Chapter 2

Discrete probability distributions

2.1 Synopsis of chapter

This chapter introduces the concept of random variables and discrete probability
distributions. These distributions are univariate, which means that they are used to
model a single numerical quantity. The concepts of expected value and variance of
discrete random variables are also discussed.

2.2 Learning outcomes

After completing this chapter, you should be able to:

formally dene a random variable and distinguish it from the values which it takes

explain the dierence between discrete and continuous random variables

summarise basic discrete distributions such as the uniform, Bernoulli, binomial,

Poisson, geometric and negative binomial.

2.3 Introduction
A random variable is a mapping of the elementary outcomes in the sample space to
real numbers. This allows us to attach probabilities to the experimental outcomes.
Hence the concept of a random variable is that of a measurement which takes a
particular value for each possible trial (experiment). Frequently, this will be a numerical
value.

Example 2.1 Suppose we sample ve people and measure their heights, hence
height is the random variable and the ve (observed) values of this random variable
are the realised measurements for the heights of these ve people.

Example 2.2 Suppose a fair die is thrown four times and we observe two 6s, a 3
and a 1. The random variable is the score on the die, and for these four trials it
takes the values 6, 6, 3 and 1. (In this case, since we do not know the true order in
which the values occurred, we could also say that the results were 1, 6, 3 and 6, or 1,
3, 6 and 6, or   .)

37
2. Discrete probability distributions

An example of an experiment with non-numerical outcomes would be a coin toss, for

which recall S = H, T . We can use a random variable, X, to convert the sample space
elements to real numbers: {
1 if heads
X=
0 if tails.
The value of any of the above variables will typically vary from sample to sample, hence
the name random variable.
So each experimental random variable has a collection of possible outcomes, and a
numerical value associated with each outcome. We have already encountered the term
sample space which here is the set of all possible numerical values of the random
variable.

Example 2.3 Examples of random variables include the following:

Experiment Random variable Sample space

Die is thrown Value on top face 1, 2, 3, 4, 5, 6
Coin is tossed ve times Number of heads 0, 1, 2, 3, 4, 5
Twenty people sampled Number with blue eyes 0, 1, 2,    , 19, 20
Machine operates for a day Number of breakdowns 0, 1, 2,   
One adult sampled Height in cm [100cm, 200cm] (roughly)

2.4 Probability distribution

A natural question to ask is what is the probability of any of these values?. That is, we
are interested in the probability distribution of the experimental random variable.
Be aware that random variables come in two varieties – discrete and continuous.1

Discrete and continuous random variables

Discrete: Synonymous with count data, that is, as far as this course is
concerned, random variables which take non-negative integer values, such as
0, 1, 2,   . For example, the number of heads in n coin tosses.

Continuous: Synonymous with measured data such as the real line, R =

(−∞, ∞), or some subset of R, for example the unit interval [0, 1]. For example,
the height of adults in centimetres.

The mathematical treatment of probability distributions depends on whether we are

dealing with discrete or continuous random variables. This chapter will explore the
former, while Chapter 3 will explore the latter.
In Example 2.3, the sample spaces of various experiments are shown. In most cases
there will be a higher chance of the random variable taking some sample space values
1
For completeness, be aware that mixture distributions (with discrete and continuous components)
exist, although they will not be considered in this course.

38
 2.4. Probability distribution

relative to others. Our objective is to express these chances using an associated

probability distribution. In the discrete case, we can associate with each point in the
sample space a probability which represents the chance of the random variable being
equal to that particular value. (The probability is typically non-zero, although
sometimes we need to use a probability of zero to identify impossible events.)
To summarise, a probability distribution is the complete set of sample space values with
their associated probabilities which, by axiom 2, must sum to 1 for discrete random
variables.2 The probability distribution can be represented diagrammatically by plotting
the probabilities against sample space values.
Finally, before we proceed, let us spend a moment to briey discuss some important
issues with regard to the notation associated with random variables. For notational
eciency reasons, we often use a capital letter to represent the random variable. The
letter X is often adopted, but it is perfectly legitimate to use any other letter: Y , Z etc.
In contrast, a lower case letter denotes a particular value of the random variable.

Example 2.4 Let X = the score of a fair die. If the die results in a 3, then this is
written as x = 3.
The probability distribution of X is:

X=x 1 2 3 4 5 6
P (X = x) 1/6 1/6 1/6 1/6 1/6 1/6

This is an example of the (discrete) uniform distribution.3 For discrete random

variables, we talk about a mass of probability at each respective sample space value.
In the discrete uniform case this mass is the same, i.e. 1/6, and this is plotted to
show the probability distribution of X, as shown in Figure 2.1.

Discrete uniform distribution

A random variable X has a discrete uniform distribution if it has k possible

outcomes, all of which are equally likely. If the distinct outcomes are 1, 2,    , k, then
the fact that k is nite means the distribution is discrete, and it is uniform because
all the probabilities are equal. The probability function is:
{
1k for x = 1, 2,    , k
P (X = x) = p(x) =
0 otherwise

For a fair die, k = 6.

2
When dealing with continuous random variables the analogous condition is integrating, rather than
summing, to 1. More of this in Chapter 3.
3
At school, uniforms are worn, i.e. all pupils wear the same clothes (possibly slight dierences across
genders), hence when the term uniform is applied to a probability distribution, we have the same
probability of occurrence for each sample space value.

39
2. Discrete probability distributions

Probability distribution: Score on die

1.0
0.8
0.6
Probability

0.4
0.2
0.0

1 2 3 4 5 6

Score

Figure 2.1: Probability distribution for the score on a fair die in Example 2.4.

Example 2.5 Let X = the number of heads when ve fair coins are tossed. The
probability distribution of X is:

X=x 0 1 2 3 4 5
P (X = x) 0.03 0.16 0.31 0.31 0.16 0.03
= (05)5 = 5 × (05)5 = 10 × (05)5 = 10 × (05)5 = 5 × (05)5 = (05)5

This is an example of the binomial distribution (discussed shortly) and can be

represented as:  
5
p(x) = × (05)5 for x = 0, 1, 2,    , 5
x
and 0 otherwise. The probability distribution of X is shown in Figure 2.2.

A probability distribution has a natural frequency interpretation – if the experiment is

repeated a very large number of times, then the probability of any particular value of
the random variable is equal to the limit of its relative frequency as the number of
experiments becomes innitely large.
There are many important probability distributions which describe the chances of
real-life events, and these form the basis of statistical inference and data analysis. The
binomial and Poisson distributions (both about counting) are discussed in this chapter,
while the normal and other important continuous distributions are covered in the
following chapter.

40
 2.5. Binomial distribution

Probability distribution: Number of heads for 5 tosses

1.0
0.8
0.6
Probability

0.4
0.2
0.0

0 1 2 3 4 5

Number of heads

Figure 2.2: Probability distribution for the number of heads when ve fair coins are
tossed.

2.5 Binomial distribution

The binomial distribution is a series of n independent Bernoulli trials. Hence it
makes sense to dene a Bernoulli trial rst of all. In fact, we have already seen an
example of this – the single coin toss. Key features of a Bernoulli trial are as follows.

A Bernoulli trial has only two possible outcomes (i.e. it is dichotomous) which are
typically called success and failure – such as heads and tails. We usually code
a success as 1 and a failure as 0.

There is a xed probability of success, π, and, therefore, a xed probability of

failure, 1 − π. So, for a fair coin, π = 050 – repeatedly tossing the same coin will
not change π.

Consequently, given a constant π, then successive Bernoulli trials are independent.

Bernoulli distribution

The probability distribution for a Bernoulli trial is:

X=x 0 1
P (X = x) 1−π π
and the Bernoulli distribution can be expressed with the following probability
function: {
π x (1 − π)1−x for x = 0, 1
P (X = x) =
0 otherwise.

41
2. Discrete probability distributions

Example 2.6 Other potential examples of Bernoulli trials are: (i.) the sex of
new-born babies (male or female), (ii.) the classication of factory output (defective
or not defective), and (iii.) voters supporting a candidate (support or not support).

In fact, many sampling situations become Bernoulli trials if we are only interested in
classifying the result categorically in one of two ways – for example, heights of people if
we are only interested in whether or not each person is taller than 180 cm, say.
Extending this idea, if we have n successive Bernoulli trials, then we dene the binomial
distribution.

Binomial distribution

Let X = the number of successes in a sequence of n independent and identically

distributed Bernoulli trials, then:4

X ∼ Bin(n, π)

where the terms n and π are called parameters, since the values of these dene
which specic binomial distribution we have. Its probability function is:
{ 
n
x
π x (1 − π)n−x for x = 0, 1, 2,    , n
P (X = x) = (2.1)
0 otherwise.

n is the number of Bernoulli trials, π is the (constant) probability of success for each
trial, P (X = x) is the probability that the number of successes in the n trials is equal
to x. That is, we are seeking to count the number of successes, and each P (X = x)
is the probability that the discrete (count) random variable X takes the value x.

(2.1) can be used to calculate probabilities for any binomial distribution, provided n
and π are both specied. Note that a binomial random variable can take n + 1 dierent
values, not n, since the variable measures the number of successes. The smallest number
of successes in n trials is zero (i.e. if all trials resulted in failure); the largest number of
successes is n (i.e. if all trials resulted in success); with the intervening number of
successes being 1, 2,    , n − 1. Therefore, there are n + 1 dierent values in total.

Necessary conditions to apply the binomial distribution

Each trial has only two possible outcomes – success and failure.

Fixed probability of success, π.

Fixed number of trials, n.

All trials are statistically independent.

4
Read ∼ as is distributed as.

42
 2.6. Cumulative distribution functions

2.6 Cumulative distribution functions

A probability function can be used to compute p(x) = P (X = x), i.e. the probability of
a single value x of the random variable. Of course, we may wish to know the probability
that the random variable X is less than or equal to x. We call such a probability a
cumulative probability, denoted by the cumulative distribution function (cdf ):

F (x) = P (X ≤ x) (2.2)

Cumulative distribution function for discrete random variables

For discrete random variables taking non-negative integer values, the cumulative
distribution function (cdf) is:5

F (x) = P (X = 0) + P (X = 1) + P (X = 2) + · · · + P (X = x)
= p(0) + p(1) + p(2) + · · · + p(x)

It follows that we can easily nd the probability function from the cumulative
distribution function, or vice versa, using this relationship. Specically, note that:

P (X = x) = F (x) − F (x − 1)

Example 2.7 Consider ten test tubes of bacterial solution and let us suppose that
the probability of any single test tube showing bacterial growth is 0.2. Let X denote
the number of test tubes showing bacterial growth. Hence:
 
10
P (exactly 4 show growth) = P (X = 4) = × (02)4 × (08)6
4

and:

P (more than 1 show growth) = 1 − F (1)

= 1 − P (X = 0) − P (X = 1)
 
10 10
= 1 − (08) − × (02)1 × (08)9
1
= 1 − 01074 − 02684
= 06242

Note this technique also illustrates the advantage of computing the probability of an
event by calculating the probability of it not happening and subtracting this from 1.6
5
Note you can use either form of notation p(x) or P (X = x), whichever you prefer.
6
Recall P (A) = 1 − P (Ac ).

43
2. Discrete probability distributions

2.6.1 Cumulative distribution functions – another point of view

X ∼ Bin(n, π) is one of many distributions which take only non-negative integer values,
and are typically about counting. It does not normally make sense to ask questions like:
What is P (X = −3), P (X = 15) or P (X = π)? for the obvious reason that all such
values are impossible,7 so the relevant probabilities are all zero.
However, it makes a lot more sense to ask questions like What is P (X ≤ −3),
P (X ≤ 15) or P (X ≤ π)?. The answers are 0, F (1) and F (0), respectively (using the
notation in (2.2)). This illustrates why we can dene the cdf of a counting-style random
variable even for values of a continuous argument x as:8

F (x) = P (X ≤ x)

With this convention, F (x) is a function whose graph is a step function.

Example 2.8 If X ∼ Bin(2, π), then each time x reaches an integer value in the
range [0, 2], the cdf jumps by P (X = x), until the sum of the P (X = x)s reaches 1.
This is shown in Figure 2.3.

F(x)
1

2
2π (1-π ) + (1-π )

2
(1-π )

0 1 2
x

Figure 2.3: Step function showing the cdf of X ∼ Bin(2, π) in Example 2.8.

The same pattern shown in Figure 2.3 applies to any version of Bin(n, π), or indeed to
any other distribution for which there is a largest possible integer value. For
distributions like the Poisson (discussed next) which count, but which do not have a
largest possible value, the pattern is similar, but the value 1 is never reached.

7
Technically speaking for n = 1, if π = 0 or 1 then P (X = π) = 1, but this means a success is
impossible or certain, respectively. Hence we no longer have two possible outcomes, but one certain
outcome, i.e. a failure or success, respectively.
8
Note that the argument x is not the same as the random variable X, nor is it the same as a realisation
or value of X. This is because X is not continuous, but takes only (selected) integer values. The x value
simply tells us the range of values in which we are interested.

44
 2.7. Poisson distribution

2.7 Poisson distribution

The Poisson distribution applies to random points occurring in a continuous
medium such as time, distance, area or volume. The discussion here will concentrate
mainly on one-dimensional cases, such as time or distance. In all cases, we are dealing
with random points which have the following properties.

Properties of random points in a Poisson process

Each point is equally likely to occur anywhere in the medium.

The position taken by each point is completely independent of the occurrence

or non-occurrence of all the other points.

In this situation the random variable X is the number of points in a particular unit of
the medium.

Poisson probability function

The probability function for the Poisson distribution is:

{
e−λ λx x! for x = 0, 1, 2,   
P (X = x) =
0 otherwise

where λ is the average number of points per unit of the medium, and is known as
the rate parameter. Note that, unlike the binomial distribution, there is no upper
bound on the value of x.

Example 2.9 Examples of a Poisson process include (i) machine breakdowns per
unit of time, (ii) arrivals at an airport per unit of time, and (iii) aws along a rope
per unit of length.

Example 2.10 Consider a machine which breaks down, on average, 3.2 times per
week, hence λ = 32 per week. The probability that it will break down exactly once
next week is:
e−32 (32)1
P (X = 1) = = 01304
1!
The probability that it will break down exactly four times in the next two weeks
(hence λ is now 6.4) is:

e−64 (64)4
P (X = 4) = = 01162
4!
Note that if we know λ for one unit of time (here, per week) and we want to look at
k units of time (in this example, k = 2), then we need to proportionally change λ to
reect this, i.e. the revised rate parameter is k × λ (hence in this example the revised
λ for a two-week period is 2 × 32 = 64).

45
2. Discrete probability distributions

2.8 Poisson approximation to the binomial

Try entering 70! in your calculator. I suspect you will encounter a computer says no
moment. (If not, try 71!, 72! etc.) In such cases we will have diculty in computing
 
binomial probabilities for large values of n, due to the use of factorials in the nx
component of the probability function.
Hence it would be useful to have a suitable approximation to the binomial distribution
when direct computation of binomial probabilities proves problematic.9 There are two
commonly-used such approximations to the binomial – one using the normal
distribution (covered in Chapter 3), and another using the Poisson distribution. Clearly,
any approximation is just that, an approximation, so we should only use this
approximating procedure when we obtain a good (i.e. close) approximation.

Conditions for using a Poisson approximation to the binomial

To justify use of the Poisson as an approximating distribution to the binomial, the

following conditions should hold.

n greater than 30.

π suciently extreme such that nπ < 10.

The approximation is only good for small values of x, relative to n.

Set the Poisson rate parameter λ = nπ.

Example 2.11 Suppose we sample 100 items at random from a production line
which is providing, on average, 2% defective items. What is the probability of
exactly 3 defective items in our random sample?
First, we have to check that the relevant criteria for using the Poisson approximation
are satised. Indeed they are. n = 100 > 30, π = 002 is suciently small such that
nπ = 2 < 10 and x = 3 is small relative to n. Hence:
e−2 23
P (X = 3) = 100 C3 × (002)3 × (098)97 = 01823 ≈ = 01804
   3!
true binomial probability

2.9 Expected value of a discrete random variable

Certain important properties of distributions arise if we consider probability-weighted
averages of random variables, and of functions of random variables.10 For example, we
might want to know the average value of a random variable.
9
The Windows scientic calculator can handle
√ large factorials, though I suspect this uses an
approximation such as Stirlings formula: n! ≈ 2nπnn e−n .
10
A function, f (X), of a random variable X is, of course, a new random variable, say Y = f (X).

46
 2.9. Expected value of a discrete random variable

It would be foolish to simply take the arithmetic average of all the values taken by the
random variable, as this would mean that very unlikely values (those with small
probabilities of occurrence) would receive the same weighting as very likely values
(those with large probabilities of occurrence). The obvious approach is to use the
probability-weighted average of the sample space values, known as the expected
value of X.

Expectation of a discrete random variable

If x1 , x2 ,    , xN are the possible values of the random variable X, with corresponding

probabilities p1 , p2 ,    , pN , then:
N
∑
E(X) = µ = xi p i = x1 p 1 + x2 p 2 + · · · + xN p N 
i=1

Note that the expected value is also referred to as the population mean, which can be
written as E(X) (in words the expectation of the random variable X), or µ (in words
the (population) mean of X). Also, note the distinction between the sample mean, x̄,
(introduced in ST104a Statistics 1) based on observed sample values, and the
population mean, µ, based on the theoretical probability distribution.

Example 2.12 If the random variable X happens to be a constant, k, then

x1 = k, and p1 = 1, so trivially E(X) = k × 1 = k.

Example 2.13 If X ∼ Bin(n, π), then:

n
∑
E(X) = x P (X = x)
x=0
     
n n 1 n−1 n 2 n−2 n n
= 0 × (1 − π) + 1 × π (1 − π) +2× π (1 − π) + ··· + n × π
1 2 n
= nπ

Why this reduces to nπ is beyond the scope of this course, but the fact that
E(X) = nπ for the binomial distribution is a useful result!

Example 2.14 If X ∼ Pois(λ), then:

∞
∑
E(X) = x P (X = x)
x=0

e−λ λ0 e−λ λ1 e−λ λ2 e−λ λk

=0× +1× +2× + ··· + k × + ···
0! 1! 2! k!
= λ

Again, why this reduces to λ is beyond the scope of this course, but again is a useful
result.

47
2. Discrete probability distributions

2.9.1 New random variables

Above we have labelled the population mean as the expectation of the random
variable and introduced
 the expectation operator, E(·). This operator, like the
summation operator , is a linear operator and hence this property can be used to nd
the expectation of a new random variable, be it a transformation of a single random
variable or a linear combination of two (or more) random variables.

Example 2.15 Suppose X is a random variable and α is a non-zero constant.

Dene W = αX to be a new random variable. What is the mean of W ? We have:
N
∑
E(W ) = E(αX) = (αxi )pi
i=1

= αx1 p1 + αx2 p2 + · · · + αxN pN

= α(x1 p1 + x2 p2 + · · · + xN pN )
= α E(X)

That is, E(αX) = α E(X).

Example 2.16 Suppose X and Y are random variables. Let Z = X ± Y be a new,

but clearly related, random variable. What is the mean of Z? To obtain this, simply
exploit the linear property of the expectation operator. Hence:

E(Z) = E(X ± Y ) = E(X) ± E(Y )

We can combine these two approaches.

Expectation of linear combinations of random variables

Given random variables X and Y , and constants α and β (both non-zero), dene
T = αX ± βY . It follows that:

E(T ) = E(αX ± βY ) = α E(X) ± β E(Y )

2.10 Variance of a discrete random variable

The concept of a probability-weighted average (or expected value) can be extended to
functions of the random variable.

48
 2.10. Variance of a discrete random variable

Example 2.17 If X takes the values x1 , x2 ,    , xN with corresponding

probabilities p1 , p2 ,    , pN , then:
  ∑N
1 1
E = pi for all xi 6= 0
X x
i=1 i

N
∑
E(ln(X)) = ln(xi ) pi for all xi > 0
i=1

N
∑
2
E(X ) = x2i pi 
i=1

One very important average associated with a distribution is the expected value of the
square of the deviation11 of the random variable from its mean, µ. This can be seen to
be a measure – not the only one, but the most widely used by far – of the dispersion of
the distribution and is known as the (population) variance of the random variable.

Variance of a discrete random variable

If X takes the values x1 , x2 ,    , xN with corresponding probabilities p1 , p2 ,    , pN ,

then the (population) variance of a discrete random variable is:
N
∑
2 2
σ = E((X − µ) ) = (xi − µ)2 pi 
i=1

The (positive) square root of the variance is known as the standard deviation and,
given the variance is typically denoted by σ 2 , is denoted by σ.

Example 2.18 Let X represent the value shown when a fair die is thrown once.
We now compute the mean and variance of X as follows.

X=x 1 2 3 4 5 6 Total

P (X = x) 1/6 1/6 1/6 1/6 1/6 1/6 1

x P (X = x) 1/6 2/6 3/6 4/6 5/6 6/6 216 = 35 = µ

(x − µ)2 25/4 9/4 1/4 1/4 9/4 25/4

(x − µ)2 P (X = x) 25/24 9/24 1/24 1/24 9/24 25/24 7024 = 292

√
Hence µ = E(X) = 35, σ 2 = E((X − µ)2 ) = 292 and σ = 292 = 171. This
tabular format has some advantages. Specically, note the following.
11
Which roughly means distance with sign.

49
2. Discrete probability distributions

It helps to have (and to calculate) the Total column since, for example, if a
probability, P (X = x), has been miscalculated or miscopied then the row total
will not be 1 (recall axiom 2). Therefore, this would highlight an error so, with a
little work, could be identied.

It is often useful to do a group of calculations as fractions over the same

denominator (as here in the nal row of the table), rather than to cancel or to
work with them as decimals, because important patterns can be more obvious,
and calculations can be easier.

2.10.1 Alternative expression for the variance

What follows is an extremely useful expression for the variance – worth remembering!

σ 2 = E((X − µ)2 ) = E(X 2 − 2µX + µ2 )

= E(X 2 ) − 2µ E(X) + µ2
= E(X 2 ) − 2µ2 + µ2
= E(X 2 ) − µ2 

In words, the (population) variance is equal to the mean of the square minus the square
of the mean. Rearranging gives:

E(X 2 ) = σ 2 + µ2 

This representation is useful since we often want to know E(X 2 ), but start by knowing
the usual details of a distribution, i.e. µ and σ 2 .

Example 2.19 Continuing with Example 2.18:

X=x 1 2 3 4 5 6 Total

P (X = x) 1/6 1/6 1/6 1/6 1/6 1/6 1

x P (X = x) 1/6 2/6 3/6 4/6 5/6 6/6 216 = 35 = µ

x2 P (X = x) 1/6 4/6 9/6 16/6 25/6 36/6 91/6

Hence µ = E(X) = 35, E(X 2 ) = 916, so the variance is 916 − (35)2 = 292, as
before. However, this method is usually easier.

2.10.2 Limits and special cases

A useful interpretation of population properties is to think of them as the limiting
equivalents of the corresponding sample statistics (as were introduced in ST104a
Statistics 1). Suppose we sample n values of random variable X and get x1 , x2 ,    , xn ,
then, as n → ∞:

50
 2.11. Distributions related to the binomial distribution

the sample mean, x̄, tends to the population mean, µ, i.e. x̄ → µ

the sample variance, s2 , tends to the population variance, σ 2 , i.e. s2 → σ 2 .

Variance of binomial and Poisson distributions

If X ∼ Bin(n, π), then:

Var(X) = nπ(1 − π)

If X ∼ Pois(λ), then:
Var(X) = λ

Note that for the Poisson distribution the mean and variance are equal.

2.10.3 New random variables (again)

As in the case of the expected value, we might want to look at linear combinations of
random variables.

Variance of functions of random variables

Given random variables X and Y and non-zero constants α and β, by dening two
new random variables U = αX and T = αX ± βY , then:12

Var(U ) = Var(αX) = α2 Var(X)

Assuming independence of X and Y , then:13

Var(T ) = Var(αX ± βY ) = α2 Var(X) + β 2 Var(Y )

2.11 Distributions related to the binomial distribution

There are many useful distributions which are related to the binomial distribution. Two
of these are summarised below. Each arises in similar contexts to the binomial
distribution which, of course, is when we have a xed number of independent Bernoulli
trials with constant probability of success π.

12
One way to remember this is to think of Var as a homogeneous function of degree 2, like the
Cobb–Douglas utility and production functions which crop up in economics.
13
We have already met independent events, but we do not yet know what it means for random variables
to be independent. This will be covered later.

51
2. Discrete probability distributions

2.11.1 Geometric distribution

The geometric distribution is used when we perform a series of Bernoulli trials until
we get the rst success. The random variable X is the trial number on which we obtain
this rst success. Hence x = 1, 2,   . So x = 1 corresponds to the rst success on the
rst trial, x = 2 corresponds to the rst success on the second trial etc.
If the rst success occurs on the xth trial, then there must have been x − 1 failures prior
to this, each with probability of occurrence of 1 − π. Given independence of the
Bernoulli trials, we can derive the probability function.

Probability function of the geometric distribution

If: {
(1 − π)x−1 π for x = 1, 2,   
P (X = x) =
0 otherwise
then X has a geometric distribution, denoted X ∼ Geo(π). It can be shown that
for the geometric distribution E(X) = 1π and Var(X) = (1 − π)π 2 .

2.11.2 Negative binomial distribution

The negative binomial distribution extends the geometric distribution in that the
Bernoulli trials are continued until the rth success is achieved. Hence the geometric
distribution is a special case of this, i.e. when r = 1.
Dene X to be the trial number of the rth success, then the smallest number of trials is
r, that is we obtain r consecutive successes from the very beginning. Of course, if x is
the trial number of the rth success, then this means we have previously incurred r − 1
successes, S, and x − r failures, F , which could have occurred in any order. Again, given
independence of the Bernoulli trials, we can derive the probability function, noting that:

x−1
P (X = x) = Cr−1 π r−1 (1 − π)x−r π
  
P (r−1 Ss & x−r F s)

Probability function of the negative binomial distribution

If: { 
x−1
r−1
π r (1 − π)x−r for x = r, r + 1, r + 2,   
P (X = x) =
0 otherwise
then X has a negative binomial distribution, denoted X ∼ Neg. Bin(r, π). It
can be shown that for the negative binomial distribution E(X) = rπ and
Var(X) = r(1 − π)π 2 .

52
 2.12. Overview of chapter

Example 2.20 Suppose we are conducting independent Bernoulli trials with

success probability π = 16 (for example, we might be rolling a fair die and need to
throw a 5 for success). If we want to know the probabilities for it to take k trials
(throws) to get seven successes, then the negative binomial distribution gives us:
 7
1
P (X = 7) =
6
   7  1
7 1 5
P (X = 8) = × ×
6 6 6
   7  2
8 1 5
P (X = 9) = × ×
6 6 6
..
.
   7  k−7
k−1 1 5
P (X = k) = × × 
6 6 6

2.12 Overview of chapter

This chaper has introduced discrete random variables. In particular, some common
families of probability distributions have been presented. In addition to the functional
form of each of these distributions, important properties (such as the expected value
and variance) have been studied.

2.13 Key terms and concepts

Bernoulli trial Binomial distribution
Cumulative distribution function Discrete
Expected value Geometric distribution
Negative binomial distribution Parameter
Poisson distribution Population mean
Population variance Probability distribution
Probability function Random variable
Step function Uniform distribution
Variance

2.14 Sample examination questions

1. Find P (X ≥ 2) when X follows a binomial distribution with parameters n = 10
and π = 025.

53
2. Discrete probability distributions

2. Suppose that a particle starts at the origin of the real line and moves along the line
in jumps of one unit. For each jump, the probability is π (where 0 ≤ π ≤ 1) that
the particle will jump one unit to the left, and hence the probability is 1 − π that
the particle will jump one unit to the right. Find the expected value of the position
of the particle after n jumps.

3. Suppose X follows a Poisson distribution such that P (X = 0) = 13. Calculate

P (X ≥ 2).

2.15 Solutions to Sample examination questions

1. Since X ∼ Bin(10, 025), then we have:

P (X ≥ 2) = 1 − P (X = 0) − P (X = 1) = 1 − (075)10 − 10 × 025 × (075)9 = 07560

2. Let Xi = 1 if the ith jump of the particle is one unit to the right, and let Xi = −1
if the ith jump is one unit to the left. Therefore, for i = 1, 2,    , n, we have:

E(Xi ) = −1 × π + 1 × (1 − π) = 1 − 2π

The position of the particle after n jumps is X1 + X2 + · · · + Xn , hence:

n
∑
E(X1 + X2 + · · · + Xn ) = E(Xi ) = n(1 − 2π)
i=1

3. We have that:
e−λ λ0 1
P (X = 0) = = e−λ = ⇒ λ ≈ 110
0! 3
Therefore:
e−110 (110)1 1
P (X ≥ 2) = 1 − P (X ≤ 1) = 1 − − = 03010
1! 3

Torture numbers, and theyll confess to anything.

(Gregg Easterbrook)

54