Ch2 Probability

anafoui@iustac.
ir
i
Chapter 2
Probability
2.1 Introd uction
Suppose that an anthropologist wants to classify fossil human skulls into two classes:
male and female. The anthropologist knows from previous experience that male skulls
tend to have larger circumferences than female skulls. However, within each group of
skulls there is considerable variation in the circumferences and their ranges overlap. If
it were necessary to guess the sexes of various skulls given only their circumferences,
it would be reasonable to guess male whenever the circumference is larger than some
value. This is an example of a classification decision (or guess) based on the value of
a single feature.
In the field of pattern recognition, we are interested in techniques for making the
best decisions possible. Even if we make the best decisions based on the available
information, we may not always be correct. If we used additional measurements from
the skull, we would probably be correct more often. There are various hereditary,
nutritional, and environmental causes for the variations in the circumferences of the
skulls, and if we could know them all for each skull, we could conceivably model or
explain why these particular circumferences were produced. An even more complex
chain of cause and effect would probably be required to explain why these particular
skulls were found in our sample. Rather than requiring complete understanding, a
probability model assumes that at least some of the variability in the data is due
to chance or random variability. In this chapter, we will discuss various probability
models. They will be used in the following two chapters as the basis for decision
making techniques.
For a random occurrence with a finite number of possible outcomes, such as ran-
domly choosing a skull from some group and measuring its circumference x to the
nearest centimeter, we can define a probability model by listing all the possible out-
[email protected]
20 CHAPTER 2. PROBABILITY 2.2. PROBABILITIES OF EVENTS
21
comes and the probability that each one occurs. We denote the probability that a
particular value x occurs by P(x).. A random process can be very simple, such as
flipping a single unbiased coin. In this case the outcomes are head and tail and the
corresponding probabilities are P(head) = 1/2 and P(tail) = 1/2. If the experiment
consists of rolling a fair die, the outcomes are 1, 2, .. . , 6 and the probability of each
outcome is 1/6.
Most people have an intuitive notion of the meaning of the word probability, but
it is used in various ways. A physician might tell a patient that, according to some
tests she has made, there is a 10 percent probability that the patient has disease A.
Some people would object to the use of the word "probability" in this situation and
point out that either the patient really has the disease or he does not, so the unknown
probability is either zero or 100 percent and cannot be anything between these two
values. They would not apply the word to individual events for which the outcome has
already been determined, but only to hypothetical experiments on collections of data
or assumed models. According to this point of view, the physician could say that 10
percent of people with these test results had disease A, or that the proba,bility is 10
percent that a randomly chosen past patient with these test results had the disease,
but should not say that there is a certain probability that a particular patient has the
disease. In this text, whenever we speak of the probability that a sample belongs to
a certain class, we do not refer to some particular sample, but to a randomly selected
sample that has the same set of feature values as the sample described.
Continuous variables such as the exact circumference of a skull are considered to
have an infinity of possible values. Continuous variables can be described by probability
densities, which will be discussed in Section 2.3, or their possible values can be broken
into ranges, and the probability of lying in each range can be listed. For example,
suppose that the probability is 0.02 that the circumference x of a skull is between 39.5
and 40.5 centimeters. We express this as P(39.5 :::: x :::: 40.5) = 0.02. A feature from
any real data set is automatically broken into a finite number of ranges when it is
measured, because measurements can only be recorded with finite precision.
Although choosing a probability model is easy for simple or idealized situations
such as flipping unbiased coins or rolling dice with assumed probabilities for various
faces showing, obtaining probabilities for real world situations is usually more difficult.
Two methods for obtaining probability estimates are called the frequentist approach
and the subjective approach.
In the frequentist approach, the probability of an event is estimated by dividing the
number of occurrences of an event by the number of trials. For example, to estimate the
probability of obtaining a defective light bulb from a certain manufacturing process,
we might sample 100 light bulbs and count the number that are defective. If we obtain
three defective light bulbs, we approximate or estimate the probability of obtaining a
defective light bulb from the process as 3/100 = 0.03. We can never know the exact
probability; but, in principle, by testing a large enough number of light bulbs, we
can estimate it as closely and as confidently as desired (short of absolute certainty),
assuming that the manufacturing process remains constant.
Although the frequentist approach is easy to understand, it may be difficult or
impossible to obtain enough samples to get an estimate of the true probability. Fur-
thermore, the frequentist approach only applies to repeatable events, that is, events
for which the probability is constant over all trials. This assumption is often difficult
to verify in the real world.
The probability that a certain candidate will win the next presidential election
cannot be estimated using the frequentist approach because the event is not repeatable.
For events that are not repeatable, a subjective assessment may be the only way to
assign a probability measure.. An intuitively appealing way to quantify the process of
selecting a subjective probability is to use the notion of a fair bet. You may not be
willing to bet "even money" that it will rain in Las Vegas on the next Fourth of July.
In fact, you would prefer to bet against rain because Las Vegas has a desert climate.
However, if you were offered 200 to 1 odds, you would probably prefer to bet on rain
if you knew that it does rain there occasionally in June or July. Somewhere between
the odds of 1 to 1 and the odds of 200 to 1, there is a bet with odds, say 15 to 1, for
which you would be equally willing to take either side. Such a bet is called it fair bet.
Suppose that the probability of the event on which you are betting is P. If the event
occurs, you win the amount W, so you expect to win PW on the average per bet (not
counting losses). If the event does not occur, you lose L, and since the probability that
the event will not occur is 1- P, you expect to lose an average of (1- P)L per bet (not
counting winnings). If the bet is fair, the average amount of money you win will equal
the average amount of money you lose, so PW = (1 - P)L or P = L/(L + W). Thus
the odds of 15 to 1 against rain are equivalent to the probability of 1/(15 + 1) = 1/16
that rain will occur and 15/16 that it will not.
We will use mainly the frequentist approach in this book. The subjective approach
is, however, useful in pattern recognition, especially when subjective expert opinions
are to be incorporated into a decision.
2.2 Probabilities of Events
The term experiment is used in probability theory to describe a process for which
the outcome is not known with certainty. Examples of experiments are
1. Rolling a fair six-sided die.
2. Randomly choosing ten transistors from a lot of 1,000 new transistors.
3. Selecting a newborn child at St. Luke's Hospital.
An event is an outcome or combination of outcomes from a statistical experiment.
The theory of probability studies the relative likelihood of the events that might occur
[email protected]
22
CHAPTER 2. PROBABILITY
when an experiment is performed. We will represent events by uppercase letters such
as A, B, and C. Examples of events that might occur as a result of performing the
previously listed experiments are
1. Obtaining a 6 when a fair six-sided die is rolled.
- 2. Obtaining an even number when a fair six-sided die is rolled.
3. Finding more than three defective transistors out of ten transistors randomly
chosen from a lot of 1,000 new transistors.
4. Selecting a randomly selected newborn child at St. Luke's Hospital weighing more
than eight pounds.
The event consisting of all possible outcomes of a statistical experiment is called
the sample space. We let S denote the sample space. The sample spaces for the
previously listed experiments are
1. The ,.'3amplespace consists of the numbers 1, 2, 3, 4, 5, and 6-all possible
outcomes of rolling a fair six-sided die.
2. The sample space consists of the numbers 0, 1, . . . , IQ-all possible numbers or"
defective transistors that might be obtained when ten transistors are randomly
chosen from a lot of 1,000 new transistors.
3. The sample space consists of all numbers that represent the possible weights of
randomly selected newborn children at St. Luke's Hospital.
A useful way to visualize relationships among events is to use a Venn diagram (see
Figure 2.1 for examples). The sample space S is represented by the entire rectangular
region.
Events are represented by regions inside the rectangles, and their areas can be made
to correspond to the probabilities of events. In Figure 2.1a, the event A is represented
by the shaded region. Because exactly one of the outcomes in S must occur in a single
trial,
peS) = 1.
If A is an event, then the event not A is called the complement of A and is shown
in Figure 2.1b. The complement of A is also denoted by A. Because it is a certainty
that either A or not A occurs, peA) + P(not A) = 1, so
P(not A) = 1 - peA).
If A and B are events, the event "either A or B or both occurs" is denoted by
A or B (see Figure 2.1c). This event is sometimes denoted by A UB (read "A union
A
\
23
2.2. PROBABILITIES OF EVENTS
s
s
()
0
notA
(a)
(b)
s
s
CID
B
CID
B
A
A
(c)
(cl)
Figure 2.1: The shaded areas represent the following events: (a) A. (b) A, (c) A or B,
(d) A and B.
anal oui @iust.ac.ir
24
GG
Figure 2.2: Classes A and B are mutually exclusive.
B") or A + B. The event "both A and B occur" is denoted by A and E (shown in
Figure 2.1d). This event is sometimes denoted by An B (read "A intersection B")
or AB. The event A and B is called a joint event.
If the event A and B cannot occur-that is, A and B cannot occur simultaneously- .
then the events A and B are said to be mutually exclusive (see Figure 2.2). If A
and B are mutually exclusive, the following equation called the addition rule holds:
P(A or B) = P(A) + P(B)
(2.1)
For example, if the probability that x will be the next president is 30 percent and the
probability that y will be the next president is 20 percent, then the probability that
one of them will be the next president is 50 percent.
If A and B are not mutually exclusive, there are four possible joint events A and B,
A and B, A and 13, and A and 13, each of which is mutually exclusive (see Figure 2.3).
The addition rule thus applies and we can write
P(A or B) = P((A and B) or (A and B) or (A and B))
- P(A and B) + P(A and B) + P(A and B)
Furthermore,
P(A and B) + P(A and B) = P(A)
and
P(A and B) + P(A and B) = P(B).
These three equations can be combinedto yield
P(A or B) = P(A) + P(B) - P(A and B). (2.2)
.I
.~
2.2. PROBABILITIES OF EVENTS 25
Aand B
Figure 2.3: A Venn diagram illustrating (2.2).
This is also evident in Figure 2.3. If P(A) is the area of the event A and P(B) is the
area of B, then P(A) + P(B) includes the area of the overlapped region P(A and B)
twice, so it must be subtracted to obtain the area of the event A or B.
Equation (2.2) can be used to compute the probability of drawing an ace or a
spade or both from a deck of cards. Because P(ace) = 1/13, P(spade) = 1/4, and
P(ace and spade) = 1/52,
P(ace or spade) = 1/13 + 1/4 - 1/52 = 4/13
Conditional Probabilities
If A and B are events, then the probability of A may depend on whether B occurs.
For example, the probability of rolling a 2 with a fair die is 1/6, but if we know that
the outcome was an even number, then the probability that a 2 was rolled is 1/3.
The conditional probability of A occurring, given that B has occurred, is denoted
P(AIB) and is read "P of A given B." Since we know in advance that B has occurred,
B effectively becomes the new sample space, so P(AIB) is the fraction of the B cases
in which A occurs. Thus we obtain the formula
P(AIB) = P(A and B)
P(B) .
This conditional probability is not defined if P(B) =O. Similarly,
(2.3)
P(BIA) = P(B and A)
P(A)
(2.4)
analoui@,iustac.ir
26 CHAPTER 2. PROBABILITY
The expressions (2.3) and (2.4), whichcan be rewritten as
P(A and B) = P(B)P(AIB) (2.5)
and
P(A and B) = P(A)P(BIA),
can also be used to calculate P(A and B).
(2.6)
Example 2.1 Calculating the conditional probability of rain given that the barometric
pressure is high.
Weather records show that high barometric pressure (defined as being over 760 mm of
mercury) occurred on 160 of the 200 days in a data set, and it rained on 20 of the 160
days with high barometric pressure. If we let R denote the event "rain occurred" and
H the event "high barometric pressure occurred" and use the frequentist approach to
define the probabilities, we see that
P(H) = 160/200 =0.80
and
P(R and H) =20/200 = 0.10.
We can obtain the probability of rain, given high pressure, directly from the data:
P(RIH) =20/160 =0.125.
Since it is given that high barometric pressure occurred, only those 160 days of high
barometric pressure are under consideration and rain occurre<I on 20 of those days.
We could also obtain P(RIH) from (2.3):
P(RIH) = P(R and H)/ P(H) = 0.10/0.80 =0.125.
Thus (2.3) is consistent with the concept of probability according to the frequentist
approach.
Whether we consider these ratios to be estimates or definitions of the various
probabilities depends on the population. If we randomly select one of the 200 days
in the data set, the probability is exactly 80 percent that there was high barometric
pressure on that day. Alternatively, if we consider the population to include all possible
days, there is no way to know the exact probability that there will be rain on one of
these days. However, we can use past experience to estimate the probabilities of these
events on days not included in our data set. In most practical problems, we are
interested in estimating the probabilities of future events, given some assumptions or
some facts concerning the past or present.
.~
\
27
The Multiplication Rule
In many important cases, P(A) may not depend on whether B has occurred. We say
that the event A is independent of B if P(A) =P(AIB). An important consequence
of the definition of independence is the multiplication rule, which is obtained by
substituting P(A) for P(AIB) in (2.5) to obtain
P(A and B) = P(AIB)P(B) = P(A)P(B)
(2.7)
whenever A is independent of B. If A is independent of B, we may combine (2.4) and
(2.7) to obtain
P(BIA) = P(B and A) = P(B)P(A)
P(A) P(A) = P(B),
so if A is independent of B, then B is also independent of A. From now on, we will
express either of these conditions as "A and B are independent." If events are not
independent, the probabilities of the various possible values of one of them depend on
the value of the other.
Suppose we want to calculate the probability of rolling a pair of 1s with two dice.
The multiplication rule can be used in this case because the number obtained on one
die is independent of the number obtained on the other die. Hence the probability of
getting ones on both dice is (1/6)(1/6) = 1/36.
Example 2.2 The addition and multiplication rules.
The probability of rolling a pair of dice such that the sum of the faces showing will
equal three can be found by using the addition and multiplication rules. There are
two ways in which the sum of two dice can equal three: Either the first die shows a 2
and the second die a 1, or the first shows a 1 and the second shows a 2. Let Fl be the
event that the first die shows a 1, and let F2 be the event that the first die shows a 2.
Let Gl and G2 be the events that the second die shows a 1 or a 2, respectively. Also
let E be the event that the sum of the two rolls is 3. If we assume that the rolls of the
two dice are independent, the multiplication rule applies. In addition, the events Fl
and F2 are mutually exclusive, as are Gl and G2, so we can use the addition rule:
P(E) = P((Fl and G2) or (F2 and Gd)
= P(Ft and G2) + P(F2 and Gd = P(1)P(2) + P(2)P(1)
- (1/6)(1/6)+ (1/6)(1/6)= 1/36+ 1/36= 1/18.
[email protected]
i
CHAPTER 2. PROBABILITY 28
Example 2.3 Estimating the probabilitiesof oil recoverytest outcomes.
As an example in which two events are not independent, suppose that a company plans
to test two new techniques for improving the extraction of oil from the ground. Th<
first technique consists of setting off an explosion at the bottom of a well to fracture
. the strata and then testing seismically to determine the extent of the fracturing. The
second technique consists of injecting hot brine into the well to loosen the oil and
then pumping to measure the oil recovery. Let E be the event that the explosion
successfully fractures the strata within a radius of 100 meters, and let R be the event
that oil can be recovered at a rate of 50 barrels per day after pumping in hot brine.
In a certain region, FeE) is estimated to be 0.8. If E occurs, the probability of R
occurring is estimated to be P(RIE) = 0.9, but if the explosion is not a success, the
estimated probability of recovery is only P(RIE) = 0.3. These three assumptions are
sufficient to define the probabilities of any combination of outcomes, given any set of
constraints. The probabilities of the four possible joint events are
P(E and R) = P(E)P(RIE) = (0.8)(0.9) = 0.72
P(E and R) = P(E)P(RIE) = (0.8)(0.1) = 0.08
P(E and R) = P(E)P(RIE) =(0.2)(0.3)=0.06
P(E andR) = P(E)P(RIE) =(0.2)(0.7)=0.14.
Since exactly one of the four joint events E and R, E and R, E and R, or E and R
must occur, the sum of the four probabilities above equals 1. These calculations can be
arranged in the following table. The numbers 0.8, 0.9, and 0.3 are known, the others
are calculated from them.
FeE) =0.8
peR) =0.78 (0.9)(0.8) = 0.72
peR) =0.22 (0.1)(0.8) = 0.08
FeE) = 0.2
(0.3)(0.2) = 0.06
(0.7)(0.2) = 0.14
The probability of successful brine recovery was not assumed explicitly, but can
be calculated from the assumptions. There are two mutually exclusive ways of having
successful brine recovery: either with a successful explosion or without one. Thus,
peR) = P(E)P(RjE) + P(E)P(RIE)
- (0.8)(0.9) + (0.2)(0.3) = 0.72 + 0.06 = 0.78.
Equivalently, peR) could be obtained by summing the probabilities of the two
upper joint events in the table:
peR) = P(E and R) + P(E and R) = 0.72+ 0.06 = 0.78.
~
29
If only FeE) and peR) had been known and if we had assumed that they were inde-
pendent, we would have calculated the probability of both tests succeeding to be
P(E and R) = P(E)P(R) = (0.8)(0.78) = 0.624.
However, from past experience we know that P(E and R) = 0.72. The multiplication
rule does not apply here because the probability of R depends on the outcome of E,
so the events E and R are not independent.
Some other questions are
1. What is the probability that the explosion and the brine tests are both successful
or both fail? This equals
PE and R) or. (E and R)) = P(E and R) + P(E and R)
- 0.72+ 0.14 = 0.86.
2. What is the probability that the explosion test was successful, given the con-
straint that only one of the two tests was successful? The probabilities of the two
events that have exactly one success areP(E and R) = 0.08 and P(E and R) =
0.06, so
P(only one success) = 0.08+ 0.06 = 0.14,
and
P(E and only one success) = P(E andR) = 0.08.
Thus
P(Elonly one success) = P(E and only one success)/P(only one success)
- 0.08/0.14 =0.571.
3. What is the probability that the explosion was successful, given that recovery
using brine was successful? This is
P(EIR) - P(E and R)/[P(E and R) + P(E and R)]
= 0.72/(0.72 + 0.06) = 0.923.
4. If one or more of the tests was successful, what is the probability that the other
one was successful? The answer is
P(E and R)/[P(E and R) + P(E and R) + P(E and R)]
= 0.72/ (0.72+ 0.08 + 0.06) = 0.837.
anal [email protected]
~
30
2.3
Random Variables
An important first step in pattern recognition is making measurements or extracting
features to use for classifying the patterns. For example, one might measure the
circumference of a fossil skull that one is trying to classify or count the number of
teeth. A random variable is the outcome of a random process which outputs a
'numeric value. The output of a random variable is called a random number. An
example of a random variable is the process of randomly choosing a sample from some
population and measuring one of its features. We will denote the names and values of
random variables by lower-case letters such as x, y, and z.
An important special class of random variables are the discrete random vari-
ables which can take on a finite number of possible values or a countably infinite
number of values such as 0, 1, 2, '" or 2, 4, 6, A discrete random variable is
described by its distribution function which lists for each outcome x the probability
P(x) of x. If Xl, X2, ..., Xn are all the possible outcomes, then
n
LP(Xi) = 1
i=l
(2.8)
because some outcome must occur.
Example 2.4 The distribution /unction for the number of heads from two flips of a
com.
The random variable k is defined to be the total number of heads that occur when a
fair coin is flipped two times. This random variable k can have only the three possible
values 0, 1, and 2, so it is discrete. Its distribution function can be obtained by noting
that there are four equally likely outcomes from the two flips: (T, T), (H, T), (T, H),
(H, H). One of these outcomes produces no heads, two produce one head, and one
produces two heads. The probabilities P(k) of the possible values of k are thus
"
2.3. RANDOM VARIABLES
31
The Binomial Distribution
A fixed length sequence of events or trials where each event has exactly two possible
outcomes can be modeled by a binomial distribution. One of the outcomes is
generally called success and the other is called failure. In the binomial distribution,
we assume that the probability of success is the same for each trial and we denote
this probability bye. The total number of successes k obtained in n trials is called a
binomial random variable. Its distribution function is given by
P(k) = (~) ek(l - et-k,
k =O,...,n (2.9)
where
(
n
)
n!
k = k!(n - k)!"
The symbol (~) is called the binomial coefficient.
To see how the binomial distribution is obtained, notice that the n trials will gener-
ate a sequence of successes and failures of length n. We want to find the probability of
obtaining a sequence that contains k successes and n - k failures. There may be many
different sequences that contain k successes and n - k failures. For any particular one
of these sequences, we can use the multiplication rule to find the probability of the
sequence, as all the trials are independent. From the multiplication rule, we see that
the probability of obtaining any particular sequencecontaining k successesand n - k
failures is
ek(l - et-k.
Furthermore since the number of different sequences that contain k successes and n - k
failures is the number of ways of choosing a set of k samples from a set of n, or (~),
the probability that a random sequence will contain k successes and n - k failures can
be found from the distribution function (2.9). The number of heads k in Example 2.4
is a binomial random variable because each flip has only two possible outcomes. Its
parameters are n = 2 and e = 0.5. In Example 2.5, k is a binomial random variable,
where n = 6 and e = 0.6. Its mean ILk is ne and its variance (see Section 2.5) a~ is
ne(l - e).
Example 2.5 The binomial distribution.
An anthropologist knows from records kept at a dig site that the probability that a
recovered fossil human skull is female is 0.6. The probability that out of six skulls,
exactly four will be female is given by substituting into the binomial distribution
function
k
P(k)
0
1/4
1
1/2
2
1/4
analoui@ius~
32
I
(
33
C')
0
C
iL
N
0
-
0
0
0
2 3 4 5 6
0 1 2 3 4 5 6 7 8 9 10
k
Figure 2.4: The binomial distribution function P(k) for n =6 and () = 0.6.
n
Figure 2.5: The Poisson distribution: pen) = e->.).n/n!, for). = 1.5. pen) > 0 for all
n, but the values of P( n) for n > 6 are too small to see in the graph.
P( 4) = (:)(0.6)\0.4)2 = 0.311.
The graph of the binomial distribution function P( k) when n =6 and () = 0.6 is shown
in Figure 2.4.
that two events cannot occur at exactly the same time. The Poisson distribution can
be used to model the number of arrivals of input or output requests at a telephone
exchange, of automobiles arriving at a tollbooth, of radioactive particles entering a
Geiger counter, and of many other processes occurring in scientific and engineering
applications.
Like the binomial distribution, the Poisson distribution assumes that only integer
values are possible, but unlike the binomial distribution, there is no upper bound on the
number of events that can occur in a given time interval for the Poisson distribution.
There is a finite but very small probability that any extremely large value of n will
occur. It can be shown that both the mean and variance of the Poisson distribution
equal the parameter). (see Section 2.5).
The Poisson Distribution
Although the discrete random variables we have discussed so far were defined to have
only a finite number of possible values, there are random variables that may have
a countably infinite number of outcomes. A set of outcomes is countably infinite if
the outcomes can be listed in an infinite sequence. For example, the set of nonnega-
tive integers is count ably infinite because they can be listed as the infinite sequence
{O,1,2,. . .}. The Poisson distribution (see Figure 2.5)
Continuous Random Variables
e->.).n
Pen) = -,- (2.10) n.
where n = 0,1,2,... is an example of such a discrete distribution. The value pen) is
interpreted as the probability that exactly n events will occur in a fixed time interval
and ). is the number of events that Occur in that length of time on the average, so ). is
the mean of the random variable n. We assume that the events occur randomly and
independently with a constant probability of occurring in any small time interval, and
Discrete random variables can take on only a finite number of values or at most a
countable infinity of values (such as the Poisson distribution). However, many random
variables observed in everyday life are not discrete but can take on any value in some
finite or infinite range. For example, the highest temperature on a given day, the length
of a part being inspected, the average intensity of light falling on a given region of an
object, or the waiting time for a radioactive count are all random variables that are
continuous rather than discrete.
A ~ontinuous random variable is described by a probability density func-
tion. This function is used to obtain the probability that the .value of a continuous
random variable is in a given interval. If the random variable is x and its density
0
C')
0
0
N
0
iL
0
-
0
0
0
0
[email protected]
34
function is p(x), then by definition
. rh
P(a ~ x ~ b) =10 p(x)dx.
(2.11)
An example is shown in Figure 2.6. The shaded area equals the probability that x lies
between a and b. Density functions are sometimes also called distributions.
While distributions of discrete random variablesmust sum to 1 (see(2.8)), densities
of continuous random variables must integrate to 1, because the probability is 1 that
x lies between -00 and 00:
L:p(x) dx = 1. (2.12)
In this text, we will always use uppercase P to represent probabilities such as P(k),
and lowercase p to represent densities such as p(x).
A random variable x can also be defined by its cumulative distribution function
C which is given by
C(a) = P(x ~ a).
Thus C(a) is the probability that x is less than or equal to a. Using (2.11),
C(a) = L~P(X)dX.
(2.13)
For any cumulativedistributionfunction,C(-00) = 0 and C(00) = 1. Substituting
(2.13) into (2.11) gives
P(a~x~b)= lhp(x)dx= L~P(X)dX- L~P(X)dx=C(b)-C(a).
(2.14)
Even if p(x) cannot be integrated analytically, P(a ~ x ~ b) can still be obtained
numerically using (2.14) if a table of C(x) is available.
Equation (2.11) shows that the probability of occurrence of a particular value of a
continuous random variable x, say a, is
P(x = a) = P(a ~ x ~ a) = 10p(x) dx = O.
Thus the probability of outcome a is 0, even though the outcome x = a is technically
possible. However, when using a continuous random variable x such as the highest
temperature on a given day, one is not interested in knowing the probability that x
attains some exact value with infinite precision such as P(x = 85.0035741...) but
rather that x falls in a finite range such as P(84.01 :=:; x < 84.02). Such probabilities
are computed by (2.11).
J
,
35
~
'i
v
)(
it'
11
~
0
a b a b
(a) (b) x a
Figure 2.6: (a) A density function. The shaded area is the probability that x lies
between a and b. (b) The cumulative distribution function for (a).
H the cumulative distribution function is known and can be differentiated every-
where, the probability density function can be obtained by
d
I
d
l
a
p(x) = da C(a) a=z = da -00 p(x) dx.
This formula is a consequence of the fundamental theorem of calculus which asserts
that differentiation and integration are inverse operations.
The cumulative distribution function is also defined for a sample of real data. To
produce Figure 2.7, 25 16-digit numbers in the range 0 to 1 were generated by a
computer using a random number generator. They were then sorted from smallest to
largest. For any value of x less than the smallest value, the cumulative distribution
of the sample is O. For x lying between the smallest and next smallest, C(x) = 1/25.
At each sample value, the cumulative distribution function increases by 1/25. Most of
the cumulative distribution functions for the continuous random variables discussed in
this chapter can be differentiated to yield densities, but the distributions for finite data
sets, such as the function in Figure 2.7, are discontinuous and cannot be differentiated
to yield density functions. In fact, after the data points have been obtained, they
are not described by a density but by a discrete distribution, to which each sample
contributes a probability of 1/25.
The Uniform Density
A uniform or rectangular random variable x is equally likely to attain any value
in an interval [a, b). (This interval is also called the range of x.) The density of x is
anal [email protected]
36
0.0 0.2 0.4 0.6 0.8 1.0
Sorted Data
Figure 2.7: Cumulative distribution function for a sample of 25 random numbers chosen
from a uniform density.
given by
{
0 ifx<a
p(x) = 1/(b - a) if a ~ x ~ b
0 if x > b.
The denominator b - a makes the total area under the density 1.
density in (2.15) gives the cumulative probability distribution
{
0 ifx<a
C(x) = (x - a)/(b - a) if a ~ x ~ b
1 if x > b.
The graphs of (2.15) and (2.16) for a = 0 and b = 1 are shown in Figure 2.8.
(2.15)
Integrating the
(2.16)
Example 2.6 The uniform density.
To obtain the probability that x is between 0.2 and 0.7, if x is uniformly distributed
in the interval [0,1], we can use (2.15):
1
0.7 1
P(0.2 ~ x ~ 0.7) = - dx = 0.5,
0.2 1- 0
or we can use (2.16):
P(0.2~ x ~ 0.7)= 0(0.7) - C(0.2) =0.7 - 0.2 = 0.5.
J
"'
\
37
Figure 2.8: The uniform distribution for a = 0 and b = 1. (a) Density function. (b)
Cumulative distribution function.
The Exponential Density
A continuous density that is important in areas such as the study of the reliability of
systems and the decay of radioactive particles is the exponential density which has
the form
pet) =(3e-{3t , t20
where t is the time at whichan event occurs, such asthe decayof a radioactive atom; it
could be any value from 0 to 00. The parameter (3is the rate constant, whichcontrols
the rate of decay. The cumulative distribution function
C(t) = l (3e-{3t dt = 1 - e-{3t
is the probability (at time 0) that the event will occur before time t. The graphs of
the density and cumulative distribution for the exponential distribution are shown in
Figure 2.9.
Example 2.7 The exponential density.
Suppose that the lifetime of a particular type of radioactive atom in years has an
exponential density with (3 = 0.01. The probability that the atom will decay within
50 years after year 0 is
{50
P{O~ t ~ 50} = 10 O.Ole-o.Oltdt = 1 - e-0.5 = 0.39.
C!
co
as
ci
'1ii
0 ID
'0
ci
c:
0 """
ci
LL
N
ci
0
ci
1 I
0
c:
N
0
:g
c:
LI? .a LI?
c:
"0;
,2
c:
Cl> :;
"0
C!
.c
:LL
'0
i
!
1: '0
C)
Cl>
'Qj
Il)
>
:I:
.,
ci as
"3
E
0
::>
U
ci
-1.0 0.0 0.5 1.0 1.5 2.0 -1.0 0.0 0.5 1.0 1.5 2.0
(a)
x
(b)
x
[email protected]
38
CHAPTER 2. PROBABILITY 2.3. RANDOM VARIABLES
39
----------
----
--
,-
,
q
IX)
ci
"
,
,
,
,
,
,
,
,
,
,
,
,
,
..
c:
I!
N <0
0' ci
0
~ V
N .
c: 0
C\I
ci
2
0
ci
3 4
x
-4
(b)
-2 0 2 4
t
Figure 2.9: The exponential distribution when /3 = 1. The density function p(t) =
/3e-{3t is the solid curve, and the cumulative distribution function C(t) =,1 - e-{3t is
dashed.
(a) z
Figure 2.10: (a) A normal density with mean J.tand standard deviation a. (b) The
standard normal density function p(z) = e-z2 /2/...tiir (solid curve), and its cumulative
distribution function (dashed curve).
In general, the probability that a given atom still remains at time t is e-{3t, so the.
probability that it decays before t is 1- e-{3t.
Instead of expressing the lifetime distribution of a radioactive element in terms of
the parameter /3, it is traditional to express it by its half-life '\0, which is defined as
the time for half of a given sample to decay, or
Section 2.5. For a normal distribution, J.tis located at the center of symmetry of the
normal density, and it can be shown that a is the distance from the mean to one of
the inflection points of the density.
The normal cumulative distribution function is defined by
P(O ~ t ~ '\0) = 1/2.
1
8 1 1
(
~
)
2
C(s) = "'-.e-2" dx,
-00 ay 211"
which cannot be integrated in closed form. However, its values are available in math-
ematical tables and on some scientific calculators, where it is often called the "cumu-
lative error function." Fortunately, a separate table is not required for every value of
J.t and a. The special case when J.t = 0 and a = 1 is called the standard normal
density:
Since
(AO
P(O ~ t ~ '\0) = 10 /3e-{3t dt = 1- e-{3AO= 1/2,
we obtain'\o = (In2)j/3 or /3 = (In2)j'\0. In this example, the half-life of the atomic
isotope to which the atom belongs is '\0 = (In2)jO.0l = 69.31 years.
p(z) = ~e-z2/2.
...tiir
Its graph is shown in Figure 2.1Ob. Any normally distributed random variable x can
be transformed to a random variable z which has the standard normal density by using
the substitution z = ~. This reduces P( a ~ x ~ b) to P (",;;a ~ Z ~ "';;b), which is
(2.18)
The Normal Density
A very important density is the normal density (sometimes called the Gaussian
density). Its equation is
1 1
(
'!'..::E-
)
2
p(x) = -e-2 " (2.17)
aV'f,i
as shown in Figure 2.10. The parameter J.tdenotes the mean of x, while a denotes the
standard deviation of x. Both the mean and the standard deviation are defined in
lb p(x) dx l
b 1 _1
(
'!'..::E-
)
2
l
(b-I-t}/O' 1 -z2
/
2
= -e 2 " dx = -e dz
a a.../2i (a-I-t}/O'...tiir
- CC:J.t)-C(a:/L).
q
IX)
ci
-....- <0
0' ci
0
8
v
a. ci
C\I
ci
0
ci
0
ana1oui(@,ius~
Figure 2.11: Areas under the standard normal curye from -CXJto z, where z 2: O.
41
Figure 2.12: Areas under the standard normal curve from -CXJto z, where z ::; O. For
z < -5, C(z) ~ (z2 - l)e-Z2/2/( -z3J21r) with an error of less than 1 percent.
z 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
0.0 0.5000 0.5040 0.5080 0.5120 0.5160 0.5199 0.5239 0.5279 0.5319 0.5359
0.1 0.5398 0.5438 0.5478 0.5517 0.5557 0.5596 0.5636 0.5675 0.5714 0.5753
0.2 0.5793 0.5832 0.5871 0.5910 0.5948 0.5987 0.6026 0.6064 0.6103 0.6141
0.3 0.6179 0.6217 0.6255 0.6293 0.6331 0.6368 0.6406 0.6443 0.6480 0.6517
0.4 -0.6554 0.6591 0.6628 0.6664 0.6700 0.6736 0.6772 0.6808 0.6844 0.6879
0.5 0.6915 0.6950 0.6985 0.7019 0.7054 0.7088 0.7123 0.7157 0.7190 0.7224
0.6 0.7257 0.7291 0.7324 0.7357 0.7389 0.7422 0.7454 0.7486 0.7517 0.7549
0.7 0.7580 0.7611 0.7642 0.7673 0.7703 0.7734 0.7764 0.7794 0.7823 0.7852
0.8 0.7881 0.7910 0.7939 0.7967 0.7995 0.8023 0.8051 0.8078 0.8106 0.8133
0.9 0.8159 0.8186 0.8212 0.8238 0.8264 0.8289 0.8315 0.8340 0.8365 0.8389
1.0 0.8413 0.8438 0.8461 0.8485 0.8508 0.8531 0.8554 0.8577 0.8599 0.8621
1.1 0.8643 0.8665 0.8686 0.8708 0.8729 0.8749 0.8770 0.8790 0.8810 0.8830
1.2 0.8849 0.8869 0.8888 0.8907 0.8925 0.8944 0.8962 0.8980 0.8997 0.9015
1.3 0.9032 0.9049 0.9066 0.9082 0.9099 0.9115 0.9131 0.9147 0.9162 0.9177
1.4 0.9192 0.9207 0.9222 0.9236 0.9251 0.9265 0.9279 0.9292 0.9306 0.9319
1.5 0.9332 0.9345 0.9357 0.9370 0.9382 0.9394 0.9406 0.9418 0.9429 0.9441
1.6 0.9452 0.9463 0.9474 0.9484 0.9495 0.9505 0.9515 0.9525 0.9535 0.9545
1.7 0.9554 0.9564 0.9573 0.9582 0.9591 0.9599 0.9608 0.9616 0.9625 0.0033
1.8 0.9641 0.9649 0.9656 0.9664 0.9671 0.9678 0.9686 0.9693 0.9699 0.9706
1.9 0.9713 0.9719 0.9726 0.9732 0.9738 0.9744 0.9750 0.9756 0.9761 0.9767
2.0 0.9772 0.9778 0.9783 0.9788 0.9793 0.9798 0.9803 0.9808 0.9812 0.9817
2.1 0.9821 0.9826 0.9830 0.9834 0.9838 0.9842 0.9846 0.9850 0.9854 0.9857
2.2 0.9861 0.9864 0.9868 0.9871 0.9875 0.9878 0.9881 0.9884 0.9887 0.9890
2.3 0.9893 0.9896 0.9898 0.9901 0.9904 0.9906 0.9909 0.9911 0.9913 0.9916
2.4 0.9918 0.9920 0.9922 0.9925 0.9927 0.9929 0.9931 0.9932 0.9934 0.9936
2.5 0.9938 0.9940 0.9941 0.9943 0.9945 0.9946 0.9948 0.9949 0.9951 0.9952
2.6 0.9953 0.9955 0.9956 0.9957 0.9959 0.9960 0.9961 0.9962 0.9963 0.9964
2.7 0.9965 0.9966 0.9967 0.9968 0.9969 0.9970 0.9971 0.9972 0.9973 0.9974
2.8 0.9974 0.9975 0.9976 0.9977 0.9977 0.9978 0.9979 0.9979 0.9980 0.9981
2.9 0.9981 0.9982 0.9982 0.9983 0.9984 0.9984 0.9985 0.9985 0.9986 0.9986
3.0 0.998650
3.5 0.9998674
C(z)=r ....Le-z2/2dz
.
4.00.99996833
- 00 v'21r
4.5 0.999996602
5.0 0.9999997133
5.5 0.99999998101
6.0 0.999999999013
'
6.5 0.9999999999588
-4 -3 -2 -1 0 1 2 3 '
z
7.0 0.99999999999872
z 0.00 -0.01 -0.02 -0.03 -0.04 -0.05 -0.06 -0.07 -0.08 -0.09
0.0 0.5000
0.4960 0.4920 0.4880 0.4840 0.4801 0.4761
0.4721 0.4681 0.4641
-0.1 0.4602
0.4562 0.4522 0.4483 0.4443 0.4404
0.4364 0.4325 0.4286 0.4247
-0.2 0.4207 0.4168 0.4129 0.4090 0.4052 0.4013 0.3974 0.3936 0.3897 0.3859
-0.3 0.3821 0.3783 0.3745 0.3707 0.3669 0.3632 0.3594 0.3557 0.3520 0.3483
-0.4 0.3446 0.3409 0.3372 0.3336 0.3300 0.3264 0.3228
0.3192 0.3156 0.3121
-0.5 0.3085 0.3050 0.3015 0.2981 0.2946 0.2912 0.2877 0.2843 0.2810 0.2776
-0.6 0.2743
0.2709 0.2676 0.2643 0.2611 0.2578 0.2546
0.2514 0.2483 0.2451
-0.7 0.2420 0.2389 0.2358 0.2327 0.2296 0.2266 0.2236
0.2206 0.2177 0.2148
-0.8
0.2119 0.2090 0.2061 0.2033 0.2005 0.1977 0.1949 0.1922 0.1894 0.1867
-0.9 0.1841 0.1814 0.1788 0.1762 0.1736 0.1711 0.1685 0.1660 0.1635 0.1611
-1.0 0.1587 0.1562 0.1539 0.1515 0.1492 0.1469 0.1446
0.1423 0.1401 0.1379
-1.1 0.1357
0.1335 0.1314 0.1292 0.1271 0.1251 0.1230
0.1210 0.1190 0.1170
-1.2 0.1151 0.1131 0.1112 0.1093 0.1075 0.1056 0.1038 0.1020 0.1003 0.0985
-1.3 0.0968 0.0951 0.0934 0.0918 0.0901 0.0885 0.0869 0.0853 0.0838 0.0823
-1.4 0.0808 0.0793 0.0778 0.0764 0.0749 0.0735 0.0721 0.0708 0.0694 0.0681
-1.5 0.0668 0.0655 0.0643 0.0630 0.0618 0.0606 0.0594 0.0582
0.0571 0.0559
-1.6 0.0548 0.0537 0.0526 0.0516 0.0505 0.0495 0.0485
0.0475 0.0465 0.0455
-1.7 0.0446 0.0436 0.0427 0.0418 0.0409 0.0401 0.0392 0.0384 0.0375 0.0367
-1.8 0.0359 0.0351 0.0344 0.0336 0.0329 0.0322 0.0314 0.0307 0.0301 0.0294
-1.9 0.0287 0.0281 0.0274 0.0268 0.0262 0.0256 0.0250 0.0244
0.0239 0.0233
-2.0 0.0228 0.0222 0.0217 0.0212 0.0207 0.0202 0.0197 0.0192 0.0188 0.0183
-2.1 0.0179 0.0174 0.0170 0.0166 0.0162 0.0158 0.0154 0.0150 0.0146 0.0143
-2.2 0.0139 0.0136 0.0132 0.0129 0.0125 0.0122 0.0119 0.0116 0.0113 0.0110
-2.3 0.0107 0.0104 0.0102 0.0099 0.0096 0.0094
0.0091 0.0089 0.0087 0.0084
-2.4 0.0082 0.0080 0.0078 0.0075 0.0073 0.0071 0.0069
0.0068 0.0066 0.0064
-2.5 0.0062 0.0060 0.0059 0.0057 0.0055 0.0054 0.0052 0.0051 0.0049 0.0048
-2.6 0.0047 0.0045 0.0044 0.0043 0.0041 0.0040 0.0039
0.0038 0.0037 0.0036
-2.7 0.0035 0.0034 0.0033 0.0032 0.0031 0.0030
0.0029 0.0028 0.0027 0.0026
-2.8 0.0026 0.0025 0.0024 0.0023 0.0023 0.0022
0.0021 0.0021 0.0020 0.0019
-2.9 0.0019 0.0018 0.0018 0.0017 0.0016
0.0016 0.0015 0.0015 0.0014 0.0014
-3.0 1.350x 10 -3
-3.5 2.326x 10-4
C('H: '-/'d' A
-4.0 3.167 x 10-5
-4.5 3.398 x 10-6
-5.0 2.867 x 10-7
-5.5 1.800 x 10-8
-6.0 9.866 x 10-10
'
-6.5 4.016 x 10-11
-4 -3 -2 -1 0 1 2 3 4
-7.0 1.280 x 10-12
Z
[email protected]
42
2.3. RANDOM VARIABLES 43
Tables of C(z) for z ~ 0 and z :$ 0 are given in Figures 2.11 and 2.12. (Actually only
one of these tables is needed; the normal density is symmetric and has unit area, so
C(z) = 1- C( -z) for z < 0.) 'The graph of C(z) is shown as a dashed curve in Figure
2.lOb.
The transformation that standardizes the normal density
x-p,
z=-
q
(2.19)
Moreover, the normal model assumes that the feature value can range from -00 to +00
which is usually unrealistic. If a feature is normally distributed, any linear function
of that feature will also be normally distributed, but nonlinear functions of it will not
be. For example, if particle "size" is considered to be important, the most pertinent
features could be the diameters, the surface areas, or the volumes of the particles, de-
pending on the application. If the surface areas of a collection of spheres are normally
distributed, then neither their diameters nor their volumes can be normally distributed
because they are both nonlinear functions of the area. The area is proportional to the
sqtiare of the diameter and the volume is proportional to its cube. Thus the normality
of the data would depend on which of these measures we used, and we should not
expect most features to be normal.
As this example suggests, a feature that is not normally distributed can sometimes
be made approximately normal 'by a suitable nonlinear transformation. However, one
must be careful not to overfit the data by using a transformation that is too compli-
cated. This transformation may cause the next transformed data set to depart even
farther from a normal distribution. In addition, making a multivariate data set normal
in each feature still does not guarantee that the set of features taken together will be
multivariate normal.
In some cases where a normal distribution does not fit well, densities that have
thicker tails may be useful. A density p(x) is said to have thicker tails than a
density q(x) if p(x) > q(x) whenever x approaches 00 or -00. An extreme example of
such a density is the Cauchy density given by
is known as the standardizing transformation.
Example 2.8 The normal density.
Consider an electrical circuit in which the voltage is normally distributed with mean
120 and standard deviation 3. What is the probability that the next reading will be
between 119and 121volts?
The z-transformation is
z =(x - 120)/3,
so we obtain from Figure 2.11 or 2.12
= C C21 ; 120) - C (119; 120)
= C(0.333) - C( -0.333)
= 0.631 - (1 - 0.631) = 0.631- 0.369= 0.262.
P(119 :$x :$ 121)
The value 0.631 was obtained by linear interpolation between C(0.33) and C(0.34) in
Figure 2.11.
1
2 ' p(x) =
(
X a
) 7rb 1+ ~
(2.20)
This density has been used to model the scattering of electrons as they pass through
thin metal foil. Another example is the double exponential density given by
1
p(x) = 2be-I(x-a)/bl.
Choosing a Probability Distribution
In some situations, the nature of the data source determines what type of density
function is appropriate to the problem. In other cases, theory offers no guide and simply
looking at a histogram of the data may suggest a model. Any assumed distribution
can be tested using the chi-squared test or the Kolmogorov-Smirnov test to see how
likely it would be to produce data that deviate from their expected values as much as
do the data actually obtained. Further discussion of these topics is available in many
statistics texts.
Although by far the most frequently used continuous density is the normal density,
there is usually no particular reason to believe that a given feature will be even ap-
proximately normally distributed. The real world never fits any simple model exactly.
The normal, Cauchy, and double exponential distributions are compared in Figure
2.13a, where they have been normalized to contain the same area (0.68) between -1
and 1. These densities are discussed in the problems at the end of this chapter. The
corresponding cumulative distributions are compared in Figure 2.13b.
A popular graphical test for determining whether a data set is approximately nor-
mally distributed is based on a normal plot. This plot is formed by plotting the
sorted sample values versus the expected z-values, that are the values of z which
divide the area under the standard normal curve into n + 1 equal areas, where n is the
number of samples. If there are only five samples, we might expect them to lie near
the five boundaries between six equal areas on the average, if they actually are drawn
from a normal distribution. (See Section 2.6 for further discussion,)
[email protected]
44
C!
'"
!";
~
j'\,
, ,
</ ;!\.~
. / ..' '. .
-4 -2 0 2 4
(a)
x
-2 0 2 4
x
Figure 2.13: A comparison of the normal (solid), double exponential (dashed), and
Cauchy (dotted) densities. (a) Densities and (b) cumulative distribution functions.
For example, to construct the normal plot for the samples 3, 1, 2, 9, and 5, sort
the samples to get 1, 2, 3, 5, and 9 and plot them versus the points -0.967, -0.430, .
0.000, 0.430, and 0.967 which divide the standard normal curve in Figure 2.14a into
5+ 1 equal areas of 1/6 each. For example, C( -0.967) = 1/6 and C( -0.430) = 2/6, as
found in the tables of Figures 2.11 and 2.12. This produces the normal plot in Figure
2.14b. If a large data set is normally distributed, we can expect that 1/6 of the data
will havez < -0.967, 2/6 will have z <-0.430, 3/6 will have z < 0, 4/6 will have
z < 0.430, and 5/6 will have z < 0.967. For a small data set these z values will only
be approximate.
If the data are approximately normally distributed, the points on the normal plot
will be close to a straight line as in Figure 2.15a. The slope and intercept of the line
will depend on the mean and standard deviation of the distribution of the data. The
points in the normal plot in Figures 2.15b and 2.15c are not well fit by a line, but the
data in Figure 2.15a, which came from a normal density, are well fit by a line. If the
samples are drawn from a Cauchy distribution, which has thicker tails than a normal
distribution, the normal plot shown in Figure 2.15b turns down at the left end and up
at the right end. The opposite is true in Figure 2.15c, where the samples are taken
from a uniform density, which has thinner tails than a normal density.
Note that the normal plot ca/?-be used to compare data with any mean and variance
to a normal distribution, even though the mean and variance of the standard normal
distribution are 0 and 1 respectively, because the ability to fit the points in the normal
plot with a straight line is independent of the scale of the vertical axis. Although the
vertical axis of the normal plot will be translated or stretched, depending upon the
~
(
\ 2.4. JOINT DISTRIBUTIONS AND DENSITIES 45
10
ci
""
ci
",,/"0
"" 0
,,/,/0"
0
CO)
ci
0
C\I
ci
~
ci
0
0
ci
-3 -2 1.0 -1 0 0.5 2 3 -1.0
(b)
-0.5 0.0
(a)
Expected z-values
Figure 2.14: (a) A standard normal density divided into six equal areas. (b) A normal
plot of five numbers and a linear approximation to them.
mean and variance of the samples, the normal plot will approximate a straight line if
the data are close to normal. If s represents the ranked sample values, the mean value
of s will be at z = 0 on the fitted line. In Figure 2.14b, a line has been fit to the
data (by eye). At z = 0, s is about 4, so if the data really did come from a normal
distribution, it probably has a mean near 4.
The normal plot can be generalized to compare the sample data to any distribution,
not necessarily normal. This is done by choosing the points on the horizontal axis
(called quantiles) which divide the density function of that particular distribution
into n + 1 equal areas, and plotting the ranked values of the samples on the vertical
axis as before. Such a plot is called a quantile-quantile plot of which the normal
plot is a special case. Other tests of the goodness of fit of a theoretical distribution to
a data set are described in statistics texts.
2.4 Joint Distributions and Densities
Although it is useful to model the variation of a single feature within a class by a
random variable, it is frequently necessary to model the simultaneous variation of two
or more features. The joint random variable (x,y) signifies that, simultaneously, the
first feature has the value x and the second feature has the value y. If the random
variables x and y are discrete, the joint distribution function of the joint random
C!
IX)
ci
co
ci
'X
U
""
ci
C\I
ci
0
ci
-4
(b)
IX)
ci
co
ci
'X
c:
""
ci
C\I
ci
0
ci
0
IX)
In
CD
a.
E
co
os
m
'C
""
0
m
C\I
[email protected]
46
q
-~
.,"'--"--"'...
..
. .
-2
(b)
-1 0 2
Expected z-values
.
. n
2
.."
.
..
.'
..'
.'
.
-1 0
Figure 2.15: A normal plot of 30 numbers drawn from three distributions. (a) Normal,
(b) Cauchy, (c) uniform.
Expected z-values
\
2.4) JOINTDISTRIBUTIONSANDDENSITIES
47
variable (x, y) is the probability P(x, y) that both x and y occur. Thus, the joint
distribution gives the probability of every possible combination of outcomes of the
random variables that make up the joint random variable.
Example 2.9 The joint distribution function of the outcomes of flipping biased coins.
A biased coin A has P(head) = 0.6 and coin B has P(head) =0.3. Suppose that we
flip coin A, and if the outcome is head, we flip coin B, but if the result of the first flip
is tail, we flip coin A again. Since the probability of head on the second flip depends on
the outcome of the first flip, the two events are not independent. Let x be the outcome
of the first flip and let y be the outcome of the second flip. The joint distribution
P(x,y) is
A continuous one-dimensional random variable can take on any value in an interval,
but a continuous two-dimensional random variable can take on any value in a two-
dimensional region. As in the one-dimensional case, we can use a probability density
function to define a continuous two-dimensional random variable. If R is a region in
the xy-plane, then
I
Px,y) is in R) = JkP(x,y)dXdY,
where the integral is taken over the region R. This integral represents a volume in
xyp-space.
Example 2.10 Computing the probability of a joint event using a two-dimensional
probability density function.
Suppose that (x,y) has the probability density function
p(x,y) = { ~(l-X-Y)
if 0 ~ x, 0 ::5y, and x + y ::51
otherwise
N
.
..
.'
..
=
I
.
..
Q.
.
E
0 .'
as
....'
Cl)
..
..
';"
.'
a!
..'
.
-2 -1 0 1 2
(a)
Expected z-values
0
N
III
0
J!
Co
I
0
co
III
c:i
J!
Co
<D
c:i
Cl)
"i
.
c:i
N
c:i
0
d
-2
(c)
(x,y) P(x, y)
(head, head) (0.6)(0.3) = 0.18
(head, tail) (0.6)(0.7) = 0.42
(tail, head) (0.4)(0.6) = 0.24
(tail, tail) (0.4)(0.4) = 0.16
[email protected]
- -~
p(x,y)
y
x
Figure 2.16: The two-dimensional density from Example 2.10.
as shown in Figure 2.16. This qualifies as a density function since it is a pyramid
with base area A = 1/2, height h = 6, and volume V = (1/3)Ah = 1. To find the
probability of the joint event J, 0.2 :$ x :$ 0.3 and 0.4 :$ y :$ 0.5, the density must be
integrated over this region:
1
0.3
[1
.5
]
P(J) = 6(I-x-y)dy dx
0.2 0.4
1
0.3
- 6[(0.5 - 0.4) - (0.5x - O.4x) - (0.52 - 0.42)/2] dx = 0.018.
0.2
The Bivariate Normal Density
Probably the most used joint two-dimensionaldensity is the bivariate normal den-
sity given by
-~
[(
x_p,z)2 - 2pzy(x-P,z)(y-p,y) + (y_p,y)2
] e 2(1-pzy) Uz UzUy Uy
p(x,y) = .
27rUxUYV(1- P~y)
The three-dimensional graph for a particular example of this density is shown in
Figure 2.17a. The contours of constant probability density p(x,y) = c always form
ellipses in the xy-plane as shown in Figure 2.17b. .
The parameters /-Lx and Ux are called the mean and standard deviation of the
random variable x, and /-Ly and Uy are called the mean and standard deviation of
(2.21 )
)
2.5.MOMENTS OF RANDOM VARIABLES
49
0
~
ID
0 5 10
(a)
x
Figure 2.17: A bivariate normal density where /-Lx = 0,/-Ly = 0, Ux = 3, Uy = 2, and
Pxy = 1/2. (a) A three-dimensional plot. (b) Some contours of constant probability
density.
y. The parameters /-Lxand /-Lydenote the center of tile density in the xy-plane, and
Ux and Uy determine the spread in each direction. 'file fifth parameter Pxy is called
the coefficient of linear correlation or simply the correlation between x and y.
It influences the orientation of the density with respect to the coordinate axes. These
parameters are discussed further in the next section afld classification procedures that
use the bivariate normal density are discussed in CiH~pter 3.
2.5 Moments of Random VariableS
If x is a discrete random variable, the expected value of x, denoted by E(x), is
defined by
I
n
E(x) = LXiP(Xi) = /lx'
i=l
where XI, . . . , Xn are all the possible outcomes of x v,lld P(Xi) is the probability that
x = Xi. The expected value of X is thus a weighted average of its possible values, each
value weighted by the probability of its occurrence. If x has a count ably infinite range,
then the summation is from i = 1 to 00. The expected value E (x) is also called the
mean of x and is also denoted by /-Lor /-Lx.
E(x) is also called the first moment of the distribution. The kth moment is
defined as
n
E(xk) = I>7P(Xi)'
i=1
>- 0
"?
0
-10 -5
(b)
[email protected]
Thus the second and third moments of a distribution P(x) are the expected values
of x2 and x3 for that distribution.. The kth central moment, also called the kth
moment about the mean, is defined as
n
E[(x - tLx)k] = ~)Xi - tLx)k P(Xi)'
i=l
The average value of x in a randomly sampled data set Xl, . . . , Xn is defined as
- 1 n
X = ;;: ~::>i ,
i=l
which is discussed more fully in Section 2.6. This could also be called the mean of the
data set. The average value if of the data would precisely equal the mean tL of the
population from which the data were sampled only if the data set were exactly rep-
resentative of the probability distribution of the population. In general, the expected
difference between if and tL approaches 0 as the sample size increases.
If X is a continuous random variable, the expected value of x is
E (x) = 1.: xp(x) dx = tLx
where p(x) is the density of x. The expected value E(x) is also denoted by tLx.
If f is a function of x, then the expected value of f(x) is defined as
n
E (f(x)) = L f(xi)P(xd
k=l
when x is discrete. This is the sum of all the possible values of f(xi), each weighted
by its probability of occurrence. When x is continuous, each value of f(x) is weighted
by the density p(x) and the product is integrated
E (f(x)) = 1.: f(x)p(x) dx.
If the distribution function or density of a random variable is symmetric about
some point c, then that point is the mean of the random variable (Problem 2.51).
Example 2.11 The expected value for the number of heads when tossing a coin twice.
When a fair coin is tossed two times, the number of heads obtained x has the probability
distribution
"
\
2.5. MOMENTS OF RANDOM VARIABLES
51
The expected value of x is then
2
E(x) = L xP(x) = 0(1/4) + 1(1/2) + 2(1/4) = 1.
x=O
On the average, one head will be obtained in two tosses.
Example 2.12 The expectedvalue of a uniform mndom variable.
If x has a uniform density on the interval [a,b],its density function is p(x) =l/(b- a)
on the interval [a,b],and 0 outside of that interval. Thus its expected value is
I
b
00 b 1 1 x2 a+b
E(x) =
1
xp(x)dx=r x-dx= - - =-,
-00 la b- a b - a 2 2
a
which is, as expected, at the center of its range.
Example 2.13 The expected value of an exponential mndom variable.
The expected value of an exponential random variable t is expressed by
I
E(t) =1000t{3e-/3tdt.
Using integration by parts with u = t and dv = {3e-/3tdt, we find that
t{3e-/3t
l
oo
10
00 (3e-{3t
L
oo e-{3t
l
oo 1
E(t) = - - -dt = 0+ e-/3tdt = - =-.
-{3 0 -{3 0 -{3 {3
0 0
Thus if {3=0.01, the expected or average lifetime is 100. However, the half-life is 69.31
(see Example 2.7).
The second moment of a random variable x is defined to be E(x2), but the second
central moment or variance of a random variable x is defined by
(1~ =E[(x - E(x))2] = E[(x - tLx)2J.
(2.22)
x
P(x)
0
1/4
1
1/2
2
1/4
[email protected]
This is the expected squared deviation of x from its mean and describes the spread
of the values of the random variable x about its mean. The variance is a convenient
measure of the spread of a distribution, but it has the disadvantage that it is measured
in squared units. In order to have a measure of the spread that is measured in the
units of the original random variable, one must use the standard deviation ax which
is defined as the square root of the variance a;.
. An important property of the expectation operator E(.) is that it is a linear op-
erator which means that if x and y are random variables, and z is a new random
variable which is the weighted sum ax + by, then
E(z) = E(ax + by) = aE(x) + bE(y). (2.23)
To see this in the continuous case, let p(x, y) be the joint probability distribution
of (x, y). Also let Pl(X) = f:'oop(x, y) dy and P2(y) = f:'oop(x, y) dx be the marginal
densities of x and y respectively. (The marginal density of x is the density of the
component x in (x, y) regardless of the outcome of y.) Then
E(z) = E(ax + by) = l:l:(ax + by)p(x,y)dxdy
- aI: x [l:p(x, y)dY]dx+ bI: y [I: P(x,y)dx]dy
Nowsince
l:p(x,y)dx=P2(y) and l:p(x,Y)dY=P~(x),
E(ax + by) = a I: Xpl(X) dx + bI: YP2(y) dy = aE(x) + bE(y).
The proof in the discrete case is done by replacing integrals and densities by sum-
mations and distribution functions:
E(z) = E(ax+by)=:E~)ax+by)P(x,y)
x y
- a:E:ExP(x,y) + b:E:EyP(x,y)
x y x y
- a:Ex:EP(x,y)+b:Ey:EP(x,y)
x y y x
- a :EXP1(X) + b:EyP2(y)
x y
- aE(x) + bE(y).
Because E(.) is a linear operator, we can obtain an alternative expression of the
variance (compare with (2.22)):
a; = E[(x - E(x)?] = E[x2 - 2xE(x) + E(x)2]
)5. MOMENTS OF RANDOM VARIABLES
53
= E(x2) - 2E[xE(x)] + E[E(x)2].
But E(x) = /-Lx is a constant, so
E[xE(x)] = E(x/-Lx) = /-LxE(x) = /-L;
and
a; = E(x2) - 21L;+ /-L;= E(x2) - /-L;.
(2.24)
This form is more convenient than (2.22) for computing the variance of a data set.
Example 2.14 The variance of a uniform random variable with range a to b.
Between a and b the density is 1/(b - a) so the second moment is
E(x2) = lb x2(1/(b - a)) dx = :;b-=- :) = b2+ a; + a2.
Using (2.24),
a; = E(x2)- /-L2= b2 + ab+a2 -
(
b+ a
)
2 = b2 - 2ab+a2 = (b- a)2
x 3 2 12 12 .
Example 2.15 Find the variance of the number of heads obtained from two coin flips.
By using the distribution function from Example 2.11, and the value /-Lx = 1, weobtain
I
2
a; = E(x - /-Lx)2=:E(x- 1?P(x)
x=o
- (-1)\1/4) + 02(1/2) + 12(1/4) = 1/2.
This result could also be obtained using (2.24).
Example 2.16 The mean of a Poisson random variable.
[email protected]
54
The mean of a Poisson random variable is
00 00
/J = E(x) ~ L xP(x) = L xP(x)
",=0 ",=1
00 e->' >"" ->. 00 x>"" ->. 00 >.",-1 ->. 00 >""
= L x , = eL, = >.e L ( - )' = >.e L.
",=1 X. ",=1 X. ",=1 X 1. "'=0 x.
Since the Maclaurin series expansion of e>'is
00 >""
e>'=L.'
x.
",=0
00 >""
/J = >.e->'L . = >.e->'e>' = >..
",=0X.
Example 2.17 The mean and variance of the standard normal distribution.
In this example, we show that, for a standard normal random variable x with density
p(x) = vk.e-",2/2,
E(x) = 0 and E(x2) = 1. We calculate
L: xp(x) dx
1
00 1 2
x-e-'" /2 dx = O.
- 00 v'27r
E(x) =
(2.25)
Since the integrand f(x) = xe-",2/2 / J21r satisfies f( -x) = - f(x), its integral over the
entire range is zero.
To show that E(x2) = 1, we use integration by parts:
f u dv = uv - Jv du,
where u = x, dv = x(1/J21r)e-",2/2dx, du = dx, v = -(1/v'27r)e-",2/2, and
1
00 1 2
E(x2) = x2-e-'" /2dx
- 00 v'27r
1 ",2/
2
1
00
( 1
00 1 ",2
/
2
)
-X-e- - - -e- dx
J21r - 00 - 00 v'27r
2.5. MOMENTS OF RANDOM VARIABLES
55
L'Hopital's rule shows that the first term in the expression for E(x2) is zero at each
limit. The second term is the integral of the normal density and is equal to 1 since the
area under a probability density function must be 1 (see Problem 2.29).
Example 2.18 The mean and variance of a normal distribution.
To justify the definitions of the terms mean and variance for the parameters /J and 0-2
for the normal distribution in Section 2.3, we must verify that the expected value for a
normal distribution is actually the parameter /J and that the variance is the parameter
0-2. By making the substitution z = (x - /J)/o- or x = o-z + /J, and dz = dx/o-, we can
use the results from Example 2.17 to obtain
1
00 1 2 2
E(x) = x-e-("'-/L) /(20' ) dx
-00 o-J21r
1
00 1 2
= (o-Z+/J) fi>-.e-Z /2dz
-00 v 271"
= o-E(z)+/JE(1)=o-.O+/J.1=/J.
Also,
E(X2) =
1
00 x2~e-("'-/L)2/(2O'2) dx
-00 o-v'27r
1
00 1 2
= (o-z+ /J)2-e-z /2dz
-00 v'27r
= o-2E(z2) + 2o-/JE(z) + /J2E(1)
= 0-2.1 + 2o-/J.0+ /J2. 1 = 0-2 + /J2.
I
The variance of x is E(x2) - E(x? = (0-2+ /J2) - /J2 = 0-2,so we were justified in calling
/J and 0-2the mean and variance of the normal density.
Moments of Bivariate Distributions
Moments can also be computed for bivariate distributions. If (x, y) is a joint random
variable, then /J", and 0-",are the mean and standard deviation of x, and /Jy and o-yare
the mean and standard deviation of y. The covariance of x and y is defined as
[email protected]
)
:i.5. MOMENTS OF RANDOM VARIABLES
57
If x and y for a sample are both greater than or both less than their means, the product
(x - J1.x) (y - J1.y)will be positive. Thus the covariance indicates how much x and y tend
to vary together. The value of the covariance depends on how much each variable tends
to deviate from its mean, as represented by the standard deviations of x and y, and
also depends on the degree of association between x and y. To eliminate the scaling
effect of ax and ay, we can divide the covariance by the product of the two standard
de~iations to yield the correlation between i and y, which is usually denoted by the
Greek letter p (rho):
the covariance of Xi and Xj. (Note that the covariance aii of a random variable with
itself is actually the variance a~.)
The components of a random vector are often written in the column vector form:
X~(n
Pxy= ~ = E
[(
X - J1.x
)
(
y - J1.y
)]
.
axay ax ay
We define the mean vector as
~ ~ E(x) ~ [ )
(2.27)
It is also called the coefficient of linear correlation, or correlation coefficient. While
the covariance of x and y, axy, has units equal to the product of the units of x and
of y, the correlation is a dimensionless quantity. The correlation has the convenient
property that for any random variables x and y
The covariance matrix is symmetric (aij = aji) and is usually denoted by a capital
sigma:
-1 ~ Pxy ~ 1.
(
all a21 ... adl
) (
a~ ali .., aId
)
al2 a22 ... ad2 al2 a2 ..' a2d
~= . . . . = . . . . .
: : '.: ::'.:
aId a2d ... add aId a2d .., a~
The d-dimensional normal density can be written using matrix notation:
(x - P. )T'-l (x - p.)
2
(2.28)
If Pxy is positive, x and y are said to be positively correlated. If P = 1, they are
perfectly correlated, and y is a linear function of x. If Pxy is negative, x and y are
negatively correlated or "anticorrelated." If the correlation is 0, x and y are said to
be uncorrelated. The correlation is symmetric in the sense that Pxy= Pyx.
If a random variable z is a weighted sum of the random variables x and y, the
variance of z depends not only on the variances of x and y but also on the covariance
of x and y. Let z = ax + by. The variance a~ of z is
E[(z - J1.z)2] = E[((ax + by - (aJ1.x+ bJ1.y)?J
= a2E(x2 - 2xJ1.x+ J1.~)+ 2abE(xy - xJ1.y- yJ1.x+ J1.xJ1.y)
+b2E(y2 - 2yJ1.y+ J1.~)
22
b b
22
- a ax + 2a axy + ay,
1
p(x) = e
..jdet E(27r)d
where det denotes the determinant. The exponent in (2.29) can be expressed alge-
braically as
(2.29)
1
"2l::(Xi ~ J1.i)Sij(Xj- J1.j)
'3
where Sij is the ijth component of E-I, the inverse of the covariance matrix E.
When d = 2, I
x=(:),
(
J1.x
)
,
p. = J1.y
using (2.23). As a special case, when axy = 0,
a; = a2a; + b2a~.
(2.26)
The Multivariate Normal Density and
~=
(
a; axJj
)
=
(
a; PXya~ay
)
,
axy ay pxyaxay ay
so we obtain (2.21). Examples of classification procedures which use the multivariate
normal density will be discussed in Chapter 3.
The normal density can be generalized to any number of dimensions. When there are
more than two dimensions, using matrix notation to simplify the expressions is helpful.
The joint random variable (Xl"", Xd) is called a d-dimensional random vector and
will be denoted by x. Let J1.= (J1.!,. . . , J1.d)be the vector of means of x, and let aij be
[email protected]
2.6
Estimation of Parameters from Samples
To use the densities discussed in this chapter for the parametric classification pro-
cedures described in Chapter 3, some method must be available for estimating the
parameters of the density from the data available in the training sample. We discuss
three kinds of estimates for these parameters: method of moments estimates, maximum
likelihood estimates, and unbiased estimates. Other desirable properties of estimates
are discussed in many statistics texts.
The Method of Moments
If some particular family of densities (or probability distributions) is assumed to de-
scribe a data set, one way to choose the r adjustable parameters in the density is such
that the first r central moments of the density will equal the first r central moments
of the data. To estimate parameters by the method of moments, n independent
samples or patterns Xl, X2, ..., Xn are collected from the random variable X, which
may be continuous or discrete. If we randomly choose one of these samples in the
data set, we can consider its value to be a new discrete random variable x' called an
empirical random variable, which takes on one of the values Xl, X2, ..., Xn, each
with probability 1/n. (These values need not necessarily be distinct.) The sample
mean is defined by
n 1 n
X = E(x') = L: XiP(Xi) = - L: Xi,
i=l n i=l
which is the average of all the samples. The sample variance is the variance of x'
and is denoted in this section by u;,. It is given by
n 1 n
u;, = L:(Xi- E(x')? P(Xi) = - L:(Xi - x)2.
i=l n i=l
If the mean and variance are the density function parameters to be estimated, the
sample mean and the sample variance are simply used as the estimate for the popula-
tion. If other parameters are required, they are found from the moments as in Example
2.19. If the distribution contains more than two parameters to be estimated, higher
moments such as E(x' - x)3 and E(x' - x)4 are also used.
The method of moments can also be used to estimate the covariance uxy of a
bivariate distribution. We compute the covariance of the sample
1 n
Ux'y' = E[(x' - E(x'))(y' - E(y'))] = - L:(Xi - X)(Yi - y),
n i=l
which is then used as an estimate of the covariance of the population from which the
sample was chosen.
) .
2.6. ESTIMATION OF PARAMETERS FROM SAMPLES
59
Example 2.19 The method of moments estimate of the range of a uniform density.
Suppose we know that the sample values X = 2, 3, 5, 6, 8, 9, 11, and 12 come from a
uniform distribution, but we do not know the values of the parameters a and b that
determine the range. To estimate these parameters using the method of moments, we
compute
1 n
X = 7 and u;, = - L:(Xi - x? = 11.5.
n i=l
A uniform density on the interval [a, b] has mean I-L= (a + b)/2 and variance u2 =
(b - a?112 (see Examples 2.12 and 2.14). Thus we estimate a and b by a and b. (In
general, a "hat" over a parameter denotes an estimate of that parameter.)
a+b=7 and
2
, '
)
2
(b ~ = 11.5.
12
Solving for a and b,we obtain
a = 1.13 and b = 12.87.
Thus a uniform distribution with a range of 1.13 to 12.87 has the same mean and
variance as our data, so the method of moments estimate of the range of the distribution
is (1.13,12.87).
A disadvantage of using the method of moments to estimate the range of a uniform
distribution is that some samples may lie outside the estimated range [a, b] (see Problem
2.56), in which case the estimate is logically inconsistent with the data.
I
Maximum Likelihood Estimates
Although the method of moments is easy to understand and works well in many cases,
there are other methods for choosing parameters that are also appealing. The maxi-
mum likelihood estimate of a parameter is that value which, when substituted into
the probability distribution (or density if the random variable is continuous), produces
that distribution for which the probability (or joint density) of obtaining the entire
observed set of samples is maximized. To compute the maximum likelihood estimate,
we choose the parameter (or set of parameters) that maximizes the value of the joint
distribution function or multivariate density function for the entire data set when it is
evaluated at the set of sample points Xl,..., Xn.
[email protected]
2~) ESTIMATION OF PARAMETERS FROM SAMPLES
61
Example 2.20 The maximu!7t likelihood estimate of the probability of getting a head
when [tipping a biased coin.
Consider the discrete random variable x for which P(H) = (J,and P(T) = 1 - (J. To
e&timate the value of (J,obtain n independent outcomes Xl, X2, .. ., xn from x. Because
the outcomes are independent, the probability of this set of data is
Example 2.22 Finding the maximum likelihood estimate for /-in a normal distribu-
tion.
The density of a normal random variable with mean /-and variance a2 is
P(XI,. . . , xn) = P(XdP(X2) . . . P(xn) = (Jk(l - (Jt-k (2.30)
p(X) = ~ -~
(
X - /-
)
2
a-./2-ff ea.
where (Jis the probability of obtaining a head, k is the number of heads obtained, and
n is the number of samples collected. The maximum likelihoed estimate for (Jis the
value that maximizes (2.30).
To simplify the calculations, we note that the value of a parameter that maximizes
an expression is the same as the value that maximizes the logarithm of the expression.
The logarithm of (2.30) is
, InP(xI,...,xn)=kln(J+(n-k)ln(l-(J).
Taking the derivative with respect to (Jand setting the result to zero gives
k n-k
---=0
(J 1-(J
If all the random samples XI, X2,"', Xn are independent,
n
p(X1,...,Xn) =p(X1)...P(Xn) = TIp(x;),
;=1
where I1represents taking the product of the terms that follow it. Therefore
n 1 - (Xi - /-?
p(XI,...,xn) =p(X1)...P(Xn) = TI liCe 2a2
i=l av 27r
or (J= kin. We note that this estimate is the same as the sample mean which is also
the method of moments estimate.
1 n
1 - 2a2~)Xi - /-)2
- e ,=1
- an(27r )n/2 .
The derivative of the logarithm of (2.31) with respect to /-is
(2.31)
Example 2.21 The maximum likelihood estimate of the range of a uniform distribu-
tion.
1 n 1
[
n
]
~ ~)Xi - /-). 2 = 2" 2:>i- n/- .
2a i=l a i=l
Setting this derivative equal to 0 and solving for /-results in the maximum likelihood
estimate for /-:
1 n
P,=-LXi=X,
n i=l
This estimate for /-is the same as the method of moments estimate.
The maximum likelihood estimate a for parameter a of a uniform distribution with
range [a,b] is the smallest sample value found among the independent samples Xl, X2,
. . . , Xn. The maximum likelihood estimate bfor b is the largest sample value. Any
narrower range for [a, b] would be impossible, given the data, and any larger range
would reduce the joint density function p(XI,..., xn) = [l/(b - a)]n that produced the
samples. For example, if the samples are Xl = 9.2, X2 = 2.1, X3 = 3.7, X4 = 11.1, and
Xs = 9.2, the maximum likelihoodestimates are a = 2.1 and b= 11.1.
Although this estimate is consistent with the data, the true range [a, b] from which
the data came is probably larger than the range [a,b] of the actual data obtained, since
it is extremely unlikely that the limiting values a and b actually occurred in the data
set. The actual range is probably larger than the estimated range. Because of this,
the maximum likelihood estimate is said to be a biased estimate of the range of a
uniform density.
Example 2.23 The maximum likelihood estimate for a2 in a normal distribution.
To simplify the notation, we replace a2 by v in (2.31) to obtain
1 n
-- 'L(Xi - IL?
( )
- 1 2vi=1
P Xl,'" ,Xn - vn/2(27r)n/2e .
(2.32)
[email protected]
2.6. ESTIMATION OF PARAMETERS FROM SAMPLES
63
d n 1~ 2 -
d
Inp(x1,...,xn)=--+~L..,.(xi-J.t).
V 2v 2v i=1
th~ probability of obtaining x would be less, say P(x) = 1/(x + 1), which is smaller
'than l/x. Thus the distribution that has the highest probability of producing x is the
distribution for which N = x. The followingtable shows the probability of drawing a
card x for various assumed values of N.
The maximum likelihood estimate of v is the value that maximizes (2.32), so we dif-
ferentiate its logarithm with respect to v to obtain
n
a2=v=~L(xi-J.t?,
n i=1
Setting this to zero and solving for v produces
so the maximum likelihood estimate of the variance equals the variance of the sample
if J.tis known. If J.tand u2 are both unknown, setting the partial derivatives of (2.31)
with respect to J.tand u2 to 0 simultaneously yields the same solutions. This is the
same estimate for u2 as was obtained using the method of moments.
The probability P(x) is maximized when x = N, so the maximum likelihood estimate
of N is IV = x. (If several cards had been drawn, their largest value would be the
maximum likelihood estimate of N.)
The maximum likelihood estimate x may not seem to be a good guess for N,
because N can never be less than. x and is probably higher because it seems unlikely
that we drew the highest card in the deck. This means that this maximum likelihood
estimate is biased in the direction of choosing a value for IV that is less than the true
value. It is plausible that there are as many cards higher than the card x we drew
as there are cards that are lower than x. This reasoning leads to the concept of an
unbiased estimate, is discussed in the next section.
Although it will not be proven here, it turns out that when d features are mul- .
tivariate normally distributed, the means, variances and covariances of the sampled
data are maximum likelihood estimates of the means, variances, and covariances of the
population.
The sample mean x is the maximum likelihood estimate for the mean J.tof a normal
distribution, as shown in Example 2.22, but this is not the case for all distributions.
The sample median, not the sample mean, is the maximum likelihood estimate of the
center of the double exponential distribution. In the case of the Cauchy distribution,
the maximum likelihood estimate for a in (2.20) is neither the sample mean nor the
sample median but must be obtained numerically (see Problem 2.46). Since the Cauchy
density has an infinite variance, the method of moments cannot be used to estimate
its spread parameter b. The following example shows that a maximum likelihood esti-
mate can be quite counterintuitive. Compare this example with the unbiased estimate
described in Example 2.25.
Unbiased Estimators
Another desirable property for an estimate 8of () is that it be unbiased. An unbiased
estimate 8is an estimate for () that satisfies
E(8) = ().
An unbiased estimate is neither too low or too high on the average, because the ex-
pected value of the estimate equals the parameter being estimated.
Notice that since J.txis defined as E(x),
Example 2.24 The maximum likelihood estimate for estimating the number of cards
in a numbered deck, given only one draw.
Consider a well-shuffie~ deck of N cards which are labeled 1,2,. .. , N. The label of a
single card drawn from this deck is a discrete random variable x, where the probability
of any particular value is l/N; that is, P(l) = l/N, P(2) = l/N, ..., P(N) = l/N. If
we draw one card, we know that N is at least as large as the value x we drew. If N = x,
the probability of drawing x would be l/x. If N is larger than x, say N = x + 1, then
(
1 n
)
1 n 1
,E(x) = E - Lxi = - LE(x;) = -(nJ.tx) = J.tx.
ni=1 ni=1 n
Thus x is always an unbiased estimate of J.tx,regardless of the type of distribution or
density.
N
P(x)
1 0
2 0
x-I 0
x
I/N
x+l 1/(N + 1)
x+2 1/(N + 2)
analoui@ius~
64 CHAPTER 2. PROBABILITY 2.6. ESTIMATION OF PARAMETERS FROM SAMPLES
65
Example 2.25 Obtaining an. unbiased estimate of the number of cards in a numbered
deck from only one sample.
similar to Example 2.24, the maximum likelihood estimator for b is x. This estimate
is clearly biased. To determine the bias, we calculate
rb rbdx b
E(x) = 10 xp(x) dx = 10 x b = "2' .
As in Example 2.24, the probability of choosing any particular card x from a deck
containing N cards labeled 1, . . . ,N is 1/ N. If the drawn card x was the average card
in the deck, there are x-I cards with smaller numbers and x-I cards with larger
numbers, so the total number of cards in the deck is 2(x - 1) + 1 = 2x - 1. We will
show that N = 2x - 1 is an unbiased estimate for N. To be unbiased, the expected
value of an estimate must equal the parameter value being estimated, which in this
case means E(N) = N. We have
Thus, on the average, the maximum likelihood estimate x would only be half of the true
parameter value b. To remove the bias, we double the maximum likelihood estimator:
b= 2x is an unbiased estimate of b.
Now consider two samples Xl and X2 drawn from a uniform distribution with un-
known range [a, b]. The maximum likelihood estimator of b-a is IXl-x21 (see Problem
2.48). From Problem 2.49, an unbiased estimate of b - a is three times the maximum
likelihood estimate. On the average, the two samples divide the density into three
equal areas. In general, n samples would divide the density into n + 1 equal areas on
the average, which is the reason for choosing the expected z-values as they were in the
normal plots in Section 2.6.
In general there are many possible unbiased estimates for a parameter. For exam-
ple, any weighted average of the original samples
E(N) = E(2x- 1) =2E(x) - 1=
N 1 2 N(N + 1)
2Lx.--1=-. -l=N
.,=1 N N 2
which verifies that N= 2x -1 is an unbiased estimate for N. If N = 2x -1, then there
are as many cards higher than x as there are lower than x. If several cards were drawn,
N= 2x -1 would be an unbiased estimate of N, as can be seen by substituting x for x
in the preceding equations. As the number of cards drawn becomes large, the difference
between the maximum likelihood estimate and the unbiased estimate becomes small.
If all of the cards were drawn, these estimates would be identical, because the average
of all the card numbers from 1 to N would be (N + 1)/2, so N = 2x - 1, and the
largest number drawn would also equal N.
n
p,= LWiXi where
i=l
n
L Wi = 1
i=l
(2.33)
is an unbiased estimate for ILbecause
(
n
)
n n
E(p,) = E EWiXi = EwiE(Xi) =EWilL = IL.
The maximum likelihood estimate N = x obtained for one draw in Example 2.24
is very different from the unbiased estimate N= 2x - 1. The average value of a large
number of independent unbiased estimates of a parameter is expected to approach
the true value of the parameter, while the average of a large number of maximum
likelihood estimates of a parameter may be extremely biased. Most people prefer the
unbiased estimate because it does not consistently overestimate or underestimate the
value of a parameter.
Unlike method of moments and maximum likelihood estimators, there is no single
systematic method for obtaining an unbiased estimate for an arbitrary parameter.
Unbiased estimates are usually created by correcting for the bias in another estimate,
such as a method of moments or maximum likelihood estimator. The bias is found by
computing the expected value of the estimator and comparing it to the true value of
the parameter.
For example, if one is given a single sample x from a uniform density with range
[0, b], what is the maximum likelihood estimator of b, based on x? Using reasoning
The average ofthe first and last samples drawn, (Xl +xn)/2, and even the first sample
value Xl, are unbiased estimates for IL;but these estimates would have a much greater
expected error than would the estimate x. In fact, x is the best unbiased linear estimate
for ILin the sense that x has the smallest variance of all unbiased estimates of ILwhich
are weighted averages of the sample values of the form (2.33).
We verify this assertion for the simple case of two samples Xl and X2 by showing
that the variance of aXl + (1 - a)x2 is minimized when a = 0.5. Since Xl and X2 are
independent, (2.26) indicates that the variance of aXl + (1 - a)x2 is [a2+ (1 - a?]0-2
where 0-2is the variance of Xl or X2. Differentiating with respect to a and setting the
result to zero produces [2a - 2(1- a)]0-2= 0, or a = 0.5.
An Unbiased Estimator for the Variance
Although x = (l/n) 2:7=1Xi is an unbiased estimate for IL, it can be shown that the
variance of the sample 0";; = (l/n) 2:7=1 (Xi - x? is not an unbiased estimator for the
[email protected]
66
\
2.7. MINIMUM RISK ESTIMATORS
67
variance of the population u2. In fact, no matter from which distribution the samples
are drawn,
and
-
(
S~1 . .. S~d
)
s- : .. :
Sld ... Sdd
E
(
2
)
n - 1
UZI = -u2.
n
(2.34)
Thus, the variance of the data about its mean is expected to be less than the variance
of the population about its mean. Bias is introduced because the second moment of
the data is taken about the mean of the data, which is known to be smaller than the
secon,d moment about any other point. In particular, it is expected to be smaller than
the second moment about the true mean of the population. (If the true mean of the
population were known, the second moment of the data about the population mean
would be an unbiased estimate of the variance of the population.) A proof of (2.34) is
outlined in Problem 2.50.
Multiplying (2.34) by n/(n - 1) removes the bias:
is an unbiased estimate for
-
(
U~1 . . . U~d
)
~- : .. : '
Uld . .. Udd
where
1 n
Sjk = ~ L (Xij - Xj)(Xik - Xk).
n - i=1
The estimated correlation Pjk is the same regardless of which type of estimate of
Ujk and uJ is used because the factors n or n - 1 cancel.
~E
(
U~I
)
= E
(
~U~I
) n-l n-l
[
n 1~ 2
]
= E -. - L (Xi - x)
n - 1 n i=1
= E
[
~ t(Xi - X)2
] n - 1 i=1
n n-12 2
= -.-U =U.
n-l n
2.7 Minimum Risk Estimators
2 1 n
s- '"
- n -1 L (Xi - x)2.
0=1
(2.35)
Consider the following problem: Predict the next outcome of a random variable X
whose probability distribution function P(x) is known. The "best" prediction depends
on the penalty or loss if the prediction is wrong. We will see that the best answer can
be the mean, median, or mode of the distribution depending on which loss function
is used. In each case, the best prediction is the one that minimizes the risk, which is
defined to be the expected loss.
Initially we will assume that x is a discrete random variable. One common penalty
function is the squared error loss function, in which the penalty for guessing the
wrong value is equal (or proportional) to the square of the error
L(8,x) = (x - 8?
Thus, dividing by n - 1 instead of n produces an unbiased estimate 82 of the variance
of the population:
In practice, the unbiased estimate s2 is usually used instead of the maximum likelihood
estimator U~/. The standard statistical tables such as the chi-squared, t, and F tables
also assume that the unbiased estimate of u2 is used.
In the multivariate case,
X~(]
Here, L(8, x) is the loss incurred for predicting 8 when in fact x is the next outcome.
The risk of predicting 8 is
R(8) = E[L(8,x)] = E[(x - 8?].
is an unbiased estimate for
p~(:)
If x is a discrete random variable,
n
R(8) = L(Xi - 8)2p(Xi).
i=1
ana1oui@ius~
68
To minimizeR(O), differentiate it with respect to 0, set the result equal to 0, an'd-,sol'Ve
for 0:
d . d n
dOR(O) = dO~(Xi - O)2p(Xi)
0=1
n
- -2 ~)Xi - O)P(Xi) = 0
i=l
or
n n
LXiP(Xi) = LP(Xi).
i=l i=l
Solving for 0 and using the facts that L:~1 P(Xi) = 1 and L:~1 XiP(Xi) = E(x), we
obtain
0 = E(x) =tlx.
Thus the squared error loss function leads to the mean tlx as the minimum risk predictor
0 of the next outcome of x.
Often there is no good reason to square the error to establish a penalty, and it is
more reasonable to set the penalty to be proportional to the error. If the penalty is
equal (or proportional) to the absolute value of the amount of error, rather than its
square, we obtain the absolute error loss function
L(O,X)=lx-OI.
The risk of choosing 0 is
n
R(O) = E[L(O,x)] = L Ixi - OIP(Xi). (2.36)
i=l
The absolute value terms can be eliminated by breaking this sum into three parts:
those for whichXi > 0 (Xi - 0 is positive), those for which Xi < 0 (Xi - 0 is negative),
and those for whichXi =0 (Xi - 0 is zero)
R(O) = L (Xi - O)P(xd+ L (0 - Xi)P(Xi) + L O. P(Xi).
~~ ~d ~~
To minimize R(O), set its derivative to zero:
(2.37)
dR(O)
---;jO = - L P(Xi)+ L P(xd = o.
x.>O O<Xi
Thus 0 should be chosen so that the probability that Xi is greater than 0 is equal to
the probability that Xi is less than 0, in order to minimize the expected absolute error,
if possible. Note that this is a sufficient, but not a necessary, condition and may not
,
~
}.
2.7. MINIMUM RISK ESTIMATORS 69
be possible for an asymmetric discrete distribution. (It is possible for all symmetric
distributions.) For example, if P(O) = 1/3, P(l) = 1/2, and P(2) = 1/6, the minimum
risk guess is () = 1, even though P(x; > 1) =1= P(Xi < 1). Increasing () above 1 would
move it away from 1/3 + 1/2 = 5/6 of the probability, and decreasing it below 1 moves
it away from 1/2 + 1/6 = 2/3 of the probability, both of which would increase the
error. Thus R((}) is minimized when () divides the distribution into two parts that are
as nearly equal as possible.
Whenever a distribution can be divided into two halves of equal probability by
a value (), () is called a median of the distribution. The median of a ranked list of
numbers is the middle number, if there are an odd number of entries in the list. If
the list contains an even number of entries, any value of () between the two middle
numbers can serve as a median. By convention, the median is often taken to be the
average of the two middle numbers.
The median is often a more useful statistic than the sample mean. For example,
median annual income is more commonly reported than mean annual income. Even
though a typical citizen may earn $10,000 to $50,000, the sample mean will be strongly
influenced by those few individuals with huge annual incomes. The median is more
representative of the typical citizen; half of the people earn less than this amount and
half earn more.
Example 2.26 The best estimate using the absolute error loss function.
Consider the discrete distribution defined by
From (2.36) the risk of choosing () is
R((}) = (1/4)1() - 01+ (1/4)1()- 11+ (1/4)1()- 2\ + (1/4)1()- 31.
Figure 2.18 shows a graph of the risk using the absolute error penalty function versus
(). If () < 0, R((}) can be decreased by moving () to the right (increasing it) because it .
will be moving closer to all four of the possible values of x, thus reducing the penalty no
matter which of the four values of X occurs. If 0 < () < 1, R( (})can still be decreased by
moving () to the right because it will be moving closer to three of the points-reducing
their penalties-and away from only one of them. Similarly, if 2 < (), R((}) can be
decreased by moving () to the left. Only if 1 :::; () :::;2 is it impossible to decrease R((})
X
P(x)
0
1/4
1
1/4
2
1/4
3
1/4
[email protected]
-~
70
2.7. MINIMUM RISK ESTIMATORS
71
v
(') Example 2.27 Minimum risk estimates from a sample distribution.
as
CD
:5 (\J
if
-2 0
2
4
Even if the true distribution or density function of a random variable is unknown, the
next outcome of the random variable x can be predicted by considering a sample of
data with the values Xl, X2, ..., xn. Suppose the five samples 1, 1, 2, 5, and 6 are
obtained from a random variable x. If the squared error loss function is used, the
mean, 3, is the best guess for the next outcome of x. If the absolute error loss function
is used, the best predicted value is the median, 2. Finally if the 0-1 loss function is
used, the best guess is the mode 1.
0
theta
Figure 2.18: The risk using the absolute error loss function in Example 2.26.
by moving O. But if 1 ~ 0 ~ 2, there are as many possible values of X greater than or
equal to 0 as there are less than or equal to O. Since all values of Xi are equally likely, 0 satisfies (2.37).
Continuous Densities
If continuous densities are used, we obtain the same results as in the discrete case:
The squared error loss function leads to the mean as the minimum risk predictor of
the next outcome of the random variable; the absolute error loss function leads to
the median; and the 0-1 loss function leads to the mode. The mean of a continuous
random variable was defined in Section 3.5 as
Finally, consider situations in which there is no penalty for guessing Xi exactly (or
within a small error f) and a penalty of 1 for not getting the correct value. This leads
to the ()L1 loss function, which is defined by
/L",= E(x) = i:xp(x)dx.
{
0 if Ix - 01< f
L(O,x) = 1 otherwise.
A median of a continuous random variable is any value m that divides the density of
X into two equal areas, so that
We choose 0 to minimize P(x < m) = i:p(x) dx = J: p(x) dx = P(x > m).
R(O) = E[L(O, x)] = L P(Xi)
1"';-01;:::.
= 1 - L P(Xi)
1"';-1<'
If f is small enough, the minimum of R(O) will occur when 0 = Xs, where P(xs)
is the largest of all the P(xd. The most likely value of x is called the mode (or
dominant mode) of the distribution. Choosing 0 equal to the most likely Xi maximizes
its probability of being exactly correct (or within a small f). The mode of a list is the
most frequent number in it.
A mode is any value of x for which p(x) attains a local maximum. If there is more
than one mode, the x for which p(x) is largest is called the dominant mode. A
local maximum of a function is larger than any of the other values in a small interval
centered at the local maximum. A local maximum is sometimes called a "peak." The
dominant mode is the value of x at the highest peak.
The density function of a population consisting of a mixture of two normally dis-
tributed classes may have either one or two modes. For example, the graph of the
density
p(x) =0.75
[
~e-("'-1)2/8
]
+ 0.25
[
~e-("'-4)2/2
] 2V27r V27r
is shown in Figure 2.19a. Its mean is (0.75)(1) + (0.25)(4) = 1.75 and its median can
be found numerically to be 1.834. The density has two modes: The dominant mode
is at x = 3.52 and the secondary mode is at x = 1.12. The median and the modes of
[email protected]
--
72
2.8. PROBLEMS
73
v
0 "
, ,
, ,
, ,
, ,
, ,
, ,
, ,
, ,
,
/~.
,/ :\
" ",
,/ "<c
v
0
If he charges $1,000 for his estimate, but returns $20 to his clients for each dollar of
error in his estimate, what price should he estimate in order to maximize his expected
profit?
Since the penalty is proportional to the absolute error of the estimate, the median
minimizes the expected loss. By definition, half of the area lies to the left of the
median. The area to the left of 520 is 0.3 and the area between 520 and 550 is also
0.3, so the point that divides the area under the distribution in half is two-thirds of
the way from 520 to 550. Thus, 540 is the best guess; that is, 0.1 + 0.2 + (2/3)(0.3)
= 0.5. If the density is correct, the guess of 540 would be too high half of the time
and too low half of the time. .
It)
0 It)
0
(')
0 (')
0
C\I
0
C\I
0
0
0
0
0
0
0
-2 0 2 4 6 8 10
-2 0
(a)
2 4 6 8 10
(b)
Figure 2.19: Mixture densities (solid curves) which are the weighted sum of two normal
densities (dashed). N(/-L,a2) denotes a normal density with mean /-Land variance a2.
(a) p(x) = 0.75N(1,4) + 0.25N(4, 1) is bimodal. (b) p(x) = 0.8N(1,4) +0.2N(2, 1) is
unimodal.
2.8 Problems
this type of function must usually be found by trial and error rather than analytically.
Densities having two modes are called bimodal densities. The mixture density
p(x) =0.8
[
~e-("'-1)2/8
]
+0.2
[
~e-("'-2)2/2
] 2Vii Vii
is unimodal (has only one mode) as is shown in Figme 2.19b. Its mean, median, and
mode are 1.2, 1.39, and 1.73, respectively.
2.1. Suppose you want to set up a subjective probability model to predict the number
of inches that it will rain tomorrow based on the weather forecast for tomorrow
which says that "the probability of rain tomorrow is 30 percent." Assuming that
you believe the forecast, fill in your subjective probabilities in the following table:
Example 2.28 Predicting the price of gold.
Suppose that a commodities expert has reason to believe that the probability distri-
bution for the price of gold one year from now, G, is given by the following table, and
that the probability density is uniform within each of the five price ranges. Thus, price
is a continuous random variable.
According to your table, what is the probability that it will rain more than one
inch?
2.2. Two coins are flipped simultaneously. One has a probability of heads equal to
0.6 and the other has a probability of heads equal to 0.7. What is the probability
that the coins will be both heads or both tails?
2.3. On a campus, 1/5 of the students are freshmen and 1/10 of the students are in
engineering, but 1/4 of the engineering students are freshmen. What fraction of
the students are neither engineers nor freshmen? Draw a Venn diagram.
/2.4. For a very large group of students, the probability of receiving an A in the first
English course was 30 percent and the probability of an A in the first calculus
course was 20 percent. Also, 15 percent of the students had As in both classes.
Amount in inches
Probability
r =0.0
0.0 < r 0.5
0.5 < r 1.0
1.0 < r 1.5
1.5 < r 2.0
2.0 < r 2.5
2.5 < r
Event
P(Event)
470 G < 500 0.1
500 :<::: G < 520 0.2
520 :<::: G < 550 0.3
550 :<::: G < 570 0.3
570 :<::: G :<::: 600 0.1
[email protected]
(a) Construct a table like the one in Example 2.3.
(b) What is the probability that a randomly selected student got/exactly one
A?' \
(c) What is the probability that a student got no As?
(d) If someone got an A in calculus, what is the probability that he or she got
an A in English?
(e) Does the data precisely agree with the assumption that the event of getting
an A in English is independent of the event of getting an A in calculus?
[Ans: 0.2; 0.65; 0.75; no, 0.06 i- 0.15]
2.5. When does P[(A and B)IC] = P(AIC)P(BIC)?
2.6. In a sample of 10 parts, 7 passed an electrical test, 8 passed a mechanical test,
and 6 passed both tests. Let E and M represent passing. the electrical and
mechanical tests.
(a) Construct a table like the one in Example 2.3.
(b) What are the values of P(E and M), P(E and M), P(E and M), and
P(EJ and M) for the sample?
(c) What is the probability that a random part from the data set passed exactly
one of the tests?
(d) What are the values of P(EIM) and P(E) for the sample? Are they exactly
ettual? Are E and M independent in this sample?
(e) Given this data, is it plausible that E and M are independent in the pop-
ulation?
2.7. Classes A and B are not mutually exclusive. In a data set of 11 samples, 7
samples belonged to class A, 5 belonged to class B, 2 belonged to both, and 1
belonged to neither. What is P(AIB) for this data?
2.8. A pattern recognition system has been developed for classifying objects into six
categories. It has been correct on 80 percent of the past testing data. What
is the estimated probability that exactly four of the next five samples will be
correctly classified?
2.9. A lab technician routinely misreads 10 percent of the tests made, on the average.
Out of 12 tests, what is the probability that two or more are misread? Hint:
Calculate the probability that none are misread and that exactly one is misread.
.{2.1O. A classifier has a 30 percent error rate. What is the probability that exactly
three errors will be made in classifying 10 samples? [Ans: 0.2668]
"
2.8. PROBLEMS
75
2.11. A biased coin shows heads 60 percent of the time on the average. What is the
probability that three of five flips will show heads?
2.12. A fair coin is tossed 50 times. What is the probability that heads will appear
exactly 25 times?
2.13. A classifier has an average error rate of 3 errors per 50 samples. What is the
probability that it will make exactly 3 errors on a set of 50 samples?
2.14. The negative binomial distribution gives the probability that exactly n tri-
als will be required to attain r successes. For this- distribution the number of
successes is fixed, whereas with the binomial distribution, the number of trials
is fixed and the probability of obtaining a certain number of successes is found.
The distribution for the negative binomial is
P(n) =
(
n - I
)
Ok(1- t-k
k-l
where 0 is the probability of success on one trial.
A basketball player is an 80 percent successful free throw shooter. To make 20
free throws, what is the probability that he or she will need to shoot exactly 25
times?
2.15. Derive the distribution for the negative binomial. Hint: The last trial must be a
success.
2.16. An X-ray source emits 10 photons per second on the average, with a Poisson
distribution. What is the probability that exactly 10 photons will be emitted in
a given second?
2.17. A large fire station responds on the average to four fires per day. If the distri-
bution of the number of fires per day follows a Poisson distribution, what is the
probability that there will be fewer than three fires tomorrow?
2.18. If P(x) is the Poisson distribution, show that
00
L P(x) = 1.
x=o
2.19.
If p(x) = x/2 for x ranging from 0 to 2, and p(x) = 0 otherwise, what is the
mean of this density? Showyour calculations and a graph of the density.
If p(x) = 2x/9 for 0 < x < 3, and p(x) = 0 otherwise, what is P(O < .1"< I)?
[Ans: 1/9]
1/2.20.
[email protected]
-- _.--
76
...,.
ci q
~ 2
r/)
c:
G>
C C\I
'0 ci
1:-
0>
'Qi ~
J: ci
co
~ ci
'0
c:
G> <0
C ci
'0
- ...,.
~ '
0> 0
'Qi
J: C\I
ci
0
ci
0
ci
0
2 4 6
-2
(b)
(a)
x
-1
0
2
x
Figure 2.20: Triangular density functions for Problems 2.22 and 2.23.
2.21. Calculate Pea ~ x ~ b) where x is a double exponential random variable with
the distribution
p(x) = !e-'zl.
2
Hint: Consider the two cases: a and b are both on the sameside of 0, and a and
b are on opposite sidesof o.
2.22. Feature x has an asymmetric triangular density with a peak at x = 2 and a range
from 0 to 5 as shown in Figure 2.20a. Compute P(l < x < 2).
}2.23. Calculate P( -0.2 ~ x ~ 0.5) wherex has the triangular density
{
I-x ifO<x<l
p(x) = x + 1 if -1 ~ x ~ 0
0 otherwise
shown in Figure 2.20b. [Ans: 0.555]
2.24. Calculate P( a ~ x ~ b) where x has the Cauchy density
1
p(x) = ( 2)'
7['l+x
2.25, Feature x is normally distributed, with /-l = 3 and eT= 2. Find P( -3 < x < -2),
2.26. Use the normal table in Figure 2.11 to calculate P(0.5 ~ x ~ 1.5) if x has the
standard normal density in (2.18).
,
2.8. PROBLEMS
77
12.27. Usea z-transformation and the normal tables to calculate P( -1 ~ x ~ 2) when
x has the density
1 _(z-2)2/18.
p(x) = 3v'27/
[Ans: 0.3413]
2.28. Use the normal table in Figure 2.11 to find the probability that a normal random
variable will be more than two standard deviations above its mean.
2.29. Show that
1
00 ~e-Z2/2dz = 1.
-00 V2i
Hint: Note that
(1
00 ~e-z2/2dz
)
2 =
1
00
1
00 ~e-(;2+';)/2dYdx
-00 V2i -00 -00 27['
and change to polar coordinates where r = v'x2+ y2.
2.30. For each of the following data sets, construct a normal plot, and decide if the
data appear to be approximately normally distributed.
(a) 35, 43, 46, 48, 51, 55, 58, 65
(b) 2.0, 3.0, 3.2, 3.5, 3.7, 3.9, 4.0, 4.2, 4.4, 4.4, 4.5, 4.8, 5.0, 5.1, 5.4, 5.8, 6.1
2.31. Show that the multivariate normal density (2.21) becomes the bivariate normal
density (2.29) when d = 2.
2.32. Show that if p = 0 in (2.21) then this density can be factored into the product
of two one-dimensional normal densities.
~.33. Calculate the mean and variance of this data set: 2, 4, 4, 6. Calculate an unbiased
estimate of the variance of the population from which this data was obtained.
[Ans: 4, 2, 8/3]
2.34. Three samples have values of x of 1, 2, and 4 and are assumed to come from a
normal distribution.
(a) What are the unbiased estimates aHhe me~n and variance of x?
(b) What are the maximum likelihood estimat~"s~f the mean and variance of
x? '~
. ,~
2.35. Calculate the variance of the triangular density in Problem 2.~
.
",-."," .
analoui(@'iustac.ir
-~
78 CHAPTER 2. PROBABILIT
2.36. Show that u;+y = u; + 2uzy+ u~.
2.37. Find the variance of the density from Problem 2.20.
. .
2.38. Write an expression for uz+y,z+w in terms of Uzz, Uyz, Uzw, and Uyw.
2.39. Let x be a Poisson random variable with mean A. Showthat u; = A.
2.40. Three samples were drawn from a population that is believed to be ,niforml~
distributed with an unknown range. The values of the feature x were J3, 20, anc
27.
(a) What is the maximum likelihood estimate of the range of x?
(b) What is the method of moments estimate of the range of x?
(c) What is an unbiased estimate of the range of x?
2.41. Show that x is the maximum likelihood estimate of Aif x has the ,Poisson distri-
bution with mean A.
2.42. Let t be an exponential random variable with rate parameter (3. Show that
E(t) = 1/{3and ul = 1/{32. Hint: Use integration by parts.
2.43. Show that 1# is the maximum likelihood estimate for (Jif x has the exponential
distribution with parameter (J.
2.44. Let x be a random variable with the double exponential density
1
p(x) = -e-/z-a/
2 .
Its mean is a because the density is symmetric about a. However, show that the
maximum likelihood estimate of a is the median of the samples Xl, . . . , Xn. Hint:
The expression L:~llxi - al is minimized when a is the median of X!,..., Xn.
2.45. Let x be a random variable with the double exponential density
1
p(x) = -e-/z-al/b
2b '
where the parameter a (the mean) is known. Show that the maximum likelihood
estimate for b is
1 n
b= - EIxi - al.
n i=l
The estimate b is called the mean absolute deviation (MAD).
"
2.8. PROBLEMS
79
2.46. The joint probability density function of the samples Xl, . . . , Xn which come from
a Cauchy distribution with b = 1 is
~
.
n 1
p(XI, . . . , xn) = n ( )2'
i=l 1 + Xi - a
Plot p(Xl,'" Xn) versus a for the set of samples Xl = 0,x2 = 1, and X3= 4. Find
the value a that maximizes this function. Note that this maximum likelihood
estimate of a is nei.ther the mean nor the median of the samples, even though a
is the median of the Cauchy density. .
i
;'
2.47. Show that the Cauchy distribution does not have a finite variance.
2.48. Find the maximum likelihood estimator of the range b - a of a uniform density
on the interval [a, bJ, given two independent samples Xl and X2. Hint: Let S =
min(xl,x2) and L = max(XI,X2)' Show that the joint density of the independent
samples Xl and X2 is
p(XI,X2)= p(Xl)p(X2) =
{
l/(b - a)2 if Xl,X2 E [a,b]
0 otherwise.
Now show that p(Xl,X2) is maximized when a = S and b = L.
2.49. In Problem 2.48, show that an unbiased estimate of the width of the interval
b - a is 3(L - S) where S = min(xl, X2) and L = max(xl, X2)' Hint: Verify that
E(S) = ! k2Sp(XI, X2) dXl dX2
=
!
r XlP(Xl,X2)dXldX2+
!
r X2P(Xl,X2)dx2dxl
lZl$.Z2 lZ2$.Zl
rb rZ2 dXl dX2
l
b rZl dX2 dXl
= la la Xl (b - a)2 + a la X2 (b - a)2 = a + (1/3)(b - a).
Similarly, showthat E(L) = a+(2/3)(b-a) and then find E(L-S) = E(L)-E(S).
2.50. Let Xl,..' , Xn be samples from a density function with parameters fL and u2.
If the samples are chosen independently, the expected covariance between each
pair of random variables is zero; that is, E[(xi - fL)(Xj - fL)] = O. Show that if
2 1 n
s = n - 1 E(Xi - x)2,
0=1
then E(s2) = u2. Use the following steps:
(a) From equation (2.24), E(x;) = fL2+ u2.

Ch2 Probability

Uploaded by

Copyright:

Available Formats

Ch2 Probability

Uploaded by

Document Information

Original Description:

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

Ch2 Probability

Uploaded by

Copyright:

Available Formats

anafoui@iustac.

You might also like