,
and, therefore,
E(number of Heads) = EY
=
6
x=1
x
1
2
P(X = x)
=
6
x=1
x
1
2

1
6
=
7
4
.
Example 8.11. Here is another job interview question. You die and are presented with three
doors. One of them leads to heaven, one leads to one day in purgatory, and one leads to two
days in purgatory. After your stay in purgatory is over, you go back to the doors and pick again,
but the doors are reshued each time you come back, so you must in fact choose a door at
random each time. How long is your expected stay in purgatory?
Code the doors 0, 1, and 2, with the obvious meaning, and let N be the number of days in
purgatory.
Then
E(N[your rst pick is door 0) = 0,
E(N[your rst pick is door 1) = 1 +EN,
E(N[your rst pick is door 2) = 2 +EN.
Therefore,
EN = (1 +EN)
1
3
+ (2 +EN)
1
3
,
and solving this equation gives
EN = 3.
Covariance
Let X, Y be random variables. We dene the covariance of (or between) X and Y as
Cov(X, Y ) = E((X EX)(Y EY ))
= E(XY (EX) Y (EY ) X +EX EY )
= E(XY ) EX EY EY EX +EX EY
= E(XY ) EX EY.
To summarize, the most useful formula is
Cov(X, Y ) = E(XY ) EX EY.
Note immediately that, if X and Y are independent, then Cov(X, Y ) = 0, but the converse
is false.
Let X and Y be indicator random variables, so X = I
A
and Y = I
B
, for two events A and
B. Then, EX = P(A), EY = P(B), E(XY ) = E(I
AB
) = P(A B), and so
Cov(X, Y ) = P(A B) P(A)P(B) = P(A)[P(B[A) P(B)].
If P(B[A) > P(B), we say the two events are positively correlated and, in this case, the covariance
is positive; if the events are negatively correlated all inequalities are reversed. For general random
variables X and Y , Cov(X, Y ) > 0 intuitively means that, on the average, increasing X will
result in larger Y .
Variance of sums of random variables
Theorem 8.4. Variance-covariance formula:
E(
n
i=1
X
i
)
2
=
n
i=1
EX
2
i
+
i=j
E(X
i
X
j
),
Var(
n
i=1
X
i
) =
n
i=1
Var(X
i
) +
i=j
Cov(X
i
, X
j
).
Proof. The rst formula follows from writing the sum
_
n
i=1
X
i
_
2
=
n
i=1
X
2
i
+
i=j
X
i
X
j
and linearity of expectation. The second formula follows from the rst:
Var(
n
i=1
X
i
) = E
_
n
i=1
X
i
E(
n
i=1
X
i
)
_
2
= E
_
n
i=1
(X
i
EX
i
)
_
2
=
n
i=1
Var(X
i
) +
i=j
E(X
i
EX
i
)E(X
j
EX
j
),
which is equivalent to the formula.
Corollary 8.5. Linearity of variance for independent summands.
If X
1
, X
2
, . . . , X
n
are independent, then Var(X
1
+. . . +X
n
) = Var(X
1
)+. . . +Var(X
n
).
The variance-covariance formula often makes computing variance possible even if the random
variables are not independent, especially when they are indicators.
Example 8.12. Let S
n
be Binomial(n, p). We will now fulll our promise from Chapter 5 and
compute its expectation and variance.
The crucial observation is that S
n
=

n
i=1
I
i
, where I
i
is the indicator I
{ith trial is a success}
.
Therefore, I
i
are independent. Then, ES
n
= np and
Var(S
n
) =
n
i=1
Var(I
i
)
=
n
i=1
(EI
i
(EI
i
)
2
)
= n(p p
2
)
= np(1 p).
Example 8.13. Matching problem, revisited yet again. Recall that X is the number of people
who get their own gift. We will compute Var(X).
Recall also that X =

n
i=1
I
i
, where I
i
= I
{ith person gets own gift}
, so that EI
i
=
1
n
and
E(X
2
) = n
1
n
+
i=j
E(I
i
I
j
)
= 1 +
i=j
1
n(n 1)
= 1 + 1
= 2.
The E(I
i
I
j
) above is the probability that the ith person and jth person both get their own gifts
and, thus, equals
1
n(n1)
. We conclude that Var(X) = 1. (In fact, X is, for large n, very close
to Poisson with = 1.)
Example 8.14. Roll a die 10 times. Let X be the number of 6s rolled and Y be the number
of 5s rolled. Compute Cov(X, Y ).
Observe that X =

10
i=1
I
i
, where I
i
= I
{ith roll is 6}
, and Y =

10
i=1
J
i
, where J
i
= I
{ith roll is 5}
.
Then, EX = EY =
10
6
=
5
3
. Moreover, E(I
i
J
j
) equals 0 if i = j (because both 5 and 6 cannot
be rolled on the same roll), and E(I
i
J
j
) =
1
6
2
if i ,= j (by independence of dierent rolls).
Therefore,
EXY =
10
i=1
10
j=1
E(I
i
J
j
)
=
i=j
1
6
2
=
10 9
36
=
5
2
and
Cov(X, Y ) =
5
2

_
5
3
_
2
=
5
18
is negative, as should be expected.
Weak Law of Large Numbers
Assume that an experiment is performed in which an event Ahappens with probability P(A) = p.
At the beginning of the course, we promised to attach a precise meaning to the statement: If
you repeat the experiment, independently, a large number of times, the proportion of times A
happens converges to p. We will now do so and actually prove the more general statement
below.
Theorem 8.6. Weak Law of Large Numbers.
If X, X
1
, X
2
, . . . are independent and identically distributed random variables with nite expec-
tation and variance, then
X
1
+...+X
n
n
converges to EX in the sense that, for any xed > 0,
P
_
X
1
+. . . +X
n
n
EX

_
0,
as n .
In particular, if S
n
is the number of successes in n independent trials, each of which is
a success with probability p, then, as we have observed before, S
n
= I
1
+ . . . + I
n
, where
I
i
= I
{success at trial i}
. So, for every > 0,
P
_
S
n
n
p

_
0,
as n . Thus, the proportion of successes converges to p in this sense.
Theorem 8.7. Markov Inequality. If X 0 is a random variable and a > 0, then
P(X a)
1
a
EX.
Example 8.15. If EX = 1 and X 0, it must be that P(X 10) 0.1.
Proof. Here is the crucial observation:
I
{Xa}

1
a
X.
Indeed, if X < a, the left-hand side is 0 and the right-hand side is nonnegative; if X a, the
left-hand side is 1 and the right-hand side is at least 1. Taking the expectation of both sides,
we get
P(X a) = E(I
{Xa}
)
1
a
EX.
Theorem 8.8. Chebyshev inequality. If EX = and Var(X) =
2
are both nite and k > 0,
then
P([X [ k)

2
k
2
.
Example 8.16. If EX = 1 and Var(X) = 1, P(X 10) P([X 1[ 9)
1
81
.
Example 8.17. If EX = 1, Var(X) = 0.1,
P([X 1[ 0.5)
0.1
0.5
2
=
2
5
.
As the previous two examples show, the Chebyshev inequality is useful if either is small
or k is large.
Proof. By the Markov inequality,
P([X [ k) = P((X )
2
k
2
)
1
k
2
E(X )
2
=
1
k
2
Var(X).
We are now ready to prove the Weak Law of Large Numbers.
Proof. Denote = EX,
2
= Var(X), and let S
n
= X
1
+. . . +X
n
. Then,
ES
n
= EX
1
+. . . +EX
n
= n
and
Var(S
n
) = n
2
.
Therefore, by the Chebyshev inequality,
P([S
n
n[ n)
n
2
n
2
2
=

2
n
2
0,
as n .
A careful examination of the proof above will show that it remains valid if depends on n,
but goes to 0 slower than
1
n
. This suggests that
X
1
+...+Xn
n
converges to EX at the rate of
about
1
n
. We will make this statement more precise below.
Central Limit Theorem
Theorem 8.9. Central limit theorem.
Assume that X, X
1
, X
2
, . . . are independent, identically distributed random variables, with nite
= EX and
2
= V ar(X). Then,
P
_
X
1
+. . . +X
n
n
n
x
_
P(Z x),
as n , where Z is standard Normal.
We will not prove this theorem in full detail, but will later give good indication as to why
it holds. Observe, however, that it is a remarkable theorem: the random variables X
i
have an
arbitrary distribution (with given expectation and variance) and the theorem says that their
sum approximates a very particular distribution, the normal one. Adding many independent
copies of a random variable erases all information about its distribution other than expectation
and variance!
On the other hand, the convergence is not very fast; the current version of the celebrated
Berry-Esseen theorem states that an upper bound on the dierence between the two probabilities
in the Central limit theorem is
0.4785
E[X [
3
n
.
Example 8.18. Assume that X
n
are independent and uniform on [0, 1]. Let S
n
= X
1
+. . .+X
n
.
(a) Compute approximately P(S
200
90). (b) Using the approximation, nd n so that P(S
n

50) 0.99.
We know that EX
i
=
1
2
and Var(X
i
) =
1
3

1
4
=
1
12
.
For (a),
P(S
200
90) = P
_
_
S
200
200
1
2
_
200
1
12
90 200
1
2
_
200
1
12
_
_
P(Z
6)
= 1 P(Z
6)
1 0.993
= 0.007
For (b), we rewrite
P
_
_
S
n
n
1
2
_
n
1
12
50 n
1
2
_
n
1
12
_
_
= 0.99,
and then approximate
P
_
_
Z
n
2
50
_
n
1
12
_
_
= 0.99,
or
P
_
_
Z
n
2
50
_
n
1
12
_
_
=
_
_
n
2
50
_
n
1
12
_
_
= 0.99.
Using the fact that (z) = 0.99 for (approximately) z = 2.326, we get that
n 1.345
n 100 = 0
and that n = 115.
Example 8.19. A casino charges $1 for entrance. For promotion, they oer to the rst 30,000
guests the following game. Roll a fair die:
if you roll 6, you get free entrance and $2;
if you roll 5, you get free entrance;
otherwise, you pay the normal fee.
Compute the number s so that the revenue loss is at most s with probability 0.9.
In symbols, if L is lost revenue, we need to nd s so that
P(L s) = 0.9.
We have L = X
1
+ + X
n
, where n = 30, 000, X
i
are independent, and P(X
i
= 0) =
4
6
,
P(X
i
= 1) =
1
6
, and P(X
i
= 3) =
1
6
. Therefore,
EX
i
=
2
3
and
Var(X
i
) =
1
6
+
9
6

_
2
3
_
2
=
11
9
.
Therefore,
P(L s) = P
_
_
L
2
3
n
_
n
11
9
s
2
3
n
_
n
11
9
_
_
P
_
_
Z
s
2
3
n
_
n
11
9
_
_
= 0.9,
which gives
s
2
3
n
_
n
11
9
1.28,
and nally,
s
2
3
n + 1.28
_
n
11
9
20, 245.
Problems
1. An urn contains 2 white and 4 black balls. Select three balls in three successive steps without
replacement. Let X be the total number of white balls selected and Y the step in which you
selected the rst black ball. For example, if the selected balls are white, black, black, then
X = 1, Y = 2. Compute E(XY ).
2. The joint density of (X, Y ) is given by
_
f(x, y) = 3x if 0 y x 1,
0 otherwise.
Compute Cov(X, Y ).
3. Five married couples are seated at random in a row of 10 seats.
(a) Compute the expected number of women that sit next to their husbands.
(b) Compute the expected number of women that sit next to at least one man.
4. There are 20 birds that sit in a row on a wire. Each bird looks left or right with equal
probability. Let N be the number of birds not seen by any neighboring bird. Compute EN.
5. Recall that a full deck of cards contains 52 cards, 13 cards of each of the four suits. Distribute
the cards at random to 13 players, so that each gets 4 cards. Let N be the number of players
whose four cards are of the same suit. Using the indicator trick, compute EN.
6. Roll a fair die 24 times. Compute, using a relevant approximation, the probability that the
sum of numbers exceeds 100.
1. First we determine the joint p. m. f. of (X, Y ). We have
P(X = 0, Y = 1) = P(bbb) =
1
5
,
P(X = 1, Y = 1) = P(bwb or bbw) =
2
5
,
P(X = 1, Y = 2) = P(wbb) =
1
5
,
P(X = 2, Y = 1) = P(bww) =
1
15
,
P(X = 2, Y = 2) = P(wbw) =
1
15
,
P(X = 2, Y = 3) = P(wwb) =
1
15
,
so that
E(XY ) = 1
2
5
+ 2
1
5
+ 2
1
15
+ 4
1
15
+ 6
1
15
=
8
5
.
2. We have
EX =
_
1
0
dx
_
x
0
x 3xdy =
_
1
0
3x
3
dx =
3
4
,
EX =
_
1
0
dx
_
x
0
y 3xdy =
_
1
0
3
2
x
3
dx =
3
8
,
E(XY ) =
_
1
0
dx
_
x
0
xy 3xdy =
_
1
0
3
2
x
4
dx =
3
10
,
so that Cov(X, Y ) =
3
10

3
4

3
8
=
3
160
.
3. (a) Let the number be M. Let I
i
= I
{couple i sits together}
. Then
EI
i
=
9! 2!
10!
=
1
5
,
and so
EM = EI
1
+. . . +EI
5
= 5 EI
1
= 1.
(b) Let the number be N. Let I
i
= I
{woman i sits next to a man}
. Then, by dividing into cases,
whereby the woman either sits on one of the two end chairs or on one of the eight middle chairs,
EI
i
=
2
10

5
9
+
8
10

_
1
4
9

3
8
_
=
7
9
,
and so
EM = EI
1
+. . . +EI
5
= 5 EI
1
=
35
9
.
4. For the two birds at either end, the probability that it is not seen is
1
2
, while for any other
bird this probability is
1
4
. By the indicator trick
EN = 2
1
2
+ 18
1
4
=
11
2
.
5. Let
I
i
= I
{player i has four cards of the same suit}
,
so that N = I
1
+... +I
13
. Observe that:
the number of ways to select 4 cards from a 52 card deck is
_
52
4
_
;
the number of choices of a suit is 4; and
after choosing a suit, the number of ways to select 4 cards of that suit is
_
13
4
_
.
Therefore, for all i,
EI
i
=
4
_
13
4
_
_
52
4
_
and
EN =
4
_
13
4
_
_
52
4
_ 13.
6. Let X
1
, X
2
, . . . be the numbers on successive rolls and S
n
= X
1
+ . . . + X
n
the sum. We
know that EX
i
=
7
2
, and Var(X
i
) =
35
12
. So, we have
P(S
24
100) = P
_
_
S
24
24
7
2
_
24
35
12
100 24
7
2
_
24
35
12
_
_
P(Z 1.85) = 1 (1.85) 0.032.
Interlude: Practice Final
This practice exam covers the material from chapters 1 through 8. Give yourself 120 minutes
to solve the six problems, which you may assume have equal point score.
1. Recall that a full deck of cards contains 13 cards of each of the four suits (, , , ). Select
cards from the deck at random, one by one, without replacement.
(a) Compute the probability that the rst four cards selected are all hearts ().
(b) Compute the probability that all suits are represented among the rst four cards selected.
(c) Compute the expected number of dierent suits among the rst four cards selected.
(d) Compute the expected number of cards you have to select to get the rst hearts card.
2. Eleven Scandinavians: 2 Swedes, 4 Norwegians, and 5 Finns are seated in a row of 11 chairs
at random.
(a) Compute the probability that all groups sit together (i.e., the Swedes occupy adjacent seats,
as do the Norwegians and Finns).
(b) Compute the probability that at least one of the groups sits together.
(c) Compute the probability that the two Swedes have exactly one person sitting between them.
3. You have two fair coins. Toss the rst coin three times and let X be the number of Heads.
Then toss the second coin X times, that is, as many times as the number of Heads in the rst
coin toss. Let Y be the number of Heads in the second coin toss. (For example, if X = 0, Y is
automatically 0; if X = 2, toss the second coin twice and count the number of Heads to get Y .)
(a) Determine the joint probability mass function of X and Y , that is, write down a formula for
P(X = i, Y = j) for all relevant i and j.
(b) Compute P(X 2 [ Y = 1).
4. Assume that 2, 000, 000 single male tourists visit Las Vegas every year. Assume also that
each of these tourists, independently, gets married while drunk with probability 1/1, 000, 000.
(a) Write down the exact probability that exactly 3 male tourists will get married while drunk
next year.
(b) Compute the expected number of such drunk marriages in the next 10 years.
(c) Write down a relevant approximate expression for the probability in (a).
(d) Write down an approximate expression for the probability that there will be no such drunk
marriage during at least one of the next 3 years.
5. Toss a fair coin twice. You win $2 if both tosses comes out Heads, lose $1 if no toss comes
out Heads, and win or lose nothing otherwise.
(a) What is the expected number of games you need to play to win once?
(b) Assume that you play this game 500 times. What is, approximately, the probability that
you win at least $135?
(c) Again, assume that you play this game 500 times. Compute (approximately) the amount
of money x such that your winnings will be at least x with probability 0.5. Then, do the same
with probability 0.9.
6. Two random variables X and Y are independent and have the same probability density
function
g(x) =
_
c(1 +x) x [0, 1],
0 otherwise.
(a) Find the value of c. Here and in (b): use
_
1
0
x
n
dx =
1
n+1
, for n > 1.
(b) Find Var(X +Y ).
(c) Find P(X +Y < 1) and P(X +Y 1). Here and in (d): when you get to a single integral
involving powers, stop.
(d) Find E[X Y [.
Solutions to Practice Final
1. Recall that a full deck of cards contains 13 cards of each of the four suits (, , , ).
Select cards from the deck at random, one by one, without replacement.
(a) Compute the probability that the rst four cards selected are all hearts ().
Solution:
_
13
4
_
_
52
4
_
(b) Compute the probability that all suits are represented among the rst four cards
selected.
Solution:
13
4
_
52
4
_
(c) Compute the expected number of dierent suits among the rst four cards selected.
Solution:
If X is the number of suits represented, then X = I
+ I
+ I
+ I
, where I
=
I
{ is represented}
, etc. Then,
EI
= 1
_
39
4
_
_
52
4
_,
which is the same for the other three indicators, so
EX = 4 EI
= 4
_
1
_
39
4
_
_
52
4
_
_
.
(d) Compute the expected number of cards you have to select to get the rst hearts card.
Solution:
Label non- cards 1, . . . , 39 and let I
i
= I
{card i selected before any card}
. Then, EI
i
=
1
14
for any i. If N is the number of cards you have to select to get the rst hearts
card, then
EN = E (I
1
+ +EI
39
) =
39
14
.
2. Eleven Scandinavians: 2 Swedes, 4 Norwegians, and 5 Finns are seated in a row of 11
chairs at random.
(a) Compute the probability that all groups sit together (i.e., the Swedes occupy adjacent
seats, as do the Norwegians and Finns).
Solution:
2! 4! 5! 3!
11!
(b) Compute the probability that at least one of the groups sits together.
Solution:
Dene A
S
= Swedes sit together and, similarly, A
N
and A
F
. Then,
P(A
S
A
N
A
F
)
= P(A
S
) +P(A
N
) +P(A
F
)
P(A
S
A
N
) P(A
S
A
F
) P(A
N
A
F
)
+P(A
S
A
N
A
F
)
=
2! 10! + 4! 8! + 5! 7! 2! 4! 7! 2! 5! 6! 4! 5! 3! + 2! 4! 5! 3!
11!
(c) Compute the probability that the two Swedes have exactly one person sitting between
them.
Solution:
The two Swedes may occupy chairs 1, 3; or 2, 4; or 3, 5; ...; or 9, 11. There are exactly
9 possibilities, so the answer is
9
_
11
2
_ =
9
55
.
3. You have two fair coins. Toss the rst coin three times and let X be the number of Heads.
Then, toss the second coin X times, that is, as many times as you got Heads in the rst
coin toss. Let Y be the number of Heads in the second coin toss. (For example, if X = 0,
Y is automatically 0; if X = 2, toss the second coin twice and count the number of Heads
to get Y .)
(a) Determine the joint probability mass function of X and Y , that is, write down a
formula for P(X = i, Y = j) for all relevant i and j.
Solution:
P(X = i, Y = j) = P(X = i)P(Y = j[X = i) =
_
3
i
_
1
2
3

_
i
j
_
1
2
i
,
for 0 j i 3.
(b) Compute P(X 2 [ Y = 1).
Solution:
This equals
P(X = 2, Y = 1) +P(X = 3, Y = 1)
P(X = 1, Y = 1) +P(X = 2, Y = 1) +P(X = 3, Y = 1)
=
5
9
.
4. Assume that 2, 000, 000 single male tourists visit Las Vegas every year. Assume also that
each of these tourists independently gets married while drunk with probability 1/1, 000, 000.
(a) Write down the exact probability that exactly 3 male tourists will get married while
drunk next year.
Solution:
With X equal to the number of such drunk marriages, X is Binomial(n, p) with
p = 1/1, 000, 000 and n = 2, 000, 000, so we have
P(X = 3) =
_
n
3
_
p
3
(1 p)
n3
(b) Compute the expected number of such drunk marriages in the next 10 years.
Solution:
As X is binomial, its expected value is np = 2, so the answer is 10 EX = 20.
(c) Write down a relevant approximate expression for the probability in (a).
Solution:
We use that X is approximately Poisson with = 2, so the answer is
3
3!
e
=
4
3
e
2
.
(d) Write down an approximate expression for the probability that there will be no such
drunk marriage during at least one of the next 3 years.
Solution:
This equals 1P(at least one such marriage in each of the next 3 years), which equals
1
_
1 e
2
_
3
.
5. Toss a fair coin twice. You win $2 if both tosses comes out Heads, lose $1 if no toss comes
out Heads, and win or lose nothing otherwise.
(a) What is the expected number of games you need to play to win once?
Solution:
The probability of winning in
1
4
. The answer, the expectation of a Geometric
_
1
4
_
random variable, is 4.
(b) Assume that you play this game 500 times. What is, approximately, the probability
that you win at least $135?
Solution:
Let X be the winnings in one game and X
1
, X
2
, . . . , X
n
the winnings in successive
games, with S
n
= X
1
+. . . +X
n
. Then, we have
EX = 2
1
4
1
1
4
=
1
4
and
Var(X) = 4
1
4
+ 1
1
4

_
1
4
_
2
=
19
16
.
Thus,
P(S
n
135) = P
_
_
S
n
n
1
4
_
n
19
16
135 n
1
4
_
n
19
16
_
_
P
_
_
Z
135 n
1
4
_
n
19
16
_
_
,
where Z is standard Normal. Using n = 500, we get the answer
1
_
_
10
_
500
19
16
_
_
.
(c) Again, assume that you play this game 500 times. Compute (approximately) the
amount of money x such that your winnings will be at least x with probability 0.5.
Then, do the same with probability 0.9.
Solution:
For probability 0.5, the answer is exactly ES
n
= n
1
4
= 125. For probability 0.9, we
approximate
P(S
n
x) P
_
_
Z
x 125
_
500
19
16
_
_
= P
_
_
Z
125 x
_
500
19
16
_
_
=
_
_
125 x
_
500
19
16
_
_
,
where we have used that x < 125. Then, we use that (z) = 0.9 at z 1.28, leading
to the equation
125 x
_
500
19
16
= 1.28
and, therefore,
x = 125 1.28
_
500
19
16
.
(This gives x = 93.)
6. Two random variables X and Y are independent and have the same probability density
function
g(x) =
_
c(1 +x) x [0, 1],
0 otherwise.
(a) Find the value of c. Here and in (b): use
_
1
0
x
n
dx =
1
n+1
, for n > 1.
Solution:
As
1 = c
_
1
0
(1 +x) dx = c
3
2
,
we have c =
2
3
(b) Find Var(X +Y ).
Solution:
By independence, this equals 2Var(X) = 2(E(X
2
) (EX)
2
). Moreover,
E(X) =
2
3
_
1
0
x(1 +x) dx =
5
9
,
E(X
2
) =
2
3
_
1
0
x
2
(1 +x) dx =
7
18
,
and the answer is
13
81
.
(c) Find P(X + Y < 1) and P(X + Y 1). Here and in (d): when you get to a single
integral involving powers, stop.
Solution:
The two probabilities are both equal to
_
2
3
_
2
_
1
0
dx
_
1x
0
(1 +x)(1 +y) dy =
_
2
3
_
2
_
1
0
(1 +x)
_
(1 x) +
(1 x)
2
2
_
dx.
(d) Find E[X Y [.
Solution:
This equals
_
2
3
_
2
2
_
1
0
dx
_
x
0
(x y)(1 +x)(1 +y) dy =
_
2
3
_
2
2
_
1
0
_
x(1 +x)
_
x +
x
2
2
_
(1 +x)
_
x
2
2
+
x
3
3
__
dx.
9 CONVERGENCE IN PROBABILITY 112
9 Convergence in probability
One of the goals of probability theory is to extricate a useful deterministic quantity out of a
random situation. This is typically possible when a large number of random eects cancel each
other out, so some limit is involved. In this chapter we consider the following setting: given a
sequence of random variables, Y
1
, Y
2
, . . ., we want to show that, when n is large, Y
n
is approx-
imately f(n), for some simple deterministic function f(n). The meaning of approximately is
what we now make clear.
A sequence Y
1
, Y
2
, . . . of random variables converges to a number a in probability if, as n ,
P([Y
n
a[ ) converges to 1, for any xed > 0. This is equivalent to P([Y
n
a[ > ) 0 as
n , for any xed > 0.
Example 9.1. Toss a fair coin n times, independently. Let R
n
be the longest run of Heads,
i.e., the longest sequence of consecutive tosses of Heads. For example, if n = 15 and the tosses
come out
HHTTHHHTHTHTHHH,
then R
n
= 3. We will show that, as n ,
R
n
log
2
n
1,
in probability. This means that, to a rst approximation, one should expect about 20 consecutive
Heads somewhere in a million tosses.
To solve a problem such as this, we need to nd upper bounds for probabilities that R
n
is
large and that it is small, i.e., for P(R
n
k) and P(R
n
k), for appropriately chosen k. Now,
for arbitrary k,
P(R
n
k) = P (k consecutive Heads start at some i, 0 i n k + 1)
= P(
nk+1
_
i=1
i is the rst Heads in a succession of at least k Heads)
n
1
2
k
.
For the lower bound, divide the string of size n into disjoint blocks of size k. There are
n
k
| such blocks (if n is not divisible by k, simply throw away the leftover smaller block at the
end). Then, R
n
k as soon as one of the blocks consists of Heads only; dierent blocks are
independent. Therefore,
P(R
n
< k)
_
1
1
2
k
_
n
k
exp
_
1
2
k
_
n
k
_
_
,
using the famous inequality 1 x e
x
, valid for all x.
Below, we will use the following trivial inequalities, valid for any real number x 2: x|
x 1, x| x + 1, x 1
x
2
, and x + 1 2x.
To demonstrate that
R
n
log
2
n
1, in probability, we need to show that, for any > 0,
P(R
n
(1 +) log
2
n) 0, (1)
P(R
n
(1 ) log
2
n) 0, (2)
as
P
_
R
n
log
2
n
1

_
= P
_
R
n
log
2
n
1 + or
R
n
log
2
n
1
_
= P
_
R
n
log
2
n
1 +
_
+P
_
R
n
log
2
n
1
_
= P (R
n
(1 +) log
2
n) +P (R
n
(1 ) log
2
n) .
A little fussing in the proof comes from the fact that (1 ) log
2
n are not integers. This is
common in such problems. To prove (1), we plug k = (1 + ) log
2
n| into the upper bound to
get
P(R
n
(1 +) log
2
n) n
1
2
(1+) log
2
n1
= n
2
n
1+
=
2
n
0,
as n . On the other hand, to prove (2) we need to plug k = (1 ) log
2
n| + 1 into the
lower bound,
P(R
n
(1 ) log
2
n) P(R
n
< k)
exp
_
1
2
k
_
n
k
_
_
exp
_
1
2
k
_
n
k
1
_
_
exp
_
1
32

1
n
1

n
(1 ) log
2
n
_
= exp
_
1
32
n
(1 ) log
2
n
_
0,
as n , as n
is much larger than log

2
n.
The most basic tool in proving convergence in probability is the Chebyshev inequality: if X
is a random variable with EX = and Var(X) =
2
, then
P([X [ k)

2
k
2
,
for any k > 0. We proved this inequality in the previous chapter and we will use it to prove the
next theorem.
Theorem 9.1. Connection between variance and convergence in probability.
Assume that Y
n
are random variables and that a is a constant such
that
EY
n
a,
Var(Y
n
) 0,
as n . Then,
Y
n
a,
as n , in probability.
Proof. Fix an > 0. If n is so large that
[EY
n
a[ < /2,
then
P([Y
n
a[ > ) P([Y
n
EY
n
[ > /2)
4
Var(Y
n
)
2
0,
as n . Note that the second inequality in the computation is the Chebyshev inequality.
This is most often applied to sums of random variables. Let
S
n
= X
1
+. . . +X
n
,
where X
i
are random variables with nite expectation and variance. Then, without any inde-
pendence assumption,
ES
n
= EX
1
+. . . +EX
n
and
E(S
2
n
) =
n
i=1
EX
2
i
+
i=j
E(X
i
X
j
),
Var(S
n
) =
n
i=1
Var(X
i
) +
i=j
Cov(X
i
, X
j
).
Recall that
Cov(X
1
, X
j
) = E(X
i
X
j
) EX
i
EX
j
and
Var(aX) = a
2
Var(X).
Moreover, if X
i
are independent,
Var(X
1
+. . . +X
n
) = Var(X
1
) +. . . + Var(X
n
).
Continuing with the review, let us reformulate and prove again the most famous convergence in
probability theorem. We will use the common abbreviation i. i. d. for independent identically
distributed random variables.
Theorem 9.2. Weak law of large numbers. Let X, X
1
, X
2
, . . . be i. i. d. random variables with
with EX = and Var(X) =
2
< . Let S
n
= X
1
+. . . +X
n
. Then, as n ,
S
n
n

in probability.
Proof. Let Y
n
=
Sn
n
. We have EY
n
= and
Var(Y
n
) =
1
n
2
Var(S
n
) =
1
n
2
n
2
=

2
n
.
Thus, we can simply apply the previous theorem.
Example 9.2. We analyze a typical investment (the accepted euphemism for gambling on
nancial markets) problem. Assume that you have two investment choices at the beginning of
each year:
a risk-free bond which returns 6% per year; and
a risky stock which increases your investment by 50% with probability 0.8 and wipes it
out with probability 0.2.
Putting an amount s in the bond, then, gives you 1.06s after a year. The same amount in the
stock gives you 1.5s with probability 0.8 and 0 with probability 0.2; note that the expected value
is 0.8 1.5s = 1.2s > 1.06s. We will assume year-to-year independence of the stocks return.
We will try to maximize the return to our investment by hedging. That is, we invest, at
the beginning of each year, a xed proportion x of our current capital into the stock and the
remaining proportion 1 x into the bond. We collect the resulting capital at the end of the
year, which is simultaneously the beginning of next year, and reinvest with the same proportion
x. Assume that our initial capital is x
0
.
It is important to note that the expected value of the capital at the end of the year is
maximized when x = 1, but by using this strategy you will eventually lose everything. Let X
n
be your capital at the end of year n. Dene the average growth rate of your investment as
= lim
n
1
n
log
X
n
x
0
,
so that
X
n
x
0
e
n
.
We will express in terms of x; in particular, we will show that it is a nonrandom quantity.
Let I
i
= I
{stock goes up in year i}
. These are independent indicators with EI
i
= 0.8.
X
n
= X
n1
(1 x) 1.06 +X
n1
x 1.5 I
n
= X
n1
(1.06(1 x) + 1.5x I
n
)
and so we can unroll the recurrence to get
X
n
= x
0
(1.06(1 x) + 1.5x)
S
n
((1 x)1.06)
nS
n
,
where S
n
= I
1
+. . . +I
n
. Therefore,
1
n
log
X
n
x
0
=
S
n
n
log(1.06 + 0.44x) +
_
1
S
n
n
_
log(1.06(1 x))
0.8 log(1.06 + 0.44x) + 0.2 log(1.06(1 x)),
in probability, as n . The last expression denes as a function of x. To maximize this,
we set
d
dx
= 0 to get
0.8 0.44
1.06 + 0.44x
=
0.2
1 x
.
The solution is x =
7
22
, which gives 8.1%.
Example 9.3. Distribute n balls independently at random into n boxes. Let N
n
be the number
of empty boxes. Show that
1
n
N
n
converges in probability and identify the limit.
Note that
N
n
= I
1
+. . . +I
n
,
where I
i
= I
{ith box is empty}
, but we cannot use the weak law of large numbers as I
i
are not
independent. Nevertheless,
EI
i
=
_
n 1
n
_
n
=
_
1
1
n
_
n
,
and so
EN
n
= n
_
1
1
n
_
n
.
Moreover,
E(N
2
n
) = EN
n
+
i=j
E(I
i
I
j
)
with
E(I
i
I
j
) = P(box i and j are both empty) =
_
n 2
n
_
n
,
so that
Var(N
n
) = E(N
2
n
) (EN
n
)
2
= n
_
1
1
n
_
n
+n(n 1)
_
1
2
n
_
n
n
2
_
1
1
n
_
2n
.
Now, let Y
n
=
1
n
N
n
. We have
EY
n
e
1
,
as n , and
Var(Y
n
) =
1
n
_
1
1
n
_
n
+
n 1
n
_
1
2
n
_
n
_
1
1
n
_
2n
0 +e
2
e
2
= 0,
as n . Therefore,
Y
n
=
N
n
n
e
1
,
as n , in probability.
Problems
1. Assume that n married couples (amounting to 2n people) are seated at random on 2n seats
around a round table. Let T be the number of couples that sit together. Determine ET and
Var(T).
2. There are n birds that sit in a row on a wire. Each bird looks left or right with equal
probability. Let N be the number of birds not seen by any neighboring bird. Determine, with
proof, the constant c so that, as n ,
1
n
N c in probability.
3. Recall the coupon collector problem: sample from n cards, with replacement, indenitely, and
let N be the number of cards you need to get so that each of n dierent cards are represented.
Find a sequence a
n
so that, as n , N/a
n
converges to 1 in probability.
4. Kings and Lakers are playing a best of seven playo series, which means they play until
one team wins four games. Assume Kings win every game independently with probability p.
(There is no dierence between home and away games.) Let N be the number of games played.
Compute EN and Var(N).
5. An urn contains n red and m black balls. Select balls from the urn one by one without
replacement. Let X be the number of red balls selected before any black ball, and let Y be the
number of red balls between the rst and the second black one. Compute EX and EY .
1. Let I
i
be the indicator of the event that the ith couple sits together. Then, T = I
1
+ +I
n
.
Moreover,
EI
i
=
2
2n 1
, E(I
i
I
j
) =
2
2
(2n 3)!
(2n 1)!
=
4
(2n 1)(2n 2)
,
for any i and j ,= i. Thus,
ET =
2n
2n 1
and
E(T
2
) = ET +n(n 1)
4
(2n 1)(2n 2)
=
4n
2n 1
,
so
Var(T) =
4n
2n 1

4n
2
(2n 1)
2
=
4n(n 1)
(2n 1)
2
.
2. Let I
i
indicate the event that bird i is not seen by any other bird. Then, EI
i
is
1
2
if i = 1 or
i = n and
1
4
otherwise. It follows that
EN = 1 +
n 2
4
=
n + 2
4
.
Furthermore, I
i
and I
j
are independent if [i j[ 3 (two birds that have two or more birds
between them are observed independently). Thus, Cov(I
i
, I
j
) = 0 if [i j[ 3. As I
i
and I
j
are
indicators, Cov(I
i
, I
j
) 1 for any i and j. For the same reason, Var(I
i
) 1. Therefore,
Var(N) =
i
Var(I
i
) +
i=j
Cov(I
i
, I
j
) n + 4n = 5n.
Clearly, if M =
1
n
N, then EM =
1
n
EN
1
4
and Var(M) =
1
n
2
Var(N) 0. It follows that
c =
1
4
.
3. Let N
i
be the number of coupons needed to get i dierent coupons after having i 1 dierent
ones. Then N = N
1
+ . . . + N
n
, and N
i
are independent Geometric with success probability
ni+1
n
. So,
EN
i
=
n
n i + 1
, Var(N
i
) =
n(i 1)
(n i + 1)
2
,
and, therefore,
EN = n
_
1 +
1
2
+. . . +
1
n
_
,
Var(N) =
n
i=1
n(i 1)
(n i + 1)
2
n
2
_
1 +
1
2
2
+. . . +
1
n
2
_
n
2

2
6
< 2n
2
.
If a
n
= nlog n, then
1
a
n
EN 1,
1
a
2
n
EN 0,
as n , so that
1
a
n
N 1
in probability.
4. Let I
i
be the indicator of the event that the ith game is played. Then, EI
1
= EI
2
= EI
3
=
EI
4
= 1,
EI
5
= 1 p
4
(1 p)
4
,
EI
6
= 1 p
5
5p
4
(1 p) 5p(1 p)
4
(1 p)
5
,
EI
7
=
_
6
3
_
p
3
(1 p)
3
.
Add the seven expectations to get EN. To compute E(N
2
), we use the fact that I
i
I
j
= I
i
if
i > j, so that E(I
i
I
j
) = EI
i
. So,
EN
2
=
i
EI
i
+ 2
i>j
E(I
i
I
j
) =
i
EI
i
+ 2
i
(i 1)EI
i
=
7
i=1
(2i 1)EI
i
,
and the nal result can be obtained by plugging in EI
i
and by the standard formula
Var(N) = E(N
2
) (EN)
2
.
5. Imagine the balls ordered in a row where the ordering species the sequence in which they
are selected. Let I
i
be the indicator of the event that the ith red ball is selected before any black
ball. Then, EI
i
=
1
m+1
, the probability that in a random ordering of the ith red ball and all m
black balls, the red comes rst. As X = I
1
+. . . +I
n
, EX =
n
m+1
.
Now, let J
i
be the indicator of the event that the ith red ball is selected between the rst and
the second black one. Then, EJ
i
is the probability that the red ball is second in the ordering of
the above m+ 1 balls, so EJ
i
= EI
i
, and EY = EX.
10 MOMENT GENERATING FUNCTIONS 120
10 Moment generating functions
If X is a random variable, then its moment generating function is
(t) =
X
(t) = E(e
tX
) =
_
x
e
tx
P(X = x) in the discrete case,
_
e
tx
f
X
(x) dx in the continuous case.
Example 10.1. Assume that X is an Exponential(1) random variable, that is,
f
X
(x) =
_
e
x
x > 0,
0 x 0.
Then,
(t) =
_

0
e
tx
e
x
dx =
1
1 t
,
only when t < 1. Otherwise, the integral diverges and the moment generating function does not
exist. Have in mind that the moment generating function is meaningful only when the integral
(or the sum) converges.
Here is where the name comes from: by writing its Taylor expansion in place of e
tX
and
exchanging the sum and the integral (which can be done in many cases)
E(e
tX
) = E[1 +tX +
1
2
t
2
X
2
+
1
3!
t
3
X
3
+. . .]
= 1 +tE(X) +
1
2
t
2
E(X
2
) +
1
3!
t
3
E(X
3
) +. . .
The expectation of the k-th power of X, m
k
= E(X
k
), is called the k-th moment of x. In
combinatorial language, (t) is the exponential generating function of the sequence m
k
. Note
also that
d
dt
E(e
tX
)[
t=0
= EX,
d
2
dt
2
E(e
tX
)[
t=0
= EX
2
,
which lets us compute the expectation and variance of a random variable once we know its
moment generating function.
Example 10.2. Compute the moment generating function for a Poisson() random variable.
By denition,
(t) =
n=0
e
tn

n
n!
e
= e
n=0
(e
t
)
n
n!
= e
+e
t
= e
(e
t
1)
.
Example 10.3. Compute the moment generating function for a standard Normal random
variable.
By denition,
X
(t) =
1
2
_

e
tx
e
x
2
/2
dx
=
1
2
e
1
2
t
2
_

1
2
(xt)
2
dx
= e
1
2
t
2
,
where, from the rst to the second line, we have used, in the exponent,
tx
1
2
x
2
=
1
2
(2tx +x
2
) =
1
2
((x t)
2
t
2
).
Lemma 10.1. If X
1
, X
2
, . . . , X
n
are independent and S
n
= X
1
+. . . +X
n
, then
Sn
(t) =
X
1
(t) . . .
Xn
(t).
If X
i
is identically distributed as X, then
S
n
(t) = (
X
(t))
n
.
Proof. This follows from multiplicativity of expectation for independent random variables:
E[e
tS
n
] = E[e
tX
1
e
tX
2
. . . e
tX
n
] = E[e
tX
1
] E[e
tX
2
] . . . E[e
tX
n
].
Example 10.4. Compute the moment generating function of a Binomial(n, p) random variable.
Here we have S
n
=

n
k=1
I
k
, where the indicators I
k
are independent and I
k
= I
{success on kth trial}
,
so that
S
n
(t) = (e
t
p + 1 p)
n
.
Why are moment generating functions useful? One reason is the computation of large devi-
ations. Let S
n
= X
1
+ + X
n
, where X
i
are independent and identically distributed as X,
with expectation EX = and moment generating function . At issue is the probability that
S
n
is far away from its expectation n, more precisely, P(S
n
> an), where a > . We can,
of course, use the Chebyshev inequality to get an upper bound of order
1
n
. It turns out that
this probability is, for large n, much smaller; the theorem below gives an upper bound that is a
much better estimate.
Theorem 10.2. Large deviation bound.
Assume that (t) is nite for some t > 0. For any a > ,
P(S
n
an) exp(nI(a)),
where
I(a) = supat log (t) : t > 0 > 0.
Proof. For any t > 0, using the Markov inequality,
P(S
n
an) = P(e
tSntan
1) E[e
tSntan
] = e
tan
(t)
n
= exp (n(at log (t))) .
Note that t > 0 is arbitrary, so we can optimize over t to get what the theorem claims. We need
to show that I(a) > 0 when a > . For this, note that (t) = at log (t) satises (0) = 0
and, assuming that one can dierentiate inside the integral sign (which one can in this case, but
proving this requires abstract analysis beyond our scope),
(t) = a

(t)
(t)
= a
E(Xe
tX
)
(t)
,
and, then,
(0) = a > 0,
so that (t) > 0 for some small enough positive t.
Example 10.5. Roll a fair die n times and let S
n
be the sum of the numbers you roll. Find an
upper bound for the probability that S
n
exceeds its expectation by at least n, for n = 100 and
n = 1000.
We t this into the above theorem: observe that = 3.5 and ES
n
= 3.5 n, and that we need
to nd an upper bound for P(S
n
4.5 n), i.e., a = 4.5. Moreover,
(t) =
1
6
6
i=1
e
it
=
e
t
(e
6t
1)
6(e
t
1)
and we need to compute I(4.5), which, by denition, is the maximum, over t > 0, of the function
4.5 t log (t),
whose graph is in the gure below.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.2
0.15
0.1
0.05
0
0.05
0.1
0.15
0.2
t
It would be nice if we could solve this problem by calculus, but unfortunately we cannot (which
is very common in such problems), so we resort to numerical calculations. The maximum is at
t 0.37105 and, as a result, I(4.5) is a little larger than 0.178. This gives the upper bound
P(S
n
4.5 n) e
0.178n
,
which is about 0.17 for n = 10, 1.83 10
8
for n = 100, and 4.16 10
78
for n = 1000. The bound
35
12n
for the same probability, obtained by the Chebyshev inequality, is much too large for large
n.
Another reason why moment generating functions are useful is that they characterize the
distribution and convergence of distributions. We will state the following theorem without proof.
Theorem 10.3. Assume that the moment generating functions for random variables X, Y , and
X
n
are nite for all t.
1. If
X
(t) =
Y
(t) for all t, then P(X x) = P(Y x) for all x.
2. If
Xn
(t)
X
(t) for all t, and P(X x) is continuous in x, then P(X
n
x) P(X
x) for all x.
Example 10.6. Show that the sum of independent Poisson random variables is Poisson.
Here is the situation. We have n independent random variables X
1
, . . . , X
n
, such that:
X
1
is Poisson(
1
),
X
1
(t) = e
1
(e
t
1)
,
X
2
is Poisson(
2
),
X
2
(t) = e
2
(e
t
1)
,
.
.
.
X
n
is Poisson(
n
),
X
n
(t) = e
n(e
t
1)
.
Therefore,
X
1
+...+X
n
(t) = e
(
1
+...+
n
)(e
t
1)
and so X
1
+ . . . + X
n
is Poisson(
1
+ . . . +
n
). Similarly, one can also prove that the sum of
independent Normal random variables is Normal.
We will now reformulate and prove the Central Limit Theorem in a special case when the
moment generating function is nite. This assumption is not needed and the theorem may be
applied it as it was in the previous chapter.
Theorem 10.4. Assume that X is a random variable, with EX = and Var(X) =
2
, and
assume that
X
(t) is nite for all t. Let S
n
= X
1
+. . . +X
n
, where X
1
, . . . , X
n
are i. i. d. and
distributed as X. Let
T
n
=
S
n
n
n
.
Then, for every x,
P(T
n
x) P(Z x),
as n , where Z is a standard Normal random variable.
Proof. Let Y =
X
and Y
i
=
X
i
. Then, Y
i
are independent, distributed as Y , E(Y
i
) = 0,
Var(Y
i
) = 1, and
T
n
=
Y
1
+. . . +Y
n
n
.
To nish the proof, we show that
T
n
(t)
Z
(t) = exp(t
2
/2) as n :
Tn
(t) = E
_
e
t T
n
= E
_
e
t
n
Y
1
+...+
t
n
Y
n
_
= E
_
e
t
n
Y
1
_
E
_
e
t
n
Y
n
_
= E
_
e
t
n
Y
_
n
=
_
1 +
t
n
EY +
1
2
t
2
n
E(Y
2
) +
1
6
t
3
n
3/2
E(Y
3
) +. . .
_
n
=
_
1 + 0 +
1
2
t
2
n
+
1
6
t
3
n
3/2
E(Y
3
) +. . .
_
n
_
1 +
t
2
2
1
n
_
n
e
t
2
2
.
Problems
1. A player selects three cards at random from a full deck and collects as many dollars as the
number of red cards among the three. Assume 10 people play this game once and let X be the
number of their combined winnings. Compute the moment generating function of X.
2. Compute the moment generating function of a uniform random variable on [0, 1].
3. This exercise was in fact the original motivation for the study of large deviations by the
Swedish probabilist Harald Cram`er, who was working as an insurance company consultant in
the 1930s. Assume that the insurance company receives a steady stream of payments, amounting
to (a deterministic number) per day. Also, every day, they receive a certain amount in claims;
assume this amount is Normal with expectation and variance
2
. Assume also day-to-day
independence of the claims. Regulators require that, within a period of n days, the company
must be able to cover its claims by the payments received in the same period, or else. Intimidated
by the erce regulators, the company wants to fail to satisfy their requirement with probability
less than some small number . The parameters n, , and are xed, but is a quantity the
company controls. Determine .
4. Assume that S is Binomial(n, p). For every a > p, determine by calculus the large deviation
bound for P(S an).
5. Using the central limit theorem for a sum of Poisson random variables, compute
lim
n
e
n
n
i=0
n
i
i!
.
1. Compute the moment generating function for a single game, then raise it to the 10th power:
(t) =
_
1
_
52
3
_
__
26
3
_
+
_
26
1
__
26
2
_
e
t
+
_
26
2
__
26
1
_
e
2t
+
_
26
3
_
e
3t
_
_
10
.
2. Answer: (t) =
_
1
0
e
tx
dx =
1
t
(e
t
1).
3. By the assumption, a claim Y is Normal N(,
2
) and, so, X = (Y )/ is standard normal.
Note that Y = X +. Thus, he combined amount of claims is (X
1
+ +X
n
) +n, where
X
i
are i. i. d. standard Normal, so we need to bound
P(X
1
+ +X
n

n) e
In
.
As log (t) =
1
2
t
2
, we need to maximize, over t > 0,

t
1
2
t
2
,
and the maximum equals
I =
1
2
_

_
2
.
Finally, we solve the equation
e
In
= ,
to get
= +
_
2 log
n
.
4. After a computation, the answer we get is
I(a) = a log
a
p
+ (1 a) log
1 a
1 p
.
5. Let S
n
be the sum of i. i. d. Poisson(1) random variables. Thus, S
n
is Poisson(n) and
ES
n
= n. By the Central Limit Theorem, P(S
n
n)
1
2
, but P(S
n
n) is exactly the
expression in question. So, the answer is
1
2
.
11 COMPUTING PROBABILITIES AND EXPECTATIONS BY CONDITIONING 127
11 Computing probabilities and expectations by conditioning
Conditioning is the method we encountered before; to remind ourselves, it involves two-stage
(or multistage) processes and conditions are appropriate events at the rst stage. Recall also
the basic denitions:
Conditional probability: if A and B are two events, P(A[B) =
P(AB)
P(B)
;
Conditional probability mass function: if (X, Y ) has probability mass function p, p
X
(x[Y =
y) =
p(x,y)
p
Y
(y)
= P(X = x[Y = y);
Conditional density: if (X, Y ) has joint density f, f
X
(x[Y = y) =
f(x,y)
f
Y
(y)
.
Conditional expectation: E(X[Y = y) is either

x
xp
X
(x[Y = y) or
_
xf
X
(x[Y = y) dx
depending on whether the pair (X, Y ) is discrete or continuous.
Bayes formula also applies to expectation. Assume that the distribution of a random
variable X conditioned on Y = y is given, and, consequently, its expectation E(X[Y = y) is
also known. Such is the case of a two-stage process, whereby the value of Y is chosen at the
rst stage, which then determines the distribution of X at the second stage. This situation is
very common in applications. Then,
E(X) =
_
y
E(X[Y = y)P(Y = y) if Y is discrete,
_
E(X[Y = y)f
Y
(y) dy if Y is continuous.
Note that this applies to the probability of an event (which is nothing other than the expectation
of its indicator) as well if we know P(A[Y = y) = E(I
A
[Y = y), then we may compute
P(A) = EI
A
by Bayes formula above.
Example 11.1. Assume that X, Y are independent Poisson, with EX =
1
, EY =
2
. Com-
pute the conditional probability mass function of p
X
(x[X +Y = n).
Recall that X +Y is Poisson(
1
+
2
). By denition,
P(X = k[X +Y = n) =
P(X = k, X +Y = n)
P(X +Y = n)
=
P(X = k)P(Y = n k)
(
1
+
2
)
n
n!
e
(
1
+
2
)
=
k
1
k!
e

nk
2
(nk)!
e
2
(
1
+
2
)
n
n!
e
(
1
+
2
)
=
_
n
k
__

1
1
+
2
_
k
_

2
1
+
2
_
nk
.
Therefore, conditioned on X +Y = n, X is Binomial(n,

1
1
+
2
).
Example 11.2. Let T
1
, T
2
be two independent Exponential() random variables and let S
1
=
T
1
, S
2
= T
1
+T
2
. Compute f
S
1
(s
1
[S
2
= s
2
).
First,
P(S
1
s
1
, S
2
s
2
) = P(T
1
s
1
, T
1
+T
2
s
2
)
=
_
s
1
0
dt
1
_
s
2
t
1
0
f
T
1
,T
2
(t
1
, t
2
) dt
2
.
If f = f
S
1
,S
2
, then
f(s
1
, s
2
) =

2
s
1
s
2
P(S
1
s
1
, S
2
s
2
)
=

s
2
_
s
2
s
1
0
f
T
1
,T
2
(s
1
, t
2
) dt
2
= f
T
1
,T
2
(s
1
, s
2
s
1
)
= f
T
1
(s
1
)f
T
2
(s
2
s
1
)
= e
s
1
e
(s
2
s
1
)
=
2
e
s
2
.
Therefore,
f(s
1
, s
2
) =
_
2
e
s
2
if 0 s
1
s
2
,
0 otherwise
and, consequently, for s
2
0,
f
S
2
(s
2
) =
_
s
2
0
f(s
1
, s
2
) ds
1
=
2
s
2
e
s
2
.
Therefore,
f
S
1
(s
1
[S
2
= s
2
) =

2
e
s
2
2
s
2
e
s
2
=
1
s
2
,
for 0 s
1
s
2
, and 0 otherwise. Therefore, conditioned on T
1
+ T
2
= s
2
, T
1
is uniform on
[0, s
2
].
Imagine the following: a new lightbulb is put in and, after time T
1
, it burns out. It is then
replaced by a new lightbulb, identical to the rst one, which also burns out after an additional
time T
2
. If we know the time when the second bulb burns out, the rst bulbs failure time is
uniform on the interval of its possible values.
Example 11.3. Waiting to exceed the initial score. For the rst problem, roll a die once and
assume that the number you rolled is U. Then, continue rolling the die until you either match
or exceed U. What is the expected number of additional rolls?
Let N be the number of additional rolls. This number is Geometric, if we know U, so let us
condition on the value of U. We know that
E(N[U = n) =
6
7 n
,
and so, by Bayes formula for expectation,
E(N) =
6
n=1
E(N[U = n) P(U = n)
=
1
6
6
n=1
6
7 n
=
6
n=1
1
7 n
= 1 +
1
2
+
1
3
+. . . +
1
6
= 2.45 .
Now, let U be a uniform random variable on [0, 1], that is, the result of a call of a random
number generator. Once we know U, generate additional independent uniform random variables
(still on [0, 1]), X
1
, X
2
, . . ., until we get one that equals or exceeds U. Let n be the number of
additional calls of the generator, that is, the smallest n for which X
n
U. Determine the
p. m. f. of N and EN.
Given that U = u, N is Geometric(1 u). Thus, for k = 0, 1, 2, . . .,
P(N = k[U = u) = u
k1
(1 u)
and so
P(N = k) =
_
1
0
P(N = k[U = u) du =
_
1
0
u
k1
(1 u) du =
1
k(k + 1)
.
In fact, a slick alternate derivation shows that P(N = k) does not depend on the distribution of
random variables (which we assumed to be uniform), as soon as it is continuous, so that there
are no ties (i.e., no two random variables are equal). Namely, the event N = k happens
exactly when X
k
is the largest and U is the second largest among X
1
, X
2
, . . . , X
k
, U. All orders,
by diminishing size, of these k + 1 random numbers are equally likely, so the probability that
X
k
and U are the rst and the second is
1
k+1

1
k
.
It follows that
EN =
k=1
1
k + 1
= ,
which can (in the uniform case) also be obtained by
EN =
_
1
0
E(N[U = u) du =
_
1
0
1
1 u
du = .
As we see from this example, random variables with innite expectation are more common and
natural than one might suppose.
Example 11.4. The number N of customers entering a store on a given day is Poisson().
Each of them buys something independently with probability p. Compute the probability that
exactly k people buy something.
Let X be the number of people who buy something. Why should X be Poisson? Approxi-
mate: let n be the (large) number of people in the town and the probability that any particular
one of them enters the store on a given day. Then, by the Poisson approximation, with = n,
N Binomial(n, ) and X Binomial(n, p) Poisson(p). A more straightforward way to
see this is as follows:
P(X = k) =
n=k
P(X = k[N = n)P(N = n)
=
n=k
_
n
k
_
p
k
(1 p)
nk
n
e
n!
= e
n=k
n!
k!(n k)!
p
k
(1 p)
nk
n
n!
=
e
p
k
k
k!
n=k
(1 p)
nk
nk
(n k)!
=
e
(p)
k
k!
=0
((1 p))
!
=
e
(p)
k
k!
e
(1p)
=
e
p
(p)
k
k!
.
This is, indeed, the Poisson(p) probability mass function.
Example 11.5. A coin with Heads probability p is tossed repeatedly. What is the expected
number of tosses needed to get k successive heads?
Note: If we remove successive, the answer is
k
p
, as it equals the expectation of the sum of
k (independent) Geometric(p) random variables.
Let N
k
be the number of needed tosses and m
k
= EN
k
. Let us condition on the value of
N
k1
. If N
k1
= n, then observe the next toss; if it is Heads, then N
k
= n+1, but, if it is Tails,
then we have to start from the beginning, with n + 1 tosses wasted. Here is how we translate
this into mathematics:
E[N
k
[N
k1
= n] = p(n + 1) + (1 p)(n + 1 +E(N
k
))
= pn +p + (1 p)n + 1 p +m
k
(1 p)
= n + 1 +m
k
(1 p).
Therefore,
m
k
= E(N
k
) =
n=k1
E[N
k
[N
k1
= n]P(N
k1
= n)
=
n=k1
(n + 1 +m
k
(1 p)) P(N
k1
= n)
= m
k1
+ 1 +m
k
(1 p)
=
1
p
+
m
k1
p
.
This recursion can be unrolled,
m
1
=
1
p
m
2
=
1
p
+
1
p
2
.
.
.
m
k
=
1
p
+
1
p
2
+. . . +
1
p
k
.
In fact, we can even compute the moment generating function of N
k
by dierent condition-
ing
1
. Let F
a
, a = 0, . . . , k 1, be the event that the tosses begin with a Heads, followed by
Tails, and let F
k
the event that the rst k tosses are Heads. One of F
0
, . . . , F
k
must happen,
therefore, by Bayes formula,
E[e
tN
k
] =
k
a=0
E[e
tN
k
[F
k
]P(F
k
).
If F
k
happens, then N
k
= k, otherwise a + 1 tosses are wasted and one has to start over with
the same conditions as at the beginning. Therefore,
E[e
tN
k
] =
k1
a=0
E[e
t(N
k
+a+1)
]p
a
(1 p) +e
tk
p
k
= (1 p)E[e
tN
k
]
k1
a=0
e
t(a+1)
p
a
+e
tk
p
k
,
and this gives an equation for E[e
tN
k
] which can be solved:
E[e
tN
k
] =
p
k
e
tk
1 (1 p)
k1
a=0
e
t(a+1)
p
a
=
p
k
e
tk
(1 pe
t
)
1 pe
t
(1 p)e
t
(1 p
k
e
tk
)
.
We can, then, get EN
k
by dierentiating and some algebra, by
EN
k
=
d
dt
E[e
tN
k
][
t=0
.
1
Thanks to Travis Scrimshaw for pointing this out.
Example 11.6. Gamblers ruin. Fix a probability p (0, 1). Play a sequence of games; in each
game you (independently) win $1 with probability p and lose $1 with probability 1 p. Assume
that your initial capital is i dollars and that you play until you either reach a predetermined
amount N, or you lose all your money. For example, if you play a fair game, p =
1
2
, while, if
you bet on Red at roulette, p =
9
19
. You are interested in the probability P
i
that you leave the
game happy with your desired amount N.
Another interpretation is that of a simple random walk on the integers. Start at i, 0 i N,
and make steps in discrete time units: each time (independently) move rightward by 1 (i.e., add
1 to your position) with probability p and move leftward by 1 (i.e., add 1 to your position)
with probability 1 p. In other words, if the position of the walker at time n is S
n
, then
S
n
= i +X
1
+ +X
n
,
where X
k
are i. i. d. and P(X
k
= 1) = p, P(X
k
= 1) = 1 p. This random walk is one of the
very basic random (or, if you prefer a Greek word, stochastic) processes. The probability P
i
is
the probability that the walker visits N before a visit to 0.
We condition on the rst step X
1
the walker makes, i.e., the outcome of the rst bet. Then,
by Bayes formula,
P
i
= P(visit N before 0[X
1
= 1)P(X
1
= 1) +P(visit N before 0[X
1
= 1)P(X
1
= 1)
= P
i+1
p +P
i1
(1 p),
which gives us a recurrence relation, which we can rewrite as
P
i+1
P
i
=
1 p
p
(P
i
P
i1
).
We also have boundary conditions P
0
= 0, P
N
= 1. This is a recurrence we can solve quite
easily, as
P
2
P
1
=
1 p
p
P
1
P
3
P
2
=
1 p
p
(P
2
P
1
) =
_
1 p
p
_
2
P
1
.
.
.
P
i
P
i1
=
_
1 p
p
_
i1
P
1
, for i = 1, . . . , N.
We conclude that
P
i
P
1
=
_
_
1 p
p
_
+
_
1 p
p
_
2
+. . . +
_
1 p
p
_
i1
_
P
1
,
P
i
=
_
1 +
_
1 p
p
_
+
_
1 p
p
_
2
+. . . +
_
1 p
p
_
i1
_
P
1
.
Therefore,
P
i
=
_
_
1
1p
p
i
1
1p
p
P
1
if p ,=
1
2
,
iP
1
if p =
1
2
.
To determine the unknown P
1
we use that P
N
= 1 to nally get
P
i
=
_
_
1
1p
p
i
1
1p
p
N
if p ,=
1
2
,
i
N
if p =
1
2
.
For example, if N = 10 and p = 0.6, then P
5
0.8836; if N = 1000 and p =
9
19
, then
P
900
2.6561 10
5
.
Example 11.7. Bold Play. Assume that the only game available to you is a game in which
you can place even bets at any amount, and that you win each of these bets with probability p.
Your initial capital is x [0, N], a real number, and again you want to increase it to N before
going broke. Your bold strategy (which can be proved to be the best) is to bet everything unless
you are close enough to N that a smaller amount will do:
1. Bet x if x
N
2
.
2. Bet N x if x
N
2
.
We can, without loss of generality, x our monetary unit so that N = 1. We now dene
P(x) = P(reach 1 before reaching 0).
By conditioning on the outcome of your rst bet,
P(x) =
_
p P(2x) if x
_
0,
1
2
,
p 1 + (1 p) P(2x 1) if x
_
1
2
, 1
.
For each positive integer n, this is a linear system for P(
k
2
n
), k = 0, . . . , 2
n
, which can be solved.
For example:
When n = 1, P
_
1
2
_
= p.
When n = 2, P
_
1
4
_
= p
2
, P
_
3
4
_
= p + (1 p)p.
When n = 3, P
_
1
8
_
= p
3
, P
_
3
8
_
= p P
_
3
4
_
= p
2
+ p
2
(1 p), P
_
5
8
_
= p + p
2
(1 p),
P
_
7
8
_
= p +p(1 p) +p(1 p)
2
.
It is easy to verify that P(x) = x, for all x, if p =
1
2
. Moreover, it can be computed that
P(0.9) 0.8794 for p =
9
19
, which is not too dierent from a fair game. The gure below
displays the graphs of functions P(x) for p = 0.1, 0.25,
9
19
, and
1
2
.
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
P
(
x
)
A few remarks for the more mathematically inclined: The function P(x) is continuous, but
nowhere dierentiable on [0, 1]. It is thus a highly irregular function despite the fact that it is
strictly increasing. In fact, P(x) is the distribution function of a certain random variable Y ,
that is, P(x) = P(Y x). This random variable with values in (0, 1) is dened by its binary
expansion
Y =
j=1
D
j
1
2
j
,
where its binary digits D
j
are independent and equal to 1 with probability 1 p and, thus, 0
with probability p.
Theorem 11.1. Expectation and variance of sums with a random number of terms.
Assume that X, X
1
, X
2
, . . . is an i. i. d. sequence of random variables with nite EX = and
Var(X) =
2
. Let N be a nonnegative integer random variable, independent of all X
i
, and let
S =
N
i=1
X
i
.
Then
ES = EN,
Var(S) =
2
EN +
2
Var(N).
Proof. Let S
n
= X
1
+. . . +X
n
. We have
E[S[N = n] = ES
n
= nEX
1
= n.
Then,
ES =
n=0
nP(N = n) = EN.
For variance, compute rst
E(S
2
) =
n=0
E[S
2
[N = n]P(N = n)
=
n=0
E(S
2
n
)P(N = n)
=
n=0
((Var(S
n
) + (ES
n
)
2
)P(N = n)
=
n=0
(n
2
+n
2
2
)P(N = n)
=
2
EN +
2
E(N
2
).
Therefore,
Var(S) = E(S
2
) (ES)
2
=
2
EN +
2
E(N
2
)
2
(EN)
2
=
2
EN +
2
Var(N).
Example 11.8. Toss a fair coin until you toss Heads for the rst time. Each time you toss
Tails, roll a die and collect as many dollars as the number on the die. Let S be your total
winnings. Compute ES and Var(S).
This ts into the above context, with X
i
, the numbers rolled on the die, and N, the number
of Tails tossed before rst Heads. We know that
EX
1
=
7
2
,
Var(X
1
) =
35
12
.
Moreover, N + 1 is a Geometric(
1
2
) random variable, and so
EN = 2 1 = 1.
Var(N) =
1
1
2
_
1
2
_
2
= 2
Plug in to get ES =
7
2
and Var(S) =
59
12
.
Example 11.9. We now take another look at Example 8.11. We will rename the number of
days in purgatory as S, to t it better into the present context, and call the three doors 0, 1,
and 2. Let N be the number of times your choice of door is not door 0. This means that N is
Geometric(
1
3
)1. Any time you do not pick door 0, you pick door 1 or 2 with equal probability.
Therefore, each X
i
is 1 or 2 with probability
1
2
each. (Note that X
i
are not 0, 1, or 2 with
probability
1
3
each!)
It follows that
EN = 3 1 = 2,
Var(N) =
1
1
3
_
1
3
_
2
= 6
and
EX
1
=
3
2
,
Var(X
1
) =
1
2
+ 2
2
2

9
4
=
1
4
.
Therefore, ES = EN EX
1
= 3, which, of course, agrees with the answer in Example 8.11.
Moreover,
Var(S) =
1
4
2 +
9
4
6 = 14.
Problems
1. Toss an unfair coin with probability p (0, 1) of Heads n times. By conditioning on the
outcome of the last toss, compute the probability that you get an even number of Heads.
2. Let X
1
and X
2
be independent Geometric(p) random variables. Compute the conditional
p. m. f. of X
1
given X
1
+X
2
= n, n = 2, 3, . . .
3. Assume that the joint density of (X, Y ) is
f(x, y) =
1
y
e
y
, for 0 < x < y,
and 0 otherwise. Compute E(X
2
[Y = y).
4. You are trapped in a dark room. In front o you are two buttons, A and B. If you press A,
with probability 1/3 you will be released in two minutes;
with probability 2/3 you will have to wait ve minutes and then you will be able to press
one of the buttons again.
If you press B,
you will have to wait three minutes and then be able to press one of the buttons again.
Assume that you cannot see the buttons, so each time you press one of them at random. Compute
the expected time of your connement.
5. Assume that a Poisson number with expectation 10 of customers enters the store. For
promotion, each of them receives an in-store credit uniformly distributed between 0 and 100
dollars. Compute the expectation and variance of the amount of credit the store will give.
6. Generate a random number uniformly on [0, 1]; once you observe the value of , say
= , generate a Poisson random variable N with expectation . Before you start the random
experiment, what is the probability that N 3?
7. A coin has probability p of Heads. Alice ips it rst, then Bob, then Alice, etc., and the
winner is the rst to ip Heads. Compute the probability that Alice wins.
1. Let p
n
be the probability of an even number of Heads in n tosses. We have
p
n
= p (1 p
n1
) + (1 p)p
n1
= p + (1 2p)p
n1
,
and so
p
n
1
2
= (1 2p)(p
n1
1
2
),
and then
p
n
=
1
2
+C(1 2p)
n
.
As p
0
= 1, we get C =
1
2
and, nally,
p
n
=
1
2
+
1
2
(1 2p)
n
.
2. We have, for i = 1, . . . , n 1,
P(X
1
= i[X
1
+X
2
= n) =
P(X
1
= i)P(X
2
= n i)
P(X
1
+X
2
= n)
=
p(1 p)
i1
p(1 p)
ni1
n1
k=1
p(1 p)
k1
p(1 p)
nk1
=
1
n 1
,
so X
1
is uniform over its possible values.
3. The conditional density of X given Y = y is f
X
(x[Y = y) =
1
y
, for 0 < x < y (i.e., uniform
on [0, y]), and so the answer is
y
2
3
.
4. Let I be the indicator of the event that you press A, and X the time of your connement in
minutes. Then,
EX = E(X[I = 0)P(I = 0) +E(X[I = 1)P(I = 1) = (3 +EX)
1
2
+ (
1
3
2 +
2
3
(5 +EX))
1
2
and the answer is EX = 21.
5. Let N be the number of customers and X the amount of credit, while X
i
are independent
uniform on [0, 100]. So, EX
i
= 50 and Var(X
i
) =
100
2
12
. Then, X =

N
i=1
X
i
, so EX = 50EN =
500 and Var(X) =
100
2
12
10 + 50
2
10.
6. The answer is
P(N 3) =
_
1
0
P(N 3[ = ) d =
_
1
0
(1 (1 + +

2
2
)e
)d.
7. Let f(p) be the probability. Then,
f(p) = p + (1 p)(1 f(p))
which gives
f(p) =
1
2 p
.
to solve the four problems, which you may assume have equal point score.
1. Assume that a deck of 4n cards has n cards of each of the four suits. The cards are shued
and dealt to n players, four cards per player. Let D
n
be the number of people whose four cards
are of four dierent suits.
(a) Find ED
n
.
(b) Find Var(D
n
).
(c) Find a constant c so that
1
n
D
n
converges to c in probability, as n .
2. Consider the following game, which will also appear in problem 4. Toss a coin with probability
p of Heads. If you toss Heads, you win $2, if you toss Tails, you win $1.
(a) Assume that you play this game n times and let S
n
be your combined winnings. Compute
the moment generating function of S
n
, that is, E(e
tS
n
).
(b) Keep the assumptions from (a). Explain how you would nd an upper bound for the
probability that S
n
is more than 10% larger than its expectation. Do not compute.
(c) Now you roll a fair die and you play the game as many times as the number you roll. Let Y
be your total winnings. Compute E(Y ) and Var(Y ).
3. The joint density of X and Y is
f(x, y) =
e
x/y
e
y
y
,
for x > 0 and y > 0, and 0 otherwise. Compute E(X[Y = y).
4. Consider the following game again: Toss a coin with probability p of Heads. If you toss
Heads, you win $2, if you toss Tails, you win $1. Assume that you start with no money and you
have to quit the game when your winnings match or exceed the dollar amount n. (For example,
assume n = 5 and you have $3: if your next toss is Heads, you collect $5 and quit; if your next
toss is Tails, you play once more. Note that, at the amount you quit, your winnings will be
either n or n + 1.) Let p
n
be the probability that you will quit with winnings exactly n.
(a) What is p
1
? What is p
2
?
(b) Write down the recursive equation which expresses p
n
in terms of p
n1
and p
n2
.
(c) Solve the recursion.
1. Assume that a deck of 4n cards has n cards of each of the four suits. The cards are shued
and dealt to n players, four cards per player. Let D
n
be the number of people whose four
cards are of four dierent suits.
(a) Find ED
n
.
Solution:
As
D
n
=
n
i=1
I
i
,
where
I
i
= I
{ith player gets four dierent suits}
,
and
EI
i
=
n
4
_
4n
4
_,
the answer is
ED
n
=
n
5
_
4n
4
_.
(b) Find Var(D
n
).
Solution:
We also have, for i ,= j,
E(I
i
I
j
) =
n
4
(n 1)
4
_
4n
4
__
4n4
4
_.
Therefore,
E(D
2
n
) =
i=j
E(I
i
I
j
) +
i
EI
i
=
n
5
(n 1)
5
_
4n
4
__
4n4
4
_ +
n
5
_
4n
4
_
and
Var(D
n
) =
n
5
(n 1)
5
_
4n
4
__
4n4
4
_ +
n
5
_
4n
4
_
_
n
5
_
4n
4
_
_
2
.
(c) Find a constant c so that
1
n
D
n
converges to c in probability as n .
Solution:
Let Y
n
=
1
n
D
n
. Then
EY
n
=
n
4
_
4n
4
_ =
6n
3
(4n 1)(4n 2)(4n 3)

6
4
3
=
3
32
,
as n . Moreover,
Var(Y
n
) =
1
n
2
Var(D
n
) =
n
3
(n 1)
5
_
4n
4
__
4n4
4
_ +
n
3
_
4n
4
_
_
n
4
_
4n
4
_
_
2
6
2
4
6
+ 0
_
3
32
_
2
= 0,
as n , so the statement holds with c =
3
32
.
2. Consider the following game, which will also appear in problem 4. Toss a coin with
probability p of Heads. If you toss Heads, you win $2, if you toss Tails, you win $1.
(a) Assume that you play this game n times and let S
n
be your combined winnings.
Compute the moment generating function of S
n
, that is, E(e
tS
n
).
Solution:
E(e
tS
n
) = (e
2t
p +e
t
(1 p))
n
.
(b) Keep the assumptions from (a). Explain how you would nd an upper bound for the
probability that S
n
is more than 10% larger than its expectation. Do not compute.
Solution:
As EX
1
= 2p + 1 p = 1 +p, ES
n
= n(1 +p), and we need to nd an upper bound
for P(S
n
> n(1.1 + 1.1p). When (1.1 + 1.1p) 2, i.e., p
9
11
, this is an impossible
event, so the probability is 0. When p <
9
11
, the bound is
P(S
n
> n(1.1 + 1.1p) e
I(1.1+1.1p)n
,
where I(1.1 + 1.1p) > 0 and is given by
I(1.1 + 1.1p) = sup(1.1 + 1.1p)t log(pe
2t
+ (1 p)e
t
) : t > 0.
(c) Now you roll a fair die and you play the game as many times as the number you roll.
Let Y be your total winnings. Compute E(Y ) and Var(Y ).
Solution:
Let Y = X
1
+. . . +X
N
, where X
i
are independently and identically distributed with
P(X
1
= 2) = p and P(X
1
= 1) = 1 p, and P(N = k) =
1
6
, for k = 1, . . . , 6. We
know that
EY = EN EX
1
,
Var(Y ) = Var(X
1
) EN + (EX
1
)
2
Var(N).
We have
EN =
7
2
, Var(N) =
35
12
.
Moreover,
EX
1
= 2p + 1 p = 1 +p,
and
EX
2
1
= 4p + 1 p = 1 + 3p,
so that
Var(X
1
) = 1 + 3p (1 +p)
2
= p p
2
.
The answer is
EY =
7
2
(1 +p),
Var(Y ) = (p p
2
)
7
2
+ (1 +p)
2
35
12
.
3. The joint density of X and Y is
f(x, y) =
e
x/y
e
y
y
,
for x > 0 and y > 0, and 0 otherwise. Compute E(X[Y = y).
Solution:
We have
E(X[Y = y) =
_

0
xf
X
(x[Y = y) dx.
As
f
Y
(y) =
_

0
e
x/y
e
y
y
dx
=
e
y
y
_

0
e
x/y
dx
=
e
y
y
_
ye
x/y
_
x=
x=0
= e
y
, for y > 0,
we have
f
X
(x[Y = y) =
f(x, y)
f
Y
(y)
=
e
x/y
y
, for x, y > 0,
and so
E(X[Y = y) =
_

0
x
y
e
x/y
dx
= y
_

0
ze
z
dz
= y.
4. Consider the following game again: Toss a coin with probability p of Heads. If you toss
Heads, you win $2, if you toss Tails, you win $1. Assume that you start with no money
and you have to quit the game when your winnings match or exceed the dollar amount
n. (For example, assume n = 5 and you have $3: if your next toss is Heads, you collect
$5 and quit; if your next toss is Tails, you play once more. Note that, at the amount you
quit, your winnings will be either n or n+1.) Let p
n
be the probability that you will quit
with winnings exactly n.
(a) What is p
1
? What is p
2
?
Solution:
We have
p
1
= 1 p
and
p
2
= (1 p)
2
+p.
Also, p
0
= 1.
(b) Write down the recursive equation which expresses p
n
in terms of p
n1
and p
n2
.
Solution:
We have
p
n
= p p
n2
+ (1 p)p
n1
.
(c) Solve the recursion.
Solution:
We can use
p
n
p
n1
= (p)(p
n1
p
n2
)
= (p)
n1
(p
1
p
0
).
Another possibility is to use the characteristic equation
2
(1 p) p = 0 to get
=
1 p
_
(1 p)
2
+ 4p
2
=
1 p (1 +p)
2
=
_
1
p
.
This gives
p
n
= a +b(p)
n
,
with
a +b = 1, a bp = 1 p.
We get
a =
1
1 +p
, b =
p
1 +p
,
and then
p
n
=
1
1 +p
+
p
1 +p
(p)
n
.
12 MARKOV CHAINS: INTRODUCTION 145
12 Markov Chains: Introduction
Example 12.1. Take your favorite book. Start, at step 0, by choosing a random letter. Pick
one of the ve random procedures described below and perform it at each time step n = 1, 2, . . .
1. Pick another random letter.
2. Choose a random occurrence of the letter obtained at the previous step ((n 1)st), then
pick the letter following it in the text. Use the convention that the letter that follows the
last letter in the text is the rst letter in the text.
3. At step 1 use procedure (2), while for n 2 choose a random occurrence of the two letters
obtained, in order, in the previous two steps, then pick the following letter.
4. Choose a random occurrence of all previously chosen letters, in order, then pick the fol-
lowing letter.
5. At step n, perform procedure (1) with probability
1
n
and perform procedure (2) with
probability 1
1
n
.
Repeated iteration of procedure (1) merely gives the familiar independent experiments selec-
tion of letters is done with replacement and, thus, the letters at dierent steps are independent.
Procedure (2), however, is dierent: the probability mass function for the letter at the next
time step depends on the letter at this step and nothing else. If we call our current letter our
state, then we transition into a new state chosen with the p. m. f. that depends only on our
current state. Such processes are called Markov.
Procedure (3) is not Markov at rst glance. However, it becomes such via a natural rede-
nition of state: keep track of the last two letters; call an ordered pair of two letters a state.
Procedure (4) can be made Markov in a contrived fashion, that is, by keeping track, at the
current state, of the entire history of the process. There is, however, no natural way of making
this process Markov and, indeed, there is something dierent about this scheme: it ceases being
random after many steps are performed, as the sequence of the chosen letters occurs just once
in the book.
Procedure (5) is Markov, but what distinguishes it from (2) is that the p. m. f. is dependent
not only on the current step, but also on time. That is, the process is Markov, but not time-
homogeneous. We will only consider time-homogeneous Markov processes.
In general, a Markov chain is given by
a state space, a countable set S of states, which are often labeled by the positive integers
1, 2, . . . ;
transition probabilities, a (possibly innite) matrix of numbers P
ij
, where i and j range
over all states in S; and
an initial distribution , a probability mass function on the states.
Here is how these three ingredients determine a sequence X
0
, X
1
, . . . of random variables
(with values in S): Use the initial distribution as your random procedure to pick X
0
. Subse-
quently, given that you are at state i S at any time n, make the transition to state j S with
probability P
ij
, that is
(12.1) P
ij
= P(X
n+1
= j[X
n
= i).
The transition probabilities are collected into the transition matrix:
P =
_
_
P
11
P
12
P
13
. . .
P
21
P
22
P
23
. . .
.
.
.
_
_.
A stochastic matrix is a (possibly innite) square matrix with nonnegative entries, such that all
rows sum to 1, that is
j
P
ij
= 1,
for all i. In other words, every row of the matrix is a p. m. f. Clearly, by (12.1), every transition
matrix is a stochastic matrix (as X
n+1
must be in some state). The opposite also holds: given
any stochastic matrix, one can construct a Markov chain on positive integer states with the
same transition matrix, by using the entries as transition probabilities as in (12.1).
Geometrically, a Markov chain is often represented as oriented graph on S (possibly with
self-loops) with an oriented edge going from i to j whenever a transition from i to j is possible,
i.e., whenever P
ij
> 0; such an edge is labeled by P
ij
.
Example 12.2. A random walker moves on the set 0, 1, 2, 3, 4. She moves to the right (by
1) with probability, p, and to the left with probability 1 p, except when she is at 0 or at 4.
These two states are absorbing: once there, the walker does not move. The transition matrix is
P =
_
_
1 0 0 0 0
1 p 0 p 0 0
0 1 p 0 p 0
0 0 1 p 0 p
0 0 0 0 1
_
_
.
and the matching graphical representation is below.
1 p 1 p
1
p p
0 1 2 3
1
4
1 p
p
Example 12.3. Same as the previous example except that now 0 and 4 are reecting. From
0, the walker always moves to 1, while from 4 she always moves to 3. The transition matrix
changes to
P =
_
_
0 1 0 0 0
1 p 0 p 0 0
0 1 p 0 p 0
0 0 1 p 0 p
0 0 0 1 0
_
_
.
Example 12.4. Random walk on a graph. Assume that a graph with undirected edges is
given by its adjacency matrix, which is a binary matrix with the i, jth entry 1 exactly when i
is connected to j. At every step, a random walker moves to a randomly chosen neighbor. For
example, the adjacency matrix of the graph
3
4 1
2
is
_
_
0 1 0 1
1 0 1 1
0 1 0 1
1 1 1 0
_
_
,
and the transition matrix is
P =
_
_
0
1
2
0
1
2
1
3
0
1
3
1
3
0
1
2
0
1
2
1
3
1
3
1
3
0
_
_
.
Example 12.5. The general two-state Markov chain. There are two states 1 and 2 with
transitions:
1 1 with probability ;
1 2 with probability 1 ;
2 1 with probability ;
2 2 with probability 1 .
The transition matrix has two parameters , [0, 1]:
P =
_
1
1
_
.
Example 12.6. Changeovers. Keep track of two-toss blocks in an innite sequence of indepen-
dent coin tosses with probability p of Heads. The states represent (previous ip, current ip)
and are (in order) HH, HT, TH, and TT. The resulting transition matrix is
_
_
p 1 p 0 0
0 0 p 1 p
p 1 p 0 0
0 0 p 1 p
_
_
.
Example 12.7. Simple random walk on Z. The walker moves left or right by 1, with prob-
abilities p and 1 p, respectively. The state space is doubly innite and so is the transition
matrix:
_
_
.
.
.
. . . 1 p 0 p 0 0 . . .
. . . 0 1 p 0 p 0 . . .
. . . 0 0 1 p 0 p . . .
.
.
.
_
_
Example 12.8. Birth-death chain. This is a general model in which a population may change
by at most 1 at each time step. Assume that the size of a population is x. Birth probability
p
x
is the transition probability to x + 1, death probability q
x
is the transition to x 1. and
r
x
= 1 p
x
q
x
is the transition to x. Clearly, q
0
= 0. The transition matrix is
_
_
r
0
p
0
0 0 0 . . .
q
1
r
1
p
1
0 0 . . .
0 q
2
r
2
p
2
0 . . .
.
.
.
_
_
We begin our theory by studying n-step probabilities
P
n
ij
= P(X
n
= j[X
0
= i) = P(X
n+m
= j[X
m
= i).
Note that P
0
ij
= I, the identity matrix, and P
1
ij
= P
ij
. Note also that the condition X
0
= i
simply species a particular non-random initial state.
Consider an oriented path of length n from i to j, that is i, k
1
, . . . , k
n1
, j, for some states
k
1
, . . . , k
n1
. One can compute the probability of following this path by multiplying all transition
probabilities, i.e., P
ik
1
P
k
1
k
2
P
k
n1
j
. To compute P
n
ij
, one has to sum these products over all
paths of length n from i to j. The next theorem writes this in a familiar and neater fashion.
Theorem 12.1. Connection between n-step probabilities and matrix powers:
P
n
ij
is the i, jth entry of the nth power of the transition matrix.
Proof. Call the transition matrix P and temporarily denote the n-step transition matrix by
P
(n)
. Then, for m, n 0,
P
(n+m)
ij
= P(X
n+m
= j[X
0
= i)
=
k
P(X
n+m
= j, X
n
= k[X
0
= i)
=
k
P(X
n+m
= j[X
n
= k) P(X
n
= k[X
0
= i)
=
k
P(X
m
= j[X
0
= k) P(X
n
= k[X
0
= i)
=
k
P
(m)
kj
P
(n)
ik
.
The rst equality decomposes the probability according to where the chain is at time n, the
second uses the Markov property and the third time-homogeneity. Thus,
P
(m+n)
= P
(n)
P
(m)
,
and, then, by induction
P
(n)
= P
(1)
P
(1)
P
(1)
= P
n
.
The fact that the matrix powers of the transition matrix give the n-step probabilities makes
linear algebra useful in the study of nite-state Markov chains.
Example 12.9. For the two state Markov Chain
P =
_
1
1
_
,
and
P
2
=
_

2
+ (1 ) (1 ) + (1 )(1 )
+ (1 ) (1 ) + (1 )
2
_
gives all P
2
ij
.
Assume now that the initial distribution is given by
i
= P(X
0
= i),
for all states i (again, for notational purposes, we assume that i = 1, 2, . . .). As this must
determine a p. m. f., we have
i
0 and

i
i
= 1. Then,
P(X
n
= j) =
i
P(X
n
= j[X
0
= i)P(X
0
= i)
=
i
P
n
ij
.
Then, the row of probabilities at time n is given by [P(X
n
= i), i S] = [
1
,
2
, . . .] P
n
.
Example 12.10. Consider the random walk on the graph from Example 12.4. Choose a
starting vertex at random. (a) What is the probability mass function at time 2? (b) Compute
P(X
2
= 2, X
6
= 3, X
12
= 4).
As
P =
_
_
0
1
2
0
1
2
1
3
0
1
3
1
3
0
1
2
0
1
2
1
3
1
3
1
3
0
_
_
,
we have
_
P(X
2
= 1) P(X
2
= 2) P(X
2
= 3) P(X
2
= 4)
=
_
1
4
1
4
1
4
1
4
P
2
=
_
2
9
5
18
2
9
5
18
.
The probability in (b) equals
P(X
2
= 2) P
4
23
P
6
34
=
8645
708588
0.0122.
Problems
1. Three white and three black balls are distributed in two urns, with three balls per urn. The
state of the system is the number of white balls in the rst urn. At each step, we draw at
random a ball from each of the two urns and exchange their places (the ball that was in the rst
urn is put into the second and vice versa). (a) Determine the transition matrix for this Markov
chain. (b) Assume that initially all white balls are in the rst urn. Determine the probability
that this is also the case after 6 steps.
2. A Markov chain on states 0, 1, 2, has the transition matrix
_
_
1
2
1
3
1
6
0
1
3
2
3
5
6
0
1
6
_
_
Assume that P(X
0
= 0) = P(X
0
= 1) =
1
4
. Determine EX
3
.
3. We have two coins: coin 1 has probability 0.7 of Heads and coin 2 probability 0.6 of Heads.
You ip a coin once per day, starting today (day 0), when you pick one of the two coins with
equal probability and toss it. On any day, if you ip Heads, you ip coin 1 the next day,
otherwise you ip coin 2 the next day. (a) Compute the probability that you ip coin 1 on day
3. (b) Compute the probability that you ip coin 1 on days 3, 6, and 14. (c) Compute the
probability that you ip Heads on days 3 and 6.
4. A walker moves on two positions a and b. She begins at a at time 0 and is at a the next time
as well. Subsequently, if she is at the same position for two consecutive time steps, she changes
position with probability 0.8 and remains at the same position with probability 0.2; in all other
cases she decides the next position by a ip of a fair coin. (a) Interpret this as a Markov chain
on a suitable state space and write down the transition matrix P. (b) Determine the probability
that the walker is at position a at time 10.
1. The states are 0, 1, 2, 3. For (a),
P =
_
_
0 1 0 0
1
9
4
9
4
9
0
0
4
9
4
9
1
9
0 0 1 0
_
_
For (b), compute the fourth entry of
_
0 0 0 1
P
6
,
that is, the 4, 4th entry of P
6
.
2. The answer is given by
EX
3
=
_
P(X
3
= 0) P(X
3
= 1) P(X
3
= 2)
_
_
0
1
2
_
_
=
_
1
4
1
4
1
2
P
3
_
_
0
1
2
_
_
.
3. The state X
n
of our Markov chain, 1 or 2, is the coin we ip on day n. (a) Let
P =
_
0.7 0.3
0.6 0.4
_
.
Then,
_
P(X
3
= 1) P(X
3
= 2)
=
_
1
2
1
2
P
3
and the answer to (a) is the rst entry. (b) Answer: P(X
3
= 1) P
3
11
P
8
11
. (c) You toss Heads
on days 3 and 6 if and only if you toss Coin 1 on days 4 and 7, so the answer is P(X
4
= 1) P
3
11
.
4. (a) The states are the four ordered pairs aa, ab, ba, and bb, which we will code as 1, 2, 3, and
4. Then,
P =
_
_
0.2 0.8 0 0
0 0 0.5 0.5
0.5 0.5 0 0
0 0 0.8 0.2
_
_
.
The answer to (b) is the sum of the rst and the third entries of
_
1 0 0 0
P
9
.
The power is 9 instead of 10 because the initial time for the chain (when it is at state aa) is
time 1 for the walker.
13 MARKOV CHAINS: CLASSIFICATION OF STATES 153
13 Markov Chains: Classication of States
We say that a state j is accessible from state i, i j, if P
n
ij
> 0 for some n 0. This means
that there is a possibility of reaching j from i in some number of steps. If j is not accessible
from i, P
n
ij
= 0 for all n 0, and thus the chain started from i never visits j:
P(ever visit j[X
0
= i) = P(
_
n=0
X
n
= j [X
0
= i)
n=0
P(X
n
= j[X
0
= i) = 0.
Also, note that for accessibility the size of entries of P does not matter, all that matters is which
are positive and which are 0. For computational purposes, one should also observe that, if the
chain has m states, then j is accessible from i if and only if (P +P
2
+. . . +P
m
)
ij
> 0.
If i is accessible from j and j is accessible from i, then we say that i and j communicate,
i j. It is easy to check that this is an equivalence relation:
1. i i;
2. i j implies j i; and
3. i j and j k together imply i k.
The only nontrivial part is (3) and, to prove it, let us assume i j and j k. This means that
there exists an n 0 so that P
n
ij
> 0 and an m 0 so that P
m
jk
> 0. Now, one can get from i
to j in m+n steps by going rst to j in n steps and then from j to k in m steps, so that
P
n+m
ik
P
n
ij
P
m
jk
> 0.
(Alternatively, one can use that P
m+n
= P
n
P
m
and then
P
n+m
ik
=
P
n
i
P
m
k
P
n
ij
P
m
jk
,
as the sum of nonnegative numbers is at least as large as one of its terms.)
The accessibility relation divides states into classes. Within each class, all states commu-
nicate with each other, but no pair of states in dierent classes communicates. The chain is
irreducible if there is only one class. If the chain has m states, irreducibility means that all
entries of I +P +. . . +P
m
are nonzero.
Example 13.1. To determine the classes we may present the Markov chain as a graph, in which
we only need to depict the edges that signify nonzero transition probabilities (their precise value
is irrelevant for this purpose); by convention, we draw an undirected edge when probabilities in
both directions are nonzero. Here is an example:
3
4
5
1
2
Any state 1, 2, 3, 4 is accessible from any of the ve states, but 5 is not accessible from 1, 2, 3, 4.
So, we have two classes: 1, 2, 3, 4, and 5. The chain is not irreducible.
Example 13.2. Consider the chain on states 1, 2, 3 and
P =
_
_
1
2
1
2
0
1
2
1
4
1
4
0
1
3
2
3
_
_
.
As 1 2 and 2 3, this is an irreducible chain.
Example 13.3. Consider the chain on states 1, 2, 3, 4, and
P =
_
_
1
2
1
2
0 0
1
2
1
2
0 0
0 0
1
4
3
4
0 0 0 1
_
_
.
This chain has three classes 1, 2, 3 and 4, hence, it is not irreducible.
For any state i, denote
f
i
= P(ever reenter i[X
0
= i).
We call a state i recurrent if f
i
= 1, and transient if f
i
< 1.
Example 13.4. Back to the previous example. Obviously, 4 is recurrent, as it is an absorbing
state. The only possibility of returning to 3 is to do so in one step, so we have f
3
=
1
4
, and 3 is
transient. Moreover, f
1
= 1 because in order to never return to 1 we need to go to state 2 and
stay there forever. We stay at 2 for n steps with probability
_
1
2
_
n
0,
as n , so the probability of staying at 1 forever is 0 and, consequently, f
1
= 1. By similar
logic, f
2
= 1. We will soon develop better methods to determine recurrence and transience.
Starting from any state, a Markov Chain visits a recurrent state innitely many times or not
at all. Let us now compute, in two dierent ways, the expected number of visits to i (i.e., the
times, including time 0, when the chain is at i). First, we observe that, at every visit to i, the
probability of never visiting i again is 1 f
i
, therefore,
P(exactly n visits to i[X
0
= i) = f
n1
i
(1 f
i
).
This formula says that the number of visits to i is a Geometric(1 f
i
) random variable and so
its expectation is
E(number of visits to i[X
0
= i) =
1
1 f
i
.
A second way to compute this expectation is by using the indicator trick:
E(number of visits to i[X
0
= i) = E(
n=0
I
n
[X
0
= i),
where I
n
= I
{Xn=i}
, n = 0, 1, 2, . . .. Then,
E(
n=0
I
n
[X
0
= i) =
n=0
P(X
n
= i[X
0
= i)
=
n=0
P
n
ii
.
Thus,
1
1 f
i
=
n=0
P
n
ii
and we have proved the following theorem.
Theorem 13.1. Characterization of recurrence via n step return probabilities:
A state i is recurrent if and only if
n=1
P
n
ii
= .
We call a subset S
0
S of states closed if P
ij
= 0 for each i S
0
and j / S
0
. In plain
language, once entered, a closed set cannot be exited.
Proposition 13.2. If a closed subset S
0
has only nitely many states, then there must be at least
one recurrent state. In particular, any nite Markov chain must contain at least one recurrent
state.
Proof. Start from any state from S
0
. By denition, the chain stays in S
0
forever. If all states in
S
0
are transient, then each of them is visited either not at all or only nitely many times. This
is impossible.
Proposition 13.3. If i is recurrent and i j, then also j i.
Proof. There is an n
0
such that P
n
0
ij
> 0, i.e., starting from i, the chain can reach j in n
0
steps.
Thus, every time it is at i, there is a xed positive probability that it will be at j n
0
steps later.
Starting from i, the chain returns to i innitely many times and, every time it does so, it has
an independent chance to reach j n
0
steps later; thus, eventually the chain does reach j. Now
assume that it is not true that j i. Then, once the chain reaches j, it never returns to i, but
then, i is not recurrent. This contradiction ends the proof.
Proposition 13.4. If i is recurrent and i j, then j is also recurrent. Therefore, in any class,
either all states are recurrent or all are transient. In particular, if the chain is irreducible, then
either all states are recurrent or all are transient.
In light of this proposition, we can classify each class, as well as an irreducible Markov chain,
as recurrent or transient.
Proof. By the previous proposition, we know that also j i. We will now give two arguments
for the recurrence of j.
We could use the same logic as before: starting from j, the chain must visit i with probability
1 (or else the chain starting at i has a positive probability of no return to i, by visiting j), then
it returns to i innitely many times and, at each of those times, it has an independent chance
of getting to j at a later time so it must do so innitely often.
For another argument, we know that there exist k, m 0 so that P
k
ij
> 0, P
m
ji
> 0. Further-
more, for any n 0, one way to get from j to j in m+n +k steps is by going from j to i in m
steps, then from i to i in n steps, and then from i to j in k steps; thus,
P
m+n+k
jj
P
m
ji
P
n
ii
P
k
ij
.
If
n=0
P
ii
= , then
n=0
P
m+n+k
jj
= and, nally,
=0
P
jj
= . In short, if i is recurrent,
then so is j.
Proposition 13.5. Any recurrent class is a closed subset of states.
Proof. Let S
0
be a recurrent class, i S
0
and j , S
0
. We need to show that P
ij
= 0. Assume
the converse, P
ij
> 0. As j does not communicate with i, the chain never reaches i from j, i.e.,
i is not accessible from j. But this is a contradiction to Proposition 13.3.
For nite Markov chains, these propositions make it easy to determine recurrence and tran-
sience: if a class is closed, it is recurrent, but if it is not closed, it is transient.
Example 13.5. Assume that the states are 1, 2, 3, 4 and that the transition matrix is
P =
_
_
0 0
1
2
1
2
1 0 0 0
0 1 0 0
0 1 0 0
_
_
.
By inspection, every state is accessible from every other state and so this chain is irreducible.
Therefore, every state is recurrent.
Example 13.6. Assume now that the states are 1, . . . , 6 and
P =
_
_
0 1 0 0 0 0
0.4 0.6 0 0 0 0
0.3 0 0.4 0.2 0.1 0
0 0 0 0.3 0.7 0
0 0 0 0.5 0 0.5
0 0 0 0.3 0 0.7
_
_
.
2
1 3 4
5 6
0.6
1
0.4
0.3
0.3
0.5
0.5
0.7
0.1
0.3 0.2
0.4
0.7
We observe that 3 can only be reached from 3, therefore, 3 is in a class of its own. States 1 and
2 can reach each other and no other state, so they form a class together. Furthermore, 4, 5, 6
all communicate with each other. Thus, the division into classes is 1, 2, 3, and 4, 5, 6. As
it is not closed, 3 is a transient class (in fact, it is clear that f
3
= 0.4). On the other hand,
1, 2 and 4, 5, 6 both are closed and, therefore, recurrent.
Example 13.7. Recurrence of a simple random walk on Z. Recall that such a walker moves
from x to x + 1 with probability p and to x 1 with probability 1 p. We will assume that
p (0, 1) and denote the chain S
n
= S
(1)
n
. (The superscript indicates the dimension. We will
make use of this in subsequent examples in which the walker will move in higher dimensions.)
As such a walk is irreducible, we only have to check whether state 0 is recurrent or transient, so
we assume that the walker begins at 0. First, we observe that the walker will be at 0 at a later
time only if she makes an equal number of left and right moves. Thus, for n = 1, 2, . . .,
P
2n1
00
= 0
and
P
2n
00
=
_
2n
n
_
p
n
(1 p)
n
.
Now, we recall Stirlings formula:
n! n
n
e
n
2n
(the symbol means that the quotient of the two quantities converges to 1 as n ).
Therefore,
_
2n
n
_
=
(2n)!
(n!)
2
(2n)
2n
e
2n
22n
n
2n
e
2n
2n
=
2
2n
n
,
and, therefore,
P
2n
00
=
2
2n
n
p
n
(1 p)
n
n
(4p(1 p))
n
In the symmetric case, when p =
1
2
,
P
2n
00

1
n
,
therefore,
n=0
P
2n
00
= ,
and the random walk is recurrent.
When p ,=
1
2
, 4p(1 p) < 1, so that P
2n
00
goes to 0 faster than the terms of a convergent
geometric series,
n=0
P
2n
00
< ,
and the random walk is transient. In this case, what is the probability f
0
that the chain ever
reenters 0? We need to recall the Gamblers ruin probabilities,
P(S
n
reaches N before 0[S
0
= 1) =
1
1p
p
1
_
1p
p
_
N
.
As N , the probability
P(S
n
reaches 0 before N[S
0
= 1) = 1 P(S
n
reaches N before 0[S
0
= 1)
converges to
P(S
n
ever reaches 0[S
0
= 1) =
_
1 if p <
1
2
,
1p
p
if p >
1
2
.
Assume that p >
1
2
. Then,
f
0
= P(S
1
= 1, S
n
returns to 0 eventually) +P(S
1
= 1, S
n
returns to 0 eventually)
= p
1 p
p
+ (1 p) 1
= 2(1 p).
If p < 1/2, we may use the fact that replacing the walkers position with its mirror image replaces
p by 1 p; this gives f
0
= 2p when p <
1
2
.
Example 13.8. Is the simple symmetric random walk on Z
2
recurrent? A walker now moves on
integer points in two dimensions: each step is a distance 1 jump in one of the four directions (N,
S, E, or W). We denote this Markov chain by S
(2)
n
and imagine a drunk wandering at random
through the rectangular grid of streets of a large city. (Chicago would be a good example.) The
question is whether the drunk will eventually return to her home at (0, 0). All starting positions
in this and in the next example will be the appropriate origins. Note again that the walker can
only return in an even number of steps and, in fact, both the number of steps in the x direction
(E or W) and in the y direction (N or S) must be even (otherwise, the respective coordinate
cannot be 0).
We condition on the number N of times the walker moves in the x-direction:
P(S
(2)
2n
= (0, 0)) =
n
k=0
P(N = 2k)P(S
(2)
2n
= (0, 0)[N = 2k)
=
n
k=0
P(N = 2k)P(S
(1)
2k
= 0)P(S
(1)
2(nk)
= 0).
In order not to obscure the computation, we will not show the full details from now on; lling
in the missing pieces is an excellent computational exercise.
First, as the walker chooses to go horizontally or vertically with equal probability, N
2n
2
=
n with overwhelming probability and so we can assume that k
n
2
. Taking this into account,
P(S
(1)
2k
= 0)
n
,
P(S
(1)
2(nk)
= 0)
n
.
Therefore,
P(S
(2)
2n
= (0, 0))
2
n
n
k=0
P(N = 2k)
2
n
P(N is even)
1
n
,
as we know that (see Problem 1 in Chapter 11)
P(N is even) =
1
2
.
Therefore,
n=0
P(S
(2)
2n
= (0, 0)) =
and we have demonstrated that this chain is still recurrent, albeit barely. In fact, there is an
easier slick proof that does not generalize to higher dimensions, which demonstrates that
P(S
(2)
2n
= 0) = P(S
(1)
2n
= 0)
2
.
Here is how it goes. If we let each coordinate of a two-dimensional random walker move inde-
pendently, then the above is certainly true. Such a walker makes diagonal moves, from (x, y) to
(x +1, y +1), (x 1, y +1), (x +1, y 1), or (x 1, y 1) with equal probability. At rst, this
appears to be a dierent walk, but if we rotate the lattice by 45 degrees, scale by
1
2
, and ignore
half of the points that are never visited, this becomes the same walk as S
(2)
n
. In particular, it is
at the origin exactly when S
(2)
n
is.
Example 13.9. Is the simple symmetric random walk on Z
3
recurrent? Now, imagine a squirrel
running around in a 3 dimensional maze. The process S
(3)
n
moves from a point (x, y, z) to one
of the six neighbors (x 1, y, z), (x, y 1, z), (x, y, z 1) with equal probability. To return to
(0, 0, 0), it has to make an even number number of steps in each of the three directions. We will
condition on the number N of steps in the z direction. This time N
2n
3
and, thus,
P(S
(3)
2n
= (0, 0, 0)) =
n
k=0
P(N = 2k)P(S
(1)
2k
= 0)P(S
(2)
2(nk)
= (0, 0))
k=0
P(N = 2k)
n
3
2n
=
3
3
2
3/2
n
3/2
P(N is even)
3
4
3/2
n
3/2
.
Therefore,
n
P(S
(3)
2n
= (0, 0, 0)) <
and the three-dimensional random walk is transient, so the squirrel may never return home. The
probability f
0
= P(return to 0), thus, is not 1, but can we compute it? One approximation is
obtained by using
1
1 f
0
=
n=0
P(S
(3)
2n
= (0, 0, 0)) = 1 +
1
6
+. . . ,
but this series converges slowly and its terms are dicult to compute. Instead, one can use the
remarkable formula, derived by Fourier analysis,
1
1 f
0
=
1
(2)
3
___
(,)
3
dxdy dz
1
1
3
(cos(x) + cos(y) + cos(z))
,
which gives, to four decimal places,
f
0
0.3405.
Problems
1. For the following transition matrices, determine the classes and specify which are recurrent
and which transient.
P
1
=
_
_
0
1
2
1
2
0 0
1
2
0
1
2
0 0
1
2
1
2
0 0 0
1
2
1
4
0 0
1
4
0
1
2
1
4
1
4
0
_
_
P
2
=
_
_
0 0 0 1 0
0 0 0 1 0
1
3
1
3
0 0
1
3
0 0 1 0 0
0 0 1 0 0
_
_
P
3
=
_
_
1
2
0
1
2
0 0
1
4
1
2
1
4
0 0
1
4
1
4
1
2
0 0
0 0 0
1
2
1
2
0 0 0
1
2
1
2
_
_
P
4
=
_
_
1
2
1
2
0 0 0
1
2
1
2
0 0 0
0 0 1 0 0
0 0
1
2
1
2
0
1 0 0 0 0
_
_
2. Assume that a Markov chain X
n
has states 0, 1, 2 . . . and transitions from each i > 0 to i +1
with probability 1
1
2i
and to 0 with probability

1
2i
. Moreover, from 0 it transitions to 1

with probability 1. (a) Is this chain irreducible? (b) Assume that X
0
= 0 and let R be the
rst return time to 0 (i.e., the rst time after the initial time the chain is back at the origin).
Determine for which
1 f
0
= P(no return) = lim
n
P(R > n) = 0.
(c) Depending on , determine which classes are recurrent.
3. Consider the one-dimensional simple symmetric random walk S
n
= S
(1)
n
with p =
1
2
. As in
the Gamblers ruin problem, x an N and start at some 0 i N. Let E
i
be the expected
time at which the walk rst hits either 0 or N. (a) By conditioning on the rst step, determine
the recursive equation for E
i
. Also, determine the boundary conditions E
0
and E
N
. (b) Solve
the recursion. (c) Assume that the chain starts at 0 and let R be the rst time (after time 0)
that it revisits 0. By recurrence, we know that P(R < ) = 1; use (b) to show that ER = .
The walk will eventually return to 0, but the expected waiting time is innite!
1. Assume that the states are 1, . . . , 5. For P
1
: 1, 2, 3 recurrent, 4, 5 transient. For P
2
:
irreducible, so all states are recurrent. For P
3
: 1, 2, 3 recurrent, 4, 5 recurrent. For P
4
:
1, 2 recurrent, 3 recurrent (absorbing), 4 transient, 5 transient.
2. (a) The chain is irreducible. (b) If R > n, then the chain, after moving to 1, makes n 1
consecutive steps to the right, so
P(R > n) =
n1
i=1
_
1
1
2 i
_
.
The product converges to 0 if and only if its logarithm converges to and that holds if and
only if the series
i=1
1
2 i
diverges, which is when 1. (c) For 1, the chain is recurrent, otherwise, it is transient.
3. For (a), the walker makes one step and then proceeds from i+1 or i1 with equal probability,
so that
E
i
= 1 +
1
2
(E
i+1
+E
i
),
with E
0
= E
N
= 0. For (b), the homogeneous equation is the same as the one in the Gamblers
ruin, so its general solution is linear: Ci +D. We look for a particular solution of the form Bi
2
and we get Bi
2
= 1 +
1
2
(B(i
2
+2i +1) +B(i
2
2i +1)) = 1 +Bi
2
+B, so B = 1. By plugging
in the boundary conditions we can solve for C and D to get D = 0, C = N. Therefore,
E
i
= i(N i).
For (c), after a step the walker proceeds either from 1 or 1 and, by symmetry, the expected
time to get to 0 is the same for both. So, for every N,
ER 1 +E
1
= 1 + 1 (N 1) = N,
and so ER = .
14 BRANCHING PROCESSES 163
14 Branching processes
In this chapter we will consider a random model for population growth in the absence of spatial or
any other resource constraints. So, consider a population of individuals which evolves according
to the following rule: in every generation n = 0, 1, 2, . . ., each individual produces a random
number of ospring in the next generation, independently of other individuals. The probability
mass function for ospring is often called the ospring distribution and is given by
p
i
= P(number of ospring = i),
for i = 0, 1, 2, . . .. We will assume that p
0
< 1 and p
1
< 1 to eliminate the trivial cases. This
model was introduced by F. Galton in the late 1800s to study the disappearance of family names;
in this case p
i
is the probability that a man has i sons.
We will start with a single individual in generation 0 and generate the resulting random
family tree. This tree is either nite (when some generation produces no ospring at all) or
innite in the former case, we say that the branching process dies out and, in the latter case,
that it survives.
We can look at this process as a Markov chain where X
n
is the number of individuals in
generation n. Let us start with the following observations:
If X
n
reaches 0, it stays there, so 0 is an absorbing state.
If p
0
> 0, P(X
n+1
= 0[X
n
= k) > 0, for all k.
Therefore, by Proposition 13.5, all states other than 0 are transient if p
0
> 0; the popu-
lation must either die out or increase to innity. If p
0
= 0, then the population cannot
decrease and each generation increases with probability at least 1 p
1
, therefore it must
increase to innity.
It is possible to write down the transition probabilities for this chain, but they have a rather
complicated explicit form, as
P(X
n+1
= i[X
n
= k) = P(W
1
+W
2
+. . . +W
k
= i),
where W
1
, . . . , W
k
are independent random variables, each with the ospring distribution. This
suggests using moment generating functions, which we will indeed do. Recall that we are as-
suming that X
0
= 1.
Let
n
= P(X
n
= 0)
be the probability that the population is extinct by generation (which we also think of as time) n.
The probability
0
that the branching process dies out is, then, the limit of these probabilities:
0
= P(the process dies out) = P(X
n
= 0 for some n) = lim
n
P(X
n
= 0) = lim
n
n
.
Note that
0
= 0 if p
0
= 0. Our main task will be to compute
0
for general probabilities p
k
. We
start, however, with computing the expectation and variance of the population at generation n.
Let and
2
be the expectation and the variance of the ospring distribution, that is,
= EX
1
=
k=0
kp
k
,
and
2
= Var(X
1
).
Let m
n
= E(X
n
) and v
n
= Var(X
n
). Now, X
n+1
is the sum of a random number X
n
of
independent random variables, each with the ospring distribution. Thus, we have by Theorem
11.1,
m
n+1
= m
n
,
and
v
n+1
= m
n
2
+v
n
2
.
Together with initial conditions m
0
= 1, v
0
= 0, the two recursive equations determine m
n
and
v
n
. We can very quickly solve the rst recursion to get m
n
=
n
and so
v
n+1
=
n
2
+v
n
2
.
When = 1, m
n
= 1 and v
n
= n
2
. When ,= 1, the recursion has the general solution
v
n
= A
n
+B
2n
. The constant A must satisfy
A
n+1
=
2
n
+A
n+2
,
so that,
A =

2
(1 )
.
From v
0
= 0 we get A+B = 0 and the solution is given in the next theorem.
Theorem 14.1. Expectation m
n
and variance v
n
of the nth generation count.
We have
m
n
=
n
and
v
n
=
_
n
(1
n
)
(1)
if ,= 1,
n
2
if = 1.
We can immediately conclude that < 1 implies
0
= 1, as
P(X
n
,= 0) = P(X
n
1) EX
n
=
n
0;
if the individuals have less than one ospring on average, the branching process dies out.
Now, let be the moment generating function of the ospring distribution. It is more
convenient to replace e
t
in our original denition with s, so that
(s) =
X
1
(s) = E(s
X
1
) =
k=0
p
k
s
k
.
In combinatorics, this would be exactly the generating function of the sequence p
k
. Then, the
moment generating function of X
n
is
Xn
(s) = E[s
Xn
] =
k=0
P(X
n
= k)s
k
.
We will assume that 0 s 1 and observe that, for such s, this power series converges. Let us
get a recursive equation for
Xn
by conditioning on the population count in generation n 1:
Xn
(s) = E[s
X
n
]
=
k=0
E[s
X
n
[X
n1
= k]P(X
n1
= k)
=
k=0
E[s
W
1
+...+W
k
]P(X
n1
= k)
=
k=0
(E(s
W
1
)E(s
W
2
) E(s
W
k
)P(X
n1
= k)
=
k=0
(s)
k
P(X
n1
= k)
=
X
n1
((s)).
So,
Xn
is the nth iterate of ,
X
2
(s) = ((s)),
X
3
(s) = (((s))), . . .
and we can also write
X
n
(s) = (
X
n1
(s)).
Next, we take a closer look at the properties of . Clearly,
(0) = p
0
> 0
and
(1) =
k=0
p
k
= 1.
Moreover, for s > 0,
(s) =
k=0
kp
k
s
k1
> 0,
so is strictly increasing, with
(1) = .
Finally,
(s) =
k=1
k(k 1)p
k
s
k2
0,
so is also convex. The crucial observation is that
n
=
Xn
(0),
and so
n
is obtained by starting at 0 and computing the nth iteration of . It is also clear
that
n
is a nondecreasing sequence (because X
n1
= 0 implies that X
n
= 0). We now consider
separately two cases:
Assume that is always above the diagonal, that is, (s) s for all s [0, 1]. This
happens exactly when =
(1) 1. In this case,

n
converges to 1, and so
0
= 1. This
is shown in the right graph of the gure below.
Now, assume that is not always above the diagonal, which happens when > 1. In
this case, there exists exactly one s < 1 which solves s = (s). As
n
converges to this
solution, we conclude that
0
< 1 is the smallest solution to s = (s). This is shown in
the left graph of the gure below.
1
1
1
1
1

2
1
The following theorem is a summary of our ndings.
Theorem 14.2. Probability that the branching process dies out.
If 1,
0
= 1. If > 1, then
0
is the smallest solution on [0, 1] to s = (s).
Example 14.1. Assume that a branching process is started with X
0
= k instead of X
0
= 1.
How does this change the survival probability? The k individuals all evolve independent family
trees, so that the probability of eventual death is
k
0
. It also follows that
P(the process ever dies out[X
n
= k) =
k
0
for every n.
If is barely larger than 1, the probability
0
of extinction is quite close to 1. In the context
of family names, this means that the ones with already a large number of representatives in the
population are at a distinct advantage, as the probability that they die out by chance is much
lower than that of those with only a few representatives. Thus, common family names become
ever more common, especially in societies that have used family names for a long time. The
most famous example of this phenomenon is in Korea, where three family names (Kim, Lee, and
Park in English transcriptions) account for about 45% of the population.
Example 14.2. Assume that
p
k
= p
k
(1 p), k = 0, 1, 2, . . . .
This means that the ospring distribution is Geometric(1 p) minus 1. Thus,
=
1
1 p
1 =
p
1 p
and, if p
1
2
,
0
= 1. Now suppose that p >
1
2
. Then, we have to compute
(s) =
k=0
s
k
p
k
(1 p)
=
1 p
1 ps
.
The equation (s) = s has two solutions, s = 1 and s =
1p
p
. Thus, when p >
1
2
,
0
=
1 p
p
.
Example 14.3. Assume that the ospring distribution is Binomial(3,
1
2
). Compute
0
.
As =
3
2
> 1,
0
is given by
(s) =
1
8
+
3
8
s +
3
8
s
2
+
1
8
s
3
= s,
with solutions s = 1,
5 2, and

5 2. The one that lies in (0, 1),

5 2 0.2361, is the
probability
0
.
Problems
1. For a branching process with ospring distribution given by p
0
=
1
6
, p
1
=
1
2
, p
3
=
1
3
, determine
(a) the expectation and variance of X
9
, the population at generation 9, (b) the probability that
the branching process dies by generation 3, but not by generation 2, and (c) the probability that
the process ever dies out. Then, assume that you start 5 independent copies of this branching
process at the same time (equivalently, change X
0
to 5), and (d) compute the probability that
the process ever dies out.
2. Assume that the ospring distribution of a branching process is Poisson with parameter .
(a) Determine the expected combined population through generation 10. (b) Determine, with
the aid of computer if necessary, the probability that the process ever dies out for =
1
2
, = 1,
and = 2.
3. Assume that the ospring distribution of a branching process is given by p
1
= p
2
= p
3
=
1
3
.
Note that p
0
= 0. Solve the following problem for a = 1, 2, 3. Let Y
n
be the proportion of
individuals in generation n (out of the total number of X
n
individuals) from families of size
a. (A family consists of individuals that are ospring of the same parent from the previous
generation.) Compute the limit of Y
n
as n .
1. For (a), compute =
3
2
,
2
=
7
2

9
4
=
5
4
, and plug into the formula. Then compute
(s) =
1
6
+
1
2
s +
1
3
s
3
.
For (b),
P(X
3
= 0) P(X
2
= 0) = (((0))) ((0)) 0.0462.
For (c), we solve (s) = s, 0 = 2s
3
3s +1 = (s 1)(2s
2
+2s 1), and so
0
=
31
2
0.3660.
For (d), the answer is
5
0
.
2. For (a), = and
E(X
0
+X
1
+. . . +X
10
) = EX
0
+EX
1
+. . . +EX
10
= 1 + + +
10
=

11
1
1
,
if ,= 1, and 11 if = 1. For (b), if 1 then
0
= 1, but if > 1, then
0
is the solution for
s (0, 1) to
e
(s1)
= s.
This equation cannot be solved analytically, but we can numerically obtain the solution for = 2
to get
0
0.2032.
3. Assuming X
n1
= k, the number of families at time n is also k. Each of these has, in-
dependently, a members with probability p
a
. If k is large which it will be for large n, as
the branching process cannot die out then, with overwhelming probability, the number of
children in such families is about ap
a
k, while X
n
is about k. Then, the proportion Y
n
is about
ap
a
, which works out to be

1
6
,
1
3
, and
1
2
, for a = 1, 2, and 3.
15 MARKOV CHAINS: LIMITING PROBABILITIES 169
15 Markov Chains: Limiting Probabilities
Example 15.1. Assume that the transition matrix is given by
P =
_
_
0.7 0.2 0.1
0.4 0.6 0
0 1 0
_
_
.
Recall that the n-step transition probabilities are given by powers of P. Let us look at some
large powers of P, beginning with
P
4
=
_
_
0.5401 0.4056 0.0543
0.5412 0.4048 0.054
0.54 0.408 0.052
_
_
.
Then, to four decimal places,
P
8
_
_
0.5405 0.4054 0.0541
0.5405 0.4054 0.0541
0.5405 0.4054 0.0541
_
_
.
and subsequent powers are the same to this precision.
The matrix elements appear to converge and the rows become almost identical. Why? What
determines the limit? These questions will be answered in this chapter.
We say that a state i S has period d 1 if (1) P
n
ii
> 0 implies that d[n and (2) d is the
largest positive integer that satises (1).
Example 15.2. Simple random walk on Z, with p (0, 1). The period of any state is 2 because
the walker can return to her original position in any even number of steps, but in no odd number
of steps.
Example 15.3. Random walk on the vertices of a square. Again, the period of any state is 2
for the same reason.
Example 15.4. Random walk on the vertices of a triangle. The period of any state is 1
because the walker can return in two steps (one step out and then back) or three steps (around
the triangle).
Example 15.5. Deterministic cycle. If a chain has n states 0, 1, . . . , n1 and transitions from
i to (i + 1) mod n with probability 1, then it has period n. So, any period is possible.
However, if the following two transition probabilities are changed: P
01
= 0.9 and P
00
= 0.1,
then the chain has period 1. In fact, the period of any state i with P
ii
> 0 is trivially 1.
It can be shown that the period is the same for all states in the same class. If a state, and
therefore its class, has period 1, it is called aperiodic. If the chain is irreducible, we call the
entire chain aperiodic if all states have period 1.
For a state i S, let
R
i
= infn 1 : X
n
= i
be the rst time, after time 0, that the chain is at i S. Also, let
f
(n)
i
= P(R
i
= n[X
0
= i)
be the p. m. f. of R
i
when the starting state is i itself (in which case we may call R
i
the return
time). We can connect these to the familiar quantity
f
i
= P(ever reenter i[X
0
= i) =
n=1
f
(n)
i
,
so that the state i is recurrent exactly when

n=1
f
(n)
i
= 1. Then, we dene
m
i
= E[R
i
[X
0
= i] =
n=1
nf
(n)
i
.
If the above series converges, i.e., m
i
< , then we say that i is positive recurrent. It can be
shown that positive recurrence is also a class property: a state shares it with all members of its
class. Thus, an irreducible chain is positive recurrent if each of its states is.
It is not hard to show that a nite irreducible chain is positive recurrent. In this case, there
must exist an m 1 and an > 0 so that i can be reached from any j in at most m steps with
probability at least . Then, P(R
i
n) (1 )
n/m
, which goes to 0 geometrically fast.
We now state the key theorems. Some of these have rather involved proofs (although none
is exceptionally dicult), which we will merely sketch or omit altogether.
Theorem 15.1. Proportion of the time spent at i.
Assume that the chain is irreducible and positive recurrent. Let N
n
(i) be the number of visits to
i in the time interval from 0 through n. Then,
lim
n
N
n
(i)
n
=
1
m
i
,
in probability.
Proof. The idea is quite simple: once the chain visits i, it returns, on average, once per m
i
time
steps, hence the proportion of time spent there is 1/m
i
. We skip a more detailed proof.
A vector of probabilities,
i
, i S, such that
iS

i
= 1 is called an invariant, or stationary,
distribution for a Markov chain with transition matrix P if
iS
i
P
ij
=
j
, for all j S.
In matrix form, if we put into a row vector [
1
,
2
, . . .], then
[
1
,
2
, . . .] P = [
1
,
2
, . . .].
Thus [
1
,
2
, . . .] is a left eigenvector of P, for eigenvalue 1. More important for us is the following
probabilistic interpretation. If
i
is the p. m. f. for X
0
, that is, P(X
0
= i) =
i
, for all i S, it
is also the p. m. f. for X
1
and hence for all other X
n
, that is, P(X
n
= i) =
i
, for all n.
Theorem 15.2. Existence and uniqueness of invariant distributions.
An irreducible positive recurrent Markov chain has a unique invariant distribution, which is
given by
i
=
1
m
i
.
In fact, an irreducible chain is positive recurrent if and only if a stationary distribution exists.
The formula for should not be a surprise: if the probability that the chain is at i is always
i
, then one should expect that the proportion of time spent at i, which we already know to be
1/m
i
, to be equal to
i
. We will not, however, go deeper into the proof.
Theorem 15.3. Convergence to invariant distribution.
If a Markov chain is irreducible, aperiodic, and positive recurrent, then, for every i, j S,
lim
n
P
n
ij
=
j
.
Recall that P
n
ij
= P(X
n
= j[X
0
= i) and note that the limit is independent of the initial
state. Thus, the rows of P
n
are more and more similar to the row vector as n becomes large.
The most elegant proof of this theorem uses coupling, an important idea rst developed by
a young French probabilist Wolfgang Doeblin in the late 1930s. (Doeblins life is a romantic,
and quite tragic, story. An immigrant from Germany, he died as a soldier in the French army
in 1940, at the age of 25. He made signicant mathematical contributions during his army
service.) Start with two independent copies of the chain two particles moving from state to
state according to the transition probabilities one started from i, the other using the initial
distribution . Under the stated assumptions, they will eventually meet. Afterwards, the two
particles move together in unison, that is, they are coupled. Thus, the dierence between the
two probabilities at time n is bounded above by twice the probability that coupling does not
happen by time n, which goes to 0. We will not go into greater detail, but, as we will see in the
next example, periodicity is necessary.
Example 15.6. A deterministic cycle with a = 3 has the transition matrix
P =
_
_
0 1 0
0 0 1
1 0 0
_
_
.
This is an irreducible chain with the invariant distribution
0
=
1
=
2
=
1
3
(as it is easy to
check). Moreover,
P
2
=
_
_
0 0 1
1 0 0
0 1 0
_
_
,
P
3
= I, P
4
= P, etc. Although the chain does spend 1/3 of the time at each state, the transition
probabilities are a periodic sequence of 0s and 1s and do not converge.
Our nal theorem is mostly a summary of the results for the special, and for us the most
common, case.
Theorem 15.4. Convergence theorem for a nite state space S.
Assume that s Markov chain with a nite state space is irreducible.
1. There exists a unique invariant distribution given by
i
=
1
m
i
.
2. For every i, irrespective of the initial state,
1
n
N
n
(i)
i
,
in probability.
3. If the chain is also aperiodic, then, for all i and j,
P
n
ij

j
.
4. If the chain is periodic with period d, then, for every pair i, j S, there exists an integer
r, 0 r d 1, so that
lim
m
P
md+r
ij
= d
j
and so that P
n
ij
= 0 for all n such that n ,= r mod d.
Example 15.7. We begin by our rst example, Example 15.1. That was clearly an irreducible
and aperiodic (note that P
00
> 0) chain. The invariant distribution [
1
,
2
,
3
] is given by
0.7
1
+ 0.4
2
=
1
0.2
1
+ 0.6
2
+
3
=
2
0.1
1
=
3
This system has innitely many solutions and we need to use
1
+
2
+
3
= 1
to get the unique solution
1
=
20
37
0.5405,
2
=
15
37
0.4054,
3
=
2
37
0.0541.
Example 15.8. The general two-state Markov chain. Here S = 1, 2 and
P =
_
1
1
_
and we assume that 0 < , < 1.
1
+
2
=
1
(1 )
1
+ (1 )
2
=
2
1
+
2
=1
and, after some algebra,
1
=

1 +
,
2
=
1
1 +
.
Here are a few common follow-up questions:
Start the chain at 1. In the long run, what proportion of time does the chain spend at 2?
Answer:
2
(and this does not depend on the starting state).
Start the chain at 2. What is the expected return time to 2? Answer:
1
2
.
In the long run, what proportion of time is the chain at 2, while at the previous time it
was at 1? Answer:
1
P
12
, as it needs to be at 1 at the previous time and then make a
transition to 2 (again, the answer does not depend on the starting state).
Example 15.9. In this example we will see how to compute the average length of time a chain
remains in a subset of states, once the subset is entered. Assume that a machine can be in 4
states labeled 1, 2, 3, and 4. In states 1 and 2 the machine is up, working properly. In states 3
and 4 the machine is down, out of order. Suppose that the transition matrix is
P =
_
_
1
4
1
4
1
2
0
0
1
4
1
2
1
4
1
4
1
4
1
4
1
4
1
4
1
4
0
1
2
_
_
.
(a) Compute the average length of time the machine remains up after it goes up. (b) Compute
the proportion of time that the system is up, but down at the next time step (this is called the
breakdown rate).
We begin with computing the invariant distribution, which works out to be
1
=
9
48
,
2
=
12
48
,
3
=
14
48
,
4
=
13
48
. Then, the breakdown rate is
1
(P
13
+P
14
) +
2
(P
23
+P
24
) =
9
32
,
the answer to (b).
Now, let u be the average stretch of time the machine remains up and let d be the average
stretch of time it is down. We need to compute u to answer (a). We will achieve this by
guring out the two equations that u and d must satisfy. For the rst equation, we use that the
proportion of time the system is up is
u
u +d
=
1
+
2
=
21
48
.
For the second equation, we use that there is a single breakdown in every interval of time
consisting of the stretch of up time followed by the stretch of down time, i.e., from one repair
to the next repair. This means
1
u +d
=
9
32
,
the breakdown rate from (b). The system of two equations gives d = 2 and, to answer (a),
u =
14
9
.
Computing the invariant distribution amounts to solving a system of linear equations. Nowa-
days this is not dicult to do, even for enormous systems; still, it is worthwhile to observe that
there are cases when the invariant distribution is easy to identify.
We call a square matrix with nonnegative entries doubly stochastic if the sum of the entries
in each row and in each column is 1.
Proposition 15.5. Invariant distribution in a doubly stochastic case.
If the transition matrix for an irreducible Markov chain with a nite state space S is doubly
stochastic, its (unique) invariant measure is uniform over S.
Proof. Assume that S = 1, . . . , m, as usual. If [1, . . . , 1] is the row vector with m 1s, then
[1, . . . , 1]P is exactly the vector of column sums, thus [1, . . . , 1]. This vector is preserved by right
multiplication by P, as is
1
m
[1, . . . , 1]. This vector species the uniform p. m. f. on S.
Example 15.10. Simple random walk on a circle. Pick a probability p (0, 1). Assume that
a points labeled 0, 1, . . . , a 1 are arranged on a circle clockwise. From i, the walker moves to
i + 1 (with a identied with 0) with probability p and to i 1 (with 1 identied with a 1)
with probability 1 p. The transition matrix is
P =
_
_
0 p 0 0 . . . 0 0 1 p
1 p 0 p 0 . . . 0 0 0
0 1 p 0 p . . . 0 0 0
. . .
0 0 0 0 . . . 1 p 0 p
p 0 0 0 . . . 0 1 p 0
_
_
and is doubly stochastic. Moreover, the chain is aperiodic if a is odd and otherwise periodic
with period 2. Therefore, the proportion of time the walker spends at any state is
1
a
, which is
also the limit of P
n
ij
for all i and j if a is odd. If a is even, then P
n
ij
= 0 if (i j) and n have a
dierent parity, while if they are of the same parity, P
n
ij

2
a
.
Assume that we change the transition probabilities a little: assume that, only when the
walker is at 0, she stays at 0 with probability r (0, 1), moves to 1 with probability (1 r)p,
and to a 1 with probability (1 r)(1 p). The other transition probabilities are unchanged.
Clearly, now the chain is aperiodic for any a, but the transition matrix is no longer doubly
stochastic. What happens to the invariant distribution?
The walker spends a longer time at 0; if we stop the clock while she stays at 0, the chain
is the same as before and spends an equal an proportion of time at all states. It follows that
our perturbed chain spends the same proportion of time at all states except 0, where it spends
a Geometric(1 r) time at every visit. Therefore,
0
is larger by the factor
1
1r
than other
i
.
Thus, the row vector with invariant distributions is
1
1
1r
+a 1
_
1
1r
1 . . . 1
=
_
1
1+(1r)(a1)
1r
1+(1r)(a1)
. . .
1r
1+(1r)(a1)
_
.
Thus, we can still determine the invariant distribution if only the self-transition probabilities P
ii
are changed.
Problems
1. Consider the chain in Problem 2 of Chapter 12. (a) Determine the invariant distribution. (b)
Determine lim
n
P
n
10
. Why does it exist?
2. Consider the chain in Problem 4 of Chapter 12 with the same initial state. Determine the
proportion of time the walker spends at a.
3. Roll a fair die n times and let S
n
be the sum of the numbers you roll. Determine, with proof,
lim
n
P(S
n
mod 13 = 0).
4. Peter owns two pairs of running shoes. Each morning he goes running. He is equally likely
to leave from his front or back door. Upon leaving the house, he chooses a pair of running shoes
at the door from which he leaves or goes running barefoot if there are no shoes there. On his
return, he is equally likely to enter at either door and leaves his shoes (if any) there. (a) What
proportion of days does he run barefoot? (b) What proportion of days is there at least one pair
of shoes at the front door (before he goes running)? (c) Now, assume also that the pairs of shoes
are green and red and that he chooses a pair at random if he has a choice. What proportion of
mornings does he run in green shoes?
5. Prof. Messi does one of three service tasks every year, coded as 1, 2, and 3. The assignment
changes randomly from year to year as a Markov chain with transition matrix
_
_
7
10
1
5
1
10
1
5
3
5
1
5
1
10
2
5
1
5
_
_
.
Determine the proportion of years that Messi has the same assignment as the previous two years.
6. Consider the Markov chain with states 0, 1, 2, 3, 4, which transition from state i > 0 to one
of the states 0, . . . , i 1 with equal probability, and transition from 0 to 4 with probability 1.
Show that all P
n
ij
converge as n and determine the limits.
1. The chain is irreducible and aperiodic. Moreover, (a) =
_
10
21
,
5
21
,
6
21
and (b) the limit is
0
=
10
21
.
2. The chain is irreducible and aperiodic. Moreover, =
_
5
41
,
8
41
,
8
41
,
20
41
and the answer is
1
+
3
=
13
41
.
3. Consider S
n
mod 13. This is a Markov chain with states 0, 1 . . . , 12 and transition matrix is
_
_
0
1
6
1
6
1
6
1
6
1
6
1
6
0 . . . 0
. . .
1
6
1
6
1
6
1
6
1
6
1
6
0 0 . . . 0
_
_
.
(To get the next row, shift cyclically to the right.) This is a doubly stochastic matrix with
i
=
1
13
, for all i. So the answer is
1
13
.
4. Consider the Markov chain with states given by the number of shoes at the front door. Then
P =
_
_
3
4
1
4
0
1
4
1
2
1
4
0
1
4
3
4
_
_
.
This is a doubly stochastic matrix with
0
=
1
=
2
=
1
3
. Answers: (a)
0
1
2
+
2
1
2
=
1
3
; (b)
1
+
2
=
2
3
; (c)
0
1
4
+
2
1
4
+
1
1
2
=
1
3
.
5. Solve for the invariant distribution: =
_
6
17
,
7
17
,
4
17
. The answer is
1
P
2
11
+
2
P
2
22
+
3
P
2
33
=
19
50
.
6. As the chain is irreducible and aperiodic, P
n
ij
converges to
j
, j = 0, 1, 2, 3, 4, where is given
by =
_
12
37
,
6
37
,
4
37
,
3
37
,
12
37
.
to solve the four problems, which you may assume have equal point score.
1. Suppose that whether it rains on a day or not depends on whether it did so on the previous
two days. If it rained yesterday and today, then it will rain tomorrow with probability 0.7;
if it rained yesterday but not today, it will rain tomorrow with probability 0.4; if it did not
rain yesterday, but rained today, it will rain tomorrow with probability 0.5; if it did not rain
yesterday nor today, it will rain tomorrow with probability 0.2.
(a) Let X
n
be the Markov chain with 4 states, (R, R), (N, R), (R, N), (N, N), which code
(weather yesterday, weather today) with R = rain and N = no rain. Write down the tran-
sition probability matrix for this chain. (b) Today is Wednesday and it is raining. It also rained
yesterday. Explain how you would compute the probability that it will rain on Saturday. Do
not carry out the computation.
(c) Under the same assumption as in (b), explain how you would approximate the probability
of rain on a day exactly a year from today. Carefully justify your answer, but do not carry out
the computation.
2. Consider the Markov chain with states 1, 2, 3, 4, 5, given by the following transition matrix:
P =
_
_
1
2
0
1
2
0 0
1
4
1
2
1
4
0 0
1
2
0
1
2
0 0
0 0 0
1
2
1
2
0 0 0
1
2
1
2
_
_
.
Specify the classes and determine whether they are transient or recurrent.
3. In a branching process the number of descendants is determined as follows. An individual
rst tosses a coin that comes out Heads with probability p. If this coin comes out Tails, the
individual has no descendants. If the coin comes Heads, the individual has 1 or 2 descendants,
each with probability
1
2
.
(a) Compute
0
, the probability that the branching process eventually dies out. Your answer
will, of course, depend on the parameter p.
(b) Write down the expression for the probability that the branching process is still alive at
generation 3. Do not simplify.
4. A random walker is in one of the four states, 0, 1, 2, or 3. If she is at i at some time, she
makes the following transition. With probability
1
2
she moves from i to (i +1) mod 4 (that is, if
she is at 0 she moves to 1, from 1 she moves to 2, from 2 to 3, and from 3 to 0). With probability
1
2
, she moves to a random state among the four states, each chosen with equal probability.
(a) Show that this chain has a unique invariant distribution and compute it. (Take a good look
at the transition matrix before you start solving this).
(b) After the walker makes many steps, compute the proportion of time she spends at 1. Does
the answer depend on the chains starting point?
(c) After the walker makes many steps, compute the proportion of times she is at the same state
as at the previous time.
1. Suppose that whether it rains on a day or not depends on whether it did so on the previous
two days. If it rained yesterday and today, then it will rain tomorrow with probability 0.7;
if it rained yesterday but not today, it will rain tomorrow with probability 0.4; if it did
not rain yesterday, but rained today, it will rain tomorrow with probability 0.5; if it did
not rain yesterday nor today, it will rain tomorrow with probability 0.2.
(a) Let X
n
be the Markov chain with 4 states, (R, R), (N, R), (R, N), (N, N), which code
(weather yesterday, weather today) with R = rain and N = no rain. Write down the
transition probability matrix for this chain.
Solution:
Let (R, R) be state 1, (N, R) state 2, (R, N) state 3, and (N, N) state 4. The
transition matrix is
P =
_
_
0.7 0 0.3 0
0.5 0 0.5 0
0 0.4 0 0.6
0 0.2 0 0.8
_
_
.
(b) Today is Wednesday and it is raining. It also rained yesterday. Explain how you
would compute the probability that it will rain on Saturday. Do not carry out the
computation.
Solution: If Wednesday is time 0, then Saturday is time 3. The initial state is given by
the row [1, 0, 0, 0] and it will rain on Saturday if we end up at state 1 or 2. Therefore,
our solution is
_
1 0 0 0

P
3
_
1
1
0
0
_
_
,
that is, the sum of the rst two entries of
_
1 0 0 0
P
3
.
(c) Under the same assumption as in (b), explain how you would approximate the prob-
ability of rain on a day exactly a year from today. Carefully justify your answer, but
do not carry out the computation.
Solution:
The matrix P is irreducible since the chain makes the following transitions with
positive probability: (R, R) (R, N) (N, N) (N, R) (R, R). It is also
aperiodic because the transition (R, R) (R, R) has positive probability. Therefore,
the probability can be approximated by
1
+
2
, where [
1
,
2
,
3
,
4
] is the unique
solution to [
1
,
2
,
3
,
4
] P = [
1
,
2
,
3
,
4
] and
1
+
2
+
3
+
4
= 1.
2. Consider the Markov chain with states 1, 2, 3, 4, 5, given by the following transition
matrix:
P =
_
_
1
2
0
1
2
0 0
1
4
1
2
1
4
0 0
1
2
0
1
2
0 0
0 0 0
1
2
1
2
0 0 0
1
2
1
2
_
_
.
Specify all classes and determine whether they are transient or recurrent.
Solution:
Answer:
2 is transient;
1, 3 is recurrent;
4, 5 is recurrent.
3. In a branching process, the number of descendants is determined as follows. An individual
rst tosses a coin that comes out Heads with probability p. If this coin comes out Tails,
the individual has no descendants. If the coin comes Heads, the individual has 1 or 2
descendants, each with probability
1
2
.
(a) Compute
0
, the probability that the branching process eventually dies out. Your
answer will, of course, depend on the parameter p.
Solution:
The probability mass function for the number of descendants is
_
0 1 2
1 p
p
2
p
2
_
,
and so
E(number of descendants) =
p
2
+p =
3p
2
.
If
3p
2
1, i.e., p
2
3
, then
0
= 1. Otherwise, we need to compute (s) and solve
(s) = s. We have
(s) = 1 p +
p
2
s +
p
2
s
2
.
Then,
s = 1 p +
p
2
s +
p
2
s
2
,
0 = ps
2
+ (p 2)s + 2(1 p),
0 = (s 1)(ps 2(1 p)).
We conclude that
0
=
2(1p)
p
, if p >
2
3
.
(b) Write down the expression for the probability that the branching process is still alive
at generation 3. Do not simplify.
Solution:
The answer is 1 (((0))) and we compute
(0) = 1 p,
((0)) = 1 p +
p
2
(1 p) +
p
2
(1 p)
2
,
1 (((0))) = 1
_
1 p +
p
2
(1 p +
p
2
(1 p)+
p
2
(1 p)
2
) +
p
2
(1 p +
p
2
(1 p) +
p
2
(1 p)
2
)
2
_
.
4. A random walker is at one of the four states, 0, 1, 2, or 3. If she at i at some time, she
makes the following transition. With probability
1
2
she moves from i to (i + 1) mod 4
(that is, if she is at 0 she moves to 1, from 1 she moves to 2, from 2 to 3, and from 3 to
0). With probability
1
2
, she moves to a random state among the four states, each chosen
with equal probability.
(a) Show that this chain has a unique invariant distribution and compute it. (Take a
good look at the transition matrix before you start solving this).
Solution:
The transition matrix is
P =
_
_
1
8
5
8
1
8
1
8
1
8
1
8
5
8
1
8
1
8
1
8
1
8
5
8
5
8
1
8
1
8
1
8
_
_
.
As P is a doubly stochastic irreducible matrix, = [
1
4
,
1
4
,
1
4
,
1
4
] is the unique invariant
distribution. (Note that irreducibility is trivial, as all entries are positive).
(b) After the walker makes many steps, compute the proportion of time she spends at 1.
Does the answer depend on the chains starting point?
Solution:
The proportion equals
1
=
1
4
, independently of the starting point.
(c) After the walker makes many steps, compute the proportion of time she is at the
same state as at the previous time.
Solution:
The probability of staying at the same state is always
1
8
, which is the answer.
16 MARKOV CHAINS: REVERSIBILITY 184
16 Markov Chains: Reversibility
Assume that you have an irreducible and positive recurrent chain, started at its unique invariant
distribution . Recall that this means that is the p. m. f. of X
0
and of all other X
n
as well.
Now suppose that, for every n, X
0
, X
1
, . . . , X
n
have the same joint p. m. f. as their time-reversal
X
n
, X
n1
, . . . , X
0
. Then, we call the chain reversible sometimes it is, equivalently, also
said that its invariant distribution is reversible. This means that a recorded simulation of a
reversible chain looks the same if the movie is run backwards.
Is there a condition for reversibility that can be easily checked? The rst thing to observe
is that for the chain started at , reversible or not, the time-reversed chain has the Markov
property. This is not completely intuitively clear, but can be checked:
P(X
k
= i[X
k+1
= j, X
k+2
= i
k+2
, . . . , X
n
= i
n
)
=
P(X
k
= i, X
k+1
= j, X
k+2
= i
k+2
, . . . , X
n
= i
n
)
P(X
k+1
= j, X
k+2
= i
k+2
, . . . , X
n
= i
n
)
=

i
P
ij
P
ji
k+2
P
i
n1
in
j
P
ji
k+2
P
i
n1
i
n
=

i
P
ij
j
,
which is an expression dependent only on i and j. For reversibility, this expression must be
the same as the forward transition probability P(X
k+1
= i[X
k
= j) = P
ji
. Conversely, if both
the original and the time-reversed chain have the same transition probabilities (and we already
know that the two start at the same invariant distribution and that both are Markov), then
their p. m. f.s must agree. We have proved the following useful result.
Theorem 16.1. Reversibility condition.
A Markov chain with invariant measure is reversible if and only if
i
P
ij
=
j
P
ji
,
for all states i and j.
Another useful fact is that once reversibility is checked, invariance is automatic.
Proposition 16.2. Reversibility implies invariance. If a probability mass function
i
satises
the condition in the previous theorem, then it is invariant.
Proof. We need to check that, for every j,
j
=

i
i
P
ij
, and here is how we do it:
i
P
ij
=
j
P
ji
=
j
i
P
ji
=
j
.
We now proceed to describe random walks on weighted graphs, the most easily recognizable
examples of reversible chains. Assume that every undirected edge between vertices i and j in a
complete graph has weight w
ij
= w
ji
; we think of edges with zero weight as not present at all.
When at i, the walker goes to j with probability proportional to w
ij
, so that
P
ij
=
w
ij
k
w
ik
.
What makes such random walks easy to analyze is the existence of a simple reversible measure.
Let
s =
i,k
w
ik
be the sum of all weights and let
i
=
k
w
ik
s
.
To see why this is a reversible distribution, compute
i
P
ij
=
k
w
ik
s

w
ij
k
w
ik
=
w
ij
s
,
which clearly remains the same if we switch i and j.
We should observe that this chain is irreducible exactly when the graph with present edges
(those with w
ij
> 0) is connected. The graph can only be periodic and the period can only be
2 (because the walker can always return in two steps) when it is bipartite: the set of vertices V
is divided into two sets V
1
and V
2
with every edge connecting a vertex from V
1
to a vertex from
V
2
. Finally, we note that there is no reason to forbid self-edges: some of the weights w
ii
may be
nonzero. (However, each w
ii
appears only once in s, while each w
ij
, with i ,= j, appears there
twice.)
By far the most common examples have no self-edges and all nonzero weights equal to 1
we already have a name for these cases: random walks on graphs. The number of neighbors of
a vertex is commonly called its degree. Then, the invariant distribution is
i
=
degree of i
2 (number of all edges)
.
Example 16.1. Consider the random walk on the graph below.
1
2
3 4
5
6
What is the proportion of time the walk spends at vertex 2?
The reversible distribution is
1
=
3
18
,
2
=
4
18
,
3
=
2
18
,
4
=
3
18
,
5
=
3
18
,
6
=
3
18
,
and, thus, the answer is
2
9
.
Assume now that the walker may stay at a vertex with probability P
ii
, but when she does
move she moves to a random neighbor as before. How can we choose P
ii
so that becomes
uniform,
i
=
1
6
for all i?
We should choose the weights of self-edges so that the sum of the weights of all edges
emanating from any vertex is the same. Thus, w
22
= 0, w
33
= 2, and w
ii
= 1, for all other i.
Example 16.2. Ehrenfest chain. We have M balls distributed in two urns. Each time, pick a
ball at random, move it from the urn where it currently resides to the other urn. Let X
n
be the
number of balls in urn 1. Prove that this chain has a reversible distribution.
The nonzero transition probabilities are
P
0,1
= 1,
P
M,M1
= 1,
P
i,i1
=
i
M
,
P
i,i+1
=
M i
M
.
Some inspiration: the invariant measure puts each ball at random into one of the two urns, as
switching any ball between the two urns does not alter this assignment. Thus, is Binomial(M,
1
2
),
i
=
_
M
i
_
1
2
M
.
Let us check that this is a reversible measure. The following equalities need to be veried:
0
P
01
=
1
P
10
,
i
P
i,i+1
=
i+1
P
i+1,i
,
i
P
i,i1
=
i1
P
i1,i
,
M
P
M,M1
=
M1
P
M1,M
,
and it is straightforward to do so. Note that this chain is irreducible, but not aperiodic (it has
period 2).
Example 16.3. Markov chain Monte Carlo. Assume that we have a very large probability
space, say some subset of S = 0, 1
V
, where V is a large set of n sites. Assume also that
we have a probability measure on S given via the energy (sometimes called the Hamiltonian)
function E : S R. The probability of any conguration S is
() =
1
Z
1
T
E()
.
Here, T > 0 is the temperature, a parameter, and Z is the normalizing constant that makes
S
() = 1. Such distributions frequently occur in statistical physics and are often called
Maxwell-Boltzmann distributions. They have numerous other applications, however, especially
in optimization problems, and have yielded an optimization technique called simulated annealing.
If T is very large, the role of the energy is diminished and the states are almost equally likely.
On the other hand, if T is very small, the large energy states have a much lower probability than
the small energy ones, thus the system is much more likely to be found in the close to minimal
energy states. If we want to nd states with small energy, we merely choose some small T and
generate at random, according to P, some states, and we have a reasonable answer. The only
problem is that, although E is typically a simple function, is very dicult to evaluate exactly,
as Z is some enormous sum. (There are a few celebrated cases, called exactly solvable systems,
in which exact computations are dicult, but possible.)
Instead of generating a random state directly, we design a Markov chain, which has as
its invariant distribution. It is common that the convergence to is quite fast and that the
necessary number of steps of the chain to get close to is some small power of n. This is in
startling contrast to the size of S, which is typically exponential in n. However, the convergence
slows down at a rate exponential in T
1
when T is small.
We will illustrate this on the Knapsack problem. Assume that you are a burglar and have
just broken into a jewelry store. You see a large number n of items, with weights w
i
and values
v
i
. Your backpack (knapsack) has a weight limit b. You are faced with a question of how to ll
your backpack, that is, you have to maximize the combined value of the items you will carry out
V = V (
1
, . . . ,
n
) =
n
i=1
v
i
i
,
subject to the constraints that
i
0, 1 and that the combined weight does not exceed the
backpack capacity,
n
i=1
w
i
i
b.
This problem is known to be NP-hard; there is no known algorithm to solve it quickly.
The set S of feasible solutions = (
1
, . . . ,
n
) that satisfy the constraints above will be our
state space and the energy function E on S is given as E = V , as we want to maximize V .
The temperature T measures how good a solution we are happy with the idea of simulated
annealing is, in fact, a gradual lowering of the temperature to improve the solution. There is
give and take: higher temperature improves the speed of convergence and lower temperature
improves the quality of the result.
Finally, we are ready to specify the Markov chain (sometimes called a Metropolis algorithm,
in honor of N. Metropolis, a pioneer in computational physics). Assume that the chain is at
state at time t, i.e., X
t
= . Pick a coordinate i, uniformly at random. Let
i
be the same as
except that its ith coordinate is ipped:
i
i
= 1
i
. (This means that the status of the ith
item is changed from in to out or from out to in.) If
i
is not feasible, then X
t+1
= and the
state is unchanged. Otherwise, evaluate the dierence in energy E(
i
) E() and proceed as
follows:
if E(
i
) E() 0, then make the transition to
i
, X
t+1
=
i
;
if E(
i
) E() > 0, then make the transition to
i
with probability e
1
T
(E()E(
i
))
, or
else stay at .
Note that, in the second case, the new state has higher energy, but, in physicists terms, we
tolerate the transition because of temperature, which corresponds to the energy input from the
environment.
We need to check that this chain is irreducible on S: to see this, note that we can get from
any feasible solution to an empty backpack by removing object one by one, and then back by
reversing the steps. Thus, the chain has a unique invariant measure, but is it the right one, that
is, ? In fact, the measure on S is reversible. We need to show that, for any pair ,
S,
()P(,
) = (
)P(
, ),
and this is enough to do with
=
i
, for arbitrary i, and assume that both are feasible (as
only such transitions are possible). Note rst that the normalizing constant Z cancels out (the
key feature of this method) and so does the probability
1
n
that i is chosen. If E(
i
) E() 0,
then the equality reduces to
e
1
T
E()
= e
1
T
E(
i
)
e
1
T
(E(
i
)E())
,
and similarly in the other case.
Problems
1. Determine the invariant distribution for the random walk in Examples 12.4 and 12.10.
2. A total of m white and m black balls are distributed into two urns, with m balls per urn. At
each step, a ball is randomly selected from each urn and the two balls are interchanged. The
state of this Markov chain can, thus, be described by the number of black balls in urn 1. Guess
the invariant measure for this chain and prove that it is reversible.
3. Each day, your opinion on a particular political issue is either positive, neutral, or negative.
If it is positive today, then it is neutral or negative tomorrow with equal probability. If it is
neutral or negative, it stays the same with probability 0.5, and, otherwise, it is equally likely to
be either of the other two possibilities. Is this a reversible Markov chain?
4. A king moves on a standard 8 8 chessboard. Each time, it makes one of the available legal
moves (to a horizontally, vertically or diagonally adjacent square) at random. (a) Assuming
that the king starts at one of the four corner squares of the chessboard, compute the expected
number of steps before it returns to the starting position. (b) Now you have two kings, they both
start at the same corner square and move independently. What is, now, the expected number
of steps before they simultaneously occupy the starting position?
1. Answer: =
_
1
5
,
3
10
,
1
5
,
3
10
.
2. If you choose m balls to put into urn 1 at random, you get
i
=
_
m
i
__
m
mi
_
_
2m
m
_ ,
and the transition probabilities are
P
i,i1
=
i
2
m
2
, P
i,i+1
=
(mi)
2
m
2
, P
i,i
=
2i(mi)
m
2
.
Reversibility check is routine.
3. If the three states are labeled in the order given, 1, 2, and 3, then we have
P =
_
_
0
1
2
1
2
1
4
1
2
1
4
1
4
1
4
1
2
_
_
.
The only way to check reversibility is to compute the invariant distribution
1
,
2
,
3
, form the
diagonal matrix D with
1
,
2
,
3
on the diagonal and to check that DP is symmetric. We get
1
=
1
5
,
2
=
2
5
,
3
=
2
5
, and DP is, indeed, symmetric, so the chain is reversible.
4. This is a random walk on a graph with 64 vertices (squares) and degrees 3 (4 corner squares),
5 (24 side squares), and 8 (36 remaining squares). If i is a corner square,
i
=
3
34+524+836
,
so the answer to (a) is
420
3
. In (b), you have two independent chains, so
(i,j)
=
i
j
and the
answer is
_
420
3
_
2
.
17 THREE APPLICATIONS 190
17 Three Applications
Parrondos Paradox
This famous paradox was constructed by the Spanish physicist J. Parrondo. We will consider
three games A, B and C with ve parameters: probabilities p, p
1
, p
2
, and , and an integer
period M 2. These parameters are, for now, general so that the description of the games is
more transparent. We will choose particular values once we are nished with the analysis.
We will call a game losing if, after playing it for a long time, a players capital becomes more
and more negative, i.e., the player loses more and more money.
Game A is very simple; in fact it is an asymmetric one-dimensional simple random walk.
Win $1, i.e., add +1 to your capital, with probability p, and lose a dollar, i.e., add 1 to your
capital, with probability 1 p. This is clearly a losing game if p <
1
2
.
In game B, the winning probabilities depend on whether your current capital is divisible by
M. If it is, you add +1 with probability p
1
, and 1 with probability 1 p
1
, and, if it is not,
you add +1 with probability p
2
and 1 with probability 1 p
2
. We will determine below when
this is a losing game.
Now consider game C, in which you, at every step, play A with probability and B with
probability 1 . Is it possible that A and B are losing games, while C is winning?
The surprising answer is yes! However, this should not be so surprising as in game B your
winning probabilities depend on the capital you have and you can manipulate the proportion
of time your capital spends at unfavorable amounts by playing the combination of the two
games.
We now provide a detailed analysis. As mentioned, game A is easy. To analyze game B, take
a simple random walk which makes a +1 step with probability p
2
and 1 step with probability
1 p
2
. Assume that you start this walk at some x, 0 < x < M. Then, by the Gamblers ruin
computation (Example 11.6),
(17.1) P(the walk hits M before 0) =
1
_
1p
2
p
2
_
x
1
_
1p
2
p
2
_
M
.
Starting from a multiple of M, the probability that you increase your capital by M before either
decreasing it by M or returning to the starting point is
(17.2) p
1
1
_
1p
2
p
2
_
1
_
1p
2
p
2
_
M
.
(You have to make a step to the right and then use the formula (17.1) with x = 1.) Similarly, from
a multiple of M, the probability that you decrease your capital by M before either increasing it
by M or returning to the starting point is
(17.3) (1 p
1
)
_
1p
2
p
2
_
M1
_
1p
2
p
2
_
M
1
_
1p
2
p
2
_
M
.
(Now you have to move one step to the left and then use 1(probability in (17.1) with x =
M 1).)
The main trick is to observe that game B is losing if (17.2)<(17.3). Why? Observe your
capital at multiples of M: if, starting from kM, the probability that the next (dierent) multiple
of M you visit is (k 1)M exceeds the probability that it is (k + 1)M, then the game is losing
and that is exactly when (17.2)<(17.3). After some algebra, this condition reduces to
(17.4)
(1 p
1
)(1 p
2
)
M1
p
1
p
M1
2
> 1.
Now, game C is the same as game B with p
1
and p
2
replaced by q
1
= p + (1 )p
1
and
q
2
= p + (1 )p
2
, yielding a winning game if
(17.5)
(1 q
1
)(1 q
2
)
M1
q
1
q
M1
2
< 1.
This is easily achieved with large enough M as soon as p
2
<
1
2
and q
2
>
1
2
, but even for M = 3,
one can choose p =
5
11
, p
1
=
1
11
2
, p
2
=
10
11
, =
1
2
, to get
6
5
in (17.4) and
217
300
in (17.5).
A Discrete Renewal Theorem
Theorem 17.1. Assume that f
1
, . . . , f
N
0 are given numbers with

N
k=1
f
k
= 1. Let =
N
k=1
k f
k
. Dene a sequence u
n
as follows:
u
n
= 0 if n < 0,
u
0
= 1,
u
n
=
N
k=1
f
k
u
nk
if n > 0.
Assume that the greatest common divisor of the set k : f
k
> 0 is 1. Then,
lim
n
u
n
=
1
.
Example 17.1. Roll a fair die forever and let S
m
be the sum of outcomes of the rst m rolls.
Let p
n
= P(S
m
ever equals n). Estimate p
10,000
.
One can write a linear recursion
p
0
= 1,
p
n
=
1
6
(p
n1
+ +p
n6
) ,
and then solve it, but this is a lot of work! (Note that one should either modify the recursion
for n 5 or, more easily, dene p
n
= 0 for n < 0.) By the above theorem, however, we can
immediately conclude that p
n
converges to
2
7
.
Example 17.2. Assume that a random walk starts from 0 and jumps from x either to x + 1
or to x + 2, with probability p and 1 p, respectively. What is, approximately, the probability
that the walk ever hits 10, 000? The recursion is now much simpler:
p
0
= 1,
p
n
= p p
n1
+ (1 p) p
n2
,
and we can solve it, but again we can avoid the work by applying the theorem to get that p
n
converges to
1
2p
.
Proof. We can assume, without loss of generality, that f
N
> 0 (or else reduce N).
Dene a Markov chain with state space S = 0, 1, . . . , N 1 by
_
_
f
1
1 f
1
0 0 . . .
f
2
1f
1
0
1f
1
f
2
1f
1
0 . . .
f
3
1f
1
f
2
0 0
1f
1
f
2
f
3
1f
1
f
2
. . .
. . .
f
N
1f
1
f
N1
0 0 0 . . .
_
_
.
This is called a renewal chain: it moves to the right (from x to x + 1) on the nonnegative
integers, except for renewals, i.e., jumps to 0. At N 1, the jump to 0 is certain (note that the
matrix entry P
N1,0
is 1, since the sum of f
k
s is 1).
The chain is irreducible (you can get to N 1 from anywhere, from N 1 to 0, and from
0 anywhere) and we will see shortly that is also aperiodic. If X
0
= 0 and R
0
is the rst return
time to 0, then
P(R
0
= k)
clearly equals f
1
, if k = 1. Then, for k = 2 it equals
(1 f
1
)
f
2
1 f
1
= f
2
,
and, then, for k = 3 it equals
(1 f
1
)
1 f
1
f
2
1 f
1
f
3
1 f
1
f
2
= f
3
,
and so on. We conclude that (recall again that X
0
= 0)
P(R
0
= k) = f
k
for all k 1.
In particular, the promised aperiodicity follows, as the chain can return to 0 in k steps if f
k
> 0.
Moreover, the expected return time to 0 is
m
00
=
N
k=1
kf
k
= .
The next observation is that the probability P
n
00
that the chain is at 0 in n steps is given by the
recursion
(17.6) P
n
00
=
n
k=1
P(R
0
= k)P
nk
00
.
To see this, observe that you must return to 0 at some time not exceeding n in order to end up
at 0; either you return for the rst time at time n or you return at some previous time k and,
then, you have to be back at 0 in n k steps.
The above formula (17.6) is true for every Markov chain. In this case, however, we note that
the rst return time to 0 is, certainly, at most N, so we can always sum to N with the proviso
that P
nk
00
= 0 when k > n. So, from (17.6) we get
P
n
00
=
N
k=1
f
k
P
nk
00
.
The recursion for P
n
00
is the same as the recursion for u
n
. The initial conditions are also the
same and we conclude that u
n
= P
n
00
. It follows from the convergence theorem (Theorem 15.3)
that
lim
n
u
n
= lim
n
P
n
00
=
1
m
00
=
1
,
which ends the proof.
Patterns in coin tosses
Assume that you repeatedly toss a coin, with Heads represented by 1 and Tails represented by
0. On any toss, 1 occurs with probability p. Assume also that you have a pattern of outcomes,
say 1011101. What is the expected number of tosses needed to obtain this pattern? It should
be about 2
7
= 128 when p =
1
2
, but what is it exactly? One can compare two patterns by this
waiting game, saying that the one with the smaller expected value wins.
Another way to compare two patterns is the horse race: you and your adversary each choose
a pattern, say 1001 and 0100, and the person whose pattern appears rst wins.
Here are the natural questions. How do we compute the expectations in the waiting game
and the probabilities in the horse race? Is the pattern that wins in the waiting game more
likely to win in the horse race? There are several ways of solving these problems (a particularly
elegant one uses the so called Optional Stopping Theorem for martingales), but we will use
Markov chains.
The Markov chain X
n
we will utilize has as the state space all patterns of length . Each
time, the chain transitions into the pattern obtained by appending 1 (with probability p) or 0
(with probability 1 p) at the right end of the current pattern and by deleting the symbol at
the left end of the current pattern. That is, the chain simply keeps track of the last symbols
in a sequence of tosses.
There is a slight problem before we have tosses. For now, assume that the chain starts
with some particular sequence of tosses, chosen in some way.
We can immediately gure out the invariant distribution for this chain. At any time n 2
and for any pattern A with k 1s and k 0s,
P(X
n
= A) = p
k
(1 p)
k
,
as the chain is generated by independent coin tosses! Therefore, the invariant distribution of
X
n
assigns to A the probability
A
= p
k
(1 p)
k
.
Now, if we have two patterns B and A, denote by N
BA
the expected number of additional
tosses we need to get A provided that the rst tosses ended in B. Here, if A is a subpattern of
B, this does not count, we have to actually make A in the additional tosses, although we can
use a part of B. For example, if B = 111001 and A = 110, and the next tosses are 10, then
N
BA
= 2, and, if the next tosses are 001110, then N
BA
= 6.
Also denote
E(B A) = E(N
BA
).
Our initial example can, therefore, be formulated as follows: compute
E( 1011101).
The convergence theorem for Markov chains guarantees that, for every A,
E(A A) =
1
A
.
The hard part of our problem is over. We now show how to analyze the waiting game by the
example.
We know that
E(1011101 1011101) =
1
1011101
.
However, starting with 1011101, we can only use the overlap 101 to help us get back to 1011101,
so that
E(1011101 1011101) = E(101 1011101).
To get from to 1011101, we have to get rst to 101 and then from there to 1011101, so that
E( 1011101) = E( 101) +E(101 1011101).
We have reduced the problem to 101 and we iterate our method:
E( 101) = E( 1) +E(1 101)
= E( 1) +E(101 101)
= E(1 1) +E(101 101)
=
1
1
+
1
101
.
The nal result is
E( 1011101) =
1
1011101
+
1
101
+
1
1
=
1
p
5
(1 p)
2
+
1
p
2
(1 p)
+
1
p
,
which is equal to 2
7
+ 2
3
+ 2 = 138 when p =
1
2
.
In general, the expected time E( A) can be computed by adding to 1/
A
all the overlaps
between A and its shifts, that is, all the patterns by which A begins and ends. In the example,
the overlaps are 101 and 1. The more overlaps A has, the larger E( A) is. Accordingly, for
p =
1
2
, of all patterns of length , the largest expectation is 2
+ 2
1
+ + 2 = 2
+1
2 (for
constant patterns 11 . . . 1 and 00 . . . 0) and the smallest is 2
when there is no overlap at all (for

example, for 100 . . . 0).
Now that we know how to compute the expectations in the waiting game, we will look at
the horse race. Fix two patterns A and B and let p
A
= P(A wins) and p
B
= P(B wins). The
trick is to consider the time N, the rst time one of the two appears. Then, we can write
N
A
= N +I
{B appears before A}
N
BA
,
where N
BA
is the additional number of tosses we need to get to A after we reach B for the rst
time. In words, to get to A we either stop at N or go further starting from B, but the second
case occurs only when B occurs before A. It is clear that N
BA
has the same distribution as
N
BA
and is independent of the event that B appears before A. (At the time B appears for the
rst time, what matters for N
BA
is that we are at B and not whether we have seen A earlier.)
Taking expectations,
E( A) = E(N) +p
B
E(B A),
E( B) = E(N) +p
A
E(A B),
p
A
+p
B
= 1.
We already know how to compute E( A), E( B), E(B A), and E(A B), so this is
a system of three equations with three unknowns: p
A
, p
B
and N.
Example 17.3. Let us return to the patterns A = 1001 and B = 0100, and p =
1
2
, and compute
the winning probabilities in the horse race.
We compute E( A) = 16 + 2 = 18 and E( B) = 16 + 2 = 18. Next, we compute
E(B A) = E(0100 1001). First, we note that E(0100 1001) = E(100 1001) and,
then, E( 1001) = E( 100) + E(100 1001), so that E(0100 1001) = E(
1001) E( 100) = 18 8 = 10. Similarly, E(A B) = 18 4 = 14, and, then, the above
three equations with three unknowns give p
A
=
5
12
, p
B
=
7
12
, E(N) =
73
6
.
We conclude with two examples, each somewhat paradoxical and thus illuminating.
Example 17.4. Consider sequences A = 1010 and B = 0100. It is straightforward to verify
that E( A) = 20, E( B) = 18, while p
A
=
9
14
. So, A loses in the waiting game, but wins
in the horse race! What is going on? Simply, when A loses in the horse race, it loses by a lot,
thereby tipping the waiting game towards B.
Example 17.5. This example concerns the horse race only. Consider the relation given by
A B if P(A beats B) 0.5. Naively, one would expect that this relation is transitive, but
this is not true! The simplest example are triples 011 100 001 011, with probabilities
1
2
,
3
4
and
2
3
.
Problems
1. Start at 0 and perform the following random walk on the integers. At each step, ip 3 fair
coins and make a jump forward equal to the number of Heads (you stay where you are if you
ip no Heads). Let p
n
be the probability that you ever hit n. Compute lim
n
p
n
. (It is not
2
3
!)
2. Suppose that you have three patterns A = 0110, B = 1010, C = 0010. Compute the
probability that A appears rst among the three in a sequence of fair coin tosses.
1. The size S of the step has the p. m. f. given by P(X = 0) =
1
8
, P(X = 1) =
3
8
, P(X = 2) =
3
8
,
P(X = 3) =
1
8
. Thus,
p
n
=
1
8
p
n
+
3
8
p
n1
+
3
8
p
n2
+
1
8
p
n3
,
and so
p
n
=
8
7
_
3
8
p
n1
+
3
8
p
n2
+
1
8
p
n3
_
.
It follows that p
n
converges to the reciprocal of
8
7
_
3
8
1 +
3
8
2 +
1
8
3
_
,
that is, to E(S[S > 0)
1
. The answer is
lim
n
p
n
=
7
12
.
2. If N is the rst time one of the three appears, we have
E( A) = EN +p
B
E(B A) +p
C
E(C A)
E( B) = EN +p
A
E(A B) +p
C
E(C B)
E( C) = EN +p
A
E(A C) +p
B
E(B C)
p
A
+p
B
+p
C
= 1
and
E( A) = 18
E( B) = 20
E( C) = 18
E(B A) = 16
E(C A) = 16
E(A B) = 16
E(C B) = 16
E(A C) = 16
E(B C) = 16
The solution is EN = 8, p
A
=
3
8
, p
B
=
1
4
, and p
C
=
3
8
. The answer is
3
8
.
18 POISSON PROCESS 198
18 Poisson Process
A counting process is a random process N(t), t 0, such that
1. N(t) is a nonnegative integer for each t;
2. N(t) is nondecreasing in t; and
3. N(t) is right-continuous.
The third condition is merely a convention: if the rst two events happen at t = 2 and t = 3,
then N(2) = 1, N(3) = 2, N(t) = 1 for t (2, 3), and N(t) = 0 for t < 2. Thus, N(t) N(s)
represents the number of events in (s, t].
A Poisson process with rate (or intensity) > 0 is a counting process N(t) such that
1. N(0) = 0;
2. it has independent increments: if (s
1
, t
1
]
(s
2
, t
2
] = , then N(t
1
) N(s
1
) and N(t
2
)
N(s
2
) are independent; and
3. the number of events in any interval of length t is Poisson(t).
In particular,
P(N(t +s) N(s) = k) = e
t
(t)
k
k!
, k = 0, 1, 2, . . . ,
E(N(t +s) N(s)) = t.
Moreover, as h 0,
P(N(h) = 1) = e
h
h h,
P(N(h) 2) = O(h
2
) h.
Thus, in small time intervals, a single event happens with probability proportional to the length
of the interval; this is why is called the rate.
A denition as the above should be followed by the question whether the object in question
exists we may be wishing for contradictory properties. To demonstrate the existence, we
will outline two constructions of the Poisson process. Yes, it is unique, but it would require
some probabilistic sophistication to prove this, as would the proof (or even the formulation) of
convergence in the rst construction we are about to give. Nevertheless, it is very useful, as it
makes many properties of the Poisson process almost instantly understandable.
Construction by tossing a low-probability coin very fast. Pick a large n and assume
that you have a coin with (low) Heads probability

n
. Toss the coin at times which are positive
integer multiples of
1
n
(that is, very fast) and let N
n
(t) be the number of Heads in [0, t]. Clearly,
as n , the number of Heads in any interval (s, t] is Binomial with the number of trials
n(t s) 2 and success probability

n
; thus, it converges to Poisson(t s), as n . Moreover,
N
n
has independent increments for any n and hence the same holds in the limit. We should
note that the Heads probability does not need to be exactly

n
, instead, it suces that this
probability converges to when multiplied by n. Similarly, we do not need all integer multiples
of
1
n
; it is enough that their number in [0, t], divided by n, converges to t in probability for any
xed t.
An example of a property that follows immediately is the following. Let S
k
be the time of
the kth (say, 3rd) event (which is a random time) and let N
k
(t) be the number of additional
events within time t after time S
k
. Then, N
k
(t) is another Poisson process, with the same rate
, as starting to count the Heads afresh after the kth Heads gives us the same process as if we
counted them from the beginning we can restart a Poisson process at the time of the kth
event. In fact, we can do so at any stopping time, a random time T with the property that
T = t depends only on the behavior of the Poisson process up to time t (i.e., depends on the
past, but not on the future). The Poisson process, restarted at a stopping time, has the same
properties as the original process started at time 0; this is called the strong Markov property.
As each N
k
is a Poisson process, N
k
(0) = 0, so two events in the original Poisson N(t)
process do not happen at the same time.
Let T
1
, T
2
, . . . be the interarrival times, where T
n
is the time elapsed between (n 1)st and
nth event. A typical example would be the times between consecutive buses arriving at a station.
Proposition 18.1. Distribution of interarrival times:
T
1
, T
2
, . . . are independent and Exponential().
Proof. We have
P(T
1
> t) = P(N(t) = 0) = e
t
,
which proves that T
1
is Exponential(). Moreover, for any s > 0 and any t > 0,
P(T
2
> t[T
1
= s) = P(no events in (s, s +t][T
1
= s) = P(N(t) = 0) = e
t
,
as events in (s, s + t] are not inuenced by what happens in [0, s]. So, T
2
is independent of T
1
and Exponential(). Similarly, we can establish that T
3
is independent of T
1
and T
2
with the
same distribution, and so on.
Construction by exponential interarrival times. We can use the above Proposition 18.1
for another construction of a Poisson process, which is convenient for simulations. Let T
1
, T
2
, . . .
be i. i. d. Exponential() random variables and let S
n
= T
1
+ . . . + T
n
be the waiting time for
the nth event. We dene N(t) to be the largest n so that S
n
t.
We know that ES
n
=
n
, but we can derive its density; the distribution is called Gamma(n, ).

We start with
P(S
n
> t) = P(N(t) < n) =
n1
j=0
e
t
(t)
j
j!
,
and then we dierentiate to get
f
S
n
(t) =
n1
j=0
1
j!
(e
t
(t)
j
+e
t
j(t)
j1
)
= e
t
k1
j=0
_
(t)
j
j!
+
(t)
j1
(j 1)!
_
= e
t
(t)
n1
(n 1)!
,
and so
f
S
n
(t) = e
t
(t)
n1
(n 1)!
.
Example 18.1. Consider a Poisson process with rate . Compute (a) E(time of the 10th
event), (b) P(the 10th event occurs 2 or more time units after the 9th event), (c) P(the 10th
event occurs later than time 20), and (d) P(2 events in [1, 4] and 3 events in [3, 5]).
The answer to (a) is
10
by Proposition 18.1. The answer to (b) is e

2
, as one can restart
the Poisson process at any event. The answer to (c) is P(S
10
> 20) = P(N(20) < 10), so we
can either write the integral
P(S
10
> 20) =
_

20
e
t
(t)
9
9!
dt,
or use
P(N(20) < 10) =
9
j=0
e
20
(20)
j
9!
.
To answer (d), we condition on the number of events in [3, 4]:
2
k=0
P(2 events in [1, 4] and 3 events in [3, 5] [ k events in [3, 4]) P(k events in [3, 4])
=
2
k=0
P(2 k events in [1, 3] and 3 k events in [4, 5]) P(k events in [3, 4])
=
2
k=0
e
2
(2)
2k
(2 k)!
e

3k
(3 k)!
e
k
k!
=e
4
_
1
3
5
+
4
+
1
2
3
_
.
Theorem 18.2. Superposition of independent Poisson processes.
Assume that N
1
(t) and N
2
(t) are independent Poisson processes with rates
1
and
2
. Combine
them into a single process by taking the union of both sets of events or, equivalently, N(t) =
N
1
(t) +N
2
(t). This is a Poisson process with rate
1
+
2
.
Proof. This is a consequence of the same property for Poisson random variables.
Theorem 18.3. Thinning of a Poisson process.
Each event in a Poisson process N(t) with rate is independently a Type I event with probability
p; the remaining events are Type II. Let N
1
(t) and N
2
(t) be the numbers of Type I and Type II
events in [0, t]. These are independent Poisson processes with rates p and (1 p).
The most substantial part of this theorem is independence, as the other claims follow from
the thinning properties of Poisson random variables (Example 11.4).
Proof. We argue by discrete approximation. At each integer multiple of
1
n
, we toss two indepen-
dent coins: coin A has Heads probability
p
n
; and coin B has Heads probability
(1p)
n
/
_
1
p
n
_
.
Then call discrete Type I events the locations with coin A Heads; discrete Type II(a) events the
locations with coin A Tails and coin B Heads; and discrete Type II(b) events the locations with
coin B Heads. A location is a discrete event if it is either a Type I or a Type II(a) event.
One can easily compute that a location
k
n
is a discrete event with probability

n
. Moreover,
given that a location is a discrete event, it is Type I with probability p. Therefore, the process
of discrete events and its division into Type I and Type II(a) events determines the discrete
versions of the processes in the statement of the theorem. Now, discrete Type I and Type II(a)
events are not independent (for example, both cannot occur at the same location), but discrete
Type I and Type II(b) events are independent (as they depend on dierent coins). The proof
will be concluded by showing that discrete Type II(a) and Type II(b) events have the same limit
as n . The Type I and Type II events will then be independent as limits of independent
discrete processes.
To prove the claimed asymptotic equality, observe rst that the discrete schemes (a) and (b)
result in a dierent outcome at a location
k
n
exactly when two events occur there: a discrete Type
I event and a discrete Type II(b) event. The probability that the two discrete Type II schemes
dier at
k
n
is thus at most
C
n
2
, for some constant C. This causes the expected number of such
double points in [0, t] to be at most
Ct
n
. Therefore, by the Markov inequality, an upper bound
for the probability that there is at least one double point in [0, t] is also
Ct
n
. This probability
goes to zero, as n , for any xed t and, consequently, discrete (a) and (b) schemes indeed
result in the same limit.
Example 18.2. Customers arrive at a store at a rate of 10 per hour. Each is either male or
female with probability
1
2
. Assume that you know that exactly 10 women entered within some
hour (say, 10 to 11am). (a) Compute the probability that exactly 10 men also entered. (b)
Compute the probability that at least 20 customers have entered.
Male and female arrivals are independent Poisson processes, with parameter
1
2
10 = 5, so
the answer to (a) is
e
5
5
10
10!
.
The answer to (b) is
k=10
P(k men entered) =
k=10
e
5
5
k
k!
= 1
9
k=0
e
5
5
k
k!
.
Example 18.3. Assume that cars arrive at a rate of 10 per hour. Assume that each car will
pick up a hitchhiker with probability
1
10
. You are second in line. What is the probability that
you will have to wait for more than 2 hours?
Cars that pick up hitchhikers are a Poisson process with rate 10
1
10
= 1. For this process,
P(T
1
+T
2
> 2) = P(N(2) 1) = e
2
(1 + 2) = 3e
2
.
Proposition 18.4. Order of events in independent Poisson processes.
Assume that we have two independent Poisson processes, N
1
(t) with rate
1
and N
2
(t) with rate
2
. The probability that n events occur in the rst process before m events occur in the second
process is
n+m1
k=n
_
n +m1
k
__

1
1
+
2
_
k
_

2
1
+
2
_
n+m1k
.
We can easily extend this idea to more than two independent Poisson processes; we will not
make a formal statement, but instead illustrate by the few examples below.
Proof. Start with a Poisson process with
1
+
2
, then independently decide for each event
whether it belongs to the rst process, with probability

1
1
+
2
, or to the second process, with
probability

2
1
+
2
. The obtained processes are independent and have the correct rates. The
probability we are interested in is the probability that among the rst m + n 1 events in the
combined process, n or more events belong to the rst process, which is the binomial probability
in the statement.
Example 18.4. Assume that
1
= 5,
2
= 1. Then,
P(5 events in the rst process before 1 in the second) =
_
5
6
_
5
and
P(5 events in the rst process before 2 in the second) =
6
k=5
_
6
k
__
5
6
_
k
_
1
6
_
6k
=
11 5
5
6
6
.
Example 18.5. You have three friends, A, B, and C. Each will call you after an Exponential
amount of time with expectation 30 minutes, 1 hour, and 2.5 hours, respectively. You will go
out with the rst friend that calls. What is the probability that you go out with A?
We could evaluate the triple integral, but we will avoid that. Interpret each call as the rst
event in the appropriate one of three Poisson processes with rates 2, 1, and
2
5
, assuming the
time unit to be one hour. (Recall that the rates are inverses of the expectations.)
We will solve the general problem with rates
1
,
2
, and
3
. Start with rate
1
+
2
+
3
Poisson process, distribute the events with probability

1
1
+
2
+
3
,

2
1
+
2
+
3
, and

3
1
+
2
+
3
,
respectively. The probability of A calling rst is clearly

1
1
+
2
+
3
, which in our case works out
to be
2
2 + 1 +
2
5
=
10
17
.
Our next theorem illustrates what we can say about previous event times if we either know
that their number by time t is k or we know that the kth one happens exactly at time t.
Theorem 18.5. Uniformity of previous event times.
1. Given that N(t) = k, the conditional distribution of the interarrival times, S
1
, . . . , S
k
, is
distributed as order statistics of k independent uniform variables: the set S
1
, . . . , S
k
is
equal, in distribution, to U
1
, . . . , U
k
, where U
i
are independent and uniform on [0, t].
2. Given that S
k
= t, S
1
, . . . , S
k1
are distributed as order statistics of k 1 independent
uniform random variables on [0, t].
Proof. Again, we discretize and the discrete counterpart is as follows. Assume that we toss a
coin, with arbitrary xed Heads probability, N times in succession and that we know that the
number of Heads in these N tosses is k. Then, these Heads occur in any of the
_
N
k
_
subsets (of
k tosses out of a total of N) with equal probability, simply by symmetry. This is exactly the
statement of the theorem, in the appropriate limit.
Example 18.6. Passengers arrive at a bus station as a Poisson process with rate .
(a) The only bus departs after a deterministic time T. Let W be the combined waiting time for
all passengers. Compute EW.
If S
1
, S
2
, . . . are the arrival times in [0, T], then the combined waiting time is W = T S
1
+
T S
3
+ . . .. Recall that we denote by N(t) the number of arrivals in [0, t]. We obtain the
answer by conditioning on the value of N(T): if we know that N(T) = k, then W is the sum
of k i. i. d. uniform random variables on [0, T]. Hence, W =

N(T)
k=1
U
k
, where U
1
, U
2
, . . . are
i. i. d. uniform on [0, T] and independent of N(T). By Theorem 11.1,
EW =
T
2
2
and Var(W) =
T
3
3
.
(b) Now two buses depart, one at T and one at S < T. What is EW now?
We have two independent Poisson processes in time intervals [0, S] and [S, T], so the answer
is
S
2
2
+
(T S)
2
2
.
(c) Now assume T, the only bus departure time, is Exponential(), independent of the passen-
gers arrivals.
This time,
EW =
_

0
E(W[T = t)f
T
(t) dt =
_

0
t
2
2
f
T
(t) dt =

2
E(T
2
)
=

2
(Var(T) + (ET)
2
) =

2
2
2
=

2
.
(d) Finally, two buses depart as the rst two events in a rate Poisson process.
This makes
EW = 2

2
.
Example 18.7. You have two machines. Machine 1 has lifetime T
1
, which is Exponential(
1
),
and Machine 2 has lifetime T
2
, which is Exponential(
2
). Machine 1 starts at time 0 and Machine
2 starts at a time T.
(a) Assume that T is deterministic. Compute the probability that M
1
is the rst to fail.
We could compute this via a double integral (which is a good exercise!), but instead we
proceed thus:
P(T
1
< T
2
+T) = P(T
1
< T) +P(T
1
T, T
1
< T
2
+T)
= P(T
1
< T) +P(T
1
< T
2
+T[T
1
T)P(T
1
T)
= 1 e
1
T
+P(T
1
T < T
2
[T
1
T)e
1
T
= 1 e
1
T
+P(T
1
< T
2
)e
1
T
= 1 e
1
T
+

1
1
+
2
e
1
T
.
The key observation above is that P(T
1
T < T
2
[T
1
T) = P(T
1
< T
2
). Why does this
hold? We can simply quote the memoryless property of the Exponential distribution, but it
is instructive to make a short argument using Poisson processes. Embed the failure times into
appropriate Poisson processes. Then, T
1
T means that no events in the rst process occur
during time [0, T]. Under this condition, T
1
T is the time of the rst event of the same process
restarted at T, but this restarted process is not inuenced by what happened before T, so the
condition (which in addition does not inuence T
2
) drops out.
(b) Answer the same question when T is Exponential() (and, of course, independent of the
machines). Now, by the same logic,
P(T
1
< T
2
+T) = P(T
1
< T) +P(T
1
T, T
1
< T
2
+T)
=

1
1
+
+

1
1
+
2
1
+
2
.
Example 18.8. Impatient hitchhikers. Two people, Alice and Bob, are hitchhiking. Cars that
would pick up a hitchhiker arrive as a Poisson process with rate
C
. Alice is rst in line for a
ride. Moreover, after Exponential(
A
) time, Alice quits, and after Exponential(
B
) time, Bob
quits. Compute the probability that Alice is picked up before she quits and compute the same
for Bob.
Embed each quitting time into an appropriate Poisson process, call these A and B processes,
and call the car arrivals C process. Clearly, Alice gets picked if the rst event in the combined
A and C process is a C event:
P(Alice gets picked) =

C
A
+
C
.
Moreover,
P(Bob gets picked)
= P(at least 2 C events before a B event
at least one A event before either a B or a C event,
and then at least one C event before a B event)
= P(at least 2 C events before a B event)
+P(at least one A event before either a B or a C event,
and then at least one C event before a B event)
P(at least one A event before either a B or a C event,
and then at least two C events before a B event)
=
_

C
B
+
C
_
2
+
_

A
A
+
B
+
C
__

C
B
+
C
_
_

A
A
+
B
+
C
__

C
B
+
C
_
2
=

A
+
C
A
+
B
+
C
B
+
C
.
This leaves us with an excellent hint that there may be a shorter way and, indeed, there is:
P(Bob gets picked)
= P(rst event is either A or C, and then the next event among B and C is C).
Problems
1. An oce has two clerks. Three people, A, B, and C enter simultaneously. A and B begin
service with the two clerks, while C waits for the rst available clerk. Assume that the service
time is Exponential(). (a) Compute the probability that A is the last to nish the service. (b)
Compute the expected time before C is nished (i.e., Cs combined waiting and service time).
2. A car wash has two stations, 1 and 2, with Exponential(
1
) and Exponential(
2
) service
times. A car enters at station 1. Upon completing the service at station 1, the car proceeds to
station 2, provided station 2 is free; otherwise, the car has to wait at station 1, blocking the
entrance of other cars. The car exits the wash after the service at station 2 is completed. When
you arrive at the wash there is a single car at station 1. Compute the expected time before you
exit.
3. A system has two server stations, 1 and 2, with Exponential(
1
) and Exponential(
2
) service
times. Whenever a new customer arrives, any customer in the system immediately departs.
Customer arrivals are a rate Poisson process, and a new arrival enters the system at station
1, then goes to station 2. (a) What proportion of customers complete their service? (b) What
proportion of customers stay in the system for more than 1 time unit, but do not complete the
service?
4. A machine needs frequent maintenance to stay on. The maintenance times occur as a Poisson
process with rate . Once the machine receives no maintenance for a time interval of length h,
it breaks down. It then needs to be repaired, which takes an Exponential() time, after which it
goes back on. (a) After the machine is started, nd the probability that the machine will break
down before receiving its rst maintenance. (b) Find the expected time for the rst breakdown.
(c) Find the proportion of time the machine is on.
5. Assume that certain events (say, power surges) occur as a Poisson process with rate 3 per hour.
These events cause damage to a certain system (say, a computer), thus, a special protecting unit
has been designed. That unit now has to be removed from the system for 10 minutes for service.
(a) Assume that a single event occurring in the service period will cause the system to crash.
What is the probability that the system will crash?
(b) Assume that the system will survive a single event, but two events occurring in the service
period will cause it to crash. What is, now, the probability that the system will crash?
(c) Assume that a crash will not happen unless there are two events within 5 minutes of each
other. Compute the probability that the system will crash.
(d) Solve (b) by assuming that the protective unit will be out of the system for a time which is
exponentially distributed with expectation 10 minutes.
1. (a) This is the probability that two events happen in a rate Poisson process before a single
event in an independent rate process, that is,
1
4
. (b) First, C has to wait for the rst event in
two combined Poisson processes, which is a single process with rate 2, and then for the service
time; the answer is
1
2
+
1
=
3
2
.
2. Your total time is (the time the other car spends at station 1) + (the time you spend at
station 2)+(maximum of the time the other car spends at station 2 and the time you spend at
station 1). If T
1
and T
2
are Exponential(
1
) and Exponential(
2
), then you need to compute
E(T
1
) +E(T
2
) +E(maxT
1
, T
2
).
Now use that
maxT
1
, T
2
= T
1
+T
2
minT
1
, T
2
and that minT

1
, T
2
is Exponential(
1
+
2
), to get
2
1
+
2
1
+
2
.
3. (a) A customer needs to complete the service at both stations before a new one arrives, thus
the answer is
1
+

2
2
+
.
(b) Let T
1
and T
2
be the customers times at stations 1 and 2. The event will happen if either:
T
1
> 1, no newcomers during time 1, and a newcomer during time [1, T
1
]; or
T
1
< 1, T
1
+T
2
> 1, no newcomers by time 1, and a newcomer during time [1, T
1
+T
2
].
For the rst case, nothing will happen by time 1, which has probability e
(+
1
)
. Then, after
time 1, a newcomer has to appear before the service time at station 1, which has probability
1
+
.
For the second case, conditioned on T
1
= t < 1, the probability is
e
2
(1t)

2
+
.
Therefore, the probability of the second case is
e
2
+
_
1
0
e
2
(1t)
1
e
1
t
dt = e
2
+
e
2
e
1
1
1
,
where the last factor is 1 when
1
=
2
. The answer is
e
(+
1
)

1
+
+

1
2
+
e
(+
2
)
e
1
1
1
.
4. (a) The answer is e
h
. (b) Let W be the waiting time for maintenance such that the next
maintenance is at least time h in the future, and let T
1
be the time of the rst maintenance.
Then, provided t < h,
E(W[T
1
= t) = t +EW,
as the process is restarted at time t. Therefore,
EW =
_
h
0
(t +EW)e
t
dt =
_
h
0
te
t
dt +EW
_
h
0
e
t
dt.
Computing the two integrals and solving for EW gives
EW =
1 he
h
e
h
e
h
.
The answer to (b) is EW +h (the machine waits for h more units before it breaks down). The
answer to (c) is
EW +h
EW +h +
1
.
5. Assume the time unit is 10 minutes,
1
6
of an hour. The answer to (a) is
P(N(1/6) 1) = 1 e
1
2
,
and to (b)
P(N(1/6) 2) = 1
3
2
e
1
2
.
For (c), if there are 0 or 1 events in the 10 minutes, there will be no crash, but 3 or more events
in the 10 minutes will cause a crash. The nal possibility is exactly two events, in which case
the crash will happen with probability
P
_
[U
1
U
2
[ <
1
2
_
,
where U
1
and U
2
are independent uniform random variables on [0, 1]. By drawing a picture, this
probability can be computed to be
3
4
. Therefore,
P(crash) = P(X > 2) +P(crash[X = 2)P(X = 2)
= 1 e
1
2

1
2
e
1
2
e
1
2
_
1
2
_
2
2
+
3
4

_
1
2
_
2
2
e
1
2
= 1
49
32
e
1
2
Finally, for (d), we need to calculate the probability that two events in a rate 3 Poisson process
occur before an event occurs in a rate 6 Poisson process. This probability is
_
3
3 + 6
_
2
=
1
9
.
Interlude: Practice Final
to solve the six problems, which you may assume have equal point score.
1. You select a random number X in [0, 1] (uniformly). Your friend then keeps selecting random
numbers U
1
, U
2
, . . . in [0, 1] (uniformly and independently) until he gets a number larger than
X/2, then he stops.
(a) Compute the expected number N of times your friend selects a number.
(b) Compute the expected sum S of the numbers your friend selects.
2. You are a member of a sports team and your coach has instituted the following policy. You
begin with zero warnings. After every game the coach evaluates whether youve had a discipline
problem during the game; if so, he gives you a warning. After you receive two warnings (not
necessarily in consecutive games), you are suspended for the next game and your warnings count
goes back to zero. After the suspension, the rules are the same as at the beginning. You gure
you will receive a warning after each game you play independently with probability p (0, 1).
(a) Let the state of the Markov chain be your warning count after a game. Write down the
transition matrix and determine whether this chain is irreducible and aperiodic. Compute its
invariant distribution.
(b) Write down an expression for the probability that you are suspended for both games 10 and
15. Do not evaluate.
(c) Let s
n
be the probability that you are suspended in the nth game. Compute lim
n
s
n
.
3. A random walker on the nonnegative integers starts at 0 and then at each step adds either 2
or 3 to her position, each with probability
1
2
.
(a) Compute the probability that the walker is at 2n + 3 after making n steps.
(b) Let p
n
be the probability that the walker ever hits n. Compute lim
n
p
n
.
4. A random walker is at one of the six vertices, labeled 0, 1, 2, 3, 4, and 5, of the graph in the
picture. At each time, she moves to a randomly chosen vertex connected to her current position
by an edge. (All choices are equally likely and she never stays at the same position for two
successive steps.)
1
2
3 4
5
0
(a) Compute the proportion of time the walker spends at 0, after she makes many steps. Does
this proportion depend on the walkers starting vertex?
(b) Compute the proportion of time the walker is at an odd state (1, 3, or 5) while, previously,
she was at even state (0, 2, or 4).
(c) Now assume that the walker starts at 0. What is expected number of steps she will take
before she is back at 0?
5. In a branching process, an individual has two descendants with probability
3
4
and no descen-
dants with probability
1
4
. The process starts with a single individual in generation 0.
(a) Compute the expected number of individuals in generation 2.
(b) Compute the probability that the process ever becomes extinct.
6. Customers arrive at two service stations, labeled 1 and 2, as a Poisson process with rate
. Assume that the time unit is one hour. Whenever a new customer arrives, any previous
customer is immediately ejected from the system. A new arrival enters the service at station 1,
then goes to station 2.
(a) Assume that the service time at each station is exactly 2 hours. What proportion of entering
customers will complete the service (before they are ejected)?
(b) Assume that the service time at each station now is exponential with expectation 2 hours.
What proportion of entering customers will now complete the service?
(c) Keep the service time assumption from (b). A customer arrives, but he is now given special
treatment: he will not be ejected unless at least three or more new customers arrive during his
service. Compute the probability that this special customer is allowed to complete his service.
Solutions to Practice Final
1. You select a random number X in [0, 1] (uniformly). Your friend then keeps selecting
random numbers U
1
, U
2
, . . . in [0, 1] (uniformly and independently) until he gets a number
larger than X/2, then he stops.
(a) Compute the expected number N of times your friend selects a number.
Solution:
Given X = x, N is distributed geometrically with success probability 1
x
2
, so
E[N[X = x] =
1
1
x
2
,
and so,
EN =
_
1
0
dx
1
x
2
= 2 log(1
x
2
)[
1
0
= 2 log 2.
(b) Compute the expected sum S of the numbers your friend selects.
Solution:
Given X = x and N = n, your friend selects n 1 numbers uniformly in [0,
x
2
] and
one number uniformly in [
x
2
, 1]. Therefore,
E[S[X = x, N = n] = (n 1)
x
4
+
1
2
_
1 +
x
2
_
=
1
4
nx +
1
2
,
E[S[X = x] =
1
4

1
1
x
2
x +
1
2
=
1
2 x
,
ES =
_
1
0
1
2 x
dx = log 2.
2. You are a member of a sports team and your coach has instituted the following policy.
You begin with zero warnings. After every game the coach evaluates whether youve had
a discipline problem during the game; if so, he gives you a warning. After you receive
two warnings (not necessarily in consecutive games), you are suspended for the next game
and your warnings count goes back to zero. After the suspension, the rules are the same
as at the beginning. You gure you will receive a warning after each game you play
independently with probability p (0, 1).
(a) Let the state of the Markov chain be your warning count after a game. Write down
the transition matrix and determine whether this chain is irreducible and aperiodic.
Compute its invariant distribution.
Solution:
The transition matrix is
P =
_
_
1 p p 0
0 1 p p
1 0 0
_
_
,
and the chain is clearly irreducible (the transitions 0 1 2 0 happen with
positive probability) and aperiodic (0 0 happens with positive probability). The
invariant distribution is given by
0
(1 p) +
2
=
0
0
p +
1
(1 p) =
1
1
p =
2
and
0
+
1
+
2
= 1,
which gives
=
_
1
p + 2
,
1
p + 2
,
p
p + 2
_
.
(b) Write down an expression for the probability that you are suspended for both games
10 and 15. Do not evaluate.
Solution:
You must have 2 warnings after game 9 and then again 2 warnings after game 14:
P
9
02
P
4
02
.
(c) Let s
n
be the probability that you are suspended in the nth game. Compute lim
n
s
n
.
Solution:
As the chain is irreducible and aperiodic,
lim
n
s
n
= lim
n
P(X
n1
= 2) =
p
2 +p
.
3. A random walker on the nonnegative integers starts at 0 and then at each step adds either
2 or 3 to her position, each with probability
1
2
.
(a) Compute the probability that the walker is at 2n + 3 after making n steps.
Solution:
The walker has to make (n 3) 2-steps and 3 3-steps, so the answer is
_
n
3
_
1
2
n
.
(b) Let p
n
be the probability that the walker ever hits n. Compute lim
n
p
n
.
Solution:
The step distribution is aperiodic, as the greatest common divisor of 2 and 3 is 1, so
lim
n
p
n
=
1
1
2
2 +
1
2
3
=
2
5
.
4. A random walker is at one of six vertices, labeled 0, 1, 2, 3, 4, and 5, of the graph in the
picture. At each time, she moves to a randomly chosen vertex connected to her current
position by an edge. (All choices are equally likely and she never stays at the same position
for two successive steps.)
1
2
3 4
5
0
(a) Compute the proportion of time the walker spends at 0, after she makes many steps.
Does this proportion depend on the walkers starting vertex?
Solution:
Independently of the starting vertex, the proportion is
0
, where [
0
,
1
,
2
,
3
,
4
,
5
]
is the unique invariant distribution. (Unique because of irreducibility.) This chain
is reversible with the invariant distribution given by
1
14
[4, 2, 2, 2, 3, 1]. Therefore, the
answer is
2
7
.
(b) Compute the proportion of times the walker is at an odd state (1, 3, or 5) while,
previously, she was at even state (0, 2, or 4).
Solution:
The answer is
0
(p
03
+p
05
) +
2
p
21
+
4
(p
43
+p
41
) =
2
7

2
4
+
1
7

1
2
+
3
14

2
3
=
5
14
.
(c) Now assume that the walker starts at 0. What is expected number of steps she will
take before she is back at 0?
Solution:
The answer is
1
0
=
7
2
5. In a branching process, an individual has two descendants with probability
3
4
and no
descendants with probability
1
4
. The process starts with a single individual in generation
0.
(a) Compute the expected number of individuals in generation 2.
Solution:
As
= 2
3
4
+ 0
1
4
=
3
2
,
the answer is
2
=
9
4
.
(b) Compute the probability that the process ever goes extinct.
Solution:
As
(s) =
1
4
+
3
4
s
2
the solution to (s) = s is given by 3s
2
4s + 1 = 0, i.e., (3s 1)(s 1) = 0. The
answer is
0
=
1
3
.
6. Customers arrive at two service stations, labeled 1 and 2, as a Poisson process with rate .
Assume that the time unit is one hour. Whenever a new customer arrives, any previous
customer is immediately ejected from the system. A new arrival enters the service at
station 1, then goes to station 2.
(a) Assume that the service time at each station is exactly 2 hours. What proportion of
entering customers will complete the service (before they are ejected)?
Solution:
The answer is
P(customer served) = P(no arrival in 4 hours) = e
4
.
(b) Assume that the service time at each station now is exponential with expectation 2
hours. What proportion of entering customers will now complete the service?
Solution:
Now,
P(customer served)
= P(2 or more arrivals in rate
1
2
Poisson process
before one arrival in rate Poisson process)
=
_
1
2
1
2
+
_
2
=
1
(1 + 2)
2
.
(c) Keep the service time assumption from (b). A customer arrives, but he is now given
special treatment: he will not be ejected unless at least three or more new customers
arrive during his service. Compute the probability that this special customer is
allowed to complete his service.
Solution:
The special customer is served exactly when 2 or more arrivals in rate
1
2
Poisson
before three arrivals in rate Poisson process happen. Equivalently, among the rst
4 arrivals in rate +
1
2
Poisson process, 2 or more belong to the rate
1
2
Poisson process.
The answer is
1
_

+
1
2
_
4
4
_

+
1
2
_
3
1
2
+
1
2
= 1

4
+ 2
3
_
+
1
2
_
4
.

Lecture Notes For Introductory Probability

Uploaded by

Copyright:

Available Formats

Lecture Notes For Introductory Probability

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lecture Notes For Introductory Probability

Uploaded by

Copyright:

Available Formats

Lecture Notes for Introductory Probability

). The answer, then, is

g(x) f(x) dx,

3n, this yields

, which can be used to get (very poor) approximations for .

f(x, y) dx = 0, and so f(x, y) = 0 for every x. So, we

is much larger than log

and to 0 with probability

. Moreover, from 0 it transitions to 1

(1) 1. In this case,

, which works out to be

and (b) the limit is

and the answer is

when there is no overlap at all (for

, but we can derive its density; the distribution is called Gamma(n, ).

by Proposition 18.1. The answer to (b) is e

and that minT

You might also like