Professional Documents
Culture Documents
STATISTICS
FOR MATHEMATICS UNDERGRADUATES
N. A. RAHMAN
M.A.(Alld), M.Sc.(Stat.)(Ca1c.)
Ph.D.(Stat. )(Camb.)
Senior Lecturer in Mathematical Statistics,
University of Leicester
SECOND EDITION
Including an extensive supplement
bringing the total of exercises to over 580
Published in USA by
Macmillan Publishing Co., Inc.
866 Third Avenue, New York, N.Y. 10022
Distributed in Canada by
Collier Macmillan Canada, Ltd.
v
TO
BEAUMONT HALL
UNIVERSITY OF LEICESTER
Preface to First Edition
vii
viii PREFACE
The first two chapters provide the basis for the major Chapter 3,
which deals with a variety of statistical problems. These should give
the student an indication of the kind of manipulative mathematics
used in statistical theory, and a deepening insight into the principles.
The emphasis on estimation is based on the belief that this is perhaps
the most significant contribution of modern statistical theory, though,
unfortunately, a clear understanding of the ideas of estimation is
often not readily acquired by a beginner. The omission of confidence
intervals and power functions is due to my conviction that these ideas
require a thorough grasp of distribution theory which puts them
outside the scope of a first course. On the other hand, the notion of
efficiency in terms of relative variance is a useful concept which has
the additional merit of being intuitively acceptable to a beginner.
Multivariate regression and the general least-squares approach used
in the analysis of variance have both been excluded as too specialized
for a first course.
The last chapter is perhaps of limited interest, since there seems
to be an increasing tendency in universities to institute a course in
statistical theory before the study of complex functions. However, the
characteristic function is of such importance in theoretical statistics
that its complete exclusion was, on balance, considered unjustifiable.
The purpose of Chapter 4 is to provide a small selection of largely
simple examples to illustrate some of the ways in which the charac-
teristic function is used in statistical theory. Of course, these exam-
ples could also be solved by other methods, but it is hoped that, as
presented, they will give a beginner an insight into the analytical
power and simplicity of the method. In this sense the chapter can be
regarded as a self-contained unit.
In general, the above classification may be regarded as develop-
mental and, in fact, many of these exercises have been selectively
used in such a planned sequence with mathematics undergraduates
during different stages of the three-year degree. The general impres-
sion gathered on the basis of a few years' experience is that students
almost invariably found the exercises interesting and helpful in
clarifying their understanding of the subject, despite some difficulty in
the manipulative mathematics involved. Perhaps this limited experi-
ence might be an indicator of a more general acceptability and
usefulness of this collection. Answers and some hints on solutions
have been provided to assist such students as may be working on
their own, without the help of teachers.
The exercises have been gleaned from many different sources. My
major debt is to the standard research journals, and due acknow-
ledgement is made to published papers from which the ideas for the
exercises have been obtained. I am obliged to the publishers and
PREFACE ix
authors of
(1) An Introduction to Probability Theory and its Applications, Vol. I,
by W. FELLER (John Wiley, New York, 1952)
(2) Introduction to Probability and Random Variables, by G. P.
WADSWORTH and J. G. BRYAN (McGraw-Hill, New York, 1960)
(3) Probability Theory for Statistical Methods, by F. N. DAVID (Cam-
bridge University Press, 1951)
(4) Probability-An Intermediate Text-book, by M. T. L. BIZLEY
(Cambridge University Press, 1957)
(5) Introduction to Statistical Method, by B. C. BROOKES and
W. F. L. DICK (William Heinemann, London, 1958)
(6) The Advanced Theory of Statistics, Vol. I, by M. G. KENDALL and
A. STUART (Charles Griffin, London, 1958)
and to the publishers and trustees of
(7) Introduction to Mathematical Probability, by J. V. USPENSKY
(McGraw-Hill, New York, 1937)
for kindly permitting me to include certain instructive problems from
their works, Some exercises have also been obtained from past
examination papers of the University of Leicester, and my thanks are
due to the university authorities for permission to use them. Finally, a
number of exercises have been specially constructed for this collec-
tion to illustrate some specific points in theory, but no claim to
originality is made.
I should like to thank Professor H. E. Daniels, who initially
suggested my undertaking this compilation, and whose encourage-
ment has been a major support over several years. My thanks are
also due to Mr F. Downton, Reader in Mathematical Statistics at
the University of Birmingham, for his careful reading of a part of the
manuscript and for suggesting several interesting exercises for this
collection. Apart from this, I am also obliged to him for his intangible
help in the development of my ideas through a close professional
association. To this I would also add my thanks to Professor R. L.
Goodstein for the opportunity for experimentation with different
groups of students which has given me a first-hand experience of the
difficulties facing beginners in the field. Finally, I am much indebted
to Messrs Charles Griffin & Co., and their consultant editor, Dr M.
G. Kendall, for their helpful co-operation in the publication of this
work.
I should be grateful if users of this book would point out any
errors or obscurities in the formulation of the exercises.
LEICESTER 1966 N.A.R.
Preface to Second Edition
The continuing demand for this book has occasioned a new edition.
This is being done with a slight modification of the title which, I
believe, now more clearly reflects the nature of the contents and also
specifically distinguishes this book from my other collection-
Practical Exercises in Probability and Statistics, Griffin, 1972.
The new edition also contains a substantial supplement of 160
new exercises. Of these, the majority were initially formulated as
questions set in various undergraduate examinations that were con-
ducted over several years at the University of Leicester and, as such,
they reflect a pattern in the development of statistical education
which, it is hoped, will interest teachers elsewhere. Many of these
exercises were gleaned from published papers, but it is regretted that
it has not been possible to identify these sources explicitly. Neverthe-
less, I should like to record my indebtedness to all whose work has
been used in the exercises. Furthermore, not all the exercises in the
supplement are devoted to the traditionally defined areas of probabil-
ity and statistical theory. In fact, in keeping with my individual
approach to statistical education at the undergraduate stage, I have
included in the supplement a few exercises in ancillary mathematics
which experience here has shown to be helpful in stimulating student
interest in statistics. Personal preference apart, there is a deeper
pedagogical reason for this departure from established practice.
There is now, more than ever before, a need to stress the underlying
unity of the mathematical sciences at the undergraduate stage in
order to avoid specious compartmentalisation. I believe this unity is
intellectually necessary to provide balance in the structure of the
honours degree in mathematics and to meet the academic challenge
to the mathematical sciences that is currently being presented by the
spectacular advances in computer technology.
I am grateful to users of the earlier edition of the book who
pointed out various misprints. These have now been corrected.
LEICESTER N .A.R.
14th March, 1980
x
Contents
EXERCISES
Page
Chapter 1 143
Chapter 2 177
Chapter 3 209
Chapter 4 285
SUPPLEMENT
xi
1 Probability and discrete random variables
1 Two dice are rolled. Let A be the event that the sum of the points on the
faces shown is odd, and B the event that there is at least one 3 shown. Describe
A u B; A II B; A - B; and (A II B) u A.
2 For k = 1,2, 3,4, let Nk be the event that N has at least k aces in a bridge
deal. Let Sk, Ek, w,. be the analogous events for S, E and Wrespectively. How
many aces has W in the events
(i) Wi; (ii) N2"S2; (iii) NiISl"B i ; (iv) W2 -W3 ;
(v) Ni II Si" Ei II Wi; (vi) N3 II Wi; (vii) (N2 u S2) II E2?
3 Given the three arbitrary events A, Band C, find simpler expressions for
(i) (A u B) II (A u B); (ii) (A u B) II (A u B) II (A II B) ;
(iii) (A u B) II (B u C); (iv) (A u B) II (A u B) II (A u B).
4 Find the probability that no two of r ( < 10) digits, each chosen at random
from 0, 1,2, ... ,9, are equal.
5 Three dice are rolled. What is the probability that there is at least one 6,
assuming that the dice are unbiased? Also obtain this probability if the dice are
loaded, the probabilities of getting a 6 with them being PI' P2 and P3 respec-
tively.
6 A set of eight cards contains one joker. A and B are two players and A
chooses five cards at random, B taking the remaining three cards. What is the
probability that A has the joker?
A now discards four cards and B two cards. If it is known that the joker
has not been discarded, what is the probability that A has the joker?
7 Find the probability that two throws with three unbiased dice each will
both show the same configuration if
(i) the dice are distinguishable; and
(ii) they are not distinguishable.
8 Consider three loaded dice. Die A is certain to show a 3. Die B is twice
as likely to show a 2 as to show a 5, and will not show any other number.
Die C is twice as likely to show a 4 as a 1, and will show no other number.
The three dice are rolled once. Find the probability that
(i) A shows a larger number than B;
(ii) B shows a larger number than C;
(iii) C shows a larger number than A; and
(iv) A shows a larger number than B, and B shows a larger number than C.
9 The Security Council numbers eleven members with Great Britain, China,
France, Russia and the United States as permanent members. If at a meeting
2 EXERCISES IN PROBABILITY AND STATISTICS
the members sit down at random, find the probability that the British and
French delegates are next to each other but that the Russian and American
delegates are not, if
(i) the delegates sit in a row; .and
(ii) the delegates sit round a circular table.
10 In a row of sixteen tiles on a roof, eleven chosen at random are blown off
in a gale. Defining a "run" as a sequence of gaps or of tiles, find the probability
that the row contains
(i) 10 runs; (ii) 9 runs.
11 Two unbiased dice are rolled r times. Find the probability that each of the
six results (1, 1), (2,2), ... , (6,6) appears at least once.
12 (i) Eight castles are placed at random on a chess-board. Prove that the
probability that none can take another is
(8 !)2/64(8)
(ii) If the castles are placed randomly on the chess-board but subject to
the condition that none can take another, then prove that the prob-
ability that none stands on a square of the white diagonal is
8
L (-1)'/r!
,= 2
(iii) Hence, or otherwise, deduce that if the castles are placed completely
at random on the chess-board, then the probability that none can
take another and none stands on a square of the white diagonal is
~ (_ 1)'/r.'
!)2. L...
(8 (8)
64 ,= 2
13 Find the probability that a hand of thirteen cards from a standard pack
contains the ace and king of at least one suit.
14 For a random distribution of cards at bridge, calculate the probabilities
for the following events:
(i) A specified player (say S) has at least two aces, irrespective of the
other players' hands.
(ii) Some one player of the four has at least two aces, irrespective of the
other three players' hands.
(iii) S and no other player has at least two aces.
(iv) Exactly one player of the four has at least two aces.
15 X chooses at random an integer from 1 to m (both inclusive). Without
knowing X's choice, Y also chooses an integer from 1 to m. Find the prob-
ability that the two numbers chosen do not differ by more than n «m).
16 Suppose that each of n sticks is broken into one long and one short part.
The 211 parts are then shuffled and arranged into n pairs from which new sticks
are formed. Find the probability that
(i) the parts will be joined into their original form;
(ii) all long parts are paired with short parts; and
(iii) at least one of the original sticks is formed.
PROBABILITY AND DISCRETE RANDOM VARIABLES 3
17 The digits 1, 2, ... , n (n < 10) are written down in a random order such
that the n ! arrangements are all equally likely to give a number of n digits.
If n == 7, find the probability that the number is divisible by
(i) 2; (ii) 4; and (iii) 8.
18 The game of craps is played as follows. In a particular game one person
rolls a pair of dice. He wins on the first throw if he scores 7 or 11, and loses on
the first throw if he scores 2, 3 or 12. For any other score on the first throw,
there are two different ways of continuing the game. In the first method, the
player continues to roll the dice until he wins with a 7 or loses with the score
obtained on the first throw. In the second method, the player also continues
to roll the dice until he loses with a 7 or wins with the score obtained on the
first throw. Find the total probability of the player winning a game under
each of the two systems of play.
19 An urn contains n white and m black balls; a second urn contains N white
and M black balls. A ball is randomly transferred from the first to the second
urn, and then from the second to the first urn. If a ball is now selected randomly
from the first urn, prove that the probability that it is white is
n mN-nM
- - +-,------:-;;-----
n+m (n+m)2(N +M + 1)"
P(AIB) = P(A)-P(AB)
I-P(B) ,
where B denotes the event complementary to B, and hence deduce that
P(AB) ~ P(AHP(B)-l.
Also, show that P(A) > or < p(AIB) according as p(AIB) > or < P(A).
29 As a rather simplified example of weather forecasting at a seaside resort,
suppose that the probability that the weather (fine or wet) of one day will be
the same as that of the preceding day is a constant p. If, on the basis of past
records, it is assumed that the probability of the first of August being fine is 0,
PROBABILITY AND DISCRETE RANDOM VARIABLES 5
a constant, determine On, the probability of it being fine on the following nth
day. Show also that as n -+ 00, this probability tends to the limit!, irrespective
of the values of P and O. Interpret this result.
30 There are n similarly biased dice such that the probability of obtaining
a 6 with each one of them is the same and equal to p, (0 < P < 1). If all the dice
are rolled once, show that Pn' the probability that an odd number of 6's is
obtained, satisfies the difference equation
Pn+(2p-l)Pn-l = p,
and hence derive an explicit expression for Pn.
31 Each of n urns contains a white and b black balls which are indistinguish-
able except for their colour. One randomly selected ball is transferred from the
first urn into the second, another one from the second to the third, and so on.
Finally, a ball is drawn at random from the nth urn. Find the probability for
it to be white when it is known that the first ball transferred was white.
32 Two urns contain, respectively, a white and b black, and b white and a
black balls. A series of random drawings is made as follows:
(i) Each time, only one ball is drawn and it is returned to the same urn
from which it was obtained.
(ii) If the ball drawn is white, then the next drawing is made from the first
urn, the second urn being used in the contrary event of a black ball
drawn.
Assuming that the first ball is drawn from the first urn, find the probability
that the 11th ball drawn is white, and obtain its limiting value as 11 -+ 00.
33 Of three urns, the first contains a proportion x of white balls and a
proportion (1- x) of blue balls, the second contains blue and red balls in
the proportions y and (1 - y) respectively, and the third contains red and
white balls in the proportions z and (1- z) respectively, none of the given
proportions being zero. A ball is drawn randomly from the first urn and then
replaced. If it was white, a ball is drawn from the second urn and replaced,
but if the first ball drawn was blue, a ball is drawn from the third urn and
replaced. The process continues, a ball being drawn each time from the first,
second or third urn according as the previous ball was red, white or blue
respectively. Every ball is replaced before the next is drawn.
If rn , Wn and bn denote respectively the probabilities at the outset that the
11th ball drawn will be red, white or blue, prove that as n -+ 00 these prob-
abilities tend to fixed limits, and find the limiting values.
34 A number n of teams, all always of equal skill, compete annually for a
cup which is to be awarded outright to the team winning three years in suc-
cession. Find the probability of team X winning the cup outright if it must
be won every year and team Y won it for the first time last year.
35 Two players A and B agree to playa game of skill in which their respective
probabilities of winning a game are Po and Pb' and the chance of a drawn
game is Po. Find the probability of A winning at least (m+ 1) games out of
(2m+ 1) played, and obtain its numerical value in the particular case of m = 2
and Po =!.
Also, if A and B play a match to be decided as soon as either has won
two games, find the probability of A winning the match in n games or less,
and show that as n -+ 00, this probability tends to the limit
p;(I- Po + 2Pb)/(1- POll.
6 EXERCISES IN PROBABILITY AND STATISTICS
Hence obtain the limiting probability that the match will end ,in a win for
either player.
If 11 = 10. PII = PI> = i, prove that the probability of the match being
finished in ten or less games is 1981/2048.
36 *
In a game of skill a player has probabilities t, 152 and of scoring 0, 1
and 2 points respectively at each trial, the game terminating on the first
realization of a zero score at a trial. Assuming that the trials are independent,
prove that the probability of the player obtaining a total score of 11 points is
u" =
3(3)"
13 4(-:3 .
4 + 39 I)"
and that the expectation of his total score is ¥.
Also, suppose the rules are changed so that the game does not end on
the first realization of a zero score at a trial but the trials continue indefinitely.
In this case, show that the probability of the player obtaining a score of
exactly 11 points at some stage of play is
p"
b [ 1]".
= 1 - (a + b) 1 - a + b
38 Two urns contain, respectively, a white and b black, and c white and d
black balls. One ball is taken from the first urn and transferred into the second,
while simultaneously one ball taken from the second urn is transferred into
the first. Find the probability, p", of drawing a white ball from the first urn
after such an exchange has been repeated n times. Also obtain the limiting
value of p" as n -+ 00.
39 Two players A and B start playing a series of games with £a and £b
respectively. The stake is £1 on a game, and no game can be drawn. If the
probability of A winning any game is a constant, p, find the initial probability
of his exhausting the funds of B or his own.
Also, show that if the resources of B are unlimited, then
(i) A is certain to be ruined if p = !; and
(ii) A has an even chance of escaping ruin if p = 2 1 / a/(1 +21/a).
40 Two players A and B agree to contest a match consisting of a series of
games, the match to be won by the player who first wins three games, with the
proviso that if the players win two games each, the match is to continue
until it is won by one player winning two games more than his opponent.
The probability of A winning any given game is p and the games cannot be
dra\vn.
(i) Prove that f(p), the initial probability of A winning the match, is
given by
PROBABILITY AND DISCRETE RANDOM VARIABLES 7
(ii) Show algebraically that
df > 0
(IX) dp for 0 <- p < 1,
(fJ) f(P)+f(l-p) = 1,
and interpret these facts in terms of the data of the problem.
(iii) Show that the equation f(P) = p has five real roots, of which three
are admissible values of p. Find these three roots and explain their
significance.
(iv) Calculate the expectation of the number of games to be played for
A to win the match.
41 An internal telephone exchange in an office has m (> 2) single connections,
one for each room in the establishment; and during a working day all the
m connections are equally likely to contact the telephone operator for service.
If Pn denotes the probability that in any sequence of n calls to the operator
no room contacts the exchange more than three times consecutively, prove
that Pn = [(1-IX2)IX~-I_(1-IXI)IX~-1 V(IX I -I(2),
where IXI and IX2 are the roots of the quadratic equation in x
m 2x 2 -III(m-l)x-(m-l) = O.
Find the limiting values of Pn when
(i) II -+ 00 and m is finite;
(ii) m -+ 00 and II is finite;
and interpret their significance.
42 In a lottery m tickets are drawn at a time out of 11 tickets numbered from
1 to 11 (m ~ 11). Find the expectation and variance of the random variable S
denoting the sum of the numbers of the III tickets drawn.
43 At an office N letters are to be posted, one to each of N different addresses.
and a capricious secretary decides to distribute the letters randomly amongst
the N addressed envelopes. If all the N ! arrangements of the letters are equally
likely, show that the expectation and the variance of the number of correct
postings are both unity.
In a similar situation, another slightly less capricious secretary decides to
make independently a guess of the correct envelope for each letter to be
posted, so that the NN choices are all equally probable. Prove that in this
case the expectation and variance of the number of correct postings are 1
and (N -l)/N respectively.
44 A box contains 2" tickets among which (~) bear the number r
(r = 0, 1.2•... , n). A group of III tickets is drawn at random from the box,
and if the random variable X denotes the sum of the numbers on the tickets
drawn. show that
E(X)
11111
= 2; ( X)
var
= mil
4
[1- (2" - 1)]
(111-
1) .
respectively.
Further, show that this probability distribution can be formally obtained
from a finite sampling scheme with replacement in which at least one ball
is selected randomly at a time from a total of (n + 1).
46 A population consists of all the positive integers, and the probability
of obtaining the integer r from this population is
P(r) = k(1- Or 1, (r = 1,2,3, ... ), where 0 < 0 < 1.
Determine the constant k, and the mean and mode of this population.
Show also that if 0 = 1_(Wln, where n is a positive integer, then the
median of the distribution may be considered to be n +1.
What is the variance of the distribution?
47 A certain mathematician always carries two match-boxes, which initially
contain N match-sticks each, and every time he wants a light, he selects a
box at random. Obtain Ur , the probability that when he finds a box empty
for the first time the other box contains exactly r match-sticks, and verify
that
N
L
r=O
Ur = 1.
Also, prove that the expectation and variance of the number of matches left
in the box are
p and [(2N+2)-(1+p)(2+p)] respectively,
where
p == (2N + 1)uo-1.
48 There are n compartments into which identical balls are distributed one
by one in such a way that each ball is equally likely to fall into anyone of
the n compartments. This process is continued until every compartment has
at least one ball. Prove that the probability that every compartment is occu-
pied after t balls have been used is
mto (:)(-If(n-m)t/nt.
Hence deduce the probability that exactly t balls are needed for filling
all the n compartments, and that the expected number of balls required is
n n
L -.
m=lm
49 A box contains N varieties of objects, the number of objects of each
variety being the same. These objects are sampled one at a time with replace-
ment; and if Xr is a random variable which denotes the number of drawings
PROBABILITY AND DISCRETE RANDOM VARIABLES 9
necessary to produce any r different varieties in the sample, find the expecta-
tion and variance of X,.
Also, for large N, show that
E(X,) '" N 10g[Nj(N -r+ 1)],
and
var
(X ) N(r-l)
, '" N
-r+ 1 N 10g[Nj(N -r+ 1)].
50 In the previous example, let the N varieties be identified by being num-
bered from 1 to N. If X denotes the largest number drawn in n drawings
when random sampling with replacement is used, find the probability of
X = k. Hence obtain the mean and variance of X.
Also, show that for large N and fixed n,
E(X) "" nN nN 2
n+l and var(X) "" (n + 1)2(n + 2)'
and verify that this represents a proper probability distribution over the given
range of variation of the random variable. Hence show that
Y= X(W+ 1) 1,
w
then prove that
E(Y) = Nand var(Y) = (N+l)(N-W)(W-w+1)jw(W+2).
52 A large number (N) of persons are subject to a blood test which can be
administered in two alternative ways:
(i) Each person is tested separately, so that N tests are required.
(ii) The blood samples of k (a factor of N) persons are pooled and analysed
together. If the test is negative,-this one test suffices for the k individuals.
If the test is positive, each of the k persons must be tested separately,
and in all (k+ 1) tests are required for k persons.
Assuming that the test responses of the N persons are statistically inde-
pendent and that the probability (1- p) for a test to be positive is the same
for all individuals, find the probability that the test for a pooled sample for
k persons will be positive.
\0 EXERCISES IN PROBABILITY AND STATISTICS
If S be the number of tests required for the N persons under plan (ii),
prove that the mean and variance of S are
N[l+k(1-pk)]/k and Nkpk(l-pk) respectively;
and that the value of k which gives the minimum expected number of tests
for the N persons satisfies the equation
k 2 + l/pk logp = O.
53 In the simplest type of weather forecasting-rain or no rain in the next
24 hours-suppose the probability of raining is p (> 1), and that a forecaster
scores a point if his forecast proves correct and zero otherwise. In making n
independent forecasts of this type, a forecaster who has no genuine ability
decides to allocate at random r days to a "rain" forecast and the rest to
"no rain". Find the expectation of his total score (Sn) for the n days and show
that this attains its maximum value for r = n. What is the variance of Sn?
Devise an alternative scoring system which would ensure that under
random forecasting the expectation and variance of Sn will be 0 and 1 re-
spectively.
54 In the above example, suppose the forecaster has some ability to forecast
the weather correctly. Accordingly, if the probability of his forecasting rain
for any day is ~, and the conditional probability of raining on a day given
that a "rain" forecast has been made is p, find the respective probabilities of
the four possible outcomes for any day. Hence, assuming independence of
weather conditions for n days, obtain a scoring system such that
(i) E(Sn) = 0 for random forecasting;
(ii) E(Sn) = n if the forecaster has perfect ability; and
(iii) var(Sn) is a minimum when the marginal distribution of forecasting
is the same as the marginal distribution of rain on any day, the two
events being assumed independent.
Verify that with this scoring system
E(Sn) = n~(p - p)/pq, (p+q = 1).
55 In a sequence of Bernoulli trials with probability of success (S) p and
of failure (F) q, (p + q = 1), find the expected number of trials for the realization
of the patterns
(i) SSSFS; (ii) SF FS ; and (iii) SSSF.
56 For a sequence of n Bernoulli trials in which the probability of success
(S) is p and that of failure (F) is q, (p+q = I), show that Yn, the probability
that the pattern SF does not occur in the entire sequence, satisfies the difference
equation
Yn-Yn-l +pqYn-2 = 0 for n ~ 2.
Hence obtain explicit expressions for Yn when p :F q and when p = q.
Also, prove that the expected number of trials necessary for the realization
of r consecutive repetitions of the pattern SF is
I-p'q'
pq'(1 -pq ) ~ ~4' -1).
,
57 Random drawings are made from an urn containing b black and w white
balls. Each ball drawn is always replaced, and, in addition, c balls of the
PROBABILITY AND DISCRETE RANDOM VARIABLES II
colour drawn are added to the urn. If P(lI, k) denotes the probability of draw-
ing exactly k black balls in the first n drawings, show that P(n, k) satisfies the
recurrence relation
b+(k-l)c w+(n-k-l)c
P(lI,k)=b (1)·P(n-l,k-l)+b (1)·P(n-l,k),
+w+n- c +w+n- c
where P(n, - I) may be taken to be zero. Hence obtain P(n, n) and P(n, 0).
Also, for k < n, verify that
n) b(b+c)(b+2c)" ·{b+(k-l)c}. w(w+c)(w+2c)" '{w+(n-k-l)c}
p(/I,k) = (k . (b+w)(b+w+c)(b+w+2c)" '{b+w+(n-l)c}
1
Finally, if now n -+ 00, p -+ 0, y -+ 0 so that np -+,-t, ny -+ -, then show that
p
the limiting form of P(n, k) is
and hence that, as p -+ 00, ll(k) tends to the Poisson distribution with mean ,-t.
58 Two players A and B alternately roll a pair of unbiased dice. A wins
if on a throw he obtains exactly six points before B gets seven points, B winning
in the opposite event. If A begins the game, prove that his probability of
winning is 30/61, and that the expected number of trials needed for A's win
is approximately 6.
59 Under a newly proposed motor insurance policy, the premium is £a in
the first year. If no claim is made in the first year, the premium is £Aa in the
second year, where (0 < ,-t < 1) and ,-t is fixed. If no claim is made in the first
or second years, the premium is £,-t2(X in the third year; and, in general, if no
claim is made in any of the first r years (r ~ 1), the premium is £,-tr(X in the
(r+ l)th year.
If in any year a claim is made, the premium in that year is unaffected, but
the next year's premium reverts to £(X, and this year is then treated as if it
were the first year of the insurance for the purpose of calculating further
reductions. Assuming that the probability that no claim will arise in any
year is constant and equal to q, prove that in the nth year (n ~ 2) of the
policy, the probabilities that the premium paid is nn-l(X or nn- j-l(X,
(1 ~j ~ n-l), are qn-l and (l_q)qn- j - l respectively.
Hence calculate the expected amount of the premium payable in the nth
year and show that if this mean must always exceed k(X (k > 0), then
, k+q-l
A> kq .
12 EXERCISES IN PROBABILITY AND STATISTICS
60 A player rolls four unbiased dice, and if S is the random variable denoting
the sum of points obtained in a single throw of the dice, prove that the prob-
ability P(S = n) is the coefficient of (}n-4 in the expansion of
(1- (}6)4/6 4(1_ (})4
for all n in the range (4 ~ n ~ 24).
Hence, or otherwise, deduce that
(i) P(S = 18) = 5/81; and (ii) E(S) = 14.
61 The probability of obtaining a 6 with a biased die is p, where (0 < p < 1).
Three players A, Band C roll this die in order, A starting. The first one to
throw a 6 wins. Find the probability of winning for A, Band C.
If X is a random variable which takes the value r if the game finishes at
the rth throw, determine the probability-generating function of X and hence,
or otherwise, evaluate E(X) and var(X).
62 The six faces of an ordinary cubical die are numbered from 1 to 6. If
two such unbiased dice are rolled once, find the probability distribution of
the random variable X denoting the sum of points obtained.
Also, find an appropriate numbering of the twelve faces of two unbiased
dice which would ensure that the probability P(X = r) is the same for all r
in the range (1 ~ r ~ 12), and show that for such a pair of dice the probability-
generating function of X is
G((}) = (}(1- (}12)/12(1- ()).
63 From an urn containing (2n+ 1) tickets numbered serially, three tickets
are drawn at random without replacement. Prove that the probability that
the numbers on them are in arithmetical progression is 3n/(4n 2 -1).
Further, by considering the sample space corresponding to the possible
realizations of the arithmetical progressions, show that the common difference
of the arithmetical progression can be regarded as a discrete random variable
X with the probability distribution defined by
(2n+ 1)-2r
P(X = r) = 2 ' for r = 1,2,3, ... , n.
n
Hence deduce that the probability-generating function of X is
G( (}) = 4(} . [1 _ (2n + 1- (}n)( 1 + (}~l.
n(l- (})2 4n J
64 In a game of skill a player has probability p of winning a point and
probability q, (p+q = 1), of losing a point at each attempt. If the trials are
independent, find the probability distribution of the random variable X
giving the player's total score in n trials. Hence, or otherwise, obtain the mean
and variance of X. Also, show that if a new random variable Y is defined by
the relation
Y= (X+n)/2,
then Y has the Bernoulli distribution with probability of success p in each
of n trials.
65 At each independent trial in a game, a player has probability p of winning
a point, probability q of losing a point, and probability r for no loss or gain,
PROBABILITY AND DISCRETE RANDOM VARIABLES 13
where (p+q+r = 1). Find the probability-generating function of the random
variable X giving the player's total score in n trials, and hence deduce the
mean and variance of X. Also, show that the probability for a total score of
m (~n) points in n trials is
P(X = m) = L
(n+m)/2( n . )(m+n-J)
. . .
{p/rY (q/ry-m. r".
}=m m+n-J J
66 A particular constituency has a total of (L + C + F) registered voters of
which L are by conviction Labour supporters, C are Conservative and Fare
the floating voters. The probability that a Labour supporter will vote for
his party in a by-election is Pl' and the probability is P2 that a Conservative
will exercise his vote. The probabilities are P3 and P4 for a flQating voter to
vote either Labour or Conservative respectively.
Prove that in a straight fight between a Labour and a Conservative can-
didate, the probability of a tie is given by the coefficient of ()c+ F in the function
[(1- Pl)+ Pl()]L [(1- P2)9+ P2f [(1- P3 - P4)() + P3()2 + P4]F.
Show also that for the random variable N representing the total number of
votes cast in the by-election, the mean and variance are:
E(N) = Lpl+CP2+F(P3+P4);
var(N) = LPl(1-Pl)+CP2(1-P2)+F(P3+P4)(1-P3-P4)'
" (11)
111(r) = r~r r prqn-r.(r_np)
71 For the discrete Poisson distribution with parameter A., obtain the
moment-generating function about the mean of the distribution, and hence
PROBABILITY AND DISCRETE RANDOM VARIABLES 15
verify that the fourth, fifth and sixth central moments are
Jl.4 = 3A.2+..t; Jl.s = 1O..t2 +..t; and Jl.6 = 15..t 3 +25..t 2 +..t.
72 A discrete random variable X is defined over the values 0, ± 1, ±2,
+ 3, ... , it being equally likely to take positive or negative integral values. If
the probability that X takes the value r > 0 is pr, (0 < p < 1), prove that
the moment-generating function of X about the origin is
1 + 24>(1 + 4>)(24) - 1) sinh 2 t/2
M(t) = 1-44>(1 +4» sinh2 t/2 ,where 4> =P(X =F 0),
and hence deduce that the variance of X cannot exceed 6.
73 A discrete random variable X, taking all the non-negative integral values,
has the probability-generating function
00
exp[A{4>(t)-l} ].
v. == (N-s)! H.!n b ! 1= 1
l
Use this result to obtain explicitly the mean and variance of the distribution.
76 A random variable Y has a distribution with mean
By expressing Y in the form
Il and variance (12.
[ Y-Il]
Y = Il 1+-Il- ,
show that
if the variations in Yare sufficiently small as compared with its mean value.
If Y has a Poisson distribution with a large mean, prove that as a first
approximation
var(jY) = 1.
n For n observations Xl> X2," ., X n , the arithmetic, geometric and har-
monic means, denoted by a, g and h respectively, are defined by
1 I n 1
- = -
h
L-'
ni=1xi
g = a [1 -2S;2] ; h= a [1-:: l
and hence a-2g+h = O.
PROBABILITY AND DISCRETE RANDOM VARIABLES 17
78 A discrete random variable X takes the values -1, 0 and 1 with prob-
b'lities tp, (1- p) and tp respectively. If Xl> X2' ... ,Xn are n independent
~e~lizations of X, find the moment-generating functions of the random
variables Yand Z defined by
n n
Y= L XI and Z = L xf,
1= 1 1= 1
var(Sn) = 2m (1 +~).
[ ~] _ ![1 + (3-8b)
var y X+u - 4 8A. +
(32b 2 -52b+ 17)]
32-1. +0
(A.--3).
, 2
82 A lady declares that by tasting a c,;!p o~ tea made with milk she can
discriminate whether the milk or the tea mfuslOn was first added to the cup.
In an experiment designed to test her claim, eight cups of tea are made, four
by each method, and are presented to ~er for judgme.n~ in ~ random order.
Calculate the probability of the cups bemg correctly dIvIded mto two groups
of four if such a division is entirely random. If the experiment is repeated ten
times 'and a "success" in an individual experiment is scored if the lady
corre~tly divides the cups, show that a score of two or more "successes" in
ten trials provides evidence, significant at the 1 per cent level, for her claim.
The lady considers this definition of "success" too stringent, so an alter-
native experiment is devised. In this experiment six cups of each kind are
presented to her in random order, and she is now regarded as scoring a
"success" if she correctly distinguishes five or more of each kind. Calculate
the probability of such a "success", and also the probability of three or more
successes in ten independent experiments, assuming her claim to be false.
Explain the difference, if any, between the stringency of these two procedures
for testing the lady's claim.
83 In an unlimited sequence of Bernoulli trials with probability p of success
(S) and q of failure (F), two patterns are defined by the outcomes of three
consecutive trials as (i) SFF and (ii) FSF. Prove that the expected number of
trials needed for r (~ 1) consecutive realizations of the pattern (i) is
84 A printing machine is capable of printing any ofn characters IX" 1X 2, ••• , IXn
and is operated by an electrical impulse, each character being, in theory,
produced by a different impulse. In fact, the machine has probability p,
independent of any previous behaviour, of printing the character corresponding
to the impulse it receives, where (0 < p < 1). If it prints the wrong character,
all such characters are, independently, equally likely to be printed.
One of the n impulses, chosen at random, was fed into the machine twice
and on both occasions the character lXi was printed. Show that the prob-
ability that the impulse chosen was the one designed to produce the character
lXi is (11- 1)p2/( I - 2p + np2).
If the machine had printed lXi on the first occasion and IXj (j 1= i) on the
second, determine now the probability that the impulse chosen was that
designed to produce lXi' Does it make any difference to this probability if it is
merely known that the second character printed was not lXi?
85 Electronic tubes in continuous production at a factory are packed in
batches of N (large) units for despatch to customers. It is known from past
experience of the manufacturing process that the probability of a random
batch containing k defective tubes is Pk> where (0 ~ k ~ m). From a randomly
selected batch, a sample of n ( < N) tubes is tested and of these r ~ m are found
to be defective. Find the probability that the chosen batch had, in fact, k ~ r
defective tubes.
PROBABILITY AND DISCRETE RANDOM VARIABLES 19
If according to a customer's specification of acceptable quality, batches
with P(~r) defectives are considered to be definitely bad, prove that the total
probability of the selected batch being in fact a bad one is
l_:t:Pk(~=;)/(~)
m
j=,
(N-n)
L Pj. (N)'
I/'.
}-r '/ ' }
DiscusS briefly how this result could be used for quality control of the produc-
tion process.
86 In a big department store there are n separate service counters, and a
customer entering the store is equally likely to go to anyone of the counters
for service. Find the probabilities that a particular counter X will be the first
to have a run of r consecutive customers before any other counter receives s
successive customers when it is known that just preceding the count only the
last customer went to (i) counter X; and (ii) another counter Y.
If /I is finite and r = s large, verify that these two probabilities are both
asymptotically equal to l/n, and explain the significance of this limiting value.
87 Housewives shopping at a grocery store are equally likely to ask for
anyone of n standard brands of detergents, but due to an advertising campaign
for a new detergent X, it is expected that the probability for a housewife to
ask for X is a constant P (0 < P < 1), the demand for the standard brands
remaining equi-probable. Assuming that all housewives coming to the store
buy one and only one brand of the (n + 1) detergents, find the probability that
on any day the first r consecutive housewives will purchase X before a run of s
shoppers buys packets of anyone of the standard brands. Hence determine
the limiting values of this probability for
(i) n large and r, s finite; and
(ii) /I finite and r. s large.
88 An infinite population consists of two kinds of biological particles
which may be likened to white (W) and black (B) beads. In their initial state
the beads are found linked together in chains of four in all the five possible
distinct colour combinations. The probability that a chain has a white bead
is a constant p, and that for a black bead is q (== 1 - pl.
Each chain temporarily splits in the middle to form two doublets, and
these doublets then randomly recombine to form new chains of four beads
each. This dual process of splitting up and recombination continues in-
definitely without any other kind of change in the original population of
beads. Determine the proportions of the three colour combinations of the
doublets (WW), (WB) and (BB) immediately after the first stage of splitting.
If now only proportions xo, 2yo and Zo (xo + 2yo + Zo = 1) of the doublets
are considered further in the process, because of certain accidental destruction,
find X n, 2Yn and Zn' the corresponding proportions after the nth successive
stage of splitting, and hence obtain their limiting values as n -+ 00.
89 In an infinite population of two kinds of biological particles, which may be
symbolised by white (W) and black (B) balls, the balls exist singly and in
pairs of the three colour combinations (WW), (WB) and (BB). Initially. the
single (W) and (B) balls are in the proportions Po and qo (Po + qo = 1), whilst
the paired balls (WW), (WB), (BB) have the proportions roo 2so, to respectively
(/'0+2so+to = 1).
20 EXERCISES IN PROBABILITY AND STATISTICS
All the paired balls undergo a simultaneous disjunction, after which they
are equally likely either to combine randomly with the original single balls
only to form new pairs of the three types, or to remain single. This second
population of paired and single balls repeats the process of separation and
combination to form a third population, and this continues indefinitely.
Find the proportions of the paired and single balls after the nth stage of the
process, and hence deduce their limiting values for n -+ 00.
90 On empirical considerations, the demand for a particular brand of
breakfast cereal at a retail store may be considered to be a random variable X
having a Poisson distribution with mean A. On the first of each month the
retailer arranges to have a stock of n packets of the cereal. Find the expected
number of lost sales per month because of inadequate stock, and hence
deduce that, as a fraction of the average monthly demand, the expected
proportion of lost sales is
7t" = (l-i)[I-F(n-l)]+p(n-l),
where the probability P(X = r) := per) for all r ~ 0, and F(r) is the distribution
function of X.
Also, prove that the expectation of the monthly proportion of lost sales is
n <Xl k!
p" '" I-I k~O;:r<' [1-F(n+k)].
94 An infinite population initially consists of white (W) and black (B) balls
in proportions Po and qo, (Po + qo = 1) respectively. These balls randomly
combine to form the three possible kinds of pairs irrespective of the ordering
of the balls; but immediately on their formation, a proportion A. (0 < A. < 1)
of the (WW) pairs and a proportion Il (0 < Il < 1) of the (BB) pairs are des-
troyed. In the resulting population the pairs then separate, and the individual
balls randomly recombine to form the (WW), (WB), and (BB) pairs. Of these,
a proportion A. of the (WW) pairs and a proportion Il of the (BB) pairs are
again destroyed, and then all the other pairs separate to form a second
population of (W) and (B) balls. This population undergoes the same stages
as the first one, and the process continues indefinitely.
(i) Prove that the proportion of black balls in the successi ve populations
remains unchanged if qo = A./(A.+ Il).
(ii) If A. = 0 and Il = 1, prove that the proportion of black balls in the
successive populations --+ o.
(iii) Finally, suppose A = Il = 0, but that the (WB) pairs are completely
eliminated immediately on formation at each stage. Show that in
this case the limiting value of the proportion of black balls is unity
if qo > 1, zero if qo < 1, and t if qo = 1.
22 EXERCISES IN PROBABILITY AND STATISTICS
v(x) = I: ArVr(X),
r=O
(~) pn-r(2_py+lj2n+l.
and . 1. 1,..1. 2 ,11 are parameters such that ..1. 1 ,..1. 2 > 0 and 11 ~ o. Verify that
G(O) represents a probability-generating function of a discrete random variable
X for integral values of X ~ o. Also, prove that the probabilities P(X = r) = ar
satisfy the recurrence relations
_ Al ~ 1 (HI)
ar+1 - --1· L... -k,.f (O).ar-k for r > 1,
r+ k=O •
t (z+k-1)(r-1) = ~[(n+k)(r)-(k-l)(r)],
z=o r
and. in particular, fOI k = r-l,
L (Z+k_l)(k) = (+k)(k+1)
n
11 .
:=0 k+ 1
Use this result to prove that
n
(i) L X4 = n(I1+1)(2n+l)(3n 2+3n-1)/30;
x=l
n
(ii) L L: x(x+l)y2 = l1(n+2)(n 2 -1)(1011 2 +411-3)/180.
x= 1 ."<.'C
2 Continuous random variables
f f
Iii b
density function of X, and hence show that 1110' the mode of the distribution
is given by ,
mo = -!<n-sin-IO().
Also, prove that m, the median of the distribution, has the value
m= tan- I GIOg2),
and that 111, the mean of X, satisfies the differential equation
d2m 1
-d
0( 2+ 111 =-.
0(
7 A continuous random variable X defined in the range (0 ::;; X ::;; O() has
the distribution function
F(X = x) = 20(x/(a 2 +x 2 ).
Derive the probability density function of X, and hence show that for the
distribution of X
3a
E(X) = O((1-loge 2) '" 10
and
a2
var(X) = a 2[(n-3)-(1-loge 2)2] '" 20'
F(x) = a+c
b
a+ +c
,where F(x):= fx
f(v) dv,
VI
to the firm of V/2. Of those not retained 80 per cent are sold at price V, the
remainder being returned to the factory, where their value is again V/2. The
average cost of production per packet is x + c, where c is a constant.
Show that the value of x which maximizes profit is given by
a+CT[IOge(2~:;2)] t.
15 The cost of operating a hydroelectric plant for a specified period consists
of a fixed charge of £c I independent of the amount of electricity generated
and a cost of £C2 per kilowatt hour for generation. All the power generated
can be sold at a price which is £C3 per kilowatt hour (C2 < C3)' The possible
amount of generation is related to the rainfall in the area, and can be con-
sidered approximately to be a random variable x having the probability
density function (X2 x e - ",x in the range (0 ~ x < (0).
Show that the expected profit in the period is 2(C3 - C2)/(X - CI' Also,
prove that if initially only £C4 are available as reserves with no borrowing
capacity (c 4 < CI), then the probability of going into bankruptcy at the end
of the period is
I in length, where I ~ t, and interpret the result in the special cases (i) I =::: I.
(ii) m = 2; and (iii) m ..... 00. I
Hence evaluate the mean and variance of X, and check these results by direct
calculation of the first two moments.
28 A continuous random variable X defined in the range (- 00 < X < <XlI
has finite moments. By considering the expectation of the quantity
(X 2 +aX +b)2,
where a and b are arbitrary constants, prove that PI and P2, the Pearsonian
coefficients of skewness and kurtosis, satisfy the inequality
P2 ~ PI + I,
where PI == liVJ·d, P2 == li4/1i~, and Ii, is the rth moment of X about the
mean.
29 In a single tossing of x unbiased pennies, find the probability P(x) for
obtaining all heads. If x is regarded as a continuous variable in the range
(0 ~ x < 00), show that P(x), regarded as a function of x and suitably scaled,
determines a probability distribution. Obtain the probability density function
of this distribution and calculate its cumulant-generating function. Hence,
or otherwise, verify that the coefficients of skewness and kurtosis for this
distribution are 2 and 6 respectively.
30 There are n biased pennies such that the probability of obtaining a head
with the rth penny (r = 1,2, ... , n) is
pi +'x, where 0 < p < 1 and 0 ~ x ~ 1.
Find the probability P(x) of obtaining all heads in a single tossing of the n
pennies. Show that for fixed nand p, the function P(x), suitably scaled, deter-
mines a probability distribution of a random variable X defined in the range
(0 ~ X ~ 1).
CONTINUOUS RANDOM VARIABLES 33
lIence obtain the cumulant-generating function of X and verify that
E(X) = p-l_(eP _l)-l and var(X) = P-2- eP(e P_1)-2,
where
P(X ~ 0:)
2 [1 ~3.7.11.15 ... (4r-l)]
= (1+o:4)iB(i,i) 3+ r~i 4r(l+o:4)'(4r+3)r! .
34 From an urn containing N identical balls of which a proportion pare
white and the rest black, a random sample of r ( < N) balls is drawn without
replacement. Find the probability Yx that (x -1) of these balls are white and
the rest black, and hence show that the finite difference i\yx satisfies the
relation
i\yx _ 2[(NP-X+ l)(r-x+ l)-X(NQ-r+X)]
Yx+t - (Np-x+ l)(r-x+ l)+x(Nq-r+x) ,
where q = (l-p) and Yx+t = i(Yx+Yx+d.
If x is regarded as a realization of a continuous random variable X, deduce
that, as a limiting form, x and Y satisfy the Pearson differential equation
1 dy a-x
i dx bo +b i x+b 2 x2'
a, bo, b i , b 2 being functions of N, p and r.
Integrate this differential equation for the particular case of b i = b 2 = 0,
and show that the mean and variance of the resulting probability distribution
of X, defined in the range (- 00 < X < (0), are a and bo respectively.
34 EXERCISES IN PROBABILITY AND STATISTICS
1
.e-x.xY , forO~X<oo,
r(v+1 )
where v is a positive integer, and the discrete random variable Y has a Poisson
distribution with parameter J.l, prove that
P(X ~ J.l) = P(Y~ v).
37 If X is a discrete random variable which has the Bernoulli distribution
such that
'/I '
P(X = r) = (r)pr qn-r, where p+q = 1,
1 . yn-r-l(l_ y)'
B(r+ 1, n-r)
in the range (0 ~ y ~ 1), show that
F(r) = G(q),
where F(x) and G(y) are the distribution functions of X and Y respectively.
38 For a random variable X having a unit normal distribution prove that
1 00 (_1)'x2r+l
P(X ~ x) =!+ ~. (2
V 27t r=O r+
1)2r I'
.r .
L
(x > 0).
where
Verify that this series expansion for the distribution function of X is asymp·
totic by proving that, for any n, the remainder is less in absolute value than the
last term taken into account.
CONTINUOUS RANDOM VARIABLES 35
39 Starting from the standard asymptotic expansion for the distribution
runctio n of a unit normal random variable X, viz.,
1 e- tx2 [ ro (-I)'(2r) ']
P(X::; x) ~ 1- r-c.~-. 1+
y 2n x
2' 2, "
,= 1 • X • r .
L
obtain Schlomilch's more rapidly convergent form of the expansion given by
P(X ::; x) -
1 e- tx2
~ 1- h:.--[I-Al+A2-5A3+9A4-129As+57A6-9141A7+ ...],
y2n x
where, for any positive integer r,
A, == 1/l].1 (x 2 + 2k).
(i) ,~o
m
(-1)' r (Ill) A(m+r+ 1) = t ,~o (-1)' (m)r A(m+r), m
f(X~
1 1
D(n, s) == Jx'!X~
o
d!Xx ,
(iii) f
r=O
(_I),(m) D(2n+l,m+r)
r
= 0; and
43 If X and Yare two random variables with the joint probability density
function f(x, y), for X = x and Y = y, find the probability density function
of Z if
(i) f(x, y) = 4xye-(X2 +)l2), for 0 ~ X, Y < co
and Z = (X2 + y2)t.
(ii) f(x, y) = 1, for 0 ~ X, Y ~ 1
and Z = X+ Y ifX+Y<1
=X+Y-l ifX+Y>1.
(iii) f(x, y) = ala2 e-(OIX+02)1), for 0 ~ X, Y < co
and Z = X+ Y.
44 Each of two independent events A and B can only occur once in the future.
The probabilities of their happening in the time-interval (t, t+dt) are propor-
tional to t e-ext . dt and t 2 e- fJt . dt respectively for all t in the range (0 < t < oo~
where oc and p are positive constants. Prove that the probability that the two
events occur in the order BA is P3(4oc + P}/(oc + P)4.
45 If the independent random variables X and Y have the probability density
functions f(x) and g(y) defined respectively in the intervals (0 ~ X ~ oc) and
(0 ~ Y ~ P), (P > oc), find the probability distribution of the random variable
Z = X + Y, indicating the discontinuities in this distribution.
Hence obtain the distribution of Z in the particular case when both X
and Yare uniformly distributed, and determine its limiting form as oc -. p.
46 For the independent random variables X and Y having the probability
density functionsf(x) and g(y), both defined in the unit interval (0 ~ X, Y~ n
find the probability distribution of Z = X + Yin the following cases:
(i) f(x) = 1 and g(y) = 3(1- 2y)2 ;
(ii) f(x)
2yx
= 1r: and g(y) =
2 l-y
J-=y ;
(iii) f(x) = 3(1-2x)2 and g(y) = 3(1-2y)2.
Also, find the distribution of W = X 2 + y2 when both X and Y have a
uniform distribution in the unit interval.
47 The two random variables X and Y have, for X = x and Y = y, the
joint probability density function
1
f(x, y) = - 22' for 1 ~ X < co; llX ~ Y ~ X.
xy
Derive the marginal distributions of X and Y. Further, obtain the con-
ditional distribution of Y for X = x and also that of X given Y = y.
CONTINUOUS RANDOM VARIABLES 37
48 The continuous random variables X and Y have a joint probability
density function proportional to
y"(x- y)(l-x)P, for 0 ~ X ~ 1;0 ~ Y ~ X,
h parameter a being > - 1 and {3 a positive integer.
t e Find the proportionality factor, and hence determine the probability
distribution of the statistic u = y/Jx. Also, for any given uo, (0 < U o < 1),
shoW that
2u"0 +1 x
) _
P(u ~ Uo - B(a+3,{3+1)
Find the limiting value of Pn as n -+ 00, and c and T are finite, and explain its
significance.
S9 In a factory with It lamps lighted at the same time daily, a bulb is replaced
immediately it burns out, the cost of replacing it being a constant y. An
alternative method is to change all I! bulbs at the same time periodically
with time-period T, but still change a bulb when it burns out. When all the
I! bulbs are changed together, the total cost is a + {JI!, where a and {J are con-
stants such that a + {JI! < I'll.
If the lifetime T of a bulb is a random variable having a uniform distribu-
tion in the interval (0 ~ T ~ k), show that the expected number of burnouts
per socket is
(T < k).
Hence deduce that the equation for determining the optimum value of T
which makes the average cost per unit of time the bulb is lit a minimum is
eX (I-x) = 1-0,
f
0000
1(It,q)",~.e-th2(1+W2).
2n n=1f: 1t-2n(~.~)n-l[
dw W
1
w(1+w2)
],
f f exp[-1<A¢2+B'12-2C~'1-2a~-2b'l)}d~d'l
OX) ifJ
2~
-:;.('I -rJ)
J
00
e-YXl dx
P(Y ~ y) = l-(ex-l) I •
( +x)"
o
and verify that all the moments of Y diverge. Use these results to prove that
X and Yare uncorrelated but not independently distributed random variables.
65 The joint distribution of the random variables X and Y is defined by a
probability density function proportional to
yfl(1- xl". for 0 ~ X ~ 1; 0 ~ Y ~ X.
the parameters ex and P being each > - l. Find the marginal distributions
of X and Y. and evaluate their means and variances.
Further, determine the conditional distribution of X. given Y = y. and
that of Y, given X = x. Hence verify that
_ (ex + 1)(1- y)2 .
E(X!Y = y) = 1 +~:~ l)y; v
ar(X!
Y_
- y) - (ex+3)(ex+2)2'
where p == Plex.
Use this distribution to show that the rth moment about the origin of
w = zip is
I mil
Jl. = .
, (n + r)(m - r)'
Also, determine the probabilities P(z ~ Po > p) and P(z ~ P1 < p) for fixed
Po and Pl'
42 EXERCISES IN PROBABILITY AND STATISTICS
Use this result to show that for m =F n the distribution of v may be approxi-
mated as fJx 2, where X2 has v dJ., the constants fJ and v being evaluated as
(1+).2) 2(1+).)2
fJ = ),(1+),) and v = (1+).2)'
68 If W1 and W2 are the sdmple ranges obtained from independent samples
of size "l
and n2 respectively from the (0, a) rectangular distribution, derive
the sampling distribution of the ratio
u= WdW2'
70 If Xl' X2"'" X" are the maximum values obtained from independent
samples of equal size n from a rectangular population in the range (0, 1~
find the exact distribution of the product v" = Xl X2 ••• X". Hence, or otherwise,
prove' that - 2n fog v" is distributed as a X2 with 2k degrees of freedom.
CONTINUOUS RANDOM VARIABLES 43
Also, if the Xj are obtained from large samples of size nj (i = 1, 2, ... , k)
ectively, then show that a convenient large-sample approximation for the
re sp f.
distribution 0 Vk IS
7
-2n [ 1- 2V(n.)] logvk ,..., X2 with 2k [ 1- V(n.)]
n2' dJ.,
k
where n is the average of the nj and kV(nj) == L (nj-n)2.
j= I
where, by definition, v_ 1 = O.
From this general differential equation deduce that
(i) if!(t, x) = kcp(t), k a constant, then v, = N . (r+ l)th term of a Poisson
distribution with parameter kT; and
(ii) if !(t, x) = (b + cx)cp(t), band c being positive constants, then
Vx = N . (r+ 1)th term of a negative binomial distribution with prob·
ability-generating function
G(l}) = (l-wt'·/(1-wOt'·,
where
f
I
B considering the joint distribution of the random variables x and A., prove
t:at the marginal distribution of x has the probability-generating function
G(O) = e'/(1 +e-O)'.
Hence, or otherwise, deduce that
corr(x, A.) = (1 +e)-t.
G(T) == L F m(T),
m=l
where F m(T) is the probability that at least m lamps fail at a post in (0, T).
Hence calculate C 2, the expected cost of maintenance in (0, T) and verify
that
pbeing the ratio of the average total costs of regular and irregular replace-
ment of a lamp.
As particular eases of this general result, show that
(a) iff(t) = A. e- At , then under all circumstances Plan I is more economical
than Plan II; and
(b) if f(t) = ..1.2 e- A1 • t, then Plan II is more economical than Plan I,
provided that
°< p < i(1-e- 2AT ).
77 Electric bulbs used for street lighting have a lifetime distribution with
probability density functionf(t) for (0 ~ t ~ To), and zero for t > To. When-
ever a bulb fails in a lamp post it is immediately replaced by a new bulb, and
this method of replacement is continued indefinitely. Starting with all new
bulbs at time t = 0, the process is considered up to a stage t = T, where, in
general, (r-l)To < T < rTo for all integral values of r ;?; 1, and F~) (T) is the
probability of making at least m replacements in a given lamp post in the
interval (0, T).
CONTINUOUS RANDOM VARIABLES 47
Prove that if r = I, then F~:' (T) satisfies the integral relation
fF:':~
T
G,(T) == L F~'(T).
m=l
where p is the proportion of bulbs not replaced at time t = T and Sl(t) is the
probability that a bulb has a lifetime > t.
In the particular case whenf(t) = A. e- A1, verify that
S:z(1") = !e-·I.t(1+e-J.X).
79 In a large city a number of street lighting posts are supplied with electric
bulbs having a lifetime distributionf(t) dt for (0 ~ t < (0). Initially, at time
t = 0, all posts carry new bulbs, and in course of time whenever a bulb burns
48 EXERCISES IN PROBABILITY AND STATISTICS
out it is immediately replaced by a new bulb. In addition, all posts are in-
spected at regular intervals of time T, so that at time t = liT, (n ~ 1),
(i) all bulbs which were replaced in the interval (nT-X, nT), (X < T)
and have not burned out by t = nT continue to be used; and '
(ii) all bulbs which were replaced during the interval [(n -1)T, (nT-X)],
and have not burned out by t = nTare replaced by new bulbs.
Suppose that after the above replacements have been made, Pn is the
proportion of bulbs not replaced at t = nT, and that of these a proportion
pi have their last replacement before t = (n + 1)T in the interval [(n + 1)T- X,
(n + 1)T]. Similarly, of the proportion (1- Pn) of bulbs actually replaced at
t = nT, a proportion p have their last replacement before t = (n + 1)T in the
interval [(n + 1)T- X, (11 + 1)T]. Prove that Pn satisfies the difference equation
Pn+ I = PnP' +(1- Pn)P, for 11 ~ 0,
and, assuming that both p and pi are independent of n, find an explicit ex-
pression for Pn. Also, verify that if the difference between p and pi is small then,
as a first approximation, P2 '" Pro so that the proportion of replacements at
t = nT is effectively stabilized soon after the second T interval.
Hence, if g(t) dt, (0 < t < T), is the probability of a replacement being
made at a given post in (t, t + dt) irrespective of when the bulb was put in,
and S(t) is the probability of a bulb having a lifetime> t, show that
x
p = f o
g(T-x.)S(x.)dxI'
and
xx
pp' = f f g(T-x )g(T+X I -X2)S(xdS(x 2) dX
l I dX2·
o 0
Next, suppose that a random sample of size m is taken from another popula-
tion, also with a doubly infinite range, and that these m observations are
distributed over the (n+ 1) intervals formed by the Uj so that m l observations
are <U .. mj lie between (Ui,U i + l ) for (i = 1,2, ... ,n-1), and mn + 1 are >un•
Find the joint distribution of the mi and Ph and then prove that every random
set of realizations of the mi has, for given m and n, the probability 1/(m~").
81 A random variable X has the Laplace distribution with probability
density function proportional to e-<xlx-ml for (- 00 < X < (0), m being a
positive parameter. If c is any given positive number ~m, prove that
1
E{lx -en = -. [e-<x(m-C)+ex(m_e)].
ex
CONTINUOUS RANDOM VARIABLES 49
Hence deduce that the minimum value of the mean deviation of X is 1/1X, which
is attained for c = m.
82 Explain clearly the difference between the standard deviation and mean
deviation of a continuous random variable X. In particular, if X has a negative
xponential distribution with probability density function proportional to
: -«" for X ~ 0, and a and b are positive constants, prove that
E(X _a)2 = 1 (1IX )2
-+
1X2
--a ,
and
E{IX-bl} = b+(2e-«b-l)/IX.
Hence deduce that for this distribution the minimum value of the mean
deviation is (loge 2) X standard deviation.
83 A dam has a total capacity of M units of water for irrigation, but if there
is no rainfall during a month then the dam can only supply IXM units of water,
(0 < IX < 1). The amount x of rainfall in a month is a random variable, and
as x increases, the supply of water in the dam also increases; and since the
capacity of the dam is limited it may be assumed on empirical considerations
that this increase dies off exponentially. As a simple model, it may therefore
be assumed that if the amount of rainfall in a month is x then the total supply
or water in the dam is
prove that
E(S) = M[I-(I-IX)/(l +2{J)t].
Hence, provided P is sufficiently small for p2 to be negligible, show that the
actual supply S will be at least equal to its expectation if x ~ (1- P/2). Also,
prove that the probability for this to happen is approximately
2[1-<I>(1-P/2)],
where <I>(t) is the distribution function of a unit normal variable.
84 A random chord of length Y is drawn across a circle of radius p such that
its perpendicular distance from the centre of the circle is px, where x is assumed
to be a uniformly distributed random variable in the interval (0, If Yh Y2' n
...• Yn are It independent realizations of Y, prove that for their mean y
pn (32 - 3n 2)p2 p2
E(y) =2 ,. ., 1·57 p and var(y) = 12n ,..., 5n'
where
A(n)
_
= n
n-I (-1)'
r~o 2'(r+2)· r(n-r)
r(n+l) I{ r (r+2)}l
2 .
Hence. or otherwise. verify that for n = 5
e.
(iii) as defined in (ii). has a probability density function proportional to
e(n - e) in (0 ~ e ~ n/2).
For any given k. (0 ~ k ~ 2). suppose Ph P2 and P3 denote the probabilities
P(l ~ kp) in the three cases. Prove that PI and P3 are both ~ P2. Comment
on the results obtained.
86 If XI and X2 are independent observations from a rectangular distribution
in the (0. 1) interval. find the joint distribution of the statistics
/I = XtX2 and v = (l-x t )(l-x 2).
Hence determine the conditional distribution of II for given v. and the marginal
distribution of v. Also. find directly the marginal distributions of II and v.
n - 1)
S
2= 2[ (-11-
S + (11+11)" (k--;X )2] .
88 The tails of a normal distribution with mean III and variance (12 are
cut off at a distance ±k(1 (k > 0) from the mean. If (1~ is the variance of the
truncated distribution. show that
Also. evaluate the P2 coefficient of the truncated distribution and prove that a
sufficient condition for it to be < 3 is that k is >.)3.
3 Estimation, sampling distributions, and
inference; bivariate correlation and
regression
,1,* = L: r. nr/N
r=2
is an unbiased estimate of A.
3 In a sequence of 11 Bernoulli trials with a probability of success p, (q == 1- p),
/' successes were observed.
(i) Obtain an unbiased estimate of pq, and find its variance.
(ii) Find an unbiased estimate of pq2.
(iii) Show that p*(1- p*j2, where p* = 1'/11, is not an unbiased estimate of
pq2, but that the bias --+ 0 as 11 --+ 00.
4 In tossing a coin with a probability p for a head, a sequence ofr consecutive
heads followed by a tail is known as a run of length r for (r = 0, 1,2, ... ,11),
where (0 < p < 1). Calculate the mean and variance of the run length r.
Also show that when 11 --+ 00,
p* = -Sn- .
n+Sn
6 In a sequence of n Bernoulli trials with probability of success p, (q == 1- p),
r successes were observed. It is desired to estimate the ratio p/q, and two
estimates are proposed:
r r
and
Il-r n-r+l
Show that
E[_r] = !!.
n-r q
f Il,/q'] '
[1 +~P,=2
var
n
r]
[---=-
r
1[
= 1:
q ,=2
(t -1)
L<Xl -,-q
L Il,/q' )2] ;
.Il, - ( ,=2 00
and
r
var [ ll-r+1
] (p+A.)2[
= -A.-
<Xl (t-1)
'~2-A.-,-·Il'- '~2Il'/A.'
(<Xl )2] ,
where Il, is the tth central moment of p* == r/n, (0 < p < 1), and A. == q+ l/n.
7 From a large lake containing an unknown number N of fish, a random
sample of M fish is taken. The fish caught are marked with red spots and
released into the lake. After some time, another random sample of n fish is
drawn and it is observed that m of them are spotted. Show that P(N, m),
the probability that the second sample contains exactly m spotted fish, is
given by
By considering the ratio P(N, m)/P(N -1, m) deduce that the maximum-
likelihood estimate of N is the largest integer short of nM/m.
[It may be assumed that there was no change in the fish popUlation of
the lake in the interval between the drawing of the two samples.]
8 For the logarithmico-normal distribution defined by the probability
density function
f(X = x) = ~ exp {-2~2(10g.,X_I1l)2}
uj2ic v
ESTIMATION, SAMPLING DISTRIBUTIONS, INFERENCE, ETC. 53
. the range (0 ~ X < co), show that the maximum-likelihood estimates of
In d (12 are
/II an 1 n
m* = g and (12' = - L (log. Xj- g)2,
nj=I
where g is. the natural logarithm of the geometric mean of the random sample
observations x l' X2" •• , Xn •
9 A continuous random variable X defined in the range (0 ~ X < co) has,
for X == x, the probability density function proportional to x e- x /6, (lJ > 0).
Find the mean and variance of X.
If a random sample of n observations x .. X2"" , Xn is given from this.
population, obtain the maximum-likelihood estimate of the parameter lJ and
calculate the variance of the estimate.
Show also that
1 n
-'L
3n
xf
j= 1
is an unbiased estimate of var(X), but that the estimate ti2 has a bias of
0(11- 1), where i is the sample average.
10 A random sample of n observations Xl' X2" .. , Xn is given of a random
variable X having, for X = x, a probability density function proportional to
x'(I-x) defined in the range (0 ~ X ~ 1), where the parameter a is unknown.
Show that a*, the appropriate maximum-likelihood estimate of a, is given
by the equation
* _ (3g+2)+(g2+4)t
a - - 2g '
where g is the natural logarithm of the geometric mean of the sample.
11 A random variable X has the probability density function
-1
f(X = x) = _x__ {I
. exp - -2 (log. X - p)2, } for X ~ o.
(1Ji1c 2(1
From this population, a random sample of size 11 has mean i and variance
S2.Use the method of moments to find estimates of p and (12.
Also show that the mean of the distribution is greater than the median.
12 A continuous random variable X, defined in the range (0 ~ X ~ 1[/2),
has, for X = x, a distribution function proportional to (l_e-,.. inX), where
IX> O. Find the probability density function of X.
Given a random sample of n observations from this distribution, derive
the maximum-likelihood equation for a*, the estimate of a. Also, prove that
the large-sample variance of a* is
4a 2 sinh 2 a/2
11(4 sinh 2 a/2 _( 2 )'
13 If g(X) is a function of a random variable X having mean m and a finite
variance, prove that to a first approximation
=
E[g(X)] g(m),
and var[g(X)] = [g'(mW. var(X).
54 EXERC'lSES IN PROBABILITY AND STATISTICS
Hence prove that a sufficient condition for E(X) > mis 0'2 ~ 0'1'
Given a random sample of N observations of X, where N is sufficiently
large for assuming that the variance of the sample median x is 1/[4N(medial
ordinate)2], prove that the condition for the sample mean x and x to be of
56 EXERCISES IN PROBABILITY AND STATISTICS
equal precision is
1- 2 {_I -_ 2 2}]2 _~. (1+),)4 p 2
[pexp(-2"m/O'd+..1.exp !(m J.l) /0'2 - 2 (l+..1.)(1+Ap2)+i,jl2/O'f
where p == 0'2/0'1'
If J.l = 0, show that this condition reduces to a quartic in p which has two
positive roots PI < 1 < P2 such that if P lies outside the interval (Pi> Pl)
then
var(x) > var(x).
20 A random variable X is uniformly distributed over the interval (O'~l
where 0( is an unknown parameter. A sample of 11 independent values of i
is given, and these observations, arranged in order of increasing magnitude
are denoted by XI' X2,' •. ,Xn • Derive the sampling distribution of x" th;
rth-order statistic of the sample, and verify that
rO( r(lI - r + 1)0(2
E(x r) = (11+1) and var(x r ) = (11+1)2(11+2)'
is an unbiased estimate of p.
Also, verify that the mode of the sampling distribution of p* lies in the
range
var(p*) = p2 f if ~(r+S-l),
s= 1 / I S
and that
p*2(1_ p*)/(r-l- p*)
is an unbiased estimate of this variance.
24 The random variables X and Y have a bivariate normal distribution
with parameters (ml, m2; 0"1' 0"2; p) in standard notation. Prove that for
positive integral values of rand s
where Vk is the kth central moment of the univariate unit normal distribution.
Given a sample of n observations (Xi> Yi) for (i = 1,2, ... , n) from this
bivariate population, two possible estimates
1 (XI)
Ii = ni~1 i~1 X~~1 Yi
n n n
Yi and T2 =
are proposed for the parameter A. == mtlm2. Find the expectation of Ii and
12· Hence show that only 12 asymptotically converges to A. as 11 -+ 00, and
that for large samples
2
var(12) ,.., -A. [ (VI-PV2) 2 +(1-p 2 )V22] ,
n
VI and V2 being the coefficients of variation of X and Y respectively.
58 EXERCISES IN PROBABILITY AND STATISTICS
1
f( X = x) = -_. e- x /a . ~
ar(p)
(
a'
)P-I for °~ X <
""
00
.
Hence deduce that the ratio of the population arithmetic and geometric
means is
Also, show that (J*, the maximum-likelihood estimate of (J, is the ratio of the
sample arithmetic and geometric means.
ESTIMATION. SAMPLING DISTRIBUTIONS, INFERENCE, ETC. 59
If the parameter a is known and only p is estimated, obtain the large-
.ample variance of p, the estimate of p, and thereby prove that the large-
:~mple variance of &, the estimate of (J in this case, is
(J2[p - I - t/>'(P)J2 /Ilt/>'(P).
28 An experiment results in six independent observations Yr (r = 1,2, ... , 6)
such that
E(Yr) =- ex cos (2nr) . (2nl')
6"" +P SID 6"" ; var(Yr) = (12.
Find the least-squares estimates of ex and p, and verify that each of these
estimates has variance (12/3.
29 A chord I is drawn at random across a circle of radius p, such that it
passes through a given point on the circle and makes an angle (J with the
tangent to the circle at the given point. Find the expectation and the variance
of /, the length of the chord.
Suppose p is not known, and it is desired to obtain its estimate from
" independent measurements I" 12 , • •• ,In of chords as drawn above. Show
how these measurements can be used to obtain an unbiased estimate of p,
and also find the variance of this estimate.
30 A target P moves randomly on the arc of a quadrant of a circle of radius r
and centre 0, such that OP makes an angle (J with the horizontal, where (J is
known to have a probability density function proportional to (J(n - (J) in
the range (0 ~ (J ~ n/2). If PM denotes the perpendicular on the horizontal
lhrough 0, show that .::\, the area of the triangle OPM, has the expected
value 3r2(n 2+ 4)/8n 3 •
If /' is unknown, and a random sample .::\1' .::\2, .•. ,.::\n of 11 values of .::\ is
given, obtain an unbiased estimate of r, and hence prove that the estimated
area of the quadrant is
2n4.::\0
3(n 2 +4)'
where .::\0 is the mean of .::\" .::\2,· .. ,.::\n.
31 The 311 independent observations x" X2'" . , Xn; YI' Y2'" . ,Yn; and
ZI' Z2"'" Zn' each have the same unknown variance (12, and
E(xj) = ml ; E(Yj) = m2 ; E(zj) = ml +m 2, for i = 1,2, ... ,11.
Use the method of least squares to obtain the unbiased estimates of ml and m2'
and hence derive the best estimate of (12 based on the total available degrees
of freedom. Also, show that the mean square for testing the hypothesis
H (1111 = 1112) is
11 2
i(X- ji) ,
[(2-A)X+(1-2A)y+(1 +A)Z]jn
t = S[6(A2-A+ 1)]"1- '
where t has the t distribution with (3n - 2) dJ., S2 is the least-squares estimate
of (12, and X, y, z are the means of the x, y and z observations respectively.
33 The 3n independent observations X"X2""'Xn ; Y"Y2,""Y'
... , Zn, have the same unknown variance (12, and their expectatio~;
Z h Z 2,
depend upon three independent parameters Oh O2 , and 03 in such a way thaI
E(Xi) = 0, +0 2 +0 3 ,
E(Yi) = -0, +0 2 +03 ,
and
E(Zi) = -20 2+0 3 , for i = 1,2, ... , n.
Use the method of least squares to obtain the unbiased estimates of lJ h 02
and 03 , and also the best estimate of (12 on the available degrees of freedom.
Also, show that the mean square for testing the hypothesis H (0, = O2 = 0l)
can be put in the form
7 There are 3n independent observations XI' X2" .. ,X n ; YI' Y2" .. ,Yn; and
J z each observation having the same unknown variance ci 2 • The
; I' ;2.1' 'v';I~e~' of the observations are given by:
mea
E(Xi) = 0 1 +20 2+30 3,
E(yJ = 20 1 +30 2 +03,
and
0* = (}'IX+A2Y+';'3 Z).
(Af + A~ +A~) ,
and
Al == (l+2a+3b); A2 == (2+3a+b); A3 == (3+a+2b).
38 A sample of n independent observations y" Y2" .. ,Yn is given of a normally
distributed random variable Y such that
E(Yr) = lX+p.r(xr-x) and var(Yr) = (f2
ror (r = 1,2, ... , n), where IX and p are unknown parameters, Xr are values of a
non-random variable, and (f2 is an independent unknown parameter.
Obtain the least-squares estimates of IX and p, and hence derive the best
estimate of (f2. Also, calculate the variance of the estimate of p and deduce
the appropriate test statistic for testing the hypothesis H (P = Po).
39 Given a random sample of n pairs of observations (xj, Yi), (i = 1,2, ... , n),
show how the method of least squares may be modified for fitting a straight
line of the type
IXx+Py+l = 0
by minimizing the sum of the squares of the perpendicular distances of
(x;. Yi) from the straight line. Hence show that the estimates of IX and pare
/I and b respectively, where
b = I/(mx - y),
ESTIMATION. SAMPLING DISTRIBUTIONS, INFERENCE. ETC. 63
111* is the best estimate of m; and
where
(I +1l2-l)S2
I)
=(A.Il I +n 2)[A. i
r;1
(xr-x?+ I
r;1
(Yr-y)2] +;,1I11I2(X-y)2,
£(12) = C. ~ 2)-
The random variables X and Y have a bivariate normal distribution with
ncan s III , Ill., variances 0';, 0'; and correlation p. Given a random sample of
" paired ob~ervations (Xi' Yi), for (i = 1,2, ... , IIJl, and a further independent
~l~J11ple of 112 observa~io~s on X only, Y not being recorded, an estimate of
/II" is given by the statistic
T= YI +hdx-xd,
where XI' YI and hi are the sample means and the regression coefficient of
Yon X calculated from the first sample, and x is the mean of the X observa-
tions in both samples.
Show that T is an unbiased estimate of m)' and that
S2
11 ~ 1: - N, and n ~ L
v
If another independent sample of n observations is taken from the same
population, show that by using the two sample means another unbiased
estimate of m can be derived whose estimated variance is for large N asymp-
totically equal to v2 •
45 If X2 has the X2 distribution with II dJ., show that for large II
defined as
n-1
d2 = I (Xi+1- X j)2/2(1I-1).
i= 1
where the a's and b's are constants and the Xj are random variables such that
E(xj) = IIlj, var(xj) = q2, and cov(x j , Xj) = q2Pii' (i"# J),
find the variances of L1 and L2 and also their covariance.
Further, calculate the variance of the function (L 1 - L 2 ), and then show
that if all Pii = p, this variance reduces to
where ~, == (a, - b,) "# 0 for all r, and hence deduce that
n
L ~;
--,--'=-,,1__ ::::.; p ::::.; 1.
LL 0(,0(.
where the a's and b's are constants not all zero, are uncorrelated if
n
I ajb j = O.
j= 1
ESTIMATION, SAMPLING DISTRIBUTIONS, INFERENCE, ETC. 65
Hence, or otherwise, show that the arithmetic mean x is uncorrelated with
any deviation (Xj - x) from the mean.
Also. prove that
var(xj-x) = 0'2(1-1/11),
and
cov[(Xj-X), (xj-x)] = -a 2 /n, for i :F j.
50 If Xl> X2" •• ,Xn are random variables such that
E(xj) = mjo var(xj) = 0'2, cov(Xj, Xj) = pa 2 , for i :F j = 1,2, ... , n,
obtain the variance of the linear function
n
L = L ajX" where the aj are constants not all zero.
j= 1
If the Xj are divided into two distinct groups of V1 and V2 elements
(v. +V2 = n) to define two new random variables
VI n
SI = L
j=1
Xj, and S2 = LXi>
j=vl+1
prove that
corr(S1,S2) = P[{I+(Vl-l);}~~+(V2-1)p}r.
Also, when n -+ 00 such that V 1 -+ 00 but V2 remains finite, find the limiting
value of this correlation.
51 Of two discrete random variables, X can take only the values ±a and Y
the values ±{J, with unequal probabilities for the four possible realizations
of the pairs (X, Y). In a random sample of N observations, the frequencies
corresponding to (-a, -fJ), (a, -{J), (-a,{J) and (a,{J) were found to be
"., 112' n3 and 114 respectively, (Ll1j = N). Prove that the sample product-
moment correlation between X and Y is
P'3-A1A2)/[(1-AD(l-A~)]t,
where . 1.1
and . 1.2
are orthogonal linear contrasts of the observed relative
frequencies associated with the sample means of and Y, and is the contrast X A3
orthogonal to both and A1 . 1.2,
Hence show that the sample correlation vanishes
when
P(X = r, Y= s) = (:1)(:2)(N~~~=:2)/(Z),
and indicate the limits of variation of rand s.
66 EXERCISES IN PROBABILITY AND STATISTICS
where
I
AIso,
by using the r function approximation
r[(v + 1)/2)] ( I)
(v/2)tr(v/2) '" 1- 4v for large v,
[ 4{n+l)p(I-P)
(1-2p)2 +
I] .
68 EXERCISES IN PROBABILITY AND STATISTICS
[4>(a)]2
o0
" a
f
= 21nf e- t (X 2+ y2 )dydx, (a> 0),
-I-
fo
f a
e-tx2
X" d ' " 21 (I -e - 2a 2'1t)t .
o
Hence show that the sample median obtained from (2v + I) independelll
observations of a unit normal variable is approximately normally distributed
with zero mean and variance nl(n+4v). Verify that the asymptotic efficiency
of the median as compared with the sample mean is 21n.
60 Given a random sample of (2v + 1) observations from a unit normal
population, prove that the probability distribution of the sample median
x is proportional to
[F(x) {1- F(x)} ]" dF(x),
y = (4V;nr. x
61 If x and yare independent random variables such that x has a unit normal
distribution, and y is distributed as ·c
with n dJ., prove that the ratio
t=--
xjn
.JY
has Student's t distribution with II dJ.
ESTIMATION. SAMPLING DISTRIBUTIONS. INFERENCE. ETC. 69
Assuming that for sufficiently large n
r{(n+ O/2} ( 1)
(1I/2)tnn/2) - 1- 411 •
derive Bartlett's approximation that (n -t)loge [1 +(t2 /n)] is distributed as
-I with 1 d.f.
62 If PI' P2'· .. ,Pk are probabi~ities derived from the r~alizations :X.I' X2.··· ,.Xk
of the independent random vanables X 1> X 2.· ..• X k With probabilIty denSity
functionsfI(xl),f2(X2) •... •f,.(Xk) such that
f
X;
in the interval Y ~ JI. and zero otherwise. Obtain the joint distribution of the
largest and smallest sample observations, and hence derive the distribution of
the sample range w. Verify that for any given fixed value \1'0 the probability
P(w ~ wo) = 1_(1_e- wo / 9),,-I.
Explain how this result can be used to test a specific hypothesis about the
parameter O.
64 A random variable X has, for X = x. the probability density function
f'(x) defined in the interval (ex ~ X ~ {3). If a sample of II independent observa-
'Iions from this distribution is given. and the (algebraically) smallest and largest
of these are denoted by x I and x" respectively, derive the joint distribution of
x, and x".
If the sample range is defined by R = XII - X I. show that the marginal
distribution of R is .
g(R)dR = n(n-1) f
/l-R
2:
!(X1)!(XJ+R)[
x, +R
f
XI
!(X) dX r- 2
dxl·dR.
where 0 ~ R ~ {3-ex.
Hence derive the distribution of R in the particular case when X is
uniformly distributed in the interval (0 ~ X ~ 1).
65 A random sample of (2v+ 2) observations of a uniformly distributed
random variable X in the range (0 ~ X ~ 2a) is ordered. the (v + 1)th and
(1'+2)th observations being XI and X 2 • (x 2 > XI).
Obtain the joint probability distribution of
XI +X2 d
Y=--- an
2
70 EXERCISES IN PROBABILITY AND STATISTICS
Hence prove that for the marginal distribution of y the probability density
function is
(v+l) y (V)
a22Y B(v+ 1, v+ 1)' r~o r (-1)
y-r( l- y)2Y-2r [1 +_ (l- y)2r+l]j (2r+ 1),
ti ti
according as the range of y is (0 ::s;;; y ::s;;; a) or (a ::s;;; y ::s;;; 2a).
Also, verify by actual integration that
P(O ::s;;; y ::s;;; a) =t.
66 Starting from the joint distribution of x and
s, the sample mean and
standard deviation based on a random sample of n observations from a
normal population with mean m and variance (l2, prove that the sampling
distribution of v, the square of the sample coefficient of variation, defined by
v = S2/X 2,
is
n+2i)[ n ] (2J+3l12
e-(n/2Jl) co (2n/A.2)i r ( -2- ~ (n-l)
(n-l\J~or(2i+l)'
~ r -2-'
[ n ](n+2J)/2
1+ (n-l)v
• -n- dv
and
(l2
var(s) '" 2v' (v == n-l).
v= Cy.t(t2~2Vr,
t being the standard Student's t statistic with v d.f., has asymptotically the
unit normal distribution. Also, determine the central moments of v and so
prove that the approximation may be regarded as effectively correct to
O(v- 1 ).
68 Assuming Stirling's asymptotic formula for the r function,
(a.n ,
~
.± p
)= I
j ).
f(X = x, Y= y) = R
2nO' I 0' 2 1- p
.exp[-2(1~ P2){x:-2P~+Y:}],
0' 0' 0'2 0' 2
1 1
f(X = x, Y = y) = 1
1
. exp [ 2(I_ p 2)(x -2pxy+ y
2 21)J
2nJt- p 2
for - 00 < X, Y < 00,
verify that
of 02f
op ex cy'
Further, if two new random variables eand 11 are defined by the relations
x ,
p = 2 sin(n:).
74 A random sample of n observations (Xl. yd. (X2. Y2).' ..• (x", Yn) is given
of random variables X and Y having a bivariate normal distribution with
the parameters (mI. m2. 0' 1.0'2. p) in standard notation. Prove that
Q2 = ~
(1 p )i=l
f [(X i-m
0'1
1) 2 _2 P(X i -m 1) (Yi- m2\ + (Yi- m2)21
0'1 0'2 J 0'2
is distributed as X2 with 2n dJ. Hence infer the distribution of
R2 = (1(n~l~p) [(~)2
0'1
_2P(~)(~)+(~)2].
0'1 0'2 0'2
sr. s~ and rs s l 2 being the second-order sample moments.
ESTIMA TION, SAMPLING DISTRIBUTIONS, INFERENCE, ETC. 73
If the variances of the random variables X and Y having a joint bivariate
7S al distribution are known to be equal, show that the appropriate estimate
norm . I t' .
of the population corre a Ion p IS
2rs"s).
u-
- S2+S2'
" y
where s~, s; ~nd rs"sy are the usual second-order moments based on a sample
f n observatIOns.
o Starting from the known sampling distribution of these moments, derive
the joint distribution of u and the random variables
2sxSy
v = -2--2' W = (2
s,,+Sy2)/(I 2.
S" + s).
Hence prove that the sampling distribution of u is
1 2)(n-l)/2
( - p . (1- pu)-(n-l) . (1- u2)(n-3)/2 . du, (-1 ~ u ~ 1),
B(n;l,t)
and thereby deduce that, in this case, the appropriate Student's t statistic
for testing the hypothesis of zero correlation is
(1_p2)f I
J l-exp
+
P
[ { u}] du
p= 1t' 2(12(1-p2) 'u{p2-(1-u)2}f'
I-p
z= (X:~I)/(Y:~2),
and hence prove that the mode of the distribution of z is p.
78 If x and yare the realizations of the two independent normally dis-
tributed random variables X and Y with zero means and variances (I~ and
74 EXERCISES IN PROBABILITY AND STATISTICS
w = (1-p)s;/(1+p)s~
and
n n
(n-l)s; == L (Ui- U)2; (n-1)s~ == L (Vi-ti)2;
i= 1 i= 1
n n
_
21tlTI(12
R'(~)
1_p2 n-2
[1+ 1 2 {X:_2P~+y:}]-n,
2(n-2)(1- p) (11 (11(12 (12
where 11 == 3(P~ - ~)/(~2 - 3), P2 being the common value of the kurtosis of
the marginal dIstrIbutIOns of X and Y.
What happens when P2 -+ 3?
82 The Pearson bivariate distribution of the random variables X and Y is
defined by the probability density function
f'(X == x, Y = y) =
- 1 (n-1)
- [ 1+ 1 {X2
- - 2 pxy- +y2}]-n
-
- 27C(11(12J1- p2' n - 2 2(n - 2)(1- p2) (1f (11 (12 (1~
for (- 00 < X, Y < 00). Obtain the marginal distribution of X, and hence
show that the conditional variance of Y given X = x is
2 2 [
(12(1-p) 1
1-(2n-3)' {X2}]
1- (1f
n > 0,
in which (1f and (1~ are the variances of X and Y, and the b's are constants.
Show that in this case the equi-probability contours must be ellipses, and
obtain the marginal distributions of X and Y, indicating the permissible
range of variation of the random variables. Hence prove that the joint
probability density function has the form
76 EXERCISES IN PROBABILITY AND STATISTICS
x (::~) [1 n > 0,
obtain the marginal distribution of X, and hence show that a;.x, the con.
ditional variance of Y, given X = x, lies on the ellipse
2 2
a},.x x_I
2(1 2 (2n+4) + (2n+4)a~ - .
a2 -p) 2n+3
Further, prove that the probability is half that a random observation falls
outside the equi-probability contour
where ), is the observed relative frequency with which the first penny turned
up tails in N independent tossings of the two pennies. Verify that
var({)*) = Pl(l;;Pd.
On the other hand, if nothing is known about the second penny, derive an
unbiased linear estimate of (j in terms of the observed relative frequencies,
and show that the variance of this estimate can never exceed 1/(4N).
89 In a factory producing synthetic yarn, the amount of raw material that
can be put into the plant at anyone time is, on a certain quantitative scale,
an integral non-random variable x which can take the values x = 1,2, ... , n.
The quantity of yarn produced, Yx' depends upon the amount x of the raw
material used and an efficiency factor of the plant, so that a linear regression
relationship between Yx and x does not hold over all the possible values of x.
It is therefore assumed that Yx is a random variable such that
E(yx) = a.X+PX2, var(yx) = (12,
and it is desired to estimate the magnitude of the linear parameter a. and the
damping parameter p.
If a set of sample values Yl, Y2,' .. , YII is obtained by independent experi-
mental runs of the plant using the quantities x = 1, 2, ... , n respectively of die
78 EXERCISES IN PROBABILITY AND STATISTICS
raw material, apply the method of least squares to obtain a* and P*, t~
estimates of a and p, and show that formally
( *) S4(]'2
var a = (S2 S4 -S~)'
where AI, A,2 and A,3 are known constants, and S2 is an independent estimate
of (]'2.
Assuming that a and b are normally distributed and ns 2 /(J2 has a X2 dis.
tribution with n degrees of freedom, by considering the function a - pb, Or
otherwise, obtain the appropriate 95 per cent confidence limits for the para.
metric ratio p == a/po Hence verify that in the particular case when )'3 =: 0,
the limits are
~. [(1-92)JO;I:~I+(1-93)JO;I;~I+J9il-94)·I:~1r
so EXERCISES IN PROBABILITY AND STATISTICS
where
k-t
Ak == (n_l)k-t / }I (n+2j-3),
to show that
E[e rx . hHr 2s2)] = E(yr)
and thus deduce that efficient estimates of It and V 2 are
m = eX. h(!S2)
and
and hence derive, correct to the same order, large-sample approximations for
m and v2 •
9S In extensive sampling of a multinomial population it occasionally happens
that the expected proportion in one of the classes is exceptionally large and,
as a measure of sampling economy, it is suggested that only a known fraction
of the preponderant class be sampled, whereas the other classes are enumerated
in the usual way. In such a situation, suppose the expected proportions in the
four classes Ai> A 2 , A3 and A4 are proportional to (2+ 9), (1- 9), (1-9)
and 9 respectively, where 9 is a parameter which can theoretically have any
value in (0 < 9 < 1). If it is decided to record only a fraction p, (0 < p < n
of the At observations but otherwise the sampling is complete, suppose that
in such a censored sample of size M the observed frequencies in the four
classes are Zi> Z2' Z3 and Z4 respectively (l:Zi = M).
Derive the equation for 9*, the maximum-likelihood estimate of 9, and
find its large-sample variance. Further, prove that 9* will be as efficient an
estimate of 9 as &, the maximum-likelihood estimate of 9 obtained from an
ESTIMATION. SAMPLING DISTRIBUTIONS. INFERENCE, ETC. 81
ored sample of N observations, if
lln cens
(1 +20){2(1 + p)-(l- p)lW
M= 4{2(1+p)+(1+7p)O} .N =N.h(O,p),say,
and hence ?educe that whatever be the value of the parameter O. as a first
approximatIOn for small p
4(1 + p)(17 + 23p)(3 +5p)2
M~ (7+ 13p)(13+19p)2 .N.
96 In a plant-breeding experiment four distinct types of progeny AB, Ab, aB
d ab are possible; and in a random sample of N 1 observations, the observed
~n quencies in the four classes were found to be Xl> X2' X3 and X4 respectively,
(fe, == N 1)' If on a genetical hypothesis the corresponding expected proba-
bi'~~ies are proportional to (2 + 0), (1- 0), (1- 0) and 0, determine the equation
for 0, the maxi~um-likeli~ood ~stimate of the parameter 0, and derive the
large-sample vanance of thiS estlmate. . . . .
Since, in general, the parameter 0 hes 10 the range (0 < 0 < 1), It IS
suggested that another convenient method for its estimation is to ignore
~ompletely the preponderant AB class in sampling. Suppose, then, that a
total number of N 2 observations are taken from such a truncated distribution,
and the observed frequencies in the remaining three classes are obtained as
}' Y3 and Y4 respectively (I:Yi = N 2 )· Use the method of maximum likelihood
t~'obtain 0*, the estimate of 0 based on this second sample, and prove that for
large samples
(0 *) _ 0(1 - 0)(2 - 0)2
var - 2N .
2
Hence show that 0* must have a variance ::f var(O) if N 2 = 0·5174 N l'
Further, if 0** denotes the estimate of 0 obtained by using the joint
likelihood of the above two independent samples. prove that
20(1- 0)(2- 0)2(2+ 0)
var(O**) = (1 + 20)(2 - 0)2 N 1 +4(2+ O)N 2'
Use this result to show that if N 1 + N 2 = N, a constant, then the best allocation
of the sample sizes for minimum var(O**) is obtained for N 1 = O.
Discuss the significance of the results in the light of efficient estimation of 0
by the method of maximum likelihood.
97 Suppose the ordered observations of a random sample of size /1 from a
population having a probability density function f(x) in the range
(-00 < x < (0) are denoted by III < "2 < 113 < ... < lin' Further. given a
random sample of III from another population. let these »I observations be
distributed amongst the (/1 + I) intervals formed by the IIi such that »11 obser-
vations are <111; "'i+1 lie between (lIi,lI i +d for (i = 1,2.... 11); and finally
IIIn+ I are> lin. where. of course.
n+1
m = L mi'
i= I
If new random variables are defined by the relations
f f
Ul U;+1 Jj
find the joint distribution of PI' pz,· .. , Pn, and use this to prove that for an
i ::F j and integral rand s I
n I r!s I
E(Pi pj) = (n~r+s; !'
Hence, by considering the conditional joint distribution of the ml for fixed p
show that h
1
n
PI = f "I
-oc:.
f(x)dx; Pi+1 = f
"'+1
IIi
f(x)dx (i = 1,2, ... , n-l); Pn+1 = f
00
lin
f(x)dx,
find the joint distribution of PI' pz, . .. , Pn and m" mz,· .. , mn+ I' Use this
distribution to show that the moment-generating function of the random
variables
Zi=
1m.)
(n+l-~' for i = 1,2, ... ,n+l is
= n ! expC~: O;/(n+ I)} .f{p, e- IJ1 / m + pz e- IJ2/ m + ... +Pn+ I e- IJn + I/m}m X
X
1= I
n dpi>
where the n-dimensional integration is over the appropriate range of
PI' pz,···, Pn'
Hence, or otherwise, evaluate the mean and variance of the statistic
n+1
C z = "t... z~"
i= I
Po"" j'J(X)dX;
-00
Pi = I
Ui. I
•
f(x)dx, for i = 1,2, ... ,n-l; Pn = I
<Xl
Un
f(x) dx.
d2 = I [~- Xi]
i=O n m
Xi == ±
j=O
mj ,
is
where
k
./, = "L...
'I'k -
p.). e-t}/m.,
j=O r=j
ror all integralj and k in the range (0, n), and the multidimensional integration
is over the appropriate range of the Pi. Hence determine the mean and variance
or el 2 ; and assuming that for large m and n d2 '" PX 2 with v dJ., use the first
tlVO moments of d2 to determine p and v in terms of m and n.
N(N -1)p
var(Y)-E(Y)[1-E(Y)/N) = (1+p)3p +O(P-2).
Hence verify that in the limit X and Y have the same binomial distribution
ith parameter p/(1 + p), but that, as compared with this limiting form, the
;istribution of X for finite a. and P is leptokurtic, whereas that of Y is platy-
kurtic.
104 It is known that in a large batch containing N items of a mass-produced
article the probability for the batch to have X defective items in all is P(X),
defined for (0 ~ X ~ N), and such that the mean and variance of X are JI.
and (12 respectively. Successive random samples of size n" n2 , • •• are taken
from the batch without replacement, and the corresponding numbers of
defective items observed are XI' X2' .••• For this inspection scheme, if
k k
Xk = X-I XI and Nk = N - I ni , for k ~ 1,
i= I j= I
find the conditional distribution of Xi for given Xi-I' Use this distribution to
prove that for integral mi and Vi
= nN [u G-~)]/N(21,
I j 2- ifi ~j.
105 In non-destructive testing of an industrial product, batches consisting
of N items each are available for sampling inspection. The number of defectives
86 EXERCISES IN PROBABILITY AND STATISTICS
E=
N(N-n)
Jl(N-Jl)'x~o
a [(X+l)
n+l Pn+l(x)-N,Pn(x)
Jl]
,
f
u
p= f(x)dx.
-00
(3),,+2) -2
2()" + 2)11 + 0(11 ).
where the Cj are arbitrary constants. Derive the joint distribution of lilt
ordered observations and use it to establish that
E(x~ x~) = n! U+s-l)! (i +r+s-l)! . ar+s
I J (n+r+s)!U-1)!(i+s-l)!
for i > j, and r, s positive integers.
Hence, or otherwise, find the mean and variance of y, and then deduCt
that y is an unbiased linear estimate of a/2 with minimum variance ir
Cn = (n + 1)/2n, and Cj = 0 for i "# n. Find the distribution of this best estimate
of a/2, and then show that. as n --+ 00. the standardized unit variable defined
by
t = {y-E(Y)}/Jvar(y)
has the limiting distribution e'- 1 dt for (- 00 < t ~ 1). Also, show that for
finite n the efficiency of y as compared with the sample mean x is (n + 2)/3.
110 If XI ~ Xl ~ ... ~ Xn are ordered observations from a population
having a probability density function
(k+ l)x k /a(k+ 1) for 0 ~ X ~ a,
where a and k are positive parameters, find the joint distribution of the XI'
Hence prove that for i > j and r, s positive integers
r s
E(x.x.) =
a+s.n(k+l) r(n)rp+k:l)r(i+::~) ,
I J {n(k+ l)+s+r}' rU)r(i +k: l)r (n+ ::~)
and use it to evaluate cov(Xj, x n ).
ESTIMATION, SAMPLING DISTRIBUTIONS, INFERENCE, ETC. 89
Assuming k to be known, prove that
1I(k+ 1)+ 1
y = n(k+2) . Xn
h best linear estimate of the population mean, and that its efficiency as
I' I eared with the sample mean x of n random observations is
(limP
n(k+ 1)+2
k +3 > 1 for n ~ 2.
Q -- _1_
(1~'
J
i+lc.
e
-tlx-m)2/a 2 d
. x.
V ~/~ i-/CS
Prove that
I
1 f dO
E(Q) = (In-I)' ( (2)n/2'
In=1B 2' -2- -I 1 +n-l
k
where N == L nj.
j= 1
Hence show that
(i) E{ v'. (ji- m)2 P + I} = 0 for fixed r and all integral values of p; and
2 2)P+' rG+p+r) r(p+t)
(H) E{,'. (y - m)") ~ (-"--. (
n rmr i+p)
for all integral values of p and for all r such that the r functions on
the right-hand side have meaning.
115 Random samples of size n1 , n2' ... ,nk are taken respectively from (
univariate normal populations having the same mean m but differing varianoo
(ff, (f~, ... , (ft. If the sample observations are denoted by xij U = 1,2, ... ,IIi;
i = 1,2, ... , k), prove that m*, the maximum-likelihood estimate of m, is
obtained as a root of the equation
L
k (-
njXj-m *) =0
j=) sf+(xj-m*)2 '
where
Iii
j=1 Sj +(xj-m)
.x = I k (kI /ljXi I1 j •
j= 1 j= 1
Find the variance of x and hence verify that if the ni are equal to n, then
(i) var(x) < var(ln) if the ai are equal; but
(ii) for unequal ai' var(x) > var(IfI), provided 11 is sufficiently large.
116 Given a random sample of (2n+ 1) observations from a population
having a probability density function f(x) in (- CX) < X < CX)), prove that,
as a first approximation, the sample median is normally distributed with
mean III and variance 1/[8nf2(m)], m being the population median.
Hence deduce that
(i) if X is normally distributed then the sample median has an asymptotic
efficiency of 0·64 as compared with the average of (211 + 1) observations;
(ii) if X has the Cauchy distribution, then the asymptotic efficiency of the
sample median as compared with the maximum-likelihood estimate of
m based on a sample of size (2n + 1) is 0·81.
117 A random sample of n observations is given from a r distribution with
(he probability density function
rxV -ax ,'-1
nv).e .X , for 0 ~ X < CX),
where I' is a known positive constant and rx is a parameter. Prove that in this
case the method of moments and the method of maximum likelihood both
give the same estimate of rx. Show that this estimate is not unbiased but is
asymptotically consistent as n -4 CX). Hence obtain the modified unbiased
estimate of rx and, assuming that I1V > 2, verify that its variance is rx 2 /(nv - 2).
118 If Xl' X2" •• ,Xn are independent observations from a univariate normal
population with mean III and variance a 2 , prove that the expectation of the
statistic
1 n
~
a* = - . - ,I IXi -
2 n i=1
IS a - - ,
11
xl . F.l1-l
.x being the sample average.
92 EXERCISES IN PROBABILITY AND STATISTICS
= ~[Jn(Il-2)+sin-I(-1
1t1l n-l
)].
Hence evaluate the exact variance of a* and verify that for large n the efficiency
of a* as compared with s, the square root of the maximum-likelihood estimate
of a 2 , is (1t - 2)-1.
Finally, prove that
11-1
but
119 ffthe random variables x and y have ajoint bivariate normal distribution
with parameters mx , my, a';, a;' and p in usual notation, prove that
P{IX-l11xl ~ ly-m).1} =
_ 1 _1{2A~}
- ~. tan 1 _ A. 2 '
120 During World War II the markings and serial numbers obtained from
captured German equipment were analysed in order to obtain estimates of
German war production and capacity. As a somewhat simplified model,
let the population of serial numbers pertaining to one type of equipment
be 0+1, 0+2, ... ,0+p, where. in general, 0 and p are both unknown, and
it is desired to estimate p.
Assuming initially that 0 is known, the population may be denoted by the
p integers 1,2,3, ... , p. Suppose then the captured equipment provides a
random sample of n integers XI' Xl> ... ,Xn, (n < p) from this population, and
let g be the largest of the sample integers. If it is further assumed that all
samples of n integers have equal probability, find the sampling distribution 01
g for given n and fixed p.
Prove that there is only one function of g, f(g) say, which is an unbiased
estimate of p and that this function satisfies the difference equation
h-I
f(h) = h- L
v=n
f(v). P{g = vIp = h}/P{g = hIp = h}.
ESTIMATION, SAMPLING DISTRIBUTIONS, INFERENCE, ETC. 93
Next, for any integral k such that n + k > 0, prove that
In
P (g+k) (P+k+
n+k = n+k+ 1
1)
and hence, or otherwise, deduce that for integral r
D= LD
j= 1
j
el == ~ I
fJla,/2
e- t82 d()
fJla,fi fJla,fi
f ff
00
where ¢(()l, ()2; p) is the bivariate normal density function of ()1 and ()2 having
zero means, unit variances, and correlation p.
ESTIMATION. SAMPLING DISTRIBUTIONS. INFERENCE. ETC. 95
Given a random sample of II paired but unequal observations (X j , lj)
124( '_ 1 2 ,'" n), the first differences of the two series are defined as
for ) - , ,
D, :=: X j + I - X j and dj = lj + I - lj, U = 1, 2, .. , , II - 1).
J
, g C. :=: dPj' a correlation statistic
Sellin J
II-I
C = L Cj
j= I
r
and hence deduce that
and
96 EXERCISES IN PROBABILITY AND STATISTICS
E(g) = ~ sin - 1 p,
n
and /
127 Suppose (Xl' yd, (x 2, Y2),·· . ,(X"' Yn) are random observations from a
bivariate normal population such that the differences Ui = (Xi - Yi) have
mean JI. and variance (12 for all (j = 1,2, ... , n). Using the signs of the Uj, a
statistic
n
S = L SJn
i= 1
is defined where the Si are random variables such that Si = 1 if Uj > 0 and
Sj = 0 otherwise. Prove that
1
E(S) = !+t/>(-r) and var(S) = _{-!-t/>2(or)},
n
where
-r == JI./(1 and t/>(-r) == fo [ r
e-)·2/ 2 dy.
and
var(i)
n-l
= n(n_3)+T 2 n-3
[(n-l) -C; ] '" 2n
T2+2 for large n,
IJ _ p(T-x)
- pb+(l-p)g(b)
and
A _ P [ 2 OA2 t 2
g(O )-(1_p)v 2 .2(V+ )-(v-2)O],
A
f
00
129 Suppose XI' X2" •• , Xn are random observations from a normal popula-
lion with mean J1 and variance a 2 • If it is known that the sampled population is
lruneated at a point T so that all Xi are < T, determine the maximum-likelihood
equations for estimating the parameters a and 0 == (T - JI.)/a. Hence show
lhat for their estimates IJ and 0
(j = (T-x)/{O+g(O))
and
g(O)
A
=
2(V2 +02)tV_(V
2
2- 2)0
'
-00
98 EXERCISES IN PROBABILITY AND STATISTICS
(i) p: [1 + n ~ 1] ; (11") -
pq;
n
"') -pq
(III [1 +--
n-l M-N]
- -- ' , -pq~1 -11-1]
(IV) -- ,
n N' M-I ' n M-l
131 In an investigation into consumer preference for the type of fuel used
for central heating of homes, suppose that in a total population of N houses
the proportions using gas, oil, and various kinds of solid fuels are p, q and
(1- p - q) respectively, In order to estimate the parametric ratio p/q, a random
sample of n houses is taken, the sampling being without replacement, and it is
observed that the frequencies of gas and oil users are x and y respectively,
Find the sampling distribution of x and y, and then show that
Np [ (N-Nq-l)(nl] p
E[x/(y+l)] = Nq+I' 1- N(n) ~q[1-(l-qr] for largeN,
[(N -11)(P+q)]t,
(N-J)npq
132 Suppose x and yare random variables having means m" 1112' equal
variance (12, and correlation p > 0, and for any given two new random e
variables II and v are defined by
1/ = X cos e+ y sin e; v = - x sin e+ y cos e,
Prove that
eSTIMATION, SAMPLING DISTRIBUTIONS. INFERENCE, ETC. 99
deduce that
/-tence
') 0 " corr (u, v) ~ corr (x, y);
(I '" M
(ii) (x+ y)/./i and (y-x)/y 2 are uncorrelated random variables.
33 Apart from the usual random errors of measurement, an instrument
1 tWO independent and additive sources of bias of unknown magnitude
h~ch cannot be totally eliminated by suitable adjustment of the instrument.
~ Iwever, according to the manufacturer's specifications, it is possible to set
hO instrument so as to make the biases either positive or negative. Four sets
/n
t independent measurements each are made of a quantity by setting the
?strument in the four possible combinations for the biases. Denoting the
~ts by Xj' Yj, Zj and Wj, (i = 1,2, ... , n), it may be assumed that
E(xj) = ex + Pl - P2 E(yj) = ex - Pl + P2
for all i, where ex is the true magnitude of the quantity measured, Pl, P2 are the
unknown instrumental biases, and all the observations have the same variance
(12,
Find the least-squares estimates of the parameters, and verify that the
expectation of the error sum of squares is (4n - 3)0'2. Hence, assuming that
the observations are normally distributed, indicate, without proof, the general-
ized t statistic for testing the hypothesis that the two instrumental biases
arc equal.
e = -.log
A 1
T
(n)
-
lI-r
.
fit.td t~e joi~t distribution of the y·s'. Hence. de?uce that 2rO/e ~as a X2 dit
tnbuhon with 2r dJ.• so that the maxlmum-hkehhood formula gives the exa'
e
variance of () as 2Jr. ~
Use these results to prove that
(i) 0 is fully efficient as compared with e*. the usual maximum-likelihoOi!
e
estimate of based on a completely enumerated sample of r iternl
and
(ii) for the random variable x,
, ,
E(x,) = eL 1/(n-j+ 1); var(x,) = e2 L 1/(n-j+ If
j= 1 j= 1
136 Suppose that YI• Y2 •• •• , y" are n independent observations such thai
E(Yv ) = IX+/3IXlv+/32X2," and var(Y.) = u 2 • for (v = 1.2•...• n). where
IX. /31' /32' and u 2 are unknown parameters. and Xl' X 2 are non-random ex.
planatory variables. If f3t and /3! are the usual least-squares estimates o!
/31 and /32 respectively. prove that
var(/3n = u 2/(l- r2)S 11; var(/3!) = u 2/(I- ,.2)S 22.
where. denoting deviations from the sample means by lower-case symbols,
n
S2 = Lt1 (Yv-bIX1v-P2x2Y-(l-r2)sisll}/(n-2).
IlSTIMATION, SAMPLING DISTRIBUTIONS. INFERENCE, ETC. 101
. Jly under assumptions of normal theory, show how this estimate can
Flr;:'ro; an approximate large-sample test of significance for any hypothesis
oe use
about P2'
In the distribution of the number of accidents per worker in a factory
137 given period of time, the frequency of workers sustaining one, two or
over aaccidents is available, but the number of persons who did not have an
rno~~ent cannot be enumerated owing to the factory population fluctuating
aeci. g that period. This gives rise to a discrete distribution in which the zero
dunn d
up is unobserve .
gro If it may be assumed that this truncated distribution is Poisson with an
known mean A, and that in a sample of N the observed frequency for x
.un I'dents per worker is lx, for x ~ I, show that a simple estimate of A based
,ICC f .
on the method 0 moments IS
A* = m~_1
I ,
m,
where m~ is the rth sample moment about the origin. Prove that for large N
(A*) _ A.{(A.-k+I)(Jt',-k+l)+2Ji'd
var, - N{Jl', -k + 1)2 '
where a, a real number, and p ~ 2 are parameters. Prove that the propor-
tionality factor is l/l(p) satisfying the recurrence relation
(p+ 1)(p+2) . .
l(p + 2) = [1 + (p + 2)2] . l(p), with 1(0) == 2 smh 1[/2.
c
Use this result to derive the expectation of 2 logf(x)/ila. 2 , and hence deduce
that the large-sample variance of the maximum-likelihood estimate of IX
obtained from a random sample of II observations. of X is
1+(p+4)2
lI(p + l)(p + 2)(p + 4)"
Also, show that the sample mean .x is not an unbiased estimate of a but is
asymptotically consistent for p --+ XJ.
104 EXERCISES IN PROBABILITY AND STATISTICS
I n. == N .
•= 1
If X and Y denote the total number of particles observed in the odd and
even time-intervals respectively, prove that
E(X) = e'<E(Y), (A, == T/(J),
6= T/Iog(_ il ),
11-1
where Nil ==
.=1
f vn,.,
( ~)2.
1 +e-'<
144 In a decaying radioactive source the probability of a particle being
emitted in time (t, t + dt) is
(J-I . e- I / 9 dt, for 0 < t < 00,
deduce that for fixed 0 and N, the variance of the estimate is minimized
ilcllrCC 1.6(} whence approximately
(M - ,
1'54(}2
var(6*) ~ ~.
e= -TjIOg(I-_n_),
N-n
and that the large-sample efficiency of ()* as compared with eis
J.e- Ajp-2(l-e- A)2}.
145 In a multinomial population with four distinct classes AI, A 2 , A3 and
.1 the corresponding probabilities are proportional to (2 + ()), (1 - ()), (1- ())
;\I:(i 0, where () is an unknown parameter. A random sample of II observations
is taken from this population and the observed frequencies in the four classes
arc found to be XI' X2' X3 and X4 respectively. To estimate 0, Fisher suggested
livc possible estimates defined as follows:
(i) 0, = (XI-X2-X3+x4)/n;
(ii) (}2 = (XI- X2- X3+ 5x4)/2n;
(iii) () obtained as an appropriate root of the quadratic equation
(}(2+0) XIX4
(l - (})2 = X2 X3;
Prove that
(a) (), and O2 are both unbiased estimates of 0, and that their exact variances
are (1- (}2)jn and (I + 6() _4(2)/4n respectively;
(b) 0* is obtained as a positive root of the quartic
xf (x~ + x~) xl
(2 + 0)2 - (1 _ 0)2 + (}2 = 0; and
(e) for large n, both {j and 0* are fully efficient as compared with e.
Discuss the relative merits of these five estimates.
146 If X and yare independent normally distributed random variables with
zero means and variances I/rx and rx respectively, prove that the information
supplied by a pair of (x, y) values for the estimation of a is l/a 2 •
Further, given a random sample of n pairs of (x, y) observations (Xi> Yj),
(i == 1,2, ... , n), find a, the maximum-likelihood estimate of IX, and determine
106 EXERCISES IN PROBABILITY AND STATISTICS
11(11)
a2' 11 + 1 ;
so that the maximum-likelihood estimate of a does not extract all the in~
mation contained in the sample. Ot.
147 Two independent random variables x and y have the joint probabil't
. function
denslty . 1\
'
e-(9x+ y /9), for x, y ~ 0,
o being an unknown parameter. Prove that the information supplied by a Pal
of observations of x and y for the estimation of 0 is 2/0 2 , and hence deduct
the total information obtained from 11 independent pairs (Xi' YI), ro~
i = 1,2, ... ,11.
Further, given the sample observations, show that the maximull\,
likelihood estimate of 0 is
t = (Y/X)t,
where X and Yare the sums of the X and y sample values respectively.
Determine the joint distribution of X and Y. Use the transformation
X = u/t and Y = ut
to derive the joint distribution of II and t. Hence, or otherwise, obtain the
marginal distribution of t. Show that the amount of information to be expec.
ted from a single observation of t is
211 ( 211 )
0 2 ' 211+ 1 '
so that the maximum-likelihood estimate does not extract all the information
in the sample for the estimation of O.
Comment on this result.
148 Show that for a single random observation from the Laplace distribution
(-00 < x < 00),
f
1
Yi = In .SZi, for 2 ~ i ~ n,
where
n n
ns 2 == L (Xj_X)2 and nx == L Xj'
j= 1 j= 1
determine the joint distribution of X, sand Z2, Z3," ., Zn-1' Hence, if for
i. ~ 2
n
mA == L1 (xrx)A/n,
j=
estimate of (). Also, show that the information per observation for estimatin
() is g
1{
()2 1-
'2 -), }
A)2 • (t-:
Further, assuming there are a total of N sources under observation in the
ex~r~me":t, prove that the expected information for the whole period per
umt ttme IS
N [(I-e-),)2 - A,2 e-),]
()3 A,(l-e-),) .
Hence show that the most efficient value of the time-interval T is approxi.
mately 4().
151 There are k events E l , E 2 , •• • ,Ek , and the probability of the occurrence
of Er is Pro (0 < Pr < 1), for (r = 1,2, ... , k). In Ar independent trials the event
Er occurred Qr times, (r = 1,2, ... , k). Further, a combined event T is defined
as the joint occurrence of E l , E 2 , • •• ,Ek with probability P, and in C indepen.
dent trials T occurred c times.
Assuming that the events Er are independent, find the maximum-likelihood
estimates of Pr and P. Hence, for large samples, obtain a Xl statistic for testing
the hypothesis of independence of the events Er •
152 If Xl ~ Xl ~ ••• ~ Xn are ordered observations from a population with
probability density functionJ(x) defined for ( - 00 < X < 00), prove that for the
sample range w
f f
00 x
Further, denoting E(w) by £0, show that the rth central moment of w is
E[(w-£O)'] = -(r-I)( -(0)' +
ff
00 Xn
+r(r-I) [1-a:-(I-al)n+(an-al)"](W-£O)r-ldxldxn,
-00 -co
_E[02~;;L] = 1I[1-pJi2(1-rd];
r =
k-
1
J2n
f 00
-00
-
y2 2
e--/-dy-
(1 +A e-P)'t'
and A == (1 -pP ) e- p2 / 2 .
(I_pn-I) ,
and that for large n
pq (l- pn _qn)2
var(p) = - . -::c-:----,-------,-,:---:-'-::----;;-"---'----=-----.---::--:;~~
II [(1_p"_qn)_lIpq(pn 2+qn 2_pn 2qn 2)],
Alternatively, if for large II, p* = 1'/11 is used as a simple approximate
estimate of p, determine the bias of this estimate and verify that exactly
pq [(1- pn _ qn) _Ilpq(pn- 2 +qn- 2 _ pn- 2qn- 2)]
var(p*) = - . (1 n n)2
It -p -q
158 An insect pest lays its eggs on the fruit blossom of mango trees, but not
all eggs hatch to become larvae, as some of them are destroyed by weather
hazards. The only definite indication of the presence of the pest is in wormed
ripe mangoes, which show up when the fruit is sliced. In order to assess the
extent of damage caused by the pest in an orchard, samples of n mangoes each
were examined from different trees, and records of wormed fruit kept from
k trees in whose samples there was at least one damaged mango. The total
number of fruit obtained from a tree is large compared with the sample size
n, so that the effect of sampling from a finite population may be ignored, and
the trees in the orchard may also be treated as independent.
If the probability of obtaining a wormed mango from a tree is a para-
meter p, find the probability of the observed distribution of 1'1>1'2, ... ,r1
wormed mangoes from the k trees. Prove that the equation for p, the maximum-
likelihood estimate of p, may be put in the form
p = Po[1 - (1 - p)n],
eSTIMATION, SAMPLING DISTRIBUTIONS, INFERENCE, ETC. 111
P is the observed relative frequency of wormed fruit in the k samples.
wheresh~w that for large k
Also,
pq(l_q")2
var(p) = nk(l"-q -npq"1)' (p+q = 1).
9 for an hereditary abnormality, the probability that a particular child
~5 family is affected depends partly on his or her birth-rank. In order to
~n astigate this effect of birth-order, a complete family (i.e. without any
~~s~arriages or stillbirths) of n children was examined, and it was found
hat r of them were abnormal.
t Assuming that there were no mUltiple births in the family so that the n
children can be ranked without any ties, suppose the sum of the birth-ranks of
fer abnormals is A. Then, for constant II and r, show that, on the null
~~pothesis that birth-order has no effect on the occurrence of the abnormality,
the probability of obtaining the sum A is
the second summation extending over all pairs of genes. each pair beinl
counted only once.
161 On a certain biological hypothesis, the occurrence of a rare hereditary
human disease may be explained as due to the inheritance of either of two gene\
A and B, which are known to occur in the population in the ratio of I:A (>Oi
The genes are indistinguishable in their effects, except as regards the age 01
onset of the disease; and, as a simple model, it may be assumed that for such a
mixed population, the age of onset is a continuous random variable whose
distribution is compounded of two unequally weighted univariate normal
distributions such that
(i) the weights are in the ratio of 1 : A;
(ii) on an appropriate scale, the means are - In and m; and
(iii) both the variances are equal to unity.
Find the probability density function of the age of onset distribution, and
hence deduce that this distribution will be bimodal if, and only if,
Ilog AI < 2m(m2 _1)1 - 2 log[m + (m 2- 1)1].
As a generalization, suppose the variances of the compounding distribution
are af and a~. Prove that in this case the least possible separation between
the means of the compounding distributions for bimodality is
Gfa'" 1·840-,
L [/k+l
00
k! n (z+m) ] =-.
1
k=O m=l Z
If sampling is confined to samples of size n which have r 'I: n, use the above
result to evaluate the expectation of 1/(n - r). Hence verify that, for n suffi-
ciently large so that the probability of r = II is negligible.
and
(q = I-p),
165 It is known that a particular gene with two alleles A and a controll
melanism in lepidoptera, and it is assumed that the alleles occur in the popu.
lation in proportions (1- (J) and (J, where (J is an unknown parameter. O[
the three types of genotypes AA, Aa and aa produced under random matin~
the darker type AA is the dominant form, and, in general, it is possible to
distinguish the three genotypes phenotypically. However, in areas of high
industrial pollution, it is occasionally possible to misclassify a heterozygote
Aa as an AA. In order, therefore, to allow for this error it is decided to divide
the whole sampling region into two strata. Of these two, the first stratum
refers to those parts of the sampling region where there is, on some specified
standard, no pollution to affect materially the classification of the heterozy·
gotes. The second stratum consists of such other areas where the pollution is
known to be sufficient to give a probability p (unknown) of misclassifying a
heterozygote as a homozygous dominant (AA).
From the first stratum a sample of v observations is taken and the fre·
quencies of the three genotypes AA, Aa and aa are found to be YI' Y2 and }')
respectively. An independent sample of size n is also taken from the second
stratum, and the corresponding frequencies of the genotypes, as classified by
observation, are XI' X2 and X 3 • Use the data from the two samples to deter·
mine the equations for () and p, the maximum-likelihood estimates of (J and p.
Also, verify that for large samples
A (J( 1- 0 2 )
var(O) = ;
2[ I' + (2n + v)(J]
A (1-(J)(1-p)!
and corr(p.O) = (I +W"[O-p)+(3p-l)0+(v/nj{ 1 +(2p-l)(J}p·
ESTIMATION, SAMPLING DISTRIBUTIONS. INFERENCE, ETC. 115
A binomial distribution has (n + 1) classes and an unknown "success"
166 ter p. Suppose the first k (0 ~ k ~ n) classes of this distribution are
paraJ11~ed and that a normalized distribution of a random variable x is defined
"r
(rllnc~he integral range (k ~ x ~ n), If denotes the rth moment of x about
IIVCr rigin prove that
(hc 0 '
"1
= r.x+np and "2
= (k-l)r.x+ {I +(n-l)p}"l'
\I here
II. == ~. (kn)pk qn-k+ 1; p+ q = 1; "0 == 1 _ k~l (n) pX if-X.
"0 x=O x
1'0 determine "3 and "4'
0\ s S~ppose N independent samples, each of size n, are obtained from this
~cated distribution such that there arefx samples having exactly x successes,
~~ that r. fx = N. If for all integral r,
n
T,. == L xr fx/N,
x=k
dr (p*) ==
pq[(n-I)T,-Il(k-Ilf
- [llpq-(r, -Ilp)(" +p -k)][ll(k-I){(k-I)-(1l-3)p} +rd(n-I)(Il-3)p-(lIk- 31l+k + I)}]
eff( *) = (n-l)p(1-q")
p [(n-l)-(n-3)q][1-npq" l_ q"]'
and that for variation in p the minimum of eff (p*) is attained for a
value of q =1= 1 which is the root of the equation
np2 qn-2 [(n-l)q"-n(n- 3)q+(n-l)2] = 2(l-q"f
167 The number of insect eggs found in a nest is known to be a random
variable having a Poisson distribution with mean A. But a nest can be recog-
nized as such only if one or more eggs are present. In addition, due to faulty
e
observation, a proportion (0 < () < 1) of the nests having only one egg each
are overlooked. Thus, assuming that there are no other observational errors
or classification, the distribution actually observed is a truncated Poisson
distribution in which the zero class and a proportion of the ones are missing. e
116 EXERCISES IN PROBABILITY AND STATISTICS
and
var(X) = ..1.[1-(..1.+ 1) e - A+ 0 e - A{e - A_ (..1. 2 - A+ I)} ]/[1- e- A(1 + 0),)]2,
Hence deduce that, irrespective of the value of 0,
. 1. 2 e- A var(X) 1
1- (I-e A)2~~I-(A+I)e A'
X = r for r ~ 1, and I n. = N .
•=1
By considering a linear function of the sample class-frequencies, sholl'
that an unbiased estimate of A is given by
00
..1.* = I r.n.IN .
• =2
Prove that
var(A*) = ~. [1 + eA~ 1 J.
and hence deduce that the efficiency of ..1.* as compared with the usual maxi·
mum-likelihood estimate 1 is
ESTIMATION. SAMPLING DISTRIBUTIONS. INFERENCE. ETC. 117
. by using the same estimating technique, prove that
AgaJO,
2112+ N ),*
T= N2
L1 IIi = N.
i=
Tl = I 112r- 1 and
r= 1
Also. verify that {j and }. are asymptotically uncorrelated and that their
large-sample vari:mces are
A A. (l-e-il
var(O) = (J(1-(J)/N and var(A.) = N(J' 1-(..1.+ l)e i.'
118 EXERCISES IN PROBABILITY AND STATISTICS
whence A A 2A
NO ~ var(,l,) ~ NO'
"here
N is the total number of samples used in the whole experiment.
4 An experimenter wants to test whether two different methods oflearning
17c equally good. He has two groups of subjects, A and B, unmatched in any
.Ir and has both groups learn the same material. Group A learns by one
\\J.rhod B by the other. Assume that the two methods really do give different
n~~ults ~ith error variances O'f and O'~ respectively.
fl
S
(i) If the total number of subjects in the two groups is n (fixed), prove
that the most precise method of estimating the difference is to allocate
the number of subjects in the two groups in the ratio of 0'1 : 0'2'
(ii) Alternatively, if the experimenter wishes that the variance of the
observed difference will be a specified quantity c, and he is free to
choose the total number of subjects, prove that the minimum number
of subjects will be required if the sample sizes are again in the ratio of
111 : 112'
Evaluate the group sizes in (i) and (ii).
(iii) Also, determine, under suitable conditions, the group allocations if the
total number of subjects is n (fixed), and the experimenter wishes to
ensure that the variance of the observed difference is also a preassigned
quantity c. Hence verify that if the two standard deviations are equal
to 11, then the group allocations are in the ratio of (l ±).,) : (l =+= 2),
where
40'2]-!
), == [ 1----;;;;- ,
provided IlC ~ 40'2.
= e-I', for k = O.
lI~nce show.that the cumulant-generating function of the number of plants
p~r quadra I IS
K(t) = It[exp{ t+ ..1.(e t -1)} -1],
T= III ;:-i Xi
/,,-1
L;,i
i= I j= 0
is the best estimate of I" Find the variance of T. Hence prove that
for all 11 > 1. var(T) < }.a 2 ; and, for large 11, var(T) '" 0 or (). -1 )a 2
according as ;. < or > 1, respectively.
(iii) What is the limit of var(T) as ;. --+ I?
179 The tensile strength of a steel cable produced by a factory has an un-
known value JI. and to estimate its magnitude, a series of /I independent
assessments X I. X 2, ... ,x" were determined by an investigator A. Since these
mcasurements were. as far as possible. made under similar conditions. it may be
assumed that they have the same but unknown variance a 2 , independent of I'.
Another investigator B made a separate series of v independent measure-
ments .1'1' .1'2" •• ,y,. of the same tensile strength, and these may also be assumed
to have variance a 2 • However. the y observations appeared to be consistently
larger than the X values. It is suspected that this discrepancy was presumably
due to a wrong setting of the measuring instrument by investigator B. so that
a constant but unknown bias () (> 0) was introduced in the y measurements.
122 EXERCISES IN PROBABILITY AND STATISTICS
Show that the y values cannot provide any information about Il, but th
e
they may be used to estimate the magnitude of and to derive a more pre/'
estimate of (12. Find the estimates and, under the usual assumptions of nil(
mality, determine a small-sample test of significance for any specific hYPoth Or.
about e. Also, if it is presumed that the wrong setting of the measUring ~I.
strument by B might have affected the random errors of measurelllein.
indicate how this could be verified on the basis of the sample data availabln!
e.
180 If X, Y and Z are random variables such that
X=X1+ X2; Y=X2+X3; andZ=x3+ x4,
where the Xi are uncorrelated random variables each with mean 111 and varianCt
(12, find the product-moment correlations between (X, Y), (Y, Z) and (X 2
Comment on the results obtained. •I
Also, determine the correlation between X and Z when
(i) Y = c, a constant, and
(ii) Y = AZ, where A is a constant.
Verify that, as ;, --> 00, the limiting value of this last correlation is unit)
182 The ranking given by a judge to N paintings Alo A 2 , ••• , AN is 1,2, ... ,N
A second judge is in general agreement with this ordering, but he divides hi,
ranking into k tied groups of 11 1,11 2" " , 11k paintings, where 1:l1i = N. Thul
the paintings A l' A 2 , • •• , An, receive the same rank; the next An,H
An. + 2.' .. , All. +n, are tied, and so on, till the kth group of Ilk paintings haw
the same rank. These k ranks are in an increasing numerical order.
If
r
prove that the coefficient of rank correlation between the two rankings is
p = [ 3 J2I1ititi-dN(N2-1)r
183 Explain clearly the difference between a goodness of fit X2 for a frequent)
distribution with k classes and a X2 obtained from a 2 x k contingency table
Illustrate your answer by suitable examples.
ESTIMATION, SAMPLING DISTRIBUTIONS, INFERENCE, ETC. 123
The observed row. frequencies i.n a 2 x k contingency t~b1e are a" a 2 ,· •• , ak
,,0
d bI' b2' ···, bk• Usmg the notatIOn I1 j == aj+b j, for all (I = 1,2, ... , k)
Laj == A, Lbj == B, and A+B == N,
e that the X2 for testing independence can be expressed as
prov
X2 = ;;[Jl af/l1 j -A 2/N).
If Q r = rr. use the fact that sin 2 kt = 2 cos kt . sin kt. or otherwise, Iii
show that as /I -+ CX) the probability density function of Y tends to
!(Y=y)=t (-I~Y~I).
,,'nce calculate the first four cumulants of X. Also, if v2 denotes the weighted
'I\~ragc of the variances of the compounding distributions, prove that
1. 2 < var(X) ~ l,2 +!m 2 .
Deduce as a special case that the sample sum of squares from a unit norlll'
population has the X2 distribution with (/1-1) dJ. '
9 Given a random sample of n observations Xl> X2" •• ,Xn from a nOflll.
population with mean m and variance a2 , show that the joint characterisl'
function of Z j and is zJ I,
L zJ and L zlJn;
j =1 j =1
and by inverting this characteristic function prove that the sample mean and
the error sum of squares are independently distributed.
Obtain explicitly the distribution of the sample standard deviation and
show that PI' the coefficient of skewness for this distribution, is
l[2A 2 - 2 + (l/v)]/(l- A2)3,
where
y- (1- L+3L2)'
and then verify that for large v
1( 1 1)
PI - 2v 1 + 4v + 8v 2 •
( = (X -m..}/ax~, 1] = (Y -my)/ay~.
Hence, given a random sample of n observations from this population, prow
that the joint characteristic function of
n n n
~ I
(l_p2)n/2 l).-n/2 exp- 21). [(I-2itdd+(l-2itl1)t~ +2(P+it I2 )t l t 2],
Using the result that. for t real, (I - it) - P is the characteristic function of a
I' variable with parameter p > 0, prove that the joint probability density
funclion of the distribution of II, v, and w is
( I p2)ln-I)/2
- -i(u+v-2pw) ( 2)'n-4)/2
2n-lfir(Il;1)r(Il;2) .e . IIV-W •
lIence obtain, in the usual form, the joint distribution of sx, Sy and r.
12 For a random variable X having the negative exponential distribution
wilh probability density function
I(X = x) = A. e-).x, for X ~ 0,
oblain K(t), the cumulant-generating function of X, and show that the sth
clIl11ulant is
(s -1) !
". = -A.-'-'
p [Y =
(2r+
21)h] = P[rh ~ X ~ (r+ 1)h], for r = 0,1,2,3, ....
Prove that the cumulant-generating function of Y is
sinh (!A.h) ]
[
"h(t) = log sinh Uh(A. - it)} ,
128 EXERCISES IN PROBABILITY AND STATISTICS
[ Sin(at/2)] 2.
(at/2)
Suppose that the range of X is divided into 2m intervals each of length.
and a probability distribution of a discrete random variable Y is defined su..
that
for (1' = 1,2, ... , m). Prove that the characteristic function of Z is
h2
"'h(t) = <Ph(t)+ 4a 2 tan 2(th/4),
,hilI!' t
hat the characteristic function of X is
3(sin at - at cos at)/(at)3,
, ify that the variance of X is a2 /S.
,nd ~er ose that the range of X is divided into 2m equal intervals each of
ShuPr, and the probability distribution of another random variable Y is
::ng l ,
.!dined by
p[Y= (21'~l)h]= P[Y= _(21'~1)h] = P[(1'-l)lI:;;; X:;;; 1'11],
hlr Ir -- I , 2,. .. , m). Prove that the characteristic function of Y is of the form
112
3[eoit,(') - eoit2(t)]/(at)2 + 4a 2 . eoit3(t),
,\here
Bji' jk . .
tnt)
_
= L -.,-.. (Ity
OC;
i=2 J .J
(k = 1,2,3),
Ail == (2a)j+(2 i + l-S)hi,
A. i2 == (22i-2i)aj+(2j-3)hi,
112 ( 2112)
var(Y) = var(X)+T2 1- Sa 2 > varIX).
Prove that the characteristic function of Z has the same form as that
fur }'.
(2r-l)h]
P [Y= 2 = P[(r-l)h ~ X ~ rh], for r = L 2, ... , m.
2 [Q(at)Q(th)
ait Q2(th/2)
Q.(th) [Q(2at) -1] /Q3(th/2)
1] IIIQ(th/2)+ 2(mt)2 Q(at) ,
sinO
where Q(O) == -0-'
By a suitable expansion of this characteristic function, prove that
h2
E(Y) = E(X)+ 6a'
and
var(Y) =
h 2 ( 1 + ah 22 ) •
var(X)+ 36
Ilr'() x =~
x . g() 2 f [or(I*I'
'r ~ r
t 2)] . e-it,Xdt 1,
1t I ut l 12=0
-x
where cP(t I' t 2 ) is the joint characteristic function of X and Y.
22 Given a random sample of n observations (xv, Yv), for (v = 1,2, ... , n),
from a bivariate normal population with the probability density function
1 1
J(X = x, Y = y) = ~. exp- . (X2+ y2-2pxy),
21ty 1 - p2 2(1- pl)
derive the joint characteristic function of the statistics
n n
lienee prove that for fixed Z l' 1/; (z I)' the rth moment about the origin of the
lllllllilional distribution of Zl' is given by
II III'1\'
.1I1l1
WI = L x;/2at,
v= I
W2 = L YV/a2~;
v= I
and by inverting it show that the joint distribution of WI and 1V2 has It,
probability density function
and
CHARACTERISTIC FUNCTIONS 133
5 If XI' X 2' ... , X n are independent normal. random. variables, each ~i~h
2 . variance and means ml , 111 2" .• ,111. respectively, denve the charactenstlc
unltt'lon of the random variable
~~
Y=
.
L X~,
A= I
f 2/f
x. oc
IJ
"",/. , 7.
U =
"
L x~ and v= L" aAx A,
A=I A= ,
134 EXERCISES IN PROBABILITY AND STATISTICS
the a being constants not all zero. Hence derive the characteristic funcr
of u ~ven v = e, and thus show that the conditional distribution of U f).IOli
.
v = e IS
\'t~
<0 ( _ c52)'
e 62 L
r=O
--,-.f(u;n+2r-l)du
r.
(0 ~ u < 00),
c5 2 == 82/2 L1 a1·
).=
Verify that
v2(v+3)
var(ulv) = 2+ 3(v+ 1)·
( 1, t2 ) = ~ (itd r (it 2Y
"t L.. "rs·---
r+s=2 r!s!
11
== !["20(it 1)2 + 2" (itd (it 2) + "02(it 2)2] + I/I(itl> it 2),
prove that the characteristic function of X given Y = y is
c/>(t I ) _ exp I/I[it1, -(%z)]. exp[!{"20(it d2-z 2/"02}]
1 y - exp 1/1[0, -(%y)]. exp( _!Z2/'(02) ,
where z = y-(it 1)"1l.
Hence show that, as a first approximation,
E(Xly) = ~.y; var(XIY) = "20-"il/"02.
"02
Also, prove that if linear terms involving the third-order cumulants "21' Ktl
and "03 only are retained, then the appropriate approximations are
,12 For a ran~om var~able X having the Laplace distribution with the
probability denSity functlOn
1
J(X = x) = 2a' e-Ixlla, for -00 <X < 00,
Verify that this distribution actually integrates out to unity over the entire
range of variation of X.
33 For a random variable X, which is uniformly distributed in the range
10 ~ X ~ a), prove that the characteristic function of log X is d t ( 1 + itr I,
If x l' X2' ' , , 'X n are n independent observations of X, derive the probability
distribution of the statistic
n
LI = L logxj •
j= •
and hence show that g, the geometric mean of the sample, has the distribution
nn
---n-()' gn-. [log(a/g)]"- 1, dg (0 ~ g ~ a),
a r II
Also, deduce from this distribution that 2nlog(a/g) is distributed as a X2
I'ariable with 2n dJ,
34 For a Laplace distribution of a random variable X defined by the prob-
ability density function
1
J(X = x) = 2a,e-lxlla, for-oo < X < 00,
e
PII
-=-.,.-;-::::-:--;-::.
2ni{r(p)l"
Je"" fr(-z)l"dz.
l J
-p-i,Y.J
Evaluate the complex integral by using a suitable contour, and then deduce
that the sampling distribution of v, the geometric mean of the II observations
of X, is
IlV"p-l'f-
" (_I)"r+,,+ I _ _
[d ll - 1 VII:]
U = I logxj ,
j= I
"
1 1
. 2--;'
JiXJ
e'lZ n r(pj-z)dz.
II
n r (p)
j= 1
1tI
-ioo
j= I
CHARACTERISTIC FUNCTIONS 137
rove that the probability distribution of v = e",n is
lienee P
n 00 (_1)r+ 2 vn(r+ p.)-I
~. L L
fI
[(Pj) j=1 r=O
r(' I) J _ .
r+ k'¢j
n
r(PA-pj-r). dv (0::::; v < (0).
jd
In the previous example, if Pj = p+(j-I)/n,p > 0, for (j = 1,2, ... ,n),
39 that the characteristic function of u is
,how
n-nil r(np + nit)/r(np).
" ce obtain the probability distribution of the geometric mean v, and verify
he~ it is of the same form as the distribution of the arithmetic mean of n
I ~ependent observations from a population having the probability density
,n ,
function
f(X = x) = e- X xP-Ijr(p), for 0::::; X < 00.
40 If XI' X2,' .• , Xn are n independent realizations of a random variable X
having zero mean and finite higher moments Pr (r ~ 2), derive the character-
islic function of X, the average of the Xj' Hence, or otherwise, deduce that the
second and fourth central moments of x are
P2
- an
d P4 + 3(n -I)pi respectiVely.
n n3
Use these results to prove that
where
n
(n _1)s2 = L (Xj- X)2.
j= I
a4J]
[at2 = [I/I(tdn{ x [a4J2] ,
12=0 ~ at2 12=0
where 1/1('1:) is the characteristic function of X and 4J2(t 2) that of S2. Hence
deduce that 1/1('1:) satisfies the differential equation
2
d (dl/l)
1/1
1/1 . d'l:2 -
d'l:
2
+ U 2 1/1 2 = 0,
and
n
2a 2B = IIS2 = L (Xj_X)2,
j= I
Mn = nil (211-~-I\(-td\'(I-t2)n-I-\"
.=0 l J
Use this series representation of cP(t I, t 2) to prove that the jth moment about
the origin of the ratio AlB is
j }-<n+2 j - l l/2 j
{
1- L
r=1
't'r X TI
r=1
d't'"
-00
where
so that
jth moment about the origin of A
mj = jth moment about the origin of B'
Hence deduce finally that
2 2 211 2 2 4112 (11- 2)
E(c5 Is ) = (11-1) and var(c5 Is ) = (11-1)3 (11+ 1)'
43 Given a random sample of II observations (Xj' y), li = 1,2, ... , II), from
a bivariate normal population with unit variances and correlation P, a
correlation statistic C is defined by
n
C =L j= 1
C)n,
'" ~[(Sin-l p)'2 - [sin -1 p/{n-l)]2] + 4~ [sin -1 p +sin -1 p/{n -1)] + constant,
var(C) = H*-(~sin-lpr +
p
* =S1O
. {n{2C-l)}
2 '
prove that for large samples the efficiency of p* as compared with the usual
product-moment estimate of p is
4 I_p2
n2 ' (2.
1- -sm
-1
p
)2-+ 0 asp-+ ±1.
n
44 Given two random observations Xl and X2 from the Laplace distribution
!e-1x-m1dx (-oo<x<oo),
and a positive constant Jl, (0 ~ Jl < t), prove that the characteristic function of
Y = (I-Jl)xl +JlX2
is
cp(t) = eirm/{1+t2(I-Jl)2}{1+t2Jl2}.
Hence, or otherwise, determine the sampling distribution of y, and then show
that the information used by Y for the estimation of m is
140 EXERCI<;ES IN PROBABILITY AND STATISTICS
and
(ii) for sufficiently small /l,
List of abbreviatiol~s used in references to
journals, books and examination papers
JOURNALS
II: A/lIwls of Ellgenics (H uman Genetics)
I.IIS Annals of Mathematical Statistics
II Biometrika
III Biometrics
r Econometrica
I "
.1: .JI1 Journal of the American Statistical Association
.I(i J ol/rnal of Genetics
.IIA Joumal of the Institute of Actuaries
.IU,fS Jou/'llal of the London Mathematical Society
IRSS Joul'Ilal of the Royal Statistical Society
.IRSS(S) JO/ll'l1al of the Royal Statistical Society (Supplement)
I/N/~AS Monthly Notices of the Royal Astronomical Society
/'CPS Proceedings of the Cambridge Philosophical Society
/,UlS Proceedings of the London Mathematical Society
/,/~S Proceedings of the Royal Society, London, Series A
I'TI~S Philosophical Transactions of the Royal Society, London, Series A
S Sankhya
BOOKS
III/LEY (1957):
Probability-An Intermediate Text-book by M. T. L. Bizley, Cambridge
University Press, 1957
IIROOKES and DICK (1958):
["troduction to Statistical Method by B. C. Brookes and W. F. L. Dick,
William Heinemann, London, 1958
])"VID (1951):
Probability Theoryfor Statistical Methods by F. N. David, Cambridge
University Press, 1951
I'IILER (1952):
An Introduction to Probability Theory and its Applications (Vol. J) by
W. Feller, John Wiley, New York, 1952
l'ISIlER (1945):
The Design of Experiments by R. A. Fisher, Oliver and Boyd, Edin-
burgh, 1945
IISIlER (1946):
Statistical Methods for Research Workers by R. A. Fisher, Oliver and
Boyd, Edinburgh, 1946
141
142 EXERCISES IN PROBABILITY AND STATISTICS
FISHER (1956):
Statistical Methods and Scientific Inference by R. A. Fisher, Olive
Boyd, Edinburgh, 1956 ran:
KENDALL and STUART-I (1958):
Thl? Advanced Theory of Statistics (Volume I) by M. G. Kendall
A. Stuart, Charles Griffin, London, 1958 an
KENDALL and STUART-II (1961):
The Advanced Theory of Statistics (Volume II) by M. G. KendalJ an-
A. Stuart, Charles Griffin, London, 1961 -
USPENSKY (1937):
Introduction to Mathematical Probability by J. V. Uspensky, MeGra
Hill, New York, 1937 ~.
WADSWORTH and BRYAN (1960):
Illtroe/uct iOIl to Probability alld Raile/om Variables by G. P. Wadswofli
and J. G. Bryan, McGraw-Hill, New York, 1960
Note
To allow for possible changes in pagination, the year refers to the date of edition or PIIOI"
used.
EXAMINATION PAPERS
Chapter 1
--- :luB = A+BA = All points of A and (1,3); (3,1); (3,3); (5,3); (3,5).
An B = (2,3); (3, 2); (4,3); (3,4); (6,3); (3,6).
A-8 = (1,.2); (2,1); (1,4); (4, I); (1,6); (6,1); (2,5); (5,2); (4,5); (5,4);
15. 6);j6, 5L _
(An 8)uA = AB.
Feller (1952). Ex. 6, p. 21.
(i) None; (ii) None; (iii) Four; (iv) Two; (v) One; (vi) One; (vii) None.
I Feller (1952). Ex. 9, p. 21.
(i) A; (ii) 0; (iii) B u AC; (iv) AB.
IO!/IO'.(IO-r)!
Ii Consider the complementary event. The probabilities are
r. (-
k~O
I)k ( 6 ) ( \ _
k
~ )'.
36
12 (i) Let X be the event that no castle can take another. There are 641~,
sample points for the random placement of the eight castles. For th.
sample points corresponding to X, there a.re 64 ways of placing th~
first castle, 49 for the second, 36 for the third, ... to 1 for the eighth
Hence P(X).
(ii) There are now 8! sample points, so that no castle can take another
Let Ai be the event that a castle occupies a square of the while
diagonal in the ith row (i = \, 2, ... , 8).
Then P(A i ) = 7!j8!, P(AiA) = 6!/8!, etc. Consider intersectiolb.
and the required probability is
8
1- L (_I),-I/ r !
,~ I
(iii) If Y is the event that none of the castles is on a square of the whitt
diagonaL the required probability is P(X n Y) = P(YIX)p(XI
whence the result, using (i) and (ii).
13 Let Ai be the event that the hand contains the ith ace and king pair.
P(Ai) =
(52\3-2
-2) //(52)
13'
i (4)(52-2k)(
k=1 k 13-2k _l)k-1 / (52).
13
14 Kendall and Stuart-I (1958). Ex. 7.2, p. 183.
(ii) P2 = 1-4! e~) G~) G~) G~) / G~) (~~) G~) G~) - 0·895.
(iii) P3 = Pl-3(~)(~~) G~) e~) G~) / G~) G~) e~)(!~) - 0·190.
whence
(m-n-l)(m-n)
p(lx - YI ,.;; n) = 1- p(lx - YI > n) = 1- 2 •
m
16 Feller(1952). Ex. II, p. 45.
(i) 2". 1t!/(2n)!; (ii) (n!)22"/(2n)!
(iii) Let the original sticks be numbered 1 to n, and Ai be the event that
stick i is reformed.
P( .) = (2n - 2)! 1(2n)!
AI 2" I 2"'
i
k= I
(_l)k- I(.kn) 2k (2n - 2k)!/(2n)!
17 Bizley (1957). Ex. 2.8, p. 48, and UL (1962).
(i) Number even if last digit even:
'I'
P ro ba bIity 3.6!
= - , =-.
3
7. 7
(ii) Number divisible by 4 if last two digits are divisible by 4:
P ro ba bT
I Ity (b y enumeratIOn)
. = ~ =
1O.5! 5
21'
(iii) Number divisible by 8 if last three digits are divisible by 8:
'I' b . 22.4! 11
Pro ba b Iity ( y enumeratIOn) = ~ = 105'
2 34
= 9+ ~ pjP(X = j) = 55 '" 0·6182
J
146 EXERCISES IN PROBABILITY AND STATISTICS
since
P4 = r~o "6
00 (1) (3)r 2
4 ="3 = PiO,
Ps =
00 (1)(13)r = 5"3= P9,
r~o "6 18
P6 = Jo "6
00 (1)(25)r 6
36 = TI = Ps·
2 244
= 9+ ~ (l-Pj)P(X = j) = 495 '" 0·4929.
J
21 UL (1963).
Let PtA) = x; PCB) = y; and P(C) = z. Then
x=a;(l-x)(l-y)(l-z)=b; I-xyz=c; andp=z(l-x)(l-y).
Elimination of x, y and z from the expression for b gives the quadratic equatior,
in p. Now a > 0 and b(1-a)(l-c) > O. Hence the quadratic has positil(
roots if
(l-a)2+ab
ab-(1-a)(a+c-l) < 0 or c > (I-a) .
I
II l~'
lee the required probability = 8094240/6 10 "" 0·13386.
" H. F. Downton-Private communication.
.. Let A and B be the two observers. Denote balls by ai' a2,' .. , an, where
'S the white ball. Then Pta,) = P(a r) = l/n.
1/, I•
Iknee
_ P(al, A and B say al)
I
P(al A an
d B)
say a l - PtA an d B say al )
PtA and B say a1IadP(al)
/I
( 1 ) 21
10 II 11-1
- - -+ 1 as n -+ 00.
1 )2 1 9 ) 2 I '1 11+80
(- -+(11-1)(- (--)"
10 II .10 11-1 II
Probability of at least one of A and B telling the truth
81 19
= 1 - 100 = 100' and
19 11-1 81(20-11)
---- = >0 for 11 < 20.
100 11+80 100(11+80)
u, = t
k=O
(_1)k (m)
k
(1-'~)' -. (1-e- P)m
n
as r -. 00,
P, -
_ (;)(N -1f-' _ ( n
Nn - N
)r( 1 _.!.)n-,
N . P /r .,
I
where
,-1
P == n
k= 1
(1-k/n).
so that
r)(,-I)/2
( 1-- <P<
(r
1--
)'-1 .
n 2n
1
For N=n-.oo, P-.1 and ( 1- N )N-' -.e- 1, so that p,-.e- 1/r!
.11 t
LJ pensky (1937). Ex. 1, p. 91.
the urns be numbered 1 to n, and Pr be the probability of drawing a
.Le ball from the rth urn. Then
\dll te
a b a
as n -+ 00.
Pn = a+b + (a+b)(a+b+1)". I -+ a+b
LJspensky (1937). Ex. 2, p. 92 .
.12 If Pn is the required probability, then
a
Pn = C:b)Pn-1 +C!b)(l-Pn-l)' PI = a+b'
lienee
Pn ="2
I [
1 + (a-b)n]
a +b I
-+"2 as n -+ 00.
I\y symmetry,
lim rn = (1- y+ yz)/[3-(x+ y+z)+(xy+ yz+zx)],
n .... co
Hence
x = (11+ 1)/(11 2 +11+ 1); y = 11/(11 2 +11+ 1);
II = (11 + 2)/(11 2 + 11 + I): L' = (211 + 1)/(11 2 + 11 + 1).
35 The probability of A winning r games out of (2m + 1) is
em,~ 1) p~(I_pym+l-r.
1- f
r;O
(211l+1)p~(I_Pa)2111+I-r.
r
For m = 2, Pa = i, this probability is 53/512.
In the second part, A wins in exactly r (~ 3) games if
(i) A wins once in (r - 1) games, (r - 2) games drawn, rth win;
(ii) A wins once in (r-l) games, loses once, (r- 3) drawn, rth win.
Therefore total probability of A winning in n or less games is
L0 11. II".
11==
19 Let Un be the probability of A's final win when he has in. Then
Un = pUn+1 + (1- P)Un-l, where Uo = 0 and Ua+b = I,
"f
Un+l-un = C~p)(Un-Un-l)
l_p)n
Un+l- Un = ( -P- Ul'
,u that
(iii) The roots of f(P) = pare 0, 1, t, 1 ± .,fi12, and for the first
values the probability of A winning the match is the same as ththrt:
his winning a game. at (.I
(iv) The expectation of the number of games played for A's win is
3[(1-2pq)2+4q(1- pq)]/(1-2pq)(1-2pq+3q),
where p+q = 1 and p i= O.
41 From th~ given sequence of n calls, we derive a seque~ce of (n-l) s
cesses and fatlures. Define a success at the rth (r > 1) call If the call Co Ut·
from the same room as the (r-l)th call. In this derived sequence, the p~'
ability of a success is 11m; and if in this sequence there is no run of at lea
two successes, it follows that in the original sequence no room makes at lea::
three consecutive calls to the exchange.
Let Un be the probability of no run of at least two successes in n trials I
the derived sequence. Then o.
Un
m-1) un- l + (m-1)
= ( ----;;;- ----;;;- ;;; (1) Un -2, and Pn = UR - l , n > 2,
whence
m-1) (m-1)
Pn = ( ----;;;- Pn-l + --;;J2 Pn-2'
But E(Xi) = (n+ 1)/2; E(xf) = (n+ 1)(2n+ 1)/6; and for i i= j
r"'. r=ln(n-l) 2
= (n+ 1)(3n 2 -n-2)/12(n-l).
Hence var(xi) = (n2-1)j12 and COV(Xi'X) = -(n+1)/12.
Therefore
,\Iso.
E(Xj) = ±d
,=0
t(n\V2/1 = n/2; E(x1) = ±
,=0
t2(n)1/2/1 = n(n+1)/4;
t ~
n2 n
- 4- 4(2/1 - i)'
fhcrefore var(Xj) = n/4; cOV(Xj• Xj) = -n/4(2/1-1), whence
lienee E(r)= ± r+
,=0
r(I1+I)V(2/1+I_1) = (11+1)
1
± (11)1/(211+1-1)-1
,= 0 , '/
(11-1)2/1 + 1
- 2"+ I-I .
E(r2) =
,=0
±\r+
(11+1),2/(2/1+1_ 1)
1
= [ lI(ll+I)L
/I (11
,=1 ,
=1I) -(I1+I)L
+L II (11 +
,=0 r+ I
(2"+ 1-1)
/I
,=0 r
(")
I1)] /
= [2/1-1(11 2 -11+2)-1]/(2/1+1-1). whence
var(r) = (11+1)2/1-1[2/1+1_(11+2)]/(2"+ 1-1)2.
154 EXERCISES IN PROBABILITY AND STATISTICS
46 UL (1960).
k = 0; E(r) = 0- 1 ; var(r) = (1- 0)/0 2 •
P(r) = 0(1- 0),-1, and since (1- 0) < I, P(r) is a maximum for r::::: I .
mode = 1. II~
n
Sn == L P(r) = 1-{l-Or = t if 0 = 1-(W /n •
,= 1
Similarly,
00
u, = CNN-r)mN<!)N-" for 0 ~ r~ N.
N N
L u, = 2- N. L [coefficient of x N-, in (1 +X/2)2N-']
,=0 ,=0
. N ( X)2N-'
= 2- N • coefficient of xN 10 ,~o x' 1 +"2
= 2- N.coefficient of x N in {l+x/2)2N+I(I-x/2)-1
_ _ 2N ~ (2N +
- 2 . L.
1) .
,=0 r
But
f (2N+l)+
,=0 r
2yl (2N+l)
,=N+I r
= 22N + 1,
so that
.Ill i
. Ie (1957). Ex. 4.7, p. 121.
~I~ balls and n compartments the total dlstnbutlve posslblhties are
Wit t tal number of favourable possibilities is
. .. . "
rhe 0
L: t! III ai' where I ai = t and each ai ~ 1
:=:coefficientofx'int!(eX -lt=
m=O
±m (n)(_1)m(n_m)l,
\',I!(t-l)-f(t)}
oJ
= I f(t) = n ±~.m (n)(_l)m+1
1=0 m m= 1
, I
1
= nf[l-(I-X)n]dX = nf(1-yn)dY =
x
1
l-y
±
m=l
n/m.
o 0
49 Let Xi be the number of drawings following the selection of the ith variety
up 10 and including the drawing which shows up the (i + l)th variety, (i ~ 1).
rhen
r-1
nr = 1 + L: Xi'
i= 1
But
00
r-1
E(nr) = N L: (N - i) - 1
i=O
N
= N f.
1=.\
t- 1 '" N
-r+ J
f t- 1 dt = N 10g[N/(N -r+ 1)].
N-r+ 1
Similarly,
r-1 fN fN
var(nr) = N i~O i(N-i)-2 '" N 2 t- 2 dt-N t- 1dt
N-r+1 N-r+1
Hence N
E(X2) -- N- [N + 2-2
II II til + 1 dt- til dt] ~ nN 2/(n+2),
o 0
whence var(X).
C~ J(~=:') W-w+l
. N-n+l '
SeN, W, w) - II,.
= N-IV-w (N-n)(n-l)
W-w w-l
= (N)
W .
Hence E(S) and var(S), using the fact that Xi are uncorrelated.
53 Let Xi be a random variable such that
Xi = 1 if forecast is correct (F) for ith day
= 0 if forecast is incorrect (F) for ith day, (i = 1. 2, ... ,11).
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER I 157
::: p. Mp· c= ~) +q. G=~) ]+q. (n~r)[p. (n~ 1) +q. (II~~~ 1)].
. cl! E(S.) == np - (n - I')(P - q) < np for p > q, and var(Sn) = npq.
Ji,nDenote rain by R and no rain by R. The four mutually exclusive possibilities
(he weather-forecast combination for any day are RF, RF, RF, RF. If
:"r 's IV,I IV2' W3 and W4 respectively are associated with these outcomes, then
.,tift
E(S.) = r(wlP+w2q-w3P-w4q)+n(w3P+w4q),
Ind
II
this ==
whence
° for all /' if W 2 = - w l P/q and W4 = - w3P/q· This ensures E(X i ) =
_ np 2 PI' 2 2
Ihcrerore var(S.) - - .
q
W3 +-.
q
(WI - W3),
.md this == 1 for all I' if WI = ±W 3 • Thus the alternative scoring system is
WI = -
= (q/np)t; W4 = -"'2 = (p/nq)t.
W3
54 For the four possible outcomes RF, RF, RF, RF, the probabilities are
1/1. 11(1- p), p - (1.13 and q - (1.( 1 - 13). For random allocation of I' days to rain,
{he corresponding probabilities are
Xo = I
r
k= 1
(pq)-k and ° ~ pq ~ !.
57 Feller (1952). Ex. 33-35, p. 128.
The event "exactly k balls in n drawings" can occur if
(i) k black balls in (n-1) drawings and white ball in nth drawing·
(ii) (k-1) black balls in (n -1) drawings and black ball in nth dra~ing
Hence difference equation.
58 Huyghens' problem. UL (1963).
Probability of A's winning in a trial is 5/36 and that of B's winning is 6.;,
Hence probability of A's winning at the (2r+ l)th turn is /.
59 UL (1962).
Expected premium in nth year is
UL (19 63 ).
~I . ho ut any restriction the three numbers can be chosen in 11(411 2-1)/3
Wlt,( x is the first number in A.P. with common difference r, then
..IP·.•((211-1) and for any x, r ~ [(2n+l)-x]/2. Direct enumeration
, r :~ -; A.P.'s in all. Rearrangement of the A.P.'s according to the magnitude
I\~S "common difference gives the distribution of X, whence G(O) by sum-
; Ihe
111011.
, rhe probability-generating function of X is G(O) = (pO+qO-l)", whence
~ n == 1I(2p - 1land var( Xl = 4npq. Also.
,I.
G(O) = ±(n~t)p("+11/2q(n-'1/201'
I=-n 2
,llhat
r-I
lIil/J =L (t+l)pq'/(l-qr) = [l-qr(l+rp)]/p(1_qr).
1=0
00
Using factorial moments ')12 = (l-6pq)/npq. whence the result for }'2 ::: O.
69 Frisch, R. (1925). B, 17, 165, and UL (1962).
The cumulant-generating function of r is ,,(t) = n log(q + pe'), when"
mean and variance.
dt
L jI'l(O). t'lr!
,=0
where
or Ie, = pq. d",_ tfdp, since dpldx = pq. Successive differentiation dele:
mines "s.
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER I 161
S" = L" 1 lj
j=
00 ;'" 00
= L e-).., L etNP(SII = NiX = n)
11=0 n'N=O
= f
11=0
e-).. i:',[</J(t)]"
n •
= exp[A.{</J(t)-I}].
(_I)"(r:.u)
and adding
where
SM == L
M-, (r+ u)( .M )
(-1)" for M ~ r.
u=o r I+U
and 0 :::;; Ai :::;; min(ai' bi' s). Out of ai' Ai cards can be selected in
Now consider the B pack. The s cards are fixed, as matches occur, but tho
other (N - s) can be permuted in
(N -s) !
t
ways.
Il (hi-A j ) !
j= 1
Hence
v. = ~
1'1
[(N-S)!.TI
.= 1 (~~)/
I~I
(bj-)'i)!] = Hs·(N-s)! I.D_tl bi !
The jth factorial moment is
N IN
Il(}) = ,~j r(j) ~ ,~o~'
where
N s
= L v..s! L (-l)s-'J(r-j)!(s-r)!
.=} r=}
V. 's-} (.)
\' ~.', L...
N
-_ f... " (_l)S-j-1 S-] -_ 0 un Iess s = j.
s=j(s-J). 1=0 I
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 1 163
N
L rUlJ¥. = lj.j! and
r"'i
,nd
\ariance = [et aibir-N. it1 aibi(ai+ bi)+N2. it1 aibi ]! N 2(N-l).
76 consider jY ~ h [1 + L (Y -/1) - 8~2 (Y _/1)2] to obtain E(J}'),
,,-1 "'~.
na
f i=1
[1_(xi-a)+(xi-a)2]
a a
= a- 1(l+s2/ a2),
so that
10gM(t) '" 4m(I+~)sinh2t/2.
Differentiation gives E(Sn) = 0 and var(Sn) = 2m(1 +m/n), these being e
results since the neglected terms in log M(t) are of power higher than /a(.
80 Anscombe, F. J. (1948). B, 35, 246.
Consider terms up to [(x-A,)/(A,+b)]6 in the expansion for (X+b)1
obtain E[(x+b)t]. I,
81 UL (1960).
Since (2n+ 1) ~ r ~ (4n-l), each of these (2n-l) cases has probabilill
1/(2n-l). Hence .
4n-1
E(r) = L r/(2n-l) = 3n, and
r= 2n+ 1
4n- 1
E(r2) = (2n _1)-1 L r2
r=2n+ 1
2n-1
= (2n-l)-1 L (s2+4ns+4n2) = n(28n-1)/3,
s= 1
whence var(r).
4n-1
P{r> 3(4n-r)} = per > 311) = L(411-1)-1
r= 3n + 1
= (11-1)/(411-1) < 1.
Hence the probability of three such breaks < 1/64 which is significant allh.
2 per cent level.
82 UL (1961), and Fisher (1945), p. 11.
In the first experiment the probability of a success is 1/70, and the lol~i
probability for the lady to establish her claim is
1- P1(0)- PI (1) = 0'0085,
where PI(r) is the probability of r successes in ten trials.
In the second experiment, the probability of a success is 37/924 > Ifill
The total probability for the lady's claim to be accepted is now
I-P2(0)-P2(1)-P2(2) = 0'0063,
where P2(r) is the probability of r successes in ten trials.
The increase in the probability of an individual success in the seeon.'
experiment is more than compensated for by the requirement that the I~dl
has to achieve at least three successes instead of at least two as in the III'
experiment.
83 (i) If Xm is the expected number of additional trials needed to complell"
consecutive realizations when the first ttl letters are in agreement, (0 ::;;; /II ~ .1"
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER I 165
tor 0 ~ k ~ r-1
Ihell il
1-P)2
( 11-
p(EjIA j) = p2; P(EkIA j) = I (k :F i).
lienee
p(AjIEj) = P(EjIAj)P(A j)/ P(E j)
p(EjIAj)P(A j)
p(EjIAj)P(A j)+
P(EkIAj)P(A j) L
k*j
= (lI-l)p2/(1-2p+np2).
For the second part, let Eij be the event that the letters IXj and IX) are printed
succession. Then
III
1_P)2
p(Ej'IA k)
J
= ( -11- I
(i :F j :F k).
lienee
P(AjIEj j) = (11- 1)p/[(II- 2) + lip ].
Finally, let E j • be the event that IX j and a nOn-lXj are printed in succession. Then
p(AkIE,) = ;(E,IAk)P(A k) ,
L p(E,IA)P(A
j=,
j)
k=p
Xk
1Xk+l + (n-n-
=;;. - 1)Yl' 1~ k ~ /'-2,
1 (n-l)
X,-1 =;;+ -n- Yl;
and
1~ k ~ /'-2,
Y,-1
1
=-.x I
/1
(n-2)
+ -/1- YI'
and
YI =
11'- 1 -1
,XI + 11-
(/1-
- -1 .
2) /1,-1-1
'Yl'
(/1-1)/1'- I II' I
Yl = ( )
~ -2 ~ + ()S-l
()S ~ +P' {1- ()S-'}'
~
\\hence
= '[1+ (1-p'){1-(q/nr 1} ]
Xo P (q/n)+(q/n)S 1-2(q/n)s+p'{1-(q/nr'}·
Ii) For large n, Xo '" 1; (ii) For large rand s, Xo '" O.
118 InitiallY, the probabilities are
p(W4) = p4; P(W 3 B) = 4p3 q ; P(W2B2) = 6p2q2;
P(W B 3 ) = 4pq 3; P(B 4) = q4.
lIence after the first splitting
P(W2) = p2; P(B 2) = q2; P(WB) = 2pq.
rhe proportions at the (n + l)th stage are obtained from the algebraic product
[XiW2)+ 2Yn(WB) + zn(B2)j2,
whence
ilild so
Xn = O~+mn.1o -+ O~, as n -+ 00
89 Let the proportions of (W2), (WB) and (B2) after the (It-l)th stage be
'._1, 2sn _" tn - 1 respectively. Also, at the same stage, the proportions of (W)
ilnd (8) are Pn-l and qn-l. Then
rn = (rn- 1 +Sn-l)Pn-l; tn = (Sn-l +tn- dqn-l;
2sn = (rn-l +Sn-l)qn-l +(Sn-l +tn-1)Pn-l;
and
168 EXERCISES IN PROBABILITY AND STATISTICS
Hence
so that
Pn = PO+j[1-( _t)n]~O -+ !<PO+2Pl) as n -+ 00.
90 For demand r ~ n,
00
whence
1tn = (l-n/A)[I-F(n-I)]+P(n-I).
Also,
r-n]
Pn = E [- - = I - n L e - A. A /r . r !
00
r
r r=n
Hence,usingxn_1+Yn_l+Zn_l = I,
Xn = tp+(l-tp-r)xn-l +!<q- P)Yn-l'
Yn = tp+(I-tp-q)Yn-l +!<r- P)Xn-l'
which may be rewritten as
Xn = (l-tp-r)Xn-1+!<q-P)y"-1'
Y" = (l-tp-q)Y,,-l +!<r- p)Xn- l •
where
Xn == xn-pq/(pq+pr+qr) and Y" == Yn-pr/(pq+pr+qr).
The equations are solved by using the trial solution
Xn = A},n and Y" = BAn.
92 The difference equations form three pairs:
Xn = (I+Xn-l-Wn-l)/6; Wn = (1 +W,,-l -xn - 1 )/6;
Yn = (I + Y,,-l -vn- .)/6; vn = (1 +vn- 1- Yn-l)f6; and
Zn = (1 +zn-l -un- .)/6; Un = (1 +Un-l -z.,,-1)/6.
To solve the first pair, transform to
6Xn = X n- 1- w,,-l and 6w" = w,,-1-Xn- 1,
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 1 169
where
X ::=
"
Xn -! and w" = Wn -i· Hence, using trial solutions Xn = AAn ,
\I~ :::: B).",
Xn= [1+(xo-wo)(W- 1 ]/6
Wn = [1 +(wo - x o)(j-)n-l]/6.
Similar results hold for the other proportions, and as n -+ 00, all -+ f;.
3 The probabilities for the nine types of chains are obtained from the
~lgebraiC product
[piA2 +2Pl(1- pdAa +(1- pl)2a 2] x
x [p~B2 + 2p2(1- P2)Bb +(1- P2)2b 2].
Xn = x n- 1 +iAn- 1 ; Zn = zn-l-iAn- 1 ;
Yn = Yn-l -iAn-I; Wn = W n- 1 +iAn- l•
(iii) For A = Il = 0,
qn = q~- 1/(1- 2qn-l + 2q~_ d
or q;; 1 -1 = (q;;..! 1 _1)2, whence q;; 1 -1 = (qi) 1 _1)2"
For limiting behaviour, consider
lOge -qn).
qn
170 EXERCISES IN PROBABILITY AND STATISTICS
and, in general,
Mn-(3k-4)Mn-I+2(k-l)(k-2)Mn-2 = 0,
so that
Mn = A.n_ I M 2 +lin_ I M.
The coefficients A.n and lin are obtained from
lin = - 2(k - l)(k - 2)A.n- I ;
A.n = (3k-4)A.n_, -2(k-l)(k-2)A.n- 2 ·
The elements of the transition matrix [M/2(k-l)]n are thus obtained as:
(pq;rs) = 2(1-0"-1)/k(k-l);
(pq;ps) = (k+2)(1-0n-I)/2k(k-I)-O(l-0"-2)/k:
(pq;pq) = (l-on-I)/(k-l)-20(l-on-2)/k,
where 0 == (k - 2)/2(k - I). Hence
2( 1 - on-I) on-I [ " ] 2/k(k 1)
(a)
X12 -
_
k(k-l) + 2(k-l) j';-I Xlj
(0) ,,(0)
+ j;-2 x 2j -+ -,
as n -+ 00,
where the transition matrix T = M - 21, 1 is the (~) x (~) identity matrix and M
is as defined in the preceding exercise. Then
Tn-(3k-1O)T"-' +2(k 2 -9k+ 16)T"-2 +4(k-2)(k-4)T"-3 = 0,
so that
Tn = A.n_IT2+lin_,T+Yn_,I,
where A., Ii and I' are scalars.
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 171
rhe clements of the transition matrix [T/2(k - 2)]" are
2 48'; 20"
(pq; rs) = k(k -I) - k(k-2) + (k-I)(~-2);
20'; (k - 4)0"
(pq ; pq) = (pq; rs)+ k_ 2 + k _ 2 2;
0" ()'!
(pq; ps) = (pq; rs)+ k~2 - k~2 ;
I rc
\lIe
() 1==- (k-4)/2(k-2) and O2 == -1/(k-2). Hence
97 UL(1963).
Let Xi be a random variable associated with the weather on the ith day,
,u.h that
Xi = I, if forecast is correct for ith day
i= I
and
u(x) =
00
q2M-X'r~0 (-,1,)'.
(x
~!
ra)(r)! 00
r~o (-A.Y.
(2M
;! '
ra)(r)
x> O.
99 UL (1964).
The probability P(R.) = (1- p)p'- 1/(1_ pm). Hence G(O) by summation.
Mean and variance obtained simply from the cumulant-generating function.
~s III -. 00, rna -+ 0, so that max E(S") = 11,1,(1- p) < 11, provided that
I. < (I-p).
172 EXERCISES IN PROBABILITY AND STATISTICS
100 UL (1964).
(i) pn+(l_p)n; (ii) pk(l_p)n-k; (iii) pk(l_p)n-k+(l_p)"pn-k.,
(iv) (")pk(I_P),,-k;
k
(v) t
r=k
(n)pr(l_pr-r;
r
(vi) (n-S)pkqn_k.
k-s'
(vii) n-",("-S)
L p.+r qn-s-r.
r=k -s r
101 Woodbury, M. A. (1949). AMS, 20, 311.
P(Il+1,x+I) = px.P(n.x)+qx+1.P(n.x+l)
PI'
G(x,O) = 0(1- Oqx)-I . G(x -1,0) = Ox. G(O, 0) /fJ. (1- Oqr)'
G(x, 0) = Ox / rU (l - Oqr),
r=O
n (I-Oqj) ==
L Ar j'i'r I.
P(n,x)
1 x-I (P-+i ) .
= -, n Lx (X)
(-1)' . (q-cr)"
x . 1=0 e r=O I
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER I 173
~ re the probability-generating function of x is
I h~re,o
G(t) == L" t X Pen, x)
x=o
= coefficient of
e" in eQO[ ( 1- t) + t e - cO] - pic,
n!
J he factorial moment generating function is obtained by putting t = (1 + a).
111 I
iI(nce
,I,ll
== e[(I+c)"-l];
('
]l(2) = !!.(~+I)[(1+2e)"-2(I+e)"+I];
e (
ii, II
== e(E-l-
('e
I) (!!.+2) [(1 + 3e)" -
e_
3(1 + 2ef + 3(1 +e)"-1];
I'IJ)
== e(E+
('e
I) (!!.+ 2) (!!.+ 3) [(1 +4e)"- 4(1 + 3e)" + 6(1 + 2e)" -4(1 + ef + 1].
c e
,-or liP == ex, lie = p, we have
lim Pen, x)
n~ '"
= .n
X-I(a .) e-"'(l-e-py
1=0
-p+1 . ,
X •
= .n (a+pl) -e-"'(I-e-P)X
x-I
1=0
.
, -p- ,
X •
rherefore
If N j is the number of flowers on the ith tree which bore no fruit, then
G(e) == L arer
r=O
so that L ar = 1
r=O
from the given form of G(e). Hence G(e) will represent a probability distribllt'
if ar <I: 0 for all x. Clearly ao = G(O) > O. Again, G'(e) = A,1 G(e)f'(9). Th~o
fore, It
pH 1)(0) = r(n+l).A.~+1 i
s=O
(-A,2)s. r(s+k+2) .
r(s+ 1) r(n+s+k+2)
Thus for n = O,f(H 1)(0) > O. But for n =F 0,
pH 1)(0) = nA,~+1
s=O
i (-~2Y.B(n,s+k+2)
s.
f
1
A,1A,2
"3 = 11+ 1
[1 + 11+2
6A,2 6A,~ ]
+ (11+2)(11+3) ;
A,1A,2 [1 14A,2 36A,~ A,~]
"4 = 11+ 1 + 11+2 + (11+2)(11+3) + (11+2)(11+3)(11+4) .
The expression for 11 is obtained by eliminating A,2 from the ratios "2/K1 and
"3/"1·
105 UL (1964).
(i) (~)(~~)/( ~~) - 0·43885
. (48')/(52) -0.10971
(IV) 12 13
M(a) 1-4(~~)/[C~)-(~~)] - 0·36964
(b) 6(1~)/[(~~)-(1~)] - 0·30666.
1h~ changed probabilities are
106 UL (1964).
n n n
Chapter 2
--
I k::: 2; E(X) = j; var(X) = is; median =
, The probability density function of X is x e - ",2/2 .
l-I/.j2.
f
,,/2
m= e-fJ'an"'dx,
o
-hence the differential equation.
7 The probability density function of X is 2a(a 2-x 2)/(a 2 +x 2)2.
Use the transformation x = a tan 6 for evaluating mean and variance.
R The probability density function of X is !(x) = a.xfJ - 1•
IIcnce E(X') = a./(a + r). Therefore the central moments are:
177
178 EXERCISES IN PROBABILITY AND STATISTICS
a 2a(a - I)
J1.2 = (a+ l)2(a+2); J1.3 = - (a+ 1)3(a+2)(a+3);
3a(3a 2- a + 2)
J1.4 = (a+ 1)4(~+2)(a+3)(a+4r
Hence coefficient of skewness
= / 3/2 __ 2(a-l) (a+2)1.
YI-J1.3J1.2 - (a+3)' a '
coefficient of kurtosis
_ 2 2 6(a 3 -a 2 -6a+2)
Y2 = (J1.4- 3J1.2)1J1.2 = a(a+3)(a+4) .
E(e'X ) = t f
00
e-(1-tlY y2 dy
-1 0
The first four cumulants are 2, 3, 6 and 18 respectively. For the median,
~ 1+~
;e f
-I
e-x(l +X)2 dx = t or f e-
0
Z • Z2 dz = 1, whence result.
10 UL (1962).
E(e'Y ) = fo J 00
oo
exp[ -!x 2+t(a+px+yx 2)]dx
x fo J 00
oo
exp-t[x(1-2yt)t-tP(1-2yt)-tFdx, where 1-2yt>0,
fI~nce
x V2
f f f
x X V2
p(x) = f ajbr
a
-00
_1_. e-(t-x)'/2a 2 dt.
0=
v
"2' p+0'8(1- p)V +0·2(1- p)"2-(x+c).
V
IIcnee
dO . ~E. = __1_ and dp
-=0 gIves dx O·4V dx
dx
180 EXERCISES IN PROBABILITY AND STATISTICS
For a maximum,
d 28 0-4V ( )212 2 •
--.[(a-x)e- a -." (1]<0 If a-x<O.
dx 2 a3 J2rr
15 Wadsworth and Bryan (1960). Ex. 12, p. 225, and UL (1961).
For given x, expected profit = (c 3 -c2)E(x)-c l , whence result.
Also, bankruptcy results if profit ~ - C4' i.e. if x ~ (CI - C4)/(c3 - e2).
16 The roots of the quadratic are
XI = [a+(a 2 -4b)t]/2 and X2 = [a-(a 2 -4b)t]/2;
and b is uniformly distributed in the range (0, a 2 /4).
Hence
E(y)
12
= n3 f
,,/2
l2
"2sinO.O(n-O)dO = 12l2 /n 3 , and
o
E(i) = n123 f
,,/2
l4
"4sin20. O(n-O) dO = l4(n 2 +3)/8n2 , whence
o
var(y) = l4(n 6 + 3n4 -1152)/8n6 •
21 (i) The inequality x(2/-x) ~ [2/2 is satisfied for the regions 0, 1(1-1/j2)
and 1(1 + 1/j2), 2/. Hence probability is (1-1/j2).
(ii) The inequality x(2/- x) ~ 3/ 2 /4 is satisfied for
1 31
-~x~-,
2 2
and so the probability is t.
(iii) The inequality
is satisfied for
1-(6-n)t.
6+n
12 Gordon, R. D. (1941). AMS, 12, 115.
E[g(x)) = f00
-00
e _(X-II)2/2 • ex2/ 2 [1_ «I>(x)] dx
f
00
f f
a 00
f (y+a)f(y+a)dy+ (y+a)f(y+a)dy
- 00 0
o
f
00
= a+ f yf(y+a)dy+ yf(y+a)dy = a.
-00 0
Similarly,
00 00
= 0 for odd r.
(iii) For c > m, define
00
S(c) == f Ix-clf(x) dx
-00
f f
C 00
= (c-x)f(x)dx+ (x-c)f(x)dx
- 00 C
= f
m
-00
(c-x)f(x) dx+ I
00
m
(x-c)f(x) dx+2 I
m
c
(c-x)f(x)dx.
Hence
C
26 UL (1958).
The proportionality factor is l/(1A which is obtained by using the
Ir;lnsrormation y = log x.
00
E(X,)=_1_fx,-le-(I08X-mll/2al dx =_1_
(1fo 0 (1fo
f 00
-00
e'Ye-(y- m l l / 2c71 dy
f ,x
1
kf -~=2kf
1+x4
-00 0
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 2 185
w:== 1/(1 +X4), so that k = 2/B(i, i)·
,.Ibere
1
= 2;r!)' j
o
C t (1-wot)-tdt,
II
hence the
t
result, using term-by-term integration after expansion of
Ii -lI'olr .
\.I A ramous result due to K. Pearson.
'.1 hence
_ [(NP-X+ 1)(r-x+ 1)]
Yx+l - Yx (Nq-r+x)x
Ayx 1 dy
Ayx = Yx+l-Yx and - - '" -'-d .
Yx+t Y x
For bl = b2 = 0, the normal distribution with mean a and variance bo
I' obtained .
.IS Cadwell,1. H. (1952), B, 39, 207.
x
.11111
lIence
2X2 2X4 7x 6
4F(x)[I-F(x)] '" 1--+---+ ...
n 3n 45n
-2x 2/" 2(n-3)x4 (60-7n2)x6
'" e + 3n2 + 45n 3 + ...
'" -2x 2/"
e
[1 + {2(n-3)X
3n2 +
4 (60-7n 2)x6 ... } 2X 2/,,]
45n + 3 e ,
,I hence the stated result.
186 EXERCISES IN PROBABILITY AND STATISTICS
= B(
qn-r
1 _)' L r (r) p" qr-·.B(n-r,r-s+l)
r+ ,n r .=0 s
= qn-r . Lr
B(r+ 1, n-r) .=0 s
p' qr-. (r) f1
zr-s(1- zr-r-1 dz
o
_
-B(
qn-r
I
r+ ,n-r
).
f
1
(l-Z)
n-r-1
.(p+qz)dz,
r
o
by reversing the order of summation and integration,
q
= 1
B(r+ 1, n-r)
. f yn-r-1(1_ y)r dy, y=q(l-z)
o
= G(q).
38 P(X
~ x) =
1
1:
1
+ fo' 0
f
x
e
-x 2 /2 d
x,
and the first result is obtained by term-wise integration of the series fore-";
For the second result,
ANSWERS AND HINTS ON SOLl'TIONS: CHAPTER 2 187
successive integrations lead to the series. Finally,
',lhe Oce
IR.(x)1 =
(211 + 1) ! e - x2/2foc·
e - (1 2 + 2Ix)/2
2., fo, n! . 0 (t+x)2n+2 . dt
(2n+l)!e- x2 /2 dt f00
(2n)!e- x2 /2
< 2n.n!fo '0 (t+x)2n+2= 2n,n!fo'x2n+I'
1
(x 2 +2r)-1 = x- 2 L (-I)k(2rNx 2k
k=O
'valuate AI' A 2 " , , ,A, as power series in inverse powers of x, The stated
::~,~It is obtained by direct substitution in the expression for P(X ~ x).
.ro Direct integration gives k = (2n + 1)/2.
(i) P(X ~ 0) = (1 +sin2n+ 10)/2.
p(IXI ~ 0) = 2P(X ~ 0) = 2[ 1- P(X < 0)] = (1- sin2n+ 10).
Iii) P( -7t/6 ~ X ~ n/6) = 1t)2n+ I.
E(X) = 0 and var(X) = E(X2) as stated, where
f
71/2
fo' I
(i) Set .~
Ux = e- x2 / 2 dx.
1 f00
(I 2\111 -x 2 /2 d - 0
fo'-oo U x '4- UxJ e x - .
fo'[
(Jx
(iv) Integrate
x2n a~+ 1(I-axY" = tx2n (X~(1-(XxY" +x2n ux(!-u~y".
188 EXERCISES IN PROBABILITY AND STATISTICS
f
IX)
1
'4 cos-
1 [
-
w2 -8W+8]
w2 ' for 1 ~ W ~ 2.
(oc+l}(oc+2) «(<<+2)/2 II r:
B(oc+3,P+l),U .x .(i-x) (yx-u)dxdu.
f
=-:(IX:-+_l_}(,--OC_+_2.:....) . u«
B(IX+3,P+I)
(P)( -1)'
,=0 r
f
1
50 If Xl and x 2 are the vertical and horizontal distances of P from the bUll
eye, use the transformation Xl = R cos (}; X2 = R sin (}, whence the distrib.
tion of R. Therefore
P(R ~ r) = e-r2/2a2 and P(rl:::; R :::; r2) = e-d/2a2 -e -r~/2al.
For n trials with k in the annulus (r 1 :::; R :::; r 2), the probability is
~
j3 . e -v 2(z2- z -1)/3(12 . lid
v v. dz, (- 00 < v, z < )
00 ,
61[(1
. ce the distribution of z is
.h~n j3 dz
2;' (Z_!)2+! .
~refore mode of the distribution is z = Also, z = 1 - t, where t.
,Ih (Xt- X2 )/(X3 ~X2)' But by symmetry z and t must have the same mode,
.h~nce mode ='2'
<\ (illl has the F distribution with (2,2) d.f., i.e. du/(l + U)2, (0 ~ U < 00).
P=n(n-1)
co
f
- 00
f(Xl) dx l
00%"
f[f
Xl XI
f(X) dX r- 2
f(x n)
Xn
[l-f f(X)dXfdXn.
Xl
f f
XI Xn
U= f(x)dx, v= f(x)dx,
-00 -<X>
192 EXERCISES IN PROBABILITY AND STATISTICS
where u ~ v ~ 1; 0 ~ u ~ 1. Hence
1 1
P = n(n-1) ff (v-u)"-2[1-(v-u)j2dv.du,
a "
whence result.
56 Carter, W. H. (1933), JIA, 64, 465.
1 50
= 4N ,~ [(t+ 2)(t+ 1)t(t-1)-(t+ 1)t(t-1)(t-2)]:: I.
50
E(X) = N- 1 L (50-t)(t 3 -t)
1= 1
50
= 50-N- 1 L (t+1)t(t-1)[(t+2)-2] = 9·6.
1= 1
k J
2
(x 3 -x)dx = 1, whence k- 1 = 1561248
and
so
E(x) =k f2
(x 4 -x 2 )dx = 40·0053, so that E(X) ,.., 9·9947.
Similarly,
E(x 2 ) = 1667·0024, and var(X),.., var(x) = 66·5744.
2S
and the probability for two consecutive outcomes ,.., 0·0039 < 1 %.
ANSWERS AND HINTS ON SOLUTIONS: ('HAPTER 2 193
. rter W. H. (1933), J I A, 64, 465.
" (8 hod similar to that of the preceding exercise.
\Id
~I
= N- 1 [H H
50 I~l t(t+ 1)- I~l t(t+ l)(t+2)+2 I~l t(t+ 1) = 1.
H]
49
E(49-X) = I (49-r)P(X = r)
r= 1
49
= N- 1 I (50-t)t(t+ l)(t-1) = 28'8, and E(X) = 20·2.
1= 1
, rnilarly,
49
£[(49- X)(48- X)] = N- 1 L (50-t)(t+ l)t(t-l)(t- 2),
1= I
f
49
Iknee
E(50-X) '" E(x) = 29'7650, so that E(X) '" 20·2350.
111o,
E(x 2 ) = 987'1429, whence var(X) '" var(x) = 10H877.
f
25
P = n(n-l)
2'
1:
f
el2
e-ll/<dt
1
f
e-It
e-(n-l)12/<dt
2,
o II
., hence result.
194 EXERCISES IN PROBA B I L ITY AND STATISTICS
rr rr+ 1
kr .'-!-kr + 1 .(r+I)! '
by successive integration. Hence
<Y.l
= -1 f f e -(x'+)")/2 d x d y - 1- f f e -(x'+y2)/2 d x d y
2n 2n
o qxlh 0 qxlh
= V(h,q)+2Inf
<Y.l
o qxlh
f e-(X'+Y')/2dXdY-2~f
'" h '"
0 0
f e-(X'+Y') /2 dxdy
f
rf) tt/2
If the region of integration in the (x, y) plane having area hQJ2 is the sam.
as the area of a sector with radius R and angle a, then R 2 = IIq/a. Thercfol(
a h sec9
V(II, q) = 21n ff
o 0
e- rll2 r dr de,
21n ff
o 0
e -r2/2 r dr de,
ate by parts successively, using [_h2ze- th2 (l+z2)] as the first function
Inlegr .
'ach integratIOn.
11 t The alternative expression for the general term is obtained by putting
:h~orem.
k f'"
-000
fOO x 2e - yx 2 dx d y _ k foo
(1+lxlt -
-00
dx
(1+lxl)a-
_ 2k foo _d_
x_ - I
(1+x)a-'
0
1 x2 e- dx
YX2
f
00
dx
E(Y) = (IX-l) 2 -+ 00.
x (1 +x)a
o
196 EXERCISES IN PROBABILITY AND STATISTICS
But
65 UL (1963).
Marginal distribution of X is
(l-X)fZ XP+ 1 dx
(0 ~ X ~ 1).
B(a+ I, P+2) ,
Marginal distribution of Y is
1(1- y)fZ+ 1 dy
(0 ~ Y ~ 1).
B(a+2, P+ 1) ,
\Iake the transformation wdrx = r sin (), W2/rx = r cos (). Then, for fixed () in
In ~ 0 ~ n/4), (0 :::;; r :::;; sec ()); and for () in (n/4 :::;; () :::;; n/2), (0 :::;; r :::;; cosec ()).
Integration over r gives the distribution of (), whence the distribution of
II ~ tan (I is
l-or ". and 112 ~ 00, E(u) ~ 1 and var(u) ~ O. The limiting form is u = 1 with
probability one.
198 EXERCISES IN PROBABILITY AND STATISTICS
Therefore
(2n-1)2 (
E[-(2n-1)logw] = n(n-1) ",4 1+ 4n 2 '
1)
and
(2n-1)2(2n2-2n+ 1) ( 3)
var[-(2n-1)10gw] = n 2(n-1)2 '" 8 1+ 4n 2 .
Assume that - (2n - 1) log w '" PX 2 with v dJ. Then p and v are obtained
by equating the first and second moments of the exact and approximalin~
distributions. Thus
Therefore
3{ 1 + 4 v(n i )}]
var[ -(2i1 -1) log zd '" 8k [ 1 + 4n 2 '
~~
ee, for unequal samples. k
-2logVk = L x1/ni>
i= I
,\here xl is an independent l
with ~ dJ. The approximation follows by using
rhe lechnique of the precedIng exercise.
71 Hyrenius, H. (1.953), J ASA, 48, 534. .
(i) 0 ~ T ~ 1 If U2 ~ VI and T > 1 If U2 > VI'
Therefore, for 0 ~ T ~ 1, the conditional distribution of T given UI is
(nl -1)n2 I . H~ I (n2 -1)(1-UIY( _ T)H2-,-1 dT x
(l-ulf,+H2 ,=0 r
f
I
x (z_u.)H,+H2-,-2dz.
II,
For T> 1, the limits of integration for z for fixed UI and Tare
1 +uI(T-l)
UI ~ Z ~ T '
(lVx
Tt = f
vx-I' (t,x-l)-v x ·f(t,x),
~/ = kljJ(t)[N e- kT -v.J
VI = e- kT • NkT. In general, if
-kT (kTy-1
vr_I=Ne . (r-l)!'
then
~r + kljJ(t) . Vr = kljJ(t). Vr - I
gives
_ N _kT(kT)'
Vr - e -,-.
r.
(ii) A similar method to (i) leads to the negative binomial.
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 2 20]
Irwin, J. O. (1941). Discussion following E. G. Chambers and G. U. Yule
7~ JRSS(S), 7, 89.
'19~)idently, E(xIA) = var(xIA) = A.. The cumulant-generating function of A is
v _rlog(1-tlc), so that E(A) = ric and var(A) = rlc 2 • The marginal dis·
jll~ "
Irlbution of x IS
x . AX c'. e C' (x + r-l) !
Je-XI'
<n A,-I dA
(r-l)!
-C). •
=(l+c)x+"x!(r-l)! forx~O.
o
lIenCC
G«(J) = c'/(l+c-(J)' and E(x) = ric, var(x) = r(l+c)/c 2 •
~rom the joint distribution of x and A,
_
E(XA) -
co
x~o x
f co
e -). . AX +I c'. e - c). • A' - I dA
x! . (r-l) !
o
co c' (x+r)! 2
= x~dl +c)x+,+ I' (x-I)! (r-l) ! = r(r+ 1)/c .
whence
x+r
E(Alx) = - ,
l+c
1S Liiders, R. (1934), B, 26, 108.
(i) If in any time-interval Vk accidents occur, each of which involves j
persons, then the probability of the configuration (VI' V2" .. Vk" .. ) is
co
n e-
k= I
Yk • ('lktk/Vk !,
whence
co
P(X = r) = L n e- Yk • (Yktklvk !,
k=1
(ii) Given .the con~guration of. accidents ~Vl' V2" •• ), .the probab'r
generatmg function of the accidents sustamed by a given person :slly.
n (OPk + qk)Vk.
00
k=l
L k. Vk = r, with r -+ 00
k= 1
and
00 00
Fl (t) = f I
o
f(t) dt
and
T
G(T) = AT -t(1-e-UT),
2
and
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 2 203
C J11pbeIl, N. R. (1941), JRSS(S), 7,110.
-1 :he interval 0 < T < To, if first replacement is made in (t, t + dt), then
In t (III -1) more replacements are made in (T - t), so that
·I~as
. T T
. h interval (r- 1) To < T < rTo, (r > 1), clearly F~)(T) = 1 for all r > m.
1:11 e,( III divide the range into
i ,If r.... '
(i) from 0 to T-(r-l) To < To;
(ii) from T - (r - 1)To to To; and
/iii) from To to T.
In (i). if the first replacement is made in (t. t+dt). then since (r-l)To <
I ./ < rTo. the contribution to F~)(T) is
l' - (r - 1)1'0
,lnd the probability of a replacement being made at the post during the
Inlerval (t, t+dt) irrespective of when the bulb was put in is
d~~(t). dt.
f
T
f
T
G(T) =
m=l
f Fm(T), and dGd(t). dt
t
= g(t) dt.
f
.Y.
P = g(T-X)SI(X) dx.
o
Hence the expected lifetime of a bulb replaced at t = T is > t is
(I-P)SI(t).
Also for (ii), the probability of a replacement bulb not replaced at t ::: .,
having a lifetime > t is
x
fo
g(T-X)SI(X)SI(t+x)dx.
P
Pn -- (l-p'+p) [1-(p' _p)n] where PI = p,
,
whence
1
--:----:-:----:-;0 ,.., 1.
1-(p' _ p)2
and
pp' = [1-e-).X (1 + ~)] [1-e-).X (1 + A.:) _A.: e-).(2T-X) (l-e- 2ATl]
n! n f(uj) dUj,
j=
n
I
and that of Zj is
n
n! n dZ
1= I
h
where
f
U;
11 ! ;,
i= I
PII+ I = 1 -
j=l
L: Pj
.,1 th at the joint distribution of mi' Pi is
m! n! 11+1
n+1 . n Pi' n d Pj'
IIIi
II
n
i= I
,i=1
l1l i .
j=1
IIcnee successive integration over p", P,,_I , ... ,P2' PI gives the result.
HI UL (1964).
The proportionality factor is rx/2 obtained from
III <XI
C III
f
0()
+ (x-c)e-a(X-m)dX}
III
H2 UL (1964).
l
b 0()
HJ UL (1964).
Hence
P[S ~ E(S)] '" P[x ~ (1- /312)].
n(l-xs)"-I dx s ' (0 ~ Xs ~ 1)
and Ym = 2p(1-x;)l.
Hence
nl2
2f
r
I
(r~ 1) re': 2) = .fir: .r(r+ 1).
( 2
EYm)-4np
_ 2 r _
J(1 x )(1
2
s
_
x s)
n- 1 d _ n(n + 3) 2
x'-(n+l)(n+2).4P .
o
For large 11.
4p2
var(Ym) '" Jt(n + 2) . var(xs),
85 UL (1964).
x [1+(~sin-IkI2) (1-~sin-IkI2)];
E(I) = 48pln 3.
For 0 ~ k ~ 2, ~sin-I
n
kl2 ~ 1 so that P3 ~ P2'
86 Bartlett, M. S. (1936), PRS, 154, 124.
The Jacobian of the transformation u = X IX2, V = (l-xtl(1-X2) is
IXI -x21, where (XI -X2)2 = (l-u+v)2-4v ~ O. Combining the contribution~
for XI-X 2 ~ 0 and X2 -XI ~ 0, the joint distribution of u and v is
2[(I-u+v)2-4vrt du dv.
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 2 207
°: : ;
. ed v in (0 ::::; v ::::; I), u ::::; (1- Jv)2, and so the marginal distribu-
v is -log v dv: (0. ::::; ~ ::::; 1 ) . .
,II I1X .
In of onditional dlstn butlon or II given r IS
The C
2[(I-u+vf-4vrt(-)ogv)-1 du, 0::::; u::::; (1_Jz.)2.
c1ly• 1/ ::::: e-(Ox : , w~erle) X2 has 4 dJ., and so the marginal distribution of
2
I rl
Iknee P2 < 3, if
3ke- k2 / 2
J(k) > 3-k2
ANSWERS AND HINTS ON SOLUTIONS
Chapter 3
---
Ifr and r2 are the number of ones in the first and second n/2 observations,
~hen /== r1 +r2' and rl' r2 are independent binomial variables.
Plackett, R. L. (1953), Bs, 9, 485.
£(n r ) = NAre-A/(l-e-A)r! for r ~ 1, whence result.
Use factorial moments of a binomial distribution.
(i) Unbiased estimate of pq is Q* == r(n - r)/n(n - 1),
and var(Q*) = E(Q*2)_p2q2 = pq[(1-4pq)+2pq/(n-1)].
n
(ii) Unbiased estimate of pq2 is r(n-I')(n-r-l)/n(n-l)(n-2).
r 2r2 r3
(iii) p*(I_p*)2 = ---+-,
2 and
n n 3 n
(n-l)(n-2) (/1-1)
£[p*(l- p*)2] = pq2 . 2 + pq . - 2 - -+ pq2, as n -+ 00.
n II
() _ p(1+p-2pn)+pn+I(l-p)[n2p-(2n-l)] E2(r)
varr - (l_p)2(1_pn+l)
-+p/(l_p)2, asn-+oo.
Also, £(r2) -+ p(l +p)/(l_p)2, as n --+ 00.
5 UL (1961).
Let Xi be the number of successes in the ith turn; then Xi is an indepe d
random variable. and n eOI
But
00
_r
n-r
= ~ = e[I+(p*-P)]
I-p* q p
[1- (p*_p)]-l
q
= e [1+(P*-P\!
q pq P
f (P*-P)'q-']'
r=2
whence E(_r_).
n-r
Again.
(-
r ) 2
=P22 [ 1+ 2(P* -p ) +2
1 L (t-l+2p)(P*-p)'q-'.
C() ]
n-r q pq p ,=2
Hence var(_r_).
n-r
( r = p2 )2 [1 + 2(P+..1.)(P*-p) +
n-r+ 1..1.2 p..1.
+ p2'~2
1 (* )'
P ;p {(P+..1.)[(t+l)p+(t-l)..1.])
C() ]
which leads to the variance.
I herefore
E(xr) = or Je-X/8(~)r+
o
1 d(x/O) = ()' [,(1'+2), so that
20 2
E(x 2) = var(x)+E2(x) = -+ 40 2,
n
whence
E(~X)2 = 202( 1 + 21n).
10 The probability density function of X is
(IX + 1)(1X+2)x a (1-x) so that 1X>-1.
rhe roots of the maximum-likelihood equation are
* _ (3g+2) ± (g2+4)t
IX -
- 2g .
~ow - 00 < g :::; 0, and IX* -+ - 1 as g -+ - 00. Hence the positive root for
It.
II
whence
E(X) = eJ.l+O'2/2; var(X) = e 2J.1+O'2(e U2 _1).
Equate the mean and variance with x and S2 to obtain two simultaneous
212 EXERCISES IN PROBABILITY AND STATISTIC'S
E(X i ) = mi;
L = (~)(e-m)n(l_e-rn)N-n,
whence
m* = 10g(N/I1) and var(m*) = (l-e- rn )/N e- rn .
In the second case,
P(X = 0) = e- rn ; P(X = I) = m e- m ; P(X ~ 2) = 1-(1 +m)e- lil
•
m(e'" -1-11l)
=var(I11*).( e '" - 1)(1 -11l+m-e
2 Ill)'
1 +11 3')1
2
(.(11\+113)= N(l-p); p* = 1- (11 -----r;;- .
~-~-~+~=O
2+/3 2-/3 1-/3 1+3/3 .
Ihen p:::: (1-0)2 so that var(O) = var(j)/4p.
IrQ = ala4/a2a3' then
4 64(5+2p-4p2)
= ;~1 l/E(a;) = 3N(4- p2)(1_ p)(l + 3p)'
18 UL (1963).
The sample likelihood from the two experiments is
By direct integration,
• 2+'2'2
/,/1 E(x 2 ) = (11 /,(12 + 1l/1
E(X) = 1+..1.; 1+..1.
so that
-7:
f f(x)dx = 1-
It may be shown that
dE(X) dm
~ > d/1' if (12 ~ (11 and /1 > O.
(1+AP2)(P+).V-~(1+A)3p2 = O.
J
E[(Xr)"] = r(n+ 1)r(k+r)
0( r(r)r(n+k+ 1)'
the mean and variance of X r •
,·~nce E(Xk)
Also,
= 0("/(k+1), for k ~ 1. This gives
E(x) = 0(/2 and var(x) = 0(2/12n.
"" 2v+ 1. the median is x v + 1 so that
,Ir n
E(Xv +l) = 0(/2 and var(xv+Il = 0(2/4(n+2).
, Since families with no abnormal children are excluded, the total proba-
,I I'S (1-t/'). Therefore
. lilY
E(r,) == i
r.=1
ri (k.) pr. q"-r'/(l_q")
r,
= kp/(1-t/'), for i = 1,2, ... , n.
lherefore for the average of the rio E(r) = kp/(l-q"), whence the moment
;,Iimate of p is obtained from
_ kp*
r = 1-(l-p*t'
\Iso,
E(n/r) = L (n)
00
n=r
-1)
- (n _
r r 1
prqn-r = L
s=o
00 (r +s)
s
prqs = p-l.
Ilul
q
E(r/n) = r~.
q n=r
I: (n =11) f t
r
n- 1 dt
o
q
= r(!!.)rftr- l .
q
I: (r+S-1\
s=o s J
tSdt
o
q
whence
E[g(1,0)] = p and E[g(2, 0)) = p2.
Also,
= p2 f tf/(r + ss - 1) .
• =0
Then
-Q) - a)
ff
00 00
= 2~ exp{-!(u 2 +w2)}.u'[w(1-p2)t+pul'dudw,
-00-00
Also. since
x ( u)(1+-V)-l
-=..1. 1+-
Y ml m2
~r~fore
Hence
.fficients of variation
,(lI:
vdJn, v2IJn respectively. Therefore
E[Y
x] [1
= . 1. 1 + In(V2 -
OJ (2k)! (V2 )2k-l]
PV1)' k~l 2k. k! In -+ ..1., as n -+ 00.
\\ a first approximation,
2~ Let '11
'2 and '3 be the number of insects killed at strengths x = -1,0,1
"spcctively so that' 1 +, 2+, 3 == ,. Hence the joint likelihood of the sample is
1.=
( ")[
'1
1
1+e (a liJ
1'1[ e-(a-l)
1+e (a 1)
]"-'1 X
(n)[
'2
1
1+e- a
]'2[ e- a
1+e- a
]"-'2
X
('3n)[ 1+e 1 (<<+1)
]"[e-(a+l)]"-"
1+e (<<+1) •
_n_l__ n2+n3+ n4 = O.
2+p* 1-p* p*
I or the bias in the estimate of var(p*) assume that E(p*) '" p. Then
~[P(1-P)(2+P)] = (5+3p+6p2+4p3)
dp2 (1+2p) (1 +2p)3
.I'hence result.
218 EXERCISES IN PROBABILITY AND STATISTIC'S
28 UL (1960).
Minimize with respect to a and f3
30 UL (1961).
The probability distribution of 0 is [120(n-0)]/n 3 • dO, for 0 ~ (J ~ 11/:.
and Il = ir2 sin 20, whence E(Il).
To estimate r, minimize
n
n == L [ll -3r2(n 2 +4)/8n 3 ]2
j with respect to r.
i=1
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 3 219
.11 UL (1961).
Minimize n
n == L [(Xj-ml)2+(yj-m2)2+(zj-ml -m2)2]
j= 1
n
= L [(Xj-X)2+(yj_y)2+(Zj-Z)2]+
j= 1
,\2 UL (1962).
Minimization of
n
n == L [(xj-md2+(yj-m2)2+(zj-m1 +m2)2]
j= 1
n
= L [(Xj-X)2+(Yj_y)2+(Zj-z)2]+
i= •
(2-A)x+O-2A)y+(1 +A)Z
U=
3
Jnd var(u) = 2(A 2 - A+ 1)0"2 j3n, whence the t test.
220 EXERCISES IN PROBABILITY AND STATISTICS
33 UL (1962).
The method same as that of Ex. 31 above. The estimates of (J (J
are Or = (x- y)/2; O~ = (x+ y-2z)/6; Ot = (x+ y+z)/3; and the r~iJ and,
of squares is Ual \,"
n
L [(X/-X)2+(y/- y)2+(Zi-Z)2] with 3(n-1)d.f.
i= I
34 /I
0* = [ ± y,]/±,=
,= 1
r- 2
I
r- 1 ,
and
where
4xI = LXljk; 4X2 = LXj1k; 4X3 = LXjU'
hk hk hk
Variances of the three estimates are all equal to 30'2/8, and each of the Ihr •.
covariances is equal to - 0'2/8.
37 The solution of the normal equations gives for the estimates of 0, I'
(r=/= s),
8f-8~ = z -Kx +y + z); 8~-8t = y-i(x+y+z);
8t-8f = z-1(X+ y + z).
The residual sum of squares is the within-sample sum of squares \\;It
3(n-1) d.f.
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 3 221
the hypothesis H(8 1 = 82 /a = 83 /b) is true, then
II' hen
E(xj) = Aiel; E(yJ = A2 e l ; E(Zi) = A3el'
;L~ the appropriate sum of squares by minimization.
,\
Minimize
I n
n == L: [Yr-C(- pr(x r -x)]2
r=1
"2(
L... r Xr-X
-)2
-
n(n +
4 1)2 . (- -)2
Xw-X
r= I
Xw = JI rx,jJI r.
tlll'residual sum of squares is
19 Minimize
n
Q == (C(2 + P2)-1. L (/XXi + PYi + 1)2 with respect to C( and p.
i= 1
Sq = ~ q(2n-q+ 1) f3
'-- Ym-qrx- 2 ..
m=n-q+ 1
Minimization of S; + S; gives
_ q(2n-q+1)l:Yr-p(p+1)l:Ym.
Q- ,
(2n-p-q)pq
b = 2[pl:Ym-ql:Yr1.
(1 ~ r ~ p; n-q+ 1 ~ m ~ n).
(21l-p-q)pq ,
42 The variables Xr and Yrl.ji have the same variance. For the least-squar"
solution minimize
h( 1I I +1I2-1)dJ. Var(m*) = -1.a 2/(-1.1I1 +11 2). Hence the t statistic for testing
.11
1/1/11 "" 1110)·
,Ind
var[bl(X2-XI)] = E[br(X2- XI)2]
= EA(X2 -Xl)2 E,(bilx)]
( ~)(X2-XI)2
nl +n2 SX
a 2 b2 v2
a+b=1; N+-=2". (1)
n s
II .
" IS chosen as stated, then with a = N /(N + II) and b = n/(N + n), a + b = 1
Ind a2/N +b 2/11 = l/(N +n). Clearly (1) have an intersection since 1/(N +n) ~
224 EXERCISES IN PROBA BILITY AND STATISTICS
w = (ax+bY-II) ( -+--
a2 b2 )-t
N 11
is a normal variable with zero mean and variance (J2. Hence its unconditio .
distribution is the same. Therefore w/s has the t distribution with (N _I)~.;I
Hence
var(ax+by) = v2 (N-l)/(N-3) ~ v2 for large N.
45 Use Stirling's approximation for the r function to prove that
If x and s are the sample mean and standard deviation, and In, (J the corre\.
ponding population parameters, then
11 -
E(s/(J) = ( -2- 1) - r (11)/
t
-2-1) ~ (1- 411'
2 r (11- 1) and
E (~)
X
= AE (~)
(J
f:
r=O
In - 2r E(z2r)
Ar(i) A2r(2r)! 00
Similarly,
n-I ]
+4 i~1 zt zt+ 1+ odd power terms .
Ihcrefo re
(311- 4)0"4
whence var(d 2 ) = ( I 2
11- )
41 UL(1962).
\\hcnce
whence resul t.
49 UL (1961).
The condition for Y\ and Y2 to be uncorrelated is obtained from first
principles by evaluating
n
cov(Y1, Y2 ) = 0"2 L aibi'
i= 1
\lote that (Xi-X) and (Xj-x) are particular linear functions, whence the results
'Iatcd.
~ UL (1962).
Let S = s, +S2 so that
cov(S I> S 2) = Mvar(S) - var(S I) - var(S 2)]
0"2
= 2"[I1{I+(11-1)p}-v l {I+(v l -l)p}-v 2{1+(v 2 -1)p}]
= 0"2pV\ V2, whence corr(SI, S2)'
226 EXERCISES IN PROBABILITY AND STA TlSTlCS
51 UL (1963).
Sample mean of X values = (n2+ n4 -nl -n3)(X/N == A, 1(X.
Sample mean of Y values = (n3+ n4 -nl -n 2)f3/N == A, 213.
A,3 = (1l 1-n Z-1l3+ n4)/N.
The correlation is obtained by evaluation of the two sums of squares and Ih
sum of products. For zero correlation, t
whence result.
53 Greenwood, J. A. (1938), AMS, 9, 56.
Consider any two cards of the B pack. There are two cases. I.. Cards ali
of same type. II. Cards are of different types. I and II can occur III
. I X , )(N)
t-l. - -N-l
P(X = 0, Y= 1) = -
t 2 a
1 N-a
N-l j(N)2
P(X = 1, Y= 0) = -.--.IX,
t
1 a-I
P(X= I,Y= 1)=-.--.IX,
N-l j(N)2 .
t
for II, '
P(X = 1, Y = 1) = ~. N ~ 1 . 1X2j(~)
P(X = 1, Y= 0)
1
= (' N-a-l
N-l .1X2j(N)2
P(X = 0, Y= 1) = [~. (~~II)+e~2)(N~I)] .1X 2 /G)
P(X=O, Y=O)= [~·(Z=~)+(t~2)(N;~~I)] .1X2/(~)'
Ihe last two probabilities are obtained by considering separately the possibili-
Ii~S that the first card of the A pack mayor may not belong to the type of
Ihe second card of the B pack. Hence the joint distribution of X and Y.
Finally, let z" Z2,'" , ZN be random variables each associated with one
"f the cards of the B pack such that
md for i #- j,
corr(Zj, Zj) = corr(X, Y).
~ Put X -m, = wand Y-m2 = z. Then
X = w+ml = . 1.
Y z+m2
(1 +~)
m1
(I---=-+ Z22_"')
m2 m2
w z) (WZ WZ2)]
",..1. [1+ ( m 1- m2 - ml m2 - m~- mlm~ .
Z2 (1)
Psing the first approximation X/y", ..1.[1 + (w/ml -z/m2)] gives E(X/Y) '" . 1.
Ind the approximation for var(X/Y).
The approximation for the bias in E(X/Y) is obtained by considering the
l1~xt terms in the expansion (1) for (X/Y) and noting that E(WZ2) = 0 for a
'Ymmetrical distribution.
Hence E(X/Y)-}, '" ..1.ViV2-PVl)'
228 EXERCISES IN PROBABILITY AND STATISTICS
a
r(I1-2)
-2(11-1)1
= J~'r(I1;I)" -2- , by direct integration.
1
21t(1_p2)!'ex p
[1 2 2
2(1_p2)'(U +V -2puv) dudv,
]
(-00 < U,V<rf.1
so that
r = [p+RJ(1-p2)!]/[(1-p2)p+p2+2p(1-p2)!Rf]t,
whence the result.
y= 1
pa.+1 r(ex+l)·e -(x+m)/P ( + )a
. x m, ( ~ X <
-/11...., 00,
)
where
4(11 + I)p(l- p) 211p + 1
{J:= (1-2p)/2 > 0; ex:= (I_2p)2 > 0; /11:= 2(1-2p) > 0,
'!llhat
2(x+m) 2{r(1-2p)+(I1+ l)p}
fJ (I_2p)2
\1 hence the result.
~9 nl2 2ul,!;
[t/>(aW '" 2~ ff
o 0
e- r212 .rd,.dO,
2r
y. e
-y2/2 d _ (2r)!
y - 2r . r !.
-00
1 ~-2U2 -t (0 ~ w ~ 1),
B(n/2,!), w (l-w) dw,
and the distribution of y = - n log w is proportional to
e->'/2. y-t [1-(y/2n)+O(n- 2)]dy.
The result follows by approximating the quantity in the square brackets a\
exp( - y/211).
62 This is K. Pearson's well-known PA test. Each Pi is an independcnI
variable with a uniform distribution in (0, 1).
63 UL (1962).
Obtained by using the general result of the next exercise.
64 The joint distribution of x I and Xn is
x,
r- 2
f(xn)dxn,
(x I ~ Xn ~ 13; ex ~ xI ~ 13)·
For fixed XI' make the transformation R = Xn-X I, and then for fixed R
integrate for x I in (ex ~ x I ~ 13 - R).
If X is uniform in (0, 1), then the distribution of R is
o
Ivhence the result by expansion and term-wise integration.
Similarly for the range (a ~ y ~ 2a).
Direct integration gives
P(O ~ y ~ a) = 22 .+ I B( 1 1
v+ , v+
1)'
r=O v
±(~ r
)(-l).-r/(2V-2r+1)
- 1 . • (V) -11 JI dz
I~O
Z21
- 22'+ l B(v+1,v+1) t ( )
o
1
= 22>+ 1 B(V +
1
o
f
1, v+ 1)' (1-Z2)" dz
B(v+1,!) 1
= 2 2.+ 2 B(v+ 1, v+ 1) = 2,
Expansion of the last exponential and term-wise integration over r then gives
Ihe distribution of e as
e- n/ 2V2
~" (2n)i,2r(n +
2
j)
- - - - . L:. .cos n- 2e sini e de,
/it r(II;1) j=O vJr(j+ 1)
Rut v = [n/(n - 1)] cot2e and cot( - e) = - cot e, so that by pooling, the
distribution of e for (0 ~ e ~ n/2) is
2 e- n/ 2v2 (2njV2)i
-1 ..
00
w= C: r. 1 cot (),
the distribution of w is
In r ( n~ 1) 'L
j=O r(j+ 1) { I+ n }(n+ jl/ 2
(n -1)w 2
for 0 ~ w < \1;,
and
f n+i)
(-IY(2njV2yI2r (-2-
{ n } (j+2l/2 (1l-1)t
(n-l)w 2 -n- dw
for - 00 < w ~ O.
1 (I'.2)t . exp {I
n0'2' ] . II
- 0'2 [12t 2s 2+ v(s - O'c.) 2} s dt ds.
;. (I')t
1 .2 . exp{ - vt2c~/(t2 + 2v)} x
2 ,·e./(t 2 + 2,·)1,2
t2 3t4 ]
v = tc [ 1--+--+0(v- 3)
y 4v 32v 2 .
fherefore E( v2r + 1) = 0 and
log ( -11) 2 .
- '" - -
11-1 211-1
lienee, putting
11 = (2p +2q -l}/2q,
_. (
X .- 2p-1 )1.2/2q
or x'" (1-2st/2.,
2p+2q-1
234 EXERCISES IN PROBABILITY AND STATISTICS
(0 ~ Un ~ 1).
ff
o -a:)
f(x,y,p)dxdy = t,
. y X
(1) z=--p.-; w=
xJ1- p 2
; and
0'2 0'1 0'1
f
CXl 00
d~ d1'/
= 12 f dx dy ,f(x, y) dx dy,
-00 -00
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 3 235
ntegration by parts with respect to x and y successively,
on "
---it f f
- n2 1- p
2'
00
_ 00
00
_ 00
ex p {- 1 2·[(2-p2)(x2+i)-2PXY1}.dXdY
2(1- p )
6
:::;J4_p2'
~cnce result since for p = 0, R = O.
•~ Make the transformation
Uj = (xj-ml)/O'I; Vj = (Yj-m2)/0'2, (i = 1,2, ... ,n).
[hen, setting Wj = (Vj- pUj)(l- p2)-!,
n
Q2 = L (ul+wl)
1= I
i' distributed as X2 with 2n dJ., since uj and Wj are independent unit normal
..Iriables.
Also,
n
R2 = Q2 --1- 2 ' (u 2 +iJ 2 -2piW),
-p
.Ihere ii and iJ are the averages of the Uj and Vj' The quadratic form of the
means is distributed as l with 2 dJ. independently of R2. Hence by Cochran's
Ihcorem R2 is distributed as l with (2n - 2) dJ.
Alternatively, use characteristic functions. If q,(t) is the characteristic
function of R2, then the characteristic function of
R2 + 11(1- p2)- I(U 2+ iJ2 - 2puiJ)
1\ If - 2il) - 1 q,(t) because of the independence of the distributions of the means
,tnd second-order moments. But, directly, the characteristic function of Q2 is
11- 2irrn. Hence q,(t) = (1 - 2it)-n+ I.
75 De Lury, D. B. (1938), AMS, 9, 149.
The joint distribution of u, v and W is
(n-l)n-I { (11-1)W(I- PU)}
n2n(1- p2)(n-ll/2 nn _2) . exp 2(1- p2) x
x (v 2 - u2)(n-4)/2 . wn - 2 • v(l- v2)-! . du dv dw,
where 0 ~ w < co; 0 ~ v ~ 1; - 1 ~ u ~ 1.
But II = /'V, so that for u > 0, (u ~ v ~ 1),
and for u < 0, ( - II ~ V ~ 1).
Inlegrate the joint distribution for wand v successively, whence the distribution
"f II.
236 EXERCISES I N PROBABILITY AND STATISTICS
=
(1
-P
2)(n-1)/2 00
. L
r( + 2·)
11 J .p 2j +1 (l-t)(n-3)/2 t (2i+1)/2
f 1
_1_. exp
27t0"10"2
{_ty2(U:+~)}
0"1 0"2
.IYI dydu, ( - 00 < u, y < 00).
j=
L1 Xj and L1 Yj
j=
are independent normal variables with zero means and variances mrf ami
110"~ respectively.
For the distribution of w, set tj = xJYj. Then the joint distribution Ilf
t h t 2, ... , tn is
I = --
E(X Y= y)
b2 0'1
. - . y and E(Y X =
I x)
b2 0'2
= --b .-.X.
b, 0'2 3 0'1
p'-
00
(n-1)B(n-!,!) [ x2 ]-n+ t
f1(X) = nO"l(2n-4)t . ~ + 2(n-2)O"f '
whence the variance by direct integration. The expectation of the conditio.
variance of Y for variation in x follows since E(x 2) = O"f. nijl
The semi-axes of the equi-probability contour are 0"1",(2n-4)i ad
0"2",[(1- p2)(2n -4)]t. Hence the volume outside the contour is n
f
00
83 '"
Pearson, K. (1923), B, 15, 231.
Since n > 0, the probability will be finite if the equi-probability contour,
are ellipses. Also the bivariate density should be zero along one of the ellipse,
within which the total probability should lie. Hence the joint density funcliu/;
must be of the form
1 {X2 xy y2}]n
f(x,y) = k [1-(1 -p 2)C2 2
0"1
- 2p - + 2 '
0"10"2 0"2
fl(X). var(YIX = x) = f
-a
(y-p :: x) f(x, y) d(y-P :: x),
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 3 239
fl(X) = M:
r(n+2) [X2 2 '
. 1
]"+t
Y 2n . 0' 1(n + 2)tr(n + i) 2(n + 2)0' 1
The semi-axes of the equi-probability contour are 0' 1 I/I(2n + 4)t and
.!~[(I_ p2)(2n + 4)]t, so that the total probability outside the contour is
f
1
f(x, y) . 4n(n + 2)0' 10' 2(1- p2)tl/l dl/l = (1- 1/1 2)"+ 1.
I/!
85 VL (1961).
Let x and s be the sample mean and standard deviation obtained from the
'Irst experiment. Then oe == (x-m)/s is known. If the required number of
plots is (v+ 1), then
* 120(2n+ 1)0'2
var(fJ ) = n(n2 _ 1)(n + 2)(311 2 + 3n + 2)'
t~ = (a-pb)2/s2(A,1 +A,2p2_2pA,3)
giving two roots for p. The stated result follows for A,3 = O.
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 3 241
Geary, R. C. (1930), J RSS, 93, 442.
ql put u = x + a and v = y + b. Then the joint distribution of u and z = vlu
f f
a/a, 00
1 az-b
fo· -a/a, e- w2/ 2 dw+ -00 R(z) dz = 1, where (0"~+O"fZ2)i = w,
:J+ f
00
or I-p[lwl ~ R(z)dz = 1,
-00
= z'+,),.
~'hcre z' satisfies the conditions of z for uncorrelated variables. Therefore
az'-(b-')'a)
[O"~(1- p2)+ O"fz 2]i l
I he number of ways a triplet (r, s) can occur in the field is (M - r)(N - s).
242 EXERCISES IN PROBABILITY AND STATISTICS
and
n2P4 n3P4
E(r4) = n IP4 +-
1 -+1
-PI -PI-P2
Using these as estimating equations and solving,
rl
pf = -; *=
P2
r2(nl-rl)
.
nl nl(nl +n2- r d'
* r3(nl-rl)(nl+n2- rl- r2 ) .
P3 = ,
nl(nl +n2 -rl)(nl +n2 +n3 - r l -r2)
*-1
P4 -
* * *_
- PI - P2 - P3 -
r4(nl-rd(n l +n2- r ,-r2)
.
nl(nl +n2 -rd(nl +n2 +n3 - r, -r 2 )
The sample likelihood is
4
L = constant x (1- pd- n2 (1- PI - P2)-n 3 n pi',
i= ,
whence the same estimates of the p's. Transform L in terms of the (J's; thel
maximization gives
0*2 nl-rl. 0* _
-- r3+ r4- n 3
--' 3-
nl r2+ r 3+ r4- n 3
The asymptotic variances are:
II~oce
E[M!r 2s2)] = e[(n-I)/2n),2a 2 and E[e'XhHr 2s2)] = e'~+tr2,,2 = E(y').
ence the estimates, their efficiency being inferred from the properties of
,h 2
I and s .
By logarithmic expansion,
I gll.k---
-o
1 _ k(k-l) [1 2(k-2) 2k
---+
2 -6k+5 O(
+ n -3)] ,
11 311 3n 2
"Ilhill
k(2) 1
A.k = 1--+-[3k(4)+ 16k(3)+ 6k(2)] + O(n- 3),
n 6n 2
.• here in factorial notation kim) == k(k-l)(k-2) ... (k-m+ 1).
Therefore
lienee
var(O) = 0(1-0)(2+0)[2(I+p)-(I-p)O]2.
2M[2(1 + p)+(l + 7p)O]
244 EXERCISES IN PROBABILITY A NO STA TISTICS
Now
and
var(O) = 20( 1 - 0)(2 + 0)/( 1 + 20)N I'
For censored data, the expected proportions in Ab, aB and ab classes.
~1- 0)/(2- 0), (1- 0)/(2- 0) and 0/(2 - 0) respectively. Hence equation r~/I:~
IS
YI+Y2 Y3 N2
1 - 0* + 0* + 2 _ 0* = O.
n! n f(u
n
i= I
j) du j , ( - 00 < UI < U2 < ... < Un < 00).
Next, put
u,
Zj = f f(x) dx, for i = 1,2, ... , n,
-00
so that PI = ZI and Pj = Zj- Z j_ 1 for (2 ::;; j ::;; n). Hence the joint distribolioll
of the pj is
n! nn
j= 1
dpj,
:d
E(mjmjlp's) = m(m-l)pjpj.
Dixon, W. J. (1940), AMS, 11, 199.
.j fhe marginal distribution of the p's is as obtained in the preceding exercise,
.d the joint distribution of the p's and m's is
.
n !m ! n n dpj.
n+l p.ml
_1_.
j= mj! j=
1
n
Ir.cexpre~slion for CPt. iS obtained by reversing the order of integration and the
dltinOmla summa IOn.
[02f/J]
o()f /1'5=0 = n .'J[(_l
n+l
)2 +(~_~)p+(m-l)p~]
m n+l m
Ii dp. 1 1 J=1 J
n !m ! n _1_,
n
j=O n dpj'
p,!,1
mj. j=O
n-l
with the limits for the p's
Now
where
2cjJ ]
[aao~. i2 ( ---
_- 2"+ 1 2i) (m-l) 2
E(PI)+ - - E(PI ), PI == L PJ
i
I 00=0 n m n m j'=O
E(d4) = ±[a4~]
1=0 aO I 0'0=0
+ 2 L [ a24cjJ 2]
i<j aO I aO j 0'0=0
d (d 2) _ (2n+ 1)(8n+ 7)
an var - 180n2
[These results do not agree with those attributed to Mood by Dixon.]
100 Mood, A. M. (1943), AMS, 14,415.
P(X, X) = p(xIX)p(X) =
e)(:=:) (~) .
P(X)
,
o ~ x ~ min(n, X).
by hypergeometric summation.
Hence
n(N -n)J.I. [ It] n(2) 2
var(x) = N(2) . 1- if + N(2) . (j .
p(xIIX I _ d = (XI-l)
XI
(N I - l
nl-xi
-XI-l)! (NI-l),
nl
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 3 247
ince X, = X,_I-X" Ni = N,_l-n,.
I~ds
" p(x,IX i- 1) = p(XIIX,_ l ), 0 ~ Xi ~ min(Xi_l> N i).
= coefficientofzN in (l-z)-«-l
r=O
f (P+N-r)zN-r
N-r
= coefficient of ZN in (l-z)-«-l
1=0
f (P+t)ZI
t
= coefficient of ZN in (l-z)-«-1I-2 = (a+ P~N + 1).
248 EXERCISES IN PROBABILITY AND STATISTICS
E = NB-fi/ = N(N-n)
fi(N-fi) fi(N-fi) X=Ox=o N-n
[f i (X
-x)P(X,X)_~ P(X)]
Nx=o n ,
i
where
E =
N(N -n)
( _ ).
n [(x+
L fi] 1)
- 1 Pn+1(x)-N·P.(x) .
fiN fi x=a+l n+
If x has a discrete rectangular distribution,
P(x = r) = l/(n+ 1) == P.(x), (x = 0, 1, 2, ... , n).
Hence
E = 2(N-n)(n-a)(a+l)
N(n + l)(n + 2)
if X is also rectangular. Its maximum is attained for
=
(4N+3)!
(N !)4
f m
'" (PIP2P3P4) dpi dp2 dp3,
N
4 [04¢]
4
E(C )
2
= L -4 + L [ ~ 02 4> 2 ] . :l
i= I Ct 1'.=0 ioFj (li
i 1'0=0 (tj
4 32 (N+1)(N+5)
E(w) --+ 3(N + 2) and var(w) --+ 27' (N +2)2(N +3)(N +4)'
4 32 m-1
E(w) --+- and var(w) --+ 27 .~.
3m
1117 Mathisen, H. C. (1943), AMS, 14, 188.
The joint distribution of ml and P is
'£11+
1) ! m !
1I1!)2"'1!(m-md!'P
n+ml (1
-p
)n+m-ml d
p,
(0 ~ ~
~p~;
1 0 ~~ml~m.
~)
250 EXERCISES IN PROBABILITY AND STATISTICS
For the rth factorial moment of ml' use the marginal distribution of .
the point probabilities Iti. Wilt,
E(v r)-_(~)(~)r[_I
22
__+_r!_ ±(n+k-l)!
n+r (n+r)!'k=O k! .
2r-k]
,
so that E(v) = a/2; var(v) = a2/2(n+ l)(n+2).
For (- co < w ~ 0), the distribution of w is
n lwlfi ]n-l dw
[1- J(n+l)(n+2) --+ _1_e-lwlfi dw as n --. 'J
[2(n+l)(n+2)]t· fi '
By symmetry the same limiting form holds for (0 ~ w < co).
The characteristic function of x is
and var(y) =
[ {n}2
= a2
(n+l)(n+2)
2
i=2
L
iC j -(n+ 1) n (i-l)ci+ n i(i+ l)cf-
i=2 i=2
L L
n }{ n } {n (n + l)a 2 a2
-2 (
i -\ } ]
.L
,=2
iCi .L (j+ l)cj +2 .~ (i+ l)ci x .~ jCj
)=2 ,-3 )-2
+ 2(n+2) '4'
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 3 251
,n~ n+l n
CI = -- -
2
L ic,.
j=2
{heCCfore for minimum variance
r-I n
, _J)cr+2r L (j-l)cj+2(r-l) L jCj = (n+l)(r-l), (r = 2,3, ... ,n).
,~r j= 2 J=r+ I
The distribution of t is
1 +n(k+ 1)]
o~ y ~ [ n(k + 2) . a.
252 EXERCISES IN PROBABILITY AND STATISTICS
f f
x. ~
J
(x-lII+ko)/a
Q= fo 1
e -,,2/2 d u.
(x -III-ko)/a
Therefore
E(Qls) = r-c
y2n
1
J
00
e -v2/2
1
d v. M:
y2n
f
v/./ii + k./a
e -,,2/2 d u
- 00 vjjn _ ks/a
ks/a
fo (,,:S b:(-'C.:l)w'J·dW
Hence
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 3 253
1
::::-~.
(1I-1)(n-I)/2
(11-1)
II I 'J'_
exp[-t{(/1-1)+(}2}z2].zn- 1 dzdO,
V 2,. 2(n - 3)/2f _ _
2 -I 0
where 0: = ( 11 )'
II' --- -
11+1
,
'I'I! rl!sull.
*-
•• 1
pul
(x-m)/u =~, s/u = '/, and ct>(z) = fry:; e-·I•2 / 2 . dy.
var(Q) = (8Q)2
~~
v." ~=O
. var(~)+ (OQ) ~
vII
2
~=O
. var(I71.
1 k
- N log(ufo)- 2u 2 j~1 [(lI j -1)st+llj(Xj- ~y],
n, n,
where Xj = L1 xjllj; (lIj-1)st = L (xij-xY
j= j= 1
Ihen (0 ~ R < 00) and (-n/2 ~ () ~ n/2). Also, for given rand k,
U 2r + k
r(-)k R2r+k . k(}
V y-m = (2r+k)/2' . sm .
/1
lienee the results on integration.
var(m) =
E [.t
,=1
wt ePt(Xj, m)+
k
.L. WjWjePi(X j , m)ePj(xj , m)]
,*) 2
116 (i) L~t the ordered sa?1ple obser~ations be XI ~ X2 ~ .. ,' ::::;. x2n , I "
that X n + I IS the sample median. Assummg that X n + I > m, the dlstnbution 1,1
x n + 1 is .
(2n+l)![, 2 ]nf( )d
(n !)2 4-eP (Xn+l) Xn+1 Xn+l,
where
f
Xn+1
Also for large samples, f(xn + tl ~ f(m). Hence the approximate distribUliol,
of X n + l .
(ii) The Cauchy distribution has the probability density function
1 1
f(x) = ;'1 +(x-m)2' (-00 < X < 00)
and
00
_
E
[o2 10g f (X)] _ ~
om 2 - 7t'
f 1-(x-m)2 d =
{1 +(x_m)2}3' X
I
2'
-00
117 The estimate of ex is ex* = vlx, and the characteristic function of thl
distribution of the sample average x is
it )-nv
( 1-- .
nex
C:2f f f
coco
= ~ [{n(n-2)}t+ si n- 1
nn
(_I_)],
n-l
\ya polar transformation.
e
Lei = x-mx and" = y-my. Denote the joint density function of e and
~ by f(e,,,,
p). Then
I~I
f f f(e,,,,
co
ff (f(e,,,,
~
f
g=1I
(g+k)
lI+k
= PIli (1I+II+k)
11=0 II
p-u
= L coefficient of X O in x- u (l-X)-II-h-I
11=0
= .
coeffiCient of ,
x P - 1I In 1/(l-X)"+k+2 = (P+k+ I) .
lI+k+1
Therefore
E[(g+r-l)(r)] = (1I+I'-1)!
(/I-I)!
f
Y=II
(g+I'-I), ~(P)
11+1'-1/ 11 1
= (_11_) (p+r) ! .
11+1' p!
121 Goodman, L. A. (1952), JASA, 47, 622.
If g and s are as defined for the sample, then (11- 2) of the sample obscrvOi'
tions must lie in the range (g, s). Therefore, the joint distribution of g and .1 I'
(11 ~ g ~ p; 1 ~ s ~ g -/I + 1)
E[(d+r-l)(r)] = (p+r)I(d+r-l)(r)
p-I (d =1)/()
p -I (d+r)(r+l) (d-')/(l'l
_ p-I II
d=1I-1 11 2 11 d=II-J /I 2
[he mean and variance of R are also obtained by using this method of sum-
mation. Thus
N-I
E(R) = (N-1)-2N-1I I R"
R=I
2 (N-Ir+1 (N-1)(1l-1)
-(N-I)-NII' n+l 11+1
,lIld
= 2nR' JJ
o 0
exp {- 2(I~P2)' (x 2+ y2-2pxy) }. dx dy;: / (saYI
where x = (Xj + 1 -Xj }/j2; Y = (lj+1 -lj)/j2.
*
But J = + (l/2n) sin - I P by the medial dichotomy theorem of Shepp d
This result is proved by formally inverting the characteristic function :r
and)' to obtain '
0/
i)= ( 21 ) 2 JOO Joo exp{-1(t,+t
I 2 2 } 1 1
2+2pt,t 2) dt,dt 2 = - · _ _ _ .
p n _ 00 _ 00 2n JI=P2
Similarly (using particular suffixes 1 and 2 instead of j and j + 1),
P(C,C 2 = 1) = 2P(u 1 > 0, VI > 0; U3 > 0, V3 > 0)
+ 2P(u I > 0, VI > 0; U3 < 0, V3 < 0).
The inversion of the characteristic function gives the stated result fill
P(1I 1 > 0,1'1 > 0; 113 > 0, tl3 > 0). In the same way, after sign change U[ I,
and (4 in ¢, integration gives
1
P(Ul> 0, VI > 0; 113 < 0, V3 < 0) = 4n 2[(sin -1 pf -(sin -1 pj2)2]+
1
+-[sin- 1 p+sin- 1 p/2] + a constant.
4n
Hence cov(C" C 2), which still involves an arbitrary constant. This is evalualed
by using the condition that for p = 0, cov(C" C 2 ) = l6' (from preccdillf
exercise).
126 Chown, L. N. and Moran, P. A. P. (1951). B, 38, 464.
Use the results of Ex. 125. For C j = Cj + 1 = 1, the contribution hI
P(CjCj + 1 = 1) is
2
P(C j C j + 1 = 1) = 2" [(sin- 1 p)2-(sin-l p/2)2] + a constant.
n
Hence E(CjC j + 1), the arbitrary constant being determined by using th ..
condition that for p = 1, E(CjCj+d = 1.
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 3 259
The inequality for var(g) is established by proving that
P(u j > 0)
11
""
2n
f
= _fol . exp{ -(u j-pf/2I1 Z } dUj = t+rb(.).
o
lienee mean and variance of S. Also,
P(x > T) = fo .f en
II
e-%2/ 2 dz,
P(x < T) =
8
f
-'"
fo' e-%2/ 2 dz,
l
8
(NP)
x
(Nq)
y
(NI1-X-Y
- Np -Nq)/ (:), 11
(0:;;:; x, Y:;;:;I1;O:;;:;X+Y~Il).
Therefore
y
E [- X] = Ln n- ( x )
L - P(x, y)
y+l y=Ox=O y+l
= np f (_1)(Nq)
y=o y+l
n~Y(NP-l)(N-Nq-l-NP-l)/(N-I)
y x=l x-I n-y-I-x-I n-I
n-I( 1 )(Nq)(N-Nq-l)/(N-l)
= np Y~O y+l y+l n-y-l n-l
f (M)u (N-M)
u=O n-u
= (N)
n
Hence the moments of x and y. To obtain the coefficient of variation of
x/(" + 1) use the following result:
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 3 261
If X and Yare random variables with coefficients of variation VI> V2' and
rrelation p, then, as a first approximation, the coefficient of variation of
,0 . ( 2+ 2_2 )t
.\'If IS VI v2 PVI v2 •
31 UL (1964).
I folloWS directly from first principles.
33 UL (1964).
I The estimates are: cr:* = (x+y+z+w)/4; In = (x-y+z-w)/4;
R.",(-x+y+z-w)/4; x,y,z,w being the averages of the x, y, z, and w
pI •
observatIons. .
The error sum of squares IS
n
M == L [(Xj-X)2+(Yj- y)2+(Zj_Z)2 +(Wj-w)2]+n(x+ y- w-z)2/4,
j= I
with (4n-3)d.f.
The t statistic for testing the hypothesis H(fJ 1 = fJ 2) is
(x-y)[n(4n-3)/2M]t, with (4n-3) dJ.
134 UL (1964).
The random variable r has a binomial distribution and
Xr = L
j= 1
Yj'
and as the Yj are independent, the moments of Xr stated in (ii) follow im-
mediately.
262 EXERCISES IN PROBABILITY AND STATISTIC'S
v= I
L (Yv- b I X lv-P2X2v)2 =
v= I
n
= L (Yv-PI X lv-P2 X 2v)2+(b l -PI)2S 11 -
v= I
n
so that
Hence
E(S:) = (n-2)cr2+(1-r2)cr~Slh
whence the unbiased estimate of cr 2. For large samples, use
(P2 - P2)/ S.E. (P2) approximately as a unit normal variable.
log(l +,1.*),..., 10g{!!.~) + [m~~Ji.~] _t[m~~Jl~J2 _[111'1 ~Ji.'I] +trl1l 'l ~Ji./I]: I
\it I Ji.2 Ji.2 Ji.1 Ji.1 (I
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 3 263
<Xl
mr' - * Px*-I"IN
- 't..." x r Px, - JxI •
x~l
E[(m~-JL~)(m~-JL~)] = ![JL~+s-JL~JL~].
Therefore (1) gives E[log(1 + ,1,*)] and var[log(1 + ,1,*)].
138 This generalization follows closely the method of the preceding exercise.
k-l
If h(,l.) == L:
r~O
,1,r/r !, then
P(X = r) = ( m+r-l)
r pr q-(m+r)/(1_ q-m), for X ~ 1.
Therefore
(lII+r-1) I
E[X(S)] = Lex
I . I' prq-(m+r)/(l_q-m)
r~s (r-s) . (m-1) .
= (m+s-1)(S) pSq-(m+s) t
I ~O
(111+ t+ S-1) (p/q)I/(1_q-m),
t
whence the stated result.
264 EXERCISES IN PROBABILITY AND STATISTICS
log[l +(m+ l)p*] = log[l +(m+ l)p]+ log [1 + (V2 ;~J1~) ]-log [1 +( ~)]
where J1~ is the sth moment of X about the origin. The large-sample result.
for p* are now obtained by applying the method of Ex. 137 above. S
141 UL (1964).
Let Xl' X2 and X3 be the observed frequencies in the AA, Aa and aa classes.
Then ~ = (X2 + 2x3)/2n, and the exact variance of ~ is obtained by noting
A
f
nl2
fO-le-tIO.dt = e-(V-I).t(1_e-.t).
(v-I)T
X 2X3 = log
[0(2+0)]
(1-;- 0)2 .
where
_E[a2~~LI] = a-3E(y2) = a- 2.
The n pairs (Xi' Yi), (i = 1,2, ... , n) give
and a2/a 2has an F distribution with (n, n) dJ. Hence, if L2 is the IikeIih
of a single observation of a, then Ood
log L2 = constant - n log a - n 10g(1 + a2/( 2),
whence
alogL = 2n (!_~)/(!+~)
ae
e·et et·
Hence the information for estimating e is
4n 2[
alog L)2] = (j2 (t e)-2] = (2n+4n 21)()2·
E [( ----ae 1-4E e+t
f
1
t S(2-t)Sdt,
o
the transformation v = t(2 - t), which reduces the integral to a Beta
use •
rdO~~:~ single observation of II, the logarithm of the likelihood is
logL2 = constant-(s+ 1)!u-O!+slog{2-e- 1u - 81 },
f
1 1
and
1 1
fo
ts+2(2_£)"-2. dt
s(s-l) s(s-l)
r
= - 2(s+ 1) + (s+ l)(s+2) tS(2-t)s. dt.
J0
The expressions on the right-hand sides are obtained by integration by parts
and are readily evaluated.
Finally, the approximation for the loss of information is obtained by using
the r function approximation
r(2s+ 1) 1
2S
2 {r(s+ 1)V '" .jm'
which follows by using Stirling's formula.
149 Fisher, R. A. (1930), PRS, 130, 16.
The joint distribution of .x. sand Zi' (2 ~ i ~ n-l), is
11-1
n dz;
i=2
x (1 -Z2-
2 Z3
2 -'" 2 )!'
-Zn-l
SO that X, s and the Zi are independently distributed. Next. since the x --+ y
268 EXERCISES IN PROBABILITY AND STATISTICS
transformation is orthogonal,
n
Yi = L cij(Xj-x),
j= 1
for 2 ::;; i::;; n
so that any (x j - x)/s is a linear function of Y2/S, Y3/S, ••• , Yn/ s, i.e. of
Z3, ... ,Zn-l' Therefore X, mi and mom2 o/2 are independently distribul~'
Hence
E(m~m2oP/2) = E(m~)/E(m2P/2).
To obtain the expectations of the powers of m3 and m4' set Xi - J.I =::
so that (Xi-X) = (ui-ii). Then (ui-ii) are normally distributed rando u,
variables with zero mean, variance [(n -1)/n](j2, and III
expected number observed in the interval is N(1-e- A). Hence the expected
information from N sources per unit time. For given Nand 0, this informatioll
is maximized for A. '" 4.
151 Fisher, R. A. (1942), AE, 11, 306.
The likelihood of the sample is
L= Ii (Ar)p~ro_
r=1 a
Pr)Ar-ar (C)pcO_ pf-c,
r
X
C
P* = n P:,
,.:= I
I' k k
(C-A) fl (ar+;,)-(C-A) TI (A r+},) = 0,
r= 1 r= 1
. hieh has a real root in the range (-as, c), where as is the smallest of the a r •
,\ The goodness of fit X2 is
t
r= 1
(ar-Arp~)2
ArP:O - p:J + C P*(l -
(C-Cp*)2 _
P*) -).
'2[ t
r= 1
Ar-ar C-C]
Ar(ar + I,) + C(c - A) .
Ihis X2 has I dJ. since for any distribution of Or and c, I. = 0 if
k
P =1= fl Pro
r=1
J f
J.
,,_ 2
-- II., " ( - l)
L.. k I ( _ k _ 2) I'
k=O . II
k
.
J
co
II -
(XII
k- 2 d OC II J
Xu
k (
OC 1 xn _ Xl
) doc 1 . d';x I •
d'
Xl
x,.----+-Cf) Xt--rf.)
where
U == JIX~
XII
+ 1 dx 1 and V=
-
J OO
II
dlX
1X"- k - 2 _ " dx
dx
n'"
-00 x"
Ilr
E(w) = II!
n-2
L
k=O (k+I)!(/1-k-l)!' -
( _l)k JCG
(l-oc:- k + 1)1X~+ 1. dx n ,
:/.;
E[(w-w)']
n-2
= n !k~O k !(n-k-2)!'
(_l)k J
00
1X:- k -
[lXk+1(_W)'
2 n k+l
rW]
+k+l docn,
xn - -00
270 EXERCISES IN PROBABILITY AND STATISTIC'S
where
f a~+!(Xn-x!-w)'-!dx!.
.Tn
W==
-7.)
But
f w. a~-k-2 f
00 00 00
dan = - dV dan f dW
d '-d .dxn =
W'- V·-.dx
xn--oo -r.
a" x" -00
dx" "
and
f
00
a:-! dan = ~.
.\ " n - - co
If the hypothesis is true, then the logarithm of the likelihood of the pOll.
budget distribution is
(I-a) ]
log L = constant + nllog [I - -a- . (1- P2) +
Ar == (1 ~ p) . exp[fJ,(xr - fJ - fJ,/2)].
. elements of the information matrix are evaluated by noting that
Ir.C
f
fo _
1
00
00
2
y2 e - Y /2 d y
1 + A e I'Y = J.l('C 1 - 'C 2) ;
fo
f1
00
-00
y e- Y
2
(I+Ae-l'y)2
/2 dy
= 2J.l('C2-'C3)·
j=O
n
lienee the equation for p. may be written as
fJ, = X/[I-(N~RnJ(:JJ.
where
fi, == N e- fJ,'/t!
il and fiR == N [1-e- Ji ±
j=O
fJ,j/j!] .
,I hence var(p*).
The maximum-likelihood estimation is straightforward.
272 EXERC'lSES IN PROBA B I LITY AND ST A TISTI C'S
L= Ii
j= I
('.l)pr i q"-r i /O_ q")k,
, j
and E(I'j) = np/(I-q").
For (iii), note that for any ranking of the abnormals, the ranking of th.
normals will consist of (n-I') different integers, none >n but summing tt
tn(n+ I)-A. II
(n -; + 1) t smh(jt/2) j
r 'h (-1)'"+ I t 2m
=tl'(I'+I)t+L
j=
L1m.
I m=
2 (2 )'
m.
.B2m [(n-j+1)2m-lm].
and
E(xyIB) = m~.
Hence
E(x) = E(y) = 1tlml +1t2m2;
Bul g(x) is real only for - III < X < III, and -+ ± oc, according as x -+ += m.
lor a multimodal distribution g(x) = 0 must have more than one root in
_/II <X< III.
Now g'(x) = 0 gives x = ±(m 2 -1 )1-. Hence for III > 1, there must be three
r~al roots of g(x) = O. This will be so if
~
(12
== !(~+~);
2 (1i (1~
15 == !(~-~),
2 (1f (1~
so that 1151 < (12.
L= IJ
c [ n I .
so' IJ S { ( kS) pk qO-k/(1_ qS) }/1
5k
] ,
s-I Il nsk I. k-I
k=1
Hence
c
E(R) = L s. ns . p/(1 - qS),
_E[iJ 2
10g L].
iJt/
If families with at least two abnormals are considered, then the probabT
that such a family of size s will have k abnormals is III)
R' = L L k. n.k
.=2k=2
The total information for estimating q is
c
L s. n~ [(1- q.-t)2 - (s - Wp2q'- 2 ]/pq(l- spq·-t _ qS)2 .
•= 2
r x
But so that
n-r+l [(n+ 1)/n]{I-[nx/(n+ I)])'
E[~]
II-r
= nil (_1
r=O n-r
)(n)prqn-r/(l_pn)
r
f !(n)pll_zqz/(l_pn)
z= 1 Z Z
n'k'
= L L
00 n
..
n - z z/( 1 _ n)
k=OZ=l(Z+k+l)!(n-z)!'P q p
~ n !k ! -(k+ 1) n ~ k~l n ! k ! oj ]
= (1- Pn) - 1[
t....
k=O (n+k+ I)!
•q - P t.... t.... . ,
k=O j=O (k-)+ 1). (n+}).
, •
[ I]
E-
11-"
~
00
n..
k~O (n+k+ 1)! .
'k' (1 + O)k+l and E[-I" ]
n-r
= nE[_1
n-,.
]-1.
164 Let Xj be the number of non-defective units inspected after the (i -1)th
and up to the ith defective unit observed, so that
j= 1
[ o210gL] 2n{(I-p)+(3p-l)e} 2v
- E (102 = 0(1- O){ 1+ (2p - 1)0} + 0(1 - 0);
2
[0 log L] 2nO(I- ( 2)
-E Op2 = (l-p){I+(2p-l)e};
e - 'l,~.'
P(X = r) = [1-(1 +AO)e ).],. !' for X ~ 2.
lIence
.nd
var(X) ~ A![I-e-).(1 +,{)].
168 Cohen Jr., A. C. (1960), JASA, 55,342.
'Xl
Let /Ix be the number of nests with x eggs each, so that L nx = (N - n).
x=2
_ n + N e-1l = O.AA
(2)
~ l-e-1 (l +AO)
[ l -1 ]
x* = l 1+1-(1:l)e- 1 '
For small A,
(X) _ ~ [l-(1+A)e-).][1-(1+A.O)e-).].
var - N' [(1_e- A)2_A.2 e-A] ,
(0) = (1-0)[1-(1 +AO) e- A][(1-e-).)(1-(} e-).)-A(1-(}+A.O)e- AI
var NAe A[(1-e )')2_A 2 e-).] ----=;
and
A 0) _ (1-0)(1-A.-e-).)[1-(1+AO)e-).]
cov(A., - N[(1-e-A)2 _A2 e A]
169 Plackett, R. L. (1953), Bs, 9, 485.
00
= -1[00
L r2p,_ {OO rp, }2] ,
L ;'"
var(A*) where p, == ( A ;
N ,=2 ,=2 e -1)r!
and
_ A (l-e-)')2
var(A) = N' 1 _ (A + 1) e -).' whence eft' (A *).
1 10
Write T == -2
N
L b,n,
,=2
so that
But
'L 00
~ence
E((}*) = e- A•
co co
cov(T1• T2) = -N L L P2r-lP2s' whence var((}*).
r = I s= I
[he estimation and the variances of the estimates are direct. Set
rhen
lim I/I(l) = 1 and lim I/I(l) = 2.
A-+co A-+O
Also
d log I/I(l) e- 2A[(2 + l)(e A + 1)- 4 eA]
d}. (l-e A)[l-(t +l)e A]'
lienee t//(l) ~ 0 for l ~ O. Therefore the inequality for var(l).
171 Peto, S. (1953), Bs, 9, 320.
The probability that ni organisms administered to a guinea-pig in the ith
group do not effect a kill is
(1- p)n, ...... e-n,p.
L = Ii (~i)(e-n,pY'(l_e-nfpr·-r'.
i= 1r,
lienee Pand its variance.
The information obtained from n organisms per animal in a group of
1/1 guinea-pigs is
In = mn 2 e- np/(1-e- np ),
dod for variation in n, this is maximized for a value of n given by
"p::: 2(1- e - np ), with root np = 1· 5936. The most economical expected
proportion of survivors is e -np . . . 0·203.
173 Cochran, W. G. (1950), Bs, 6, 105.
280 EXERCISES IN PROBABILITY AND STATISTICS
(Xi == c5v/)
1 1 1·54
~- =--
".... N' (X2) N
max~l
e -
since x 2 j(e,x-l) is maximum for X = 1·5936.
174 First part is Query 78, Bs (1950), 6, 167.
Let the group means be x and ji based on nl and n2 observations res .
tively. so that n = n t +n2 and var(x- ji) = (1Un. +(1~/n2' pc\.
For (i) minimize
(12 (12
---.!.+ __ 2_ with respect to n1 .
nl n-nl
Hence n 1 = n(1d«11 +(12) and n2 = MJ«11 +(12)'
For (ii) minimize
en 2+n (12_(12)
n= 1 1 22 1 with respect to nl'
enl- (11
Hence nl = (11«11 +(12)je; n2 = (12«11 + (12)je.
For (iii) solve the simultaneous equations
(1~ (1~
and -+-=e.
nl n2
This gives the quadratic in n1
en~-(en+(1~-(1~)nl +(1~n = O.
A necessary condition for it to have positive roots is that
en + (1~ - (1~ > 0,
and the roots are real if
(en + (1~ - (1~)2 - 4ne(1~ = [en - «1~ + (1~W - 4(1~(1~ ~ O.
For (11 = (12 = (1,
Hence var(k*).
(ii) Set E(x) = kp == J.t, and E(s2) = kp(1 + p) == It2' Then
_ J.t[I+~r
k= ( [2
J.t2-J.t ) 1 +
(s - I(2) - (x - II)
-
].
(It2 - It)
Hence, using a logarithmic expansion,
var[Iogk] '" (It 2 - It)-2 [(2J.t:: Ilf var(x)+ var(s2) 2(2 It;l- It) COV(.X,S2)],
and t,h,e second summation is over all possible sets of ex's satisfying the staled
conditions, Thus Ie 1
A.1e-'L n --
-p, ,
P(k) = ~e-A' L
,=1 r! " j=l (exj-1)!
Ie -p,
= L ~e-A', A.1e-'rle-'/(k-r)!,
, =1 r!
by considering the multinomial expansion of (Xl + X2 + .,. + X,)k-, where
each Xj = 1.
The probability-generating function of the number of plants is
00
1+(1I-1)eP
and
A
Minimize F == var(T) - () it
r iii - 11 by Lagrange's method. Thus
iii = i." i (i. - 1)/(i." - I). whence
An(A -1)0- 2
var(T) = . < A0-2 for all n > 1.
1."-1
For large 11,
t = ji-x-Oo (
- nV)1- .h
WIt (n+v-2)dJ.
s n+I'
[0 lest H[var(x i) = var(Yi)), use the F statistic
F
v-I"
= n-l' i~l (Xi- X)2 i~' (Yi-ji)2
I\" with (n-l,v-l)d.f.,
This is also the sum of the products of the two rankings. Hem:e p.
284 EXERCISES IN PROBABILITY AND STATISTICS
183 UL (1964).
The result is obtained by simplification of the X2 in standard form .
E(aj) = AnJN; E(b j) = BnJN. There are (k-l)dJ. The result is due to BrWl\h
and Snedecor. and,
ANSWERS AND HINTS ON SOLUTIONS
Chapter 4
ff
'Xl ,::1)
2~ f f
C1J 00
<Pn(t) = n cos(ai)·
j= 1
2n sin(t/2n)<pn(t) = sin t
\0 that
sin t t/2n sin t
<Pn(t) = -. . (/2n) --+ - , as n --+ 00.
t sm t t
This is also the characteristic function of a uniform distribution in (-1, 1).
J Tanner, J. C. (1951), B, 38, 383.
P(X = 0) = e-). and P(X ~ x) = I-F(x) for x> o.
The density function of X isf(x) = -dF/dx. Therefore
I/!(t) = E(e iIX )
f f
l-E 00
285
286 EXERC'lSES IN PROBABILITY AND STATISTICS
1-1: L
eit(I+},) F(y) d ,]
J
<
= lim [ei/<p(x
r.~O
= e)+ A~-.t{eit(I-<)_eil<}+Aeit-.t
It
1 -<
f eit}'F(y)dy+
+Aeit-.t f eit}'F(Y)dY] ,
1+<
Hence, integrating the two integrals by parts and then taking the tirol'ls. ' ,,~
e -+ 0,
Aeit-.t A e-.t
¢(t) = - ,- , ¢(t)+e-.t--,-, as F(O) = P(X > 0) = (l-e-,t)
It It '
But
1 2 [u 2 + v2 - 2pu- 2(1- p2)it lalu - 2(1- p2)it2a2v] =
2(1-p )
1 2 [{u-pv-(1-p2)itlad2+(1-p2){v-pitlal-it20'2}2H
2(1- p )
+ Wit I al)2 + (it 2a2)2 + 2p(it I a d(it 2( 2)]'
Hence
E(e it,X +it2Y) = exp t{(itlad 2+ (it 2a2)2 + 2p(it l ad(it 2a 2)} == ¢(t 1,t2), say.
Therefore
P= II (2~r J J
00 -00-00
exp(-itlx-it2y)·<!>(tt>t2)dtldt2dxdy,
2
4n Op =
oP ff
00
2n
4>(tI,t 2 )dt l dt2 = ~'
ff
00 00
-co -Q'J
fherefore
-00
f f
00 00
But
1 foo
27t
{I 2 }' .( a 1 fOO
exp -WI +2itlX) . t{ dt l = i) ax . 2n
1 2
exp( -2tl - itlx) dtl
)J
-00 -00
6 UL (1963).
The probability density function of X is
([J == 1 ~J,
whence the power series expansion in (it)'.
The first four cumulants of X are
"1 = [Jm; "2 =
1+[J«(J2-1)+[J(1-[J)m 2 ;
K3 = [J(l- [J)m[3«(J2 -1) + (1- 2[J)m 2];
"4 = [J(1- [J)[3(q2 _1)2 + 6(1- 2[J)m 2(q2 -1)+(1- 6[J + 6[J2)m4 ].
Finally,
so that var(X) = v2+ [J(1- p)m 2, whence the inequality as 0 < [J(l- P) ~ !.
7 Using the result of the preceding exercise, the characteristic function of the
sample mean x is
= (I+A)-n £ (n)Arexp
r=O r
[_t(n-r~rq2) {t2_ 2itrmn 2}]'
n n-r+m
But
rhcrerore
(i) for A and (J constant and m --. 00,
(. . 2)] exp{-d/2(1-2it2)}
E[exp,t 1x+,t 2x = (I+A)(I-2it2)t
GJ f f
00 00
Iz 1":)
~ , I
1 ~ (n)Ar
= (2n)t(l +At' r~O ,.
f
00
IItA)-n
__
21t . "n
L.,
(n) 'r
A.
foo (1_2't)-(n- l l/2
I 2 • exp
[
-'tZ+
I 2
it 2
(1 2 2' ) 1--
( rm r)] dt 2·
r= 0 r - It2 n
-00
Hence, reversing the order of summation and integration, and using the
Inversion Theorem gives the characteristic function of Z.
The special case follows when either m = 0 or A = O.
The error sum of squares is
n
(n-l)s2 == I (X r -X)2,
r= 1
290 EXERCISES IN PROBABILITY AND STATISTICS
ff
«> '7J
(2~r (1-2it2)-1I/2 x
-Ct.) -00
00
1 e-txt _1
= __ f e- il2Y dt 2
fo' . 2n -00 (1-2it 2 )11I-1)/2'
= (1-:/)t . exp[2~ . {(1- 2it 22 )(it,)2 + (1- 2it 11)(it2)2 + 2(p + itd(it,)(iI2111
This is evaluated by a straight application of Aitken's integral (vide Chapler ~.
Ex. 63). Hence the joint characteristic function of X" Y" X 2, Y2 and 1
Invert this characteristic function to obtain formally the joint distribution (II
X" Y" u = x 2-xf, v = Y2 - Yf, and w = Z-X,Y,. Integration over II
and t2 can again be carried out by an application of Aitken's integral, M'
that the joint distribution has the density function
«>
(1- 2)"/2
(2:)4 ll1-
~
III -')/2. exp{ -i(ut11 +Vt22 +wtd} x
-00
Hence the result after an integration over X, and Y, for ( - 00 < X" Y, < (/.1
A NSWERS A ND HINTS ON SOLUTIONS: CHAPTER 4 291
Using the joint characteristic function, the probability density function of
II .
. II
,I,
and w IS
(I_p2)(n-l)/2 X
-(2n)3
)( f
00
X dIll dt22 dt l2 ,
. here ,l. == 1 - (p + it (2)2/(1- 2it 22 ) is to be regarded a constant for integration
::ver til' Therefore the joint density may be written as
(1_ p 2)(n-l)/2 e- tll . u(n-3)/2 f'~'
1 .
.
e- u22v (1-2it 22 )-(n-l)/2 x
(2n)2(n-I)/2 r(I1;) -<Xl
00
Since II > 0, the integration over t 12 is simple and gives the density function
and the inequality follows since 0 < tanh(-V-h) < V-It. Also,
(a+x)/a 2, for -a ~ X ~ 0
and (a-x)/a 2, for 0 ~ X ~ a.
Hence cp(t) by integration.
For the discrete distribution, 111ft = a, and
P[(r-l)h ~ X ~ rhl = h{2a-(2r-l)h}/2a 2, for 1 ~ r ~ Ill.
Therefore
cph(t) =
a
I~[e-ita f
r= 1
(2r-l)heith(2r-ll/2+eita
2
f
r= 1
(2r-l)he- ith (2r- l l/2]
2
=
a
h2[e-ita.~.!!..-
10t,=1
i: eith(2r-l)/2_eita.~ iJ f 10t,=1
e- ith(2,-ll/2]
.h'oee result .
\;
(2,.-1)11] 311
P [ y= 2 = 4a 3 1!(3a 2 -h 2)-h 2,.(r-l)].
Ihererore the characteristic function of Y is
-
11(120 2-112) sinmth + -311 (0)2
-
~ Ie illl(2r-ll/2 +e -ilh(2r-l l/ 2]
L
- 160 3 • sin(thI2) 4a 3 ot r= 1
h(12a 2-h 2) sin at 3h(0)2[ sinat]
= 16a 3 . sin(tIl/2) + 4a 3 ot sin(thI2)'
\\"hich may be reduced to the form
h2 [sin at thl 2] 3 [Sin 2at sin til at (thI2)3]
4a 2 ----;;t. sin(th/2) - (at)2 ~. ~ . sin at . sin 3(th/2) +
3 [Sin at sin 2 til (thI2)5 ]
+ (at)2 ~. (tll)2 . sin 5 (tIl12) .
lienee the result by using the expansion in Bernoulli numbers.
16 Kupperman, M. (1952), B, 39, 429.
1I(12a 2 -112) 3h 3 ,.2
P(Z = rh) = 16a 3 - 4a 3 ' for,. ~ O.
== ~[et/11(tl_l]
(Glt)2
+~
Glt
[et/12(t l _ 1] , where
r/!l(t) == I: Bj~j~:t)j(22j_l);
j=2 J.J.
and
00 B .(2a)j (it)j
r/! 2(t) == L:
)=2
J • .,
J.J.
' the B j being Bernoulli numbers.
r= 1 a a
= 2h.eith/2. eit.hm _l_ 211 (~.~) ;, eith(2r-l)/2
a e'th -1 a2 i at r~1 '
whence the result. The mean and variance are deduced by considering Ihl'
expansion of log QUJ) in terms of the series involving Bernoulli numbers.
./, ( ) sin(At)
'I' h t --+ ----;:t.
f f f(u, v)dudv = 1.
o 0
As
Ihcrerore
E[ (. . ) 1 2eit 1t 2{t2- t d
exp Itl u -lt 2v) = (1- 2it l)t (1 - 2it2)t + (1- it 1)3 (1- it 2)3'
<p(t I> t 2 ) f
= eit1x g(x) dx J
eit2Y p(Ylx) dy
296 EXERCISES I N PROBABILITY AN D STA T1STlCS
so that
[or ¢(t~, t 2)]
at 2 12 =0
= ;r fe il " g(X) Jl~(X) dx.
Hence the stated result by using the Inversion Theorem on the fUnction
g(x) Jl~(x).
22 Wicksell, S. D. (1933), B, 25, 121.
Since
1
2(l_p2) [x 2{1-(I- p2)itd + y2{1-(1-p2)it 2} -2pxy]
= _{l-(I- p 2)it d
2(1_p2)
[x- py
. 1-(1_p2)it.
]2_y2{(1-itl)(l-it2)-p2(itd(il2)}
2{1-(l_p2)it.} -,
therefore
E[exp(it lX2 /2 + it2y2/2)] = [(1- it d(1- it2)- p2(it l)(it 2)]-i,
whence the joint characteristic function of z. and Z2'
Use of the result of the preceding exercise gives 11~(Z d by noting that for
N >0
and
(it dO e - il,%, dt 1
(1- it .)N/2
_ I 2
2(1- p)
[(~)2+(L)2_2P(X)(L)]+it.X2+it2Y
0". 0"2 0". 0" 2
1
2( 2 [A(u-pv/A)2+B(v-C/B)2_c 2/B],
I-p )
where A == 1-2(1- p2)it.O"I; B == 1- p2/A; C == (1- p2)it 20"2'
A !':SWERS AND HINTS ON SOLUTIONS: CHAPTER 4 297
Hence
ljJ(t Ily) = f
-00
e it1x g(xIY) dx,
and
00
f
00
f
00
f e-ilZ'[~. ~I/I]
00
E(Ylz) h(z) 1
= -2 dt 2,
1t I ut 1 11: 0
-00
and
.,
E(y2Iz) h(z) = 21
1t
J
-00
e- i1z ' [~.
I
~2fJ
uti 11=0
dt 2·
and
f i1z'
00
where A =I '<=1
ai.
Hence, using Ex. 26 above, the conditional characteristic function or 1/
given v = 0 is
I/I(tllv = 8) = (1-2it 1)-(n-lI/2. exp[~: {1-(1-2it d}]
= ell- f (-(j2Y(1_2it
,=0 r!
l r(n+2,-l l1 2,
f
00 00
Je- it2Y
-00
cf>(t1' t 2) dt2 =
-00
exp{ -it 2Y+K(t 1, t2)}' dt2
J
00
00
But
f -Z"
ei~Z
dz = 0, for 0( > 0,
r
2nO("-1
- i" . (II _ 1)!, for 0( < o.
Hence the value of I(x).
32
f
00 00
-00 -00
f
00
II - 1 (II +,. - 1) 1
g(x) dx = L
r=O II
-1 2"+r-1
-00
II-I (1 +O()"+r-I
= coefficient of 0("- I in r~o -2-
2 r=O 1 5=0
1
= 2 [1- .,211- I· LI
II - (211 - 1)]
. ' whence the result.
- r=O I
f eitA dt ,
(t - i)"
-00 -00
-00 -00
ePU f-p+ioo
= 2ni[np)]n' eUz [n _·z)]n dz.
-p-i'X)
g(u) =
e PU( -nl" 1
.-
[r(p)]n 2ni
f
- p+ ioo
eUz dz
.'
{n1 + z)}" smn nz
-p~ioo
Evaluate the iptegral using the contour Iz/ = m+t, m a positive integer, and
the line x = - p. The poles of the integrand are at z = r with residue
n-n( _1)nr n- I [d
eUZ ]
Hence
ePU 00 (_1)nr+n+ 1 [d n- I eUz ]
g(u) = [r(p)]n' r~o (n-1)! n
dz - ' {n1 +zW z=;
I
37 Let
R(z) = 'IY r(Z+~)I/lIt-"Z(21t)(n-Il/2nIlZ).
,=0 V II
Then R(z) = R(z + I) so that R(z) = !~~) R(z). But for large z, Stirling's
formula gives
• i 'X
The integral is evaluated by using the contour Izl = m+t, where m is an integer.
and the imaginary axis. For fixed j, the poles of r(pj-z) are at z = Pj+r ror
r = 0,1,2, .... The residue at z = Pj+r is
(-1)'+ 1 e U('+ Pj)
nr+ 1) . IJj r(Pk-pj-r).
j= 1
()- f
~
g u - 21t
00
e
-ilu n-nilnllp+nit) d
. nnp) . t
-00
-np+ioo
=
eUPn"P1
- (-)'-2'
f (n eu/nyn-z)dz.
nr np m
-np-ir
ANSWERS AND HINTS ON SOLUTIONS: CHAPTER 4 303
fb e integral is of the type
-a+ ico
I == ~.
2m
Jwz f'( -z) dz.
-a-ioo
"/0 and w real, which is evaluated by using the contour Izl = m+t, m a
~sitive integer, and the line x = -a. The residue at the pole z = r is
1~\V)'/1tnr+ 1), so that I = e- W • Hence the distribution of v = e"/n is
1
- - . e-m'(nv)nP-'n du, (0 ~ v < 00).
r(np)
n og I [1 + ~ IlrnL...
r=2
r'
(it)']
r.
, .
By expansion
-) = -11-2
"2 (X an d
n
Again,
n
(n-l)s2 = I xJ-ni 2, so that E[(n-l}s2] = (n-1)11-2'
j= 1
Also,
n
(n - 1)2s4 -_ "L... X4j + "L... X2X2
j k-
j= 1 j*k
so that
But
a¢ = I.
-?
t2
f
S2 .exp (.1t 1X+1t
-
2 2). TIn f( X,. ) dXv>
. S
,.= 1
f(x) being the density function of X. Therefore
[ata¢]
2 12=0
= i["'(ttfn))"-1 feiIOC/n.x2f(X)dX-
_;["'(t.ln)]n-2 [feiIIXln.Xf(X)dxr
Therefore
But
n
-! L YJ+tlA+t2B
j= 1
BUI
Assuming that
An = AIZn+A2Z-n,
the constants Al = - z/(1- Z2); A2 = z/(1- Z2).
Hence An = An- 2 +S n- 1 , where S. == z'+z-', whence
Mn = t~Mn_2+~-lSn_l'
Bul z, Z - 1 are the roots of Z2 - (}z + 1 = O. Therefore
Sn-(}Sn-l +Sn-2 = 0,
and the difference equation for Mn follows.
j
To obtain the expression for mj' put t2 = L t, in
,= 1
-00
A j e/2B • fIk=l -1-.e-tY~dYk'
fo
and integrate for ( - ex) < t. ~ 0).
Also, for the given form of </>(t 1> t 2)
Hence cov(C I, C 2 ) = constant - [(lIn) sin -I p/(n -1)F whence var(C), the
constant being evaluated since var(C) = for p = 1.
For large n,
°
var( C) - :n [1 - (~ sin - I P ) 2].
which gives the large-sample efficiency of p*.
44 Bartlett, M. S. (1936), PRS, 154, 124.
E[eit(X-m)] = (1+t 2)-I, and y-m = (1-JL)(xl-m)+JL(x 2-m).
Therefore the characteristic function of (y-m) is
E[ exp{ it(1- f/)(x 1- m)}] . E[exp{ itJL(X2 - m)}]
= [1 + t 2(1- JL)2] - I [1 + t 2p2]- I
1 [ (1- JL)2 JL2]
= 1-2JL 1+t2(1-JL)2 1+t2JL2'
,0 that
,,_ 1 [1 L
. -+y~ 2 +(y_l)2 00 y' foo exp-[{(r+2){3-(r+l)lX}z].dz,
]
(1- 2JL)(I- JL) IX P r= 0
o
IP> IX). whence the result.
SUPPLEMENT TO THEORETICAL
EXERCISES IN PROBABILITY AND STATISTICS
Supplement
(i) t L
r= 1 s<r
r(r + 4)(s - l)(s + 2)
(iv) t L (r+1)(r+2)(r+4)s(s-1)(s-3)
r= 1 s=5;;r
n+l
(v) L L (r+2)(r+6)(s-2)(s-4).
r=1 s<r
r=O
(ii) r~ (")
" r r(r -l)(r + 3)x r •
6. For a series of terms Uh U2, U3, ••• , UI1 , ••• prove by induction, Or other.
Wise, that .
~her.e, in standard difference notation, A111 +1 U, = AII1 (U,+1 - u,), for all posi.
tIve mtegral values of 111.
Also, by considering the differences of the series
where
U, = A-1u,+1 - A-1u"
show that
(iii) '~1
11 U,=A-l[(1+A)I1_1]Ul='~1
11 (")
r A'-IU I ·
12 (i) If f(r) = X-I / (x: r) where X is any positive integer, prove that
(
X
x+1
-I)! = 1-~+
1 1 1
2x 2 - 2x3'
LU r =!(e-1f.
r=1
14 (i) The function f(r) = 1/(1 + r+ r2) is defined for all non-negative val·
ues of r. Prove that
2r
f(r-1)-f(r)=1 2 4' forr;;;'1.
+r +r
SUPPLEMENT 315
Hence show that for any positive integer 11
S == i 2,. 11(11 + 1)
" r= 1 1 + ,.2 + 1'4 n 2 + 11 + 1 '
so that
lim S"
.. +--GO
= 1.
P(X;;;o2) = 48e-l03.
48(e-l)
(ii) If a and (3 (a;;;o (3) are any two positive integers, prove that
at
k=O
(a-k)= (a+l).
(3 (3 + 1
(iii) If n is a positive integer and Ixl < 1, prove that the coefficient of xr
in the expansion of (1-x)-" is (n+~-l). Hence show that
t
k=O
k
(a + -1)((3 +a
k a
=k -1) (2a + (3-1),
k
=
a
where a and (3 are any two given positive integers.
16 (i) If a is a positive constant and 11 a positive integer, prove that
"
r~o (-1)' (n) (1+ra)/(1+na)'=O.
I'
evaluate the sum of the first /l terms of the series, and hence .
that the sum to infinity is unity venfy
(iii) Given that f(r)==or 2(,.-1)2(2r-1)(r 2 -r-1), prove that
r6 = tr2 + 1[f(r + 1) - fer)].
Hence show that for any non-negative integer r
II
17 (i) For any two positive integers 11 and s (s < /1), prove that
r=\ s+2
Hence, if Sr denotes the sum of the first r natural numbers, ShOll'
that
II
(ii) L [r-!(11+1)]2=-b11(112-1).
r=1
(iii) If r is a discrete random variable which takes the values 1,2, ... , /I
with probabilities 11- 1 , then by using the exact expressions for E(Jrl
and var(..r;.), prove that
( oo) If 2r+ 1
II Ur =
r 2( r+1 )2 '
.L .L.
II
1=1/<1
XjXj
1 [(
=-2 .L )2 -.L x~
1=1
II
Xj
II
1=1
]
,
prove that the sum of the products in pairs of the first n odd integers
is in(n-1)(3n 2 -n-1).
!O (0 Define the finite differences of a function [(x) tabulated for equi-
spaced values of the argument x.
If, for x = r, the value of [(x) is
[(r) = sin«(J + rq,),
where (J and q, are constants, then prove that for any non-negative
integer n
L1"[(r) = (2 sin(q,/2))" . sin[(J +!n('1T + q,) + r</J].
(ii) If
Uo, Uh U2, .•• , are the successive values of a function correspond-
ing to those of an integral variable r ~ 0, then prove that for any
non-negative integer n,
L1"uo= ~ (n)(-l)"-rUr.
r=O r
Hence, or otherwise, show that if ur =(-lY/(r+1), then
Llluo=(-1)1(2"+ 1 -1)/(n + 1).
(iii) Find the sum to n terms of the series whose rth term is
ur =(r2 +r-1)/(r+2)!, for r~1.
Also, verify that the sum to infinity of the series is !.
ProbabDity and statistical theory
11 At a university in Ruritania the probability of a first-year student
passing the June examination is t and that of receiving a res it in the
following September :&. The probability that a resit student will pass the
September examination is~. Prove that the probability of a first-year student
~ntering the second year at the university is ~.
If the probability of a second-year student ultimately getting a degree is
;~, determine the initial probability of a student receiving a degree at the
end of three years.
Hence evaluate the probability that of n randomly selected first-year
entrants to the university
(i) exactly k students will receive a degree;
(ii) at least k students will receive a degree; and
(iii) at least k students will not receive a degree, for O::s:; k ::s:; n.
11 At a game of poker, a player is dealt out a hand of five cards from a
Itandard pack of 52 cards. Find the probability that the player has
(i) a flush, that is, all cards of the same suit;
(ii) at least one spade;
(iii) at least one spade, but not a flush; and
(iv) a flush, but not of spades.
318 EXERCISES IN PROBABILITY AND STATISTICS
and
(b) the conditional probability that a second sample of v fuses from till'
N - v remaining ones is wholly good is
p= ~ [ 1 + 2k~ J.
17 In "The Final Problem", a short story by Sir Arthur Conan Doyle,
Sherlock Holmes is pursued by his old enemy Professor Moriarty with
murder in mind. It is understood that if Moriarty catches Holmes then the
latter must die. Holmes takes a train at London bound for Dover from
where he can escape to the continent. Moriarty charters a train and follows
dose behind. There is only one stop before Dover-Canterbury. If both
Holmes and Moriarty choose to alight at the same station, then Holmes
meets his end. Secondly, if Holmes chooses Dover and Moriarty Canter-
bury, then Holmes makes good his escape. Lastly, if Holmes gets off at
Canterbury and Moriarty goes to Dover, then the pursuit is a draw since
Holmes, though temporarily safe, is still within the reach of Moriarty.
If the independent probabilities of Holmes and Moriarty getting off at
Canterbury are x and y respectively, determine the probabilities of the three
distinct outcomes of the pursuit. If the probability of Holmes meeting his
death in the encounter is a (>0), and this is also the probability of his
escape, then prove that
y2-3ay+a =0.
Hence deduce that in this particular case a must satisfy the inequality
28 (i) A standard pack of playing cards consists of four colour suits, each
suit having 13 distinct cards. The 13 cards of a suit include four
court cards-an ace, a king, a queen and a jack. From the pack, a
hand of 13 cards is dealt out randomly to a player. Find Per), the
probability that the hand contains exactly r court cards, 0:;:; r:;:; 13.
Hence show that, for r ;;'1, per) satisfies the difference equation
P( )_(17-r)(14-r)p( ).
r - r(23+r) r-1, and
P(O) = 36(13)/52(13\ in standard factorial notation.
320 EXERCISES IN PROBABILITY AND STATISTICS
(ii) Two players A and B playa game of chance as follows. From the
pack of 52 cards, 13 cards are dealt out to A and then another 13
cards to B, the player having more court cards being the Winner f
the game. The game ends in a draw if both A and B have no COU~t
cards or the same number of them. Prove that the probability fo
the game to be drawn is r
t (16)(16-k)(
k=O k k 13-k 36 )(23+k)/(52)(39)
13-k 13 13'
Also, in particular, if it is known that A has all court cards, then
show that the probability of B having no court cards is ~ approxi-
mately. Verify that the percentage error in this approximation is;'!7·
29 In a game of chance, a player repeatedly rolls k (> 1) unbiased cubical
dice till at least one six shows up. If the event "at least one six" OCcurs on
the rth throw of the dice, the player wins £2 1- kr (r;;;:.l). Prove that the
probability of the player winning on the rth throw is
(~)k(r-l)[l-(~)kJ.
k "-t+ (nr-1-
r=1
1
k)prq"-r.
Furthermore, if it is known that the urn contains one black and one
white ball and the remaining n - 2 balls are equally likely to be either white
or black, prove that the probability of the sample of k balls containing only
one white ball is
k ,,-k+l (11 - k)
n(I1-1)2,,-2' r~1 r(n-r) r-1 .
31 A man plays a game of chance in which, at each turn, he rolls a pair of
unbiased cubical dice whose faces are numbered from 1 to 6. He continues
until the first "successful" turn, a successful turn being one in which at least
one of the dice shows 6. If this is his 11th turn (11 ;;;:.1), he receives a reward
of £a" (a < 1). This terminates the game and the stake for playing it is 50
pence. Prove that the player's expected reward is £l1a/(36-25a).
Hence deduce that, on the average, the player will be on the winning
side if a> ~~.
32 In a game of poker, five cards out of a standard pack of 52 are dealt out
at random to a player. There are four suits in a pack and the 13 cards in a
SUPPLEMENT 321
suit are regarded to be in an ascending order in which the ace is taken to be
Ihe lowest card and also the highest. Any five cards (not necessarily of the
same suit) which are in an unbroken sequence are said to form a run.
I-Iowever, if the five cards in the run belong to the same suit, then the
sequence is known as a running flush.
Find the probability that the player has a run but not a running flush.
Also, show that there is a 4 per cent increase in this probability if it is known
Ihat the player has the jack of hearts. What is the explanation for this
increase?
32a There are It alternatives in a multiple choice test, and the probability of
an examinee knowing the right answer is p. However, if he knows the right
answer, then it is not absolutely certain that he will give the correct response
because of the emotional tensions of the testing situation. It may therefore
be assumed that the probability of the examinee giving the correct response
when he knows the right answer is 1 - a where a > 0 is small but not
negligible. Alternatively, if the examinee does not know the right answer
Ihen he randomly selects one of the It altet:natives.
(i) Prove that the probability of the examinee knowing the right
answer when he, in fact, gives the correct response is
Itp(1-a)
l+[n(l-a)-l]p·
A B B A
B C C B
B C C B
A B B A
After initial placement, the system is moved n times. At each move, the
system is transferred to a neighbouring cell in any direction horizontally,
"ertically or diagonally, all such moves being equally likely.
\ For n;?!: 1, suppose all' {311 and 'YII respectively are the probabilities that
after It moves the system is in an A-type, B-type or C-type cell. Prove that
Ihese probabilities satisfy the following difference equations:
a" =!{311-1 +!'YII-l
{311 = ~all_1 +~{311-1 +!'YII-l
'Y" =1all -l +~{311-1 +i'YlI-l'
Show that the solution of these difference equations is
all =!-i{3",
where
10 [ 13 (-9)"] .
{3"=21 1- 90 40
Hence deduce the probabilities that after n moves the system will be in a
specific A-type, B-type or C-type cell and verify that, as n ~ 00, these
probabilities are in the ratio 3: 5 : 8.
39 Two players A and B contest a series of games. At each trial the
probabilities of A and B winning a game are p and q respectively (pi 0,
qrO), and there is a non-zero probability r for it to be drawn. It is agreed
Ihat the winner of the series will be the player who is the first to win four
more games than his opponent. Prove that, for pi q but irrespective of the
value of r,
(i) the initial probability of A winning the series is
1
1 + (q/p)4 '
and
(ij) the probability of B winning the series when he is exactly 11 (0 ~ n ~
4) games behind A is
1-(p/q)4-11
1- (p/q)8
Hence deduce the values of the probabilities (i) and (ii) when a game
cannot be drawn, the other conditions for the contest being unaffected.
40 A self-service store has a stock of nt + n undated one pint bottles of
milk of which n are fresh and 111 a day old, and it may be assumed that these
324 EXERCISES IN PROBABILITY AND STATISTICS
Hence, or otherwise, prove that the conditional probability that the secor I
customer obtains k bottles of fresh milk when it is known that at least one :'r
her bottles is fresh is I
f
r=O
(v)(I11+~-:-k)j[(m+n)_
r nr k n f
r=O
(v)(m+n~v-k)].
r nr
41 In a game of chance, a player can score one, two or three points at "
trial with constant probabilities q, pq and p2 respectively, where p + q = 1. Ii
in repeated trials of the game, Un denotes the probability that the player<
total score will be n points at some stage, then show that Un satisfies a
third-order linear difference equation with the initial conditions Uo = 1 and
Ul = U2 = q. Hence deduce that
Also, show that the same distribution of the face values of the coi .
obtained even if the number of shillings is not known. ns IS
[Note: For those unfamiliar with the pre-decimal British monetary syste
shilling or a bob was worth 12 old pence, a two-shilling piece 24 old P:' a
and a half-crown 30 old pence.] IICr
45 (i) Define <I>(x), the distribution function of the unit normal variable X
and then show that for any x > 1 .
x
1 Jexp[1
2J(2'Tr) -2 (t2-1)2J dt=<I> (X-2-
-x
-1) +<1> (x-2-
+ 1)
-1.
What is the average return per roll of film sold under the schellle
of the dealer?
51 The attendance records of workers at a large factory kept for a IOIl~
time show that the probability of a worker being absent for r (;;;'0) days in a
year was approximately given by the Poisson distribution with mean IJ.. The
absence of a worker causes dislocation in the production schedule which
results in a loss to the factory; and it was estimated that, on the average, if a
worker was absent for r days in a year the net loss to the factory was
£a/'(r+ 1) e-13r, where a and (3 are positive parameters and (3 is small. Prove
that the expected loss (in pounds) to the factory due to absenteeism of a
worker in a year is
a/L(/L e- 13 + 2)exp[ -(/L + (3) + /L e- 13 ],
Hence show that if /L and (3 are such that 1 < /L < el3, then the expected loss
must be less than 3a.
52 The Poisson distribution is occasionally referred to as a distribution
describing the occurrence of rare events. Explain this remark.
A commercial secretarial organization charges Sa pence for cutting a
foolscap stencil. The price includes the cost of the stencil and the remunera-
tion of the typist, and the residual profit of the organization is a fifth of the
charge made, provided the typist makes no typographical errors. However.
if an error is made, then the profit of the organization is diminished by the
additional cost of correction, On empirical considerations, it is known thaI
the cost in pence of correcting r errors on a stencil is 2r(3r + 2)/(r+ 1), for
r;;;' 1, It may be assumed that, on the average, the typist makes /L errors per
SUPPLEMENT
329
.Ieocil where /L is.lsmaldl. Shuggesht a sUhitable model for the distribution of
ors on a stenci an t en s ow t at the expected cost in pence of
:rtrrecting the errors made on a stencil is
.0
2[3/L -1 + /L -1(1-e- lL )].
Ileoce show that if the organization wishes to obtain, on the average, a
oftt of A.a pence per stencil, where A. is a constant such that 0 < A. < 1, and
\ assumed to be sufficiently small for terms involving /L 2 to be negligible,
;hell. as a first approximation, the organization will attain its objective if
5/L
a=I_A. .
II'here P(x)=e-ILIL"/x!
Hence, by considering the behaviour of the finite difference G(y + 1)-
G(y), prove that the optimum size of the stock which the store must order
to maximize its expected profit is Nyu, where Yu is the smallest integral value
of y for which
a-I
F(y)~--.
a-(3
54 A government wishes to control immigration by restricting the entry of
foreigners to an annual intake of a persons. (It may be assumed that the
immigrants enter the country at the commencement of each calendar year.)
Suppose initially, at the start of the first calendar year, the total population
of the country is a constant N. However, due to the natural changes in the
population, it is presumed that the actual number of persons in the country
at the end of the year will be Nx, where x is a random variable such that its
probability density function is
A. e-A(x-l), for x ~ 1 and A. > 1 is a parameter.
Accordingly, at the beginning of the second calendar year the total popula-
tion, including the annual quota of immigrants, is another random variable
}'2=a+Nx.
Prove that E(Y2)=a+N(I+A.)/A..
In general, suppose that the expected population, including the immig-
rant quota, at the beginning of the rth (r ~ 3) year is
1 + A.) E(Yr-l)'
E(Yr) = a + ( -A.-
330 EXERCISES IN PROBABILITY AND STATISTICS
Use this equation to verify that the expected population at the beginni
the (k + l)th year is ng of
(1 + "f - AkJ (1 + A)k
a[ Ak - 1 +N A .
1- L 2 (10) (l-e-"')W(e-"')lO- w•
w=O W
a +
[1
e-A(I-e-A)(2-p e- A)]
I-pe - A '
and hence deduce that this amount can never exceed 7a/4.
58 A retailer obtains his supplies of a small machine component from the
manufacturer in large batches, each batch consisting of M components.
Although the production is under statistical control, there is a proportion e
of defective components in each batch such that e may be regarded
approximately as a continuous random variable with the probability density
function 60(1- 0) in the range 0::;;; e::;;; 1.
The retailer sells each component for one shilling but, to attract custom,
he agrees to refund a shillings (1 < a < 2) to a customer for a defective
component supplied. To protect himself under such a guarantee system, the
retailer takes a sample of N (fixed and «M) components from each batch
supplied to him by the manufacturer for destructive testing, and refuses to
accept any batch containing n or more defective components. For any given
0=0, find the conditional probability of the retailer accepting the batch
after inspection, and then show that his expected net income in shillings from
such a batch is
g(n, 0) = (M - N)(I-aO) ~t~ (~W;(I- O)N-x.
Hence deduce that the expected net income of the retailer per batch for
variation in e is
(M - N)(n + 1)(2)
g(n) = 2(N+4)(4) [6(N+2)(N +4-a)-2(2n + 1){(N + l)a + N +4}+
+ 3an(n + 1)].
332 EXERCISES IN PROBABILITY AND STATISTICS
0: + (3x - (x 2 - ( 2) e --v(x-/-L)/u,
where 0:, (3 and 'Yare small positive parameters. Prove that the expectcd
cost of a packet of margarine is
a + (3/1- - (/1- - 'YU f e _~-V2.
An economic crisis forces the government to devalue its currency and
consequently, there is an increase in the import price of edible oils used ir;
the production of the margarine. To offset a part of the increase thc
manufacturer makes certain small economies in his fixed costs and also
establishes checks to lessen waste. As a result of these measures, thc
parameters of the cost function are expected to become 0:' «0:), (3' (>~)
and 'Y. The manufacturer aims that, on the average, there will be no increasc
in his cost of production; and he plans to achieve this by suitable alterations
in the parameters of the weight distribution. Assuming that the new weight
distribution is normal with mean /1-' and variance U,2, and /1-' is so chosen
that
0: + (3/1- = a' + (3' /1-',
prove that there will be no change in the expected cost of production if
/1- -/1-' = 'Y(u-u').
Hence show that the production of packets produced by the new system
which are less than the nominal weight is
<I> ['Y{-(O: - 0:') + «(3' - (3)/1- + (3'(1-/1-)J
(0: - a') - «(3' - (3)/1- + 'Y(3' u '
where <I>(z) is the distribution function of a unit normal variable z.
60 It is known from the past experience of a publishing house that a small
proportion of typographical errors remains undetected by the usual methods
of proof-reading. In order to improve upon the existing procedure, the
publishing house agrees to an additional proof-reading of a book with the
proposal that the proof-reader will be paid
(i) £0·25 for a page on which he detects no error;
(ii) £1·00 for a page on which he detects a single error; and
(iii) £1·25 for a page on which he detects more than one error.
The probability that a word on a page is typographically incorrect is a
small number a and, on an average, there are 11 words on a page. If /I is
sufficiently large compared with 0: but 110: = /1-, a finite number, give without
proof the probability that there are exactly r (~O) errors on a page. Hence,
SUPPLEMENT 333
assuming that the proof-reader detects the errors present, show that the
eXpected payment per page to the proof-reader is
£~[5 - (4+ IL) e- IL ].
The proof-reader is unwilling to accept the suggested rates of payment
because he realizes that it is unlikely that many pages will contain one or
ptO re errors per page. As a counter-proposal, the proof-reader asks for an
arrangement which would give him an expected payment of £1·00 per page
by raising the amount £0·25 in (i) when the amounts £1·00 and £1·25 in (ii)
and (iii) remain fixed. Prove that the increased payment that the publishing
house should make for each page on which the proof-reader finds no errors,
in order to meet his demand is
£M5 + IL -e- IL ].
61 A manufacturer of a duplicating machine rents out the machine to
cuslomers on an annual basis. The annual rental for a machine is fa, a fixed
charge, plus an amount which depends upon the number of copies made in a
year. Accordingly, if x copies are made on a machine in a year, then the
lolal rental payable (in pounds sterling) is
T(x) === a + bx(1-tx),
where b is another constant.
If it is known from empirical considerations that x can be regarded as a
continuous random variable with the probability density function
(A + 3)(A + 2)x
(1+X)A+4 ,forx;;a.O,
X";-I3· 0
1- s e12)
1 log (7T 2" •
E(X') = 2n-l
(3'/""
B(~+
a'
12n -1-~)
a
in standard notation for the beta function.
Also, show that the stationary values of the probability density function
of X are the roots of the equation
a '" a-I
,..x =
(2n-l)a+l .
Hence deduce that for a> 1 the probability distribution of X has a mode at
the point
a-I ] 1/",
[
X= {(2n-l)a+l}(3 .
( 21[A4 A
F x)=- l--+x+-x 2( 1-2:x
1 2)]
2 .
Hence show that the median of the probability distribution of X is a rc' I
root of the equation a
A(x 2 -1f-4x = 0,
and that this root is > or <0 according as A> or <0.
68 A continuous random variable X has the probability density function
proportional to x 2 /(1 + x 4 ), defined in the range -00 < X < 00. If k denotes
the proportionality factor, prove that
k (2r+3 1-2r) .
E(X2 r)=2 B - 4 - ' - 4 - ' provided 1-2r>0,
in standard notation of the beta function. Hence deduce that k = .f2/7r and
that none of the even moments of X exists.
Also, show that the turning values of the probability density function of
X are the roots of the equation
X4=1.
Hence prove that the probability distribution of X is bimodal and determinc
its modes.
Finally, show that the points of inflexion of the probability distribution of
X are the real roots of the equation
3x 8 -12x 4 + 1 = 0,
and determine these roots explicitly.
[Note: It may be assumed that f(t)f(l- t) = 1T/sin 1Tt.]
69 A continuous random variable X has the probability density function
ex(3x",-l
f(x) = (1 + (3x"'f' for X;;;'O,
where ex and (3 are positive parameters. Prove that if ex > r, then the rth
moment of X about the origin is given by
[ ex -1 ]11'"
x= (3(a+1) .
SUPPLEMENT 337
Finally, prove that the median of the probability distribution of X is at
the point
flen ce , for ex> 1, infer the sign of the skewness of the probability
distribution of X.
10 (i) For any two positive numbers ex and v, the incomplete gamma
function is a
nfu
where <1>(.) denotes the distribution function of a unit normal
variable and 8 is a positive constant.
71 A continuous random variable X has the probability density function
porportional to (ex + Ix!) exp- ~ lxi, defined for -oo<X <00, where ex and ~
are positive parameters. Evaluate the proportionality factor and then verify
that for any integer r > 0
E(X2r ) = (ex~ + 2r+ 1)(2r)!
(ex~ + 1)~2r
Hence deduce that X has unit variance if a =~. Prove that in this cas
mode of the distribution of X is at the origin and the maximum ordina~ th~
the distribution of X is 3/(4";;). Also, show that the coefficient of kurtC (!f
of the distribution of X is OSI,
I'z(X) = -i.
Compare these values with the corresponding ones for a unit normal
distribution and comment on the differences observed.
73 The distribution function of a continuous random variable X having th .
"logistic" distribution is defined by the equation ~
P(X:os; x) = (1 + e- x )-l, for -00 < X < 00.
Evaluate the probability density function of X. By using the transformation
y = (1 +e-X)-l, or otherwise, prove that the moment-generating function of
X about the origin is
Itl< 1.
Hence deduce that 1'1 and 1'2, the coefficients of skewness and kurtosis of X
are 0 and 6/5 respectively. .
Also, show that
(i) the semi-interquartile range of X is loge 3; and
(ii) the points of inflexion of the probability density function of X arc <II
X = ±loge (2+.J3).
[Note: It may be assumed that r(t)r(1- t) = 7T/sin 7Tt.]
78 A standard pack of playing cards has four aces and 48 other card, l
pack is well shuffled and then cards are dealt out one by one face up~. hl,
until the second ace appears. If X is a discrete random variable denotin clrt
number of cards dealt out before the appearance of the second ace pg. Ill'
that ' 111\\.
A r
p=-,
11
and that for large samples
p2q
var(p)=-.
r
E(:=~)=P,
so that p is not an unbiased estimate of p.
83 In the distribution of the number of police prosecutions per motorist'
· over a given
a Iarge cIty . . d 0 f time,
peno ' t he f requency 0 f motorists havin
. 'III
prosec~tions is 11" for r ~ 1 ~nd .2::::"=1 nr == N, b~t the number of motori~t:
~ho dId not ?ave a pro~ecutIo~ IS unknow~ owmg to the motorist popul<!_
tr~m .of t.he cIty fluctuatmg dunng that penod. Sup~ose th~t t?is obserV\:d
dlstnbutIon can be represented by a truncated POIsson dlstnbution with
mean p" that is, a Poisson distribution in which the zero class is ignored
Show that the linear function .
00
p,*=Lr.I1.1N
r=2
Give any reasons which in your opinion make the estimate p, * preferable to
the maximum-likelihood estimate ,1.
[Note: It may be assumed that for large samples
A p,(1-e-IJ.)2 ]
var(p,) = N{l-(p, + 1) e-IJ.}·
84 Explain briefly the concept of an estimate of a population parameter,
and distinguish clearly between unbiased and consistent estimates. Illustrate
your answer by suitable examples.
The probability density function of the Cauchy distribution is
1 1
-. , for -co> X <co.
7T l+(x-p,f
where p, is a parameter of the distribution. If Xl' X2> ••• ,x.. are independent
observations of the random variable X, determine the equation of ,1, the
maximum-likelihood estimate of p,.
Also, prove that the large-sample variance of ,1 is 2/n.
SUPPLEMENT
343
I~ A continuous random. variable X has. ~ normal ~istributi~n with mean (J
,od variance k(J, where k IS a known posItIve quantIty but (J IS an unknown
Jd aOleter. If Xl> ~2' •.. ,XII are n independent observations of X, show that
;'Jf equation for (J, the maximum-likelihood estimate of (J, is
·he
e(e+k)=.e+s 2,
" "
ni=Lxj and ns 2 =L(xj-if.
i=1 i=1
What is the limiting value of var(O) if, for fixed n, k ~ 0, (J ~ 00, such
,hat k6 ~ /-L, a finite number?
86 Explain, very briefly, the concept of the best estimate of a parameter in
,he theory of linear estimation.
Suppose that Xl' X2, ... ,XII are n independent observations such that
(J2
E(x,,) = /-L; var(xv ) = , for 1 ~ v ~ n,
n-v+1
Il'here /-L and (J2 are independent parameters. Determine /-L *, the least-
squares estimate of /-L, and show that
2(J2
var(/-L *) = - - -
n(n + 1)
Another linear function suggested as an estimate of /-L is
6 11
Prove that T is an unbiased estimate of /-L but that, as compared with #.t *,
2(2n+ 1)2 8
Eff(T) = 9n(n + 1) 9 for large n.
87 Explain clearly the meaning of linear regression and state the assump-
tions underlying the standard tests of significance in linear regression
analysis.
Suppose that Yl> Y2' ... ,YI1 are independent observations such that
E(yJ = a + (3(x" - i) + "(x,, - if;
where x" are values of a non-random variable, i = I~=1 xv/n, and a, (3, "
(1-0) and (J2 are independent parameters. Prove that, in general, the linear
functions
11
a*= L yJn
v=l
are biased estimates of a and {3 and find their expectations explicitly.
What happens if the x" are equispaced?
344 EXERCISES IN PROBABILITY AND STATISTICS
E(r.,) = 0;
n-1
var(r.,) = 0'2 [ n
(x., - i)2]
X .
where
.
X= L (x.,-if;
v=l
and
i
v=l
var(r.,) = (n - 2)0'2.
where a, {3 and 0'2 are the parameters in standard linear regression analysis
and i is the average of the Xi> the values of the explanatory variable. If n';
and {3* are the usual least-squares estimates of a and {3 respectively, find
(J *, the best estimate of the parametric function (J = a - {3i.
Prove that
i=l
(1+ :2y.
91 Suppose that Xl> X2, ••• ,x,. are n independent observations with mean
~ and variance 0'2. Another independent sample of n correlated observa-
lions Yl> Y2, ... ,YI1 is given such that for all i = 1,2, ... , n
E(Yi)=IL; var(Yi) = 0'2; and corr(YhYj)=P, (i=/=i).
Prove that x and y, the means of the x and Y observations respectively, are
unbiased estimates of IL and determine their variances.
If T== ax + (3y is a linear function of the sample means, find a and (3
such that E(T) = IL and var(T) is a minimum. Verify that this minimum
var(T) is
0'2
- [1 +(n -1)p]/[2+(n -1)p].
n
Also, show that
(n _1)2p2 ]
var[(x + y)/2] = [ 1 + 4{1 + (n -1)p} . min var(T).
~Yk~a+(3(xo-i)+tO.05xS 'k+-;;+
A [1 1 (xo-xf]\
X •
Mre tOoOI is the one per cent point of Student's distribution with 11 -1 d.fo
W Also, show that, irrespective of the value of ,\ in its permissible range,
,he above confidence interval for L must be greater than the interval
11 +2)~
x±too01Xs ( 2;- .
where x and S2 are the sample mean and variance of the already observed 11
values of X, and t( T); 11 - 1) is the lOOT) per cent point of Student's
distribution with 11 -1 dJ.
96 In 11 Bernoulli trials with a constant probability of success p, w
successes were observed. Show that for the probability distribution of the
random variable w the third central moment is
1L3(W)=l1pq(q-p), where p+q=1.
p* being the observed relative frequency of successes. Also, show that for
large 11
var(T)-npq(1-6pqf.
t[ai . dmiJ
i=1 mi d8 6=6
=0.
X==t a;.dm i
i=1 mi d8
E(X)=O; var(X) = t
l.. (dmi)2.
i=1 mi d8
In a particular breeding experiment with a variety of Papaver rhoeas.
there were four distinct classes with the expected frequencies
N N N N 2
4 (3-28+8 2), 4 8(2-8)'4 8(2-8), 4 (1-8).
SUPPLEMENT 349
I'erify that in this case
2N[1 + 2(1- 6)2]
var(X)= 6(2-6)[2+(1-6)2]"
100 The serially correlated observations Xl> X2, ••• ,x.. are such that
E(Xj) = 0, var(Xj) = u 2, for i = 1, 2, ... ,n;
L vx",
xw = n ( n+ 1) 1'=1
Ihen verify that
2u 2
var(xw ) = 3n(n+ 1) [2n(1 +2p)+(1-4p)].
Hence deduce that, as n ~ 00, the limiting efficiency of x'v as compared with
jis i.
101 If X and Yare two correlated random variables having finite means
3nd variances, define cov(X, Y).
(i) Assuming that X and Yare positively correlated and var(X);;:;.
var(Y), prove that
cov(X, Y):s;;var(X).
(ii) Prove that as a first approximation
(Y X) _ cov(X, Y)
cov\X' - E(X) .
Hence deduce that to the same order of approximation
(D E(Y) [ cov(X, Y)]
E\X)=E(X) 1- E (X)E(Y) .
i, prove that
(F2
(i) var(i) = - [1 +(n -1)p];
n
(11") var (Yi) = (n -1)(1- p)(F2 , for i = 1,2, , , , , n; and
n
(iii) corr(Yi> Yj) = -~1'
n-
for i1= j = 1, 2" , " n,
103 If Xl> X 2 " " , XN are correlated random variables with finite m '
and variances, prove that edlJ\
and
var(xi) = (F2, for i = 1, 2, ... , n,
where p and (F2 are parameters (-1 < P < 1 and (F > 0), prove that
the variance of i, the average of the Xi is
(F2 [1
+ p _ 2p(1- pll)]
n I-p n(1-p)2 .
lOS (i) If Xl> X 2 and X3 are uncorrelated random variables, prove that
cov(Xl + X 2 , X 2+ X 3 ) = var(X2),
(ii) The serially correlated random variables XI> X2, • , , , XII' , , • are such
that for any s ~ 1
E(xs ) = 0;
\p\<l.
Given any two positive integers rand n such that r < n, the linear
functions SI> S2 and S3 are defined as
It n+r
S2=
v=r+l
LXv; S3= L
v=n+l
Xv·
SUPPLEMENT
351
prove that
a2pll-r+l(1_ pr)2
COV(SI,S3)= (l-p)2
(i) P(Y=r)= i
v=O
(nl)(
]1 r
~2
v
)Plq'i·-vp2-vq22-r+v, for 0:S;;Y:S;;n 1 +11 2;
and
(nl)( n2 ) ,
(ii) P(Xl=sl Y=r)= r s r-s p
L (n1)( r-]1
v=O
n2 )p v
]1
where p == Plq2/P2ql> and it is assumed that in both (i) and (ii) the binomial
coefficients are zero for inadmissible values of r> nl and r> n2'
Discuss very briefly how the conditional probability distribution (ii) may
be used to provide a test of significance for the null hypothesis H(PI = P2)'
110 A discrete random variable X has the probability distribution defined
by the system of equations
P(X = r) = Pn for integral values of X;;. 0,
Define the probability-generating function of X and show how this functioll
may be used to evaluate the factorial moments of X. Conversely, if /Ll/) is
the jth factorial moment of X about the origin, prove that
1 00 (-I)'
Pr = --, L -,-#L(r+,)'
r, ,=0 s,
Use this result to verify that if, in particular,
#Lej) = (n + j _l)(j)pi,
where n, a positive integer, and P (0 < P < 1) are parameters, then
Pr= ( n+r-1)
r p r(1+p)-<n+r),
SUPPLEMENT
353
1 In sampling from a binomial population with success parameter k
I~pen~ent samples are obtaine?, there bei?g ni observati(;lns in th/ith
pie (I == 1,2, ... , k). The relatIve frequencies of successes In the samples
~ pt, p~, ... ,pt. Prove that the linear functions
2 .J2;r=o(2r+l)2r .r!
Prove that the moment-generating function of IXI is
E(e tIX1) = 2 e!t 2
• <I>(t).
Hence obtain an expansion of the cumulant-generating function and veriry
that the first two cumulants of IXI are
KI = (2/7r)!; K2 = (7r - 2)/7r.
116 A continuous random variable X has the probability distribution with
probability density function
n(n-l)
~-~ x,,-2(a -x), for O~X~a,
a"
where a >0 and n, an integer >1, are parameters. If Xl and X2 arc
independent observations of X, then show that the probability density
function of the distribution of the ratio u = XI/X2 is
n(n-l)
2(211 -1) . u
n-2[ 11 - ( n - 1)]
u,
f
or
0
~
1
u~ ;
and
n(n-l)
( ) u-"[n-(n-l)/u], for l~u<oo.
22n-l
Use this distribution to verify that for (3 > 1
43-24J3
16
119 A random variable X has the Laplace distribution with the probability
density function
1
f(x) = 20- e-lxI/CT, for -oo<X <00.
Prove that the characteristic function of X is
E(ei'CT ) = (1 + t 2 0- 2 )-I.
356 EXERCISES IN PROBABILITY AND STATISTICS
120 Given that Xl, X2, ... are observations obtained in a temporal order
such that
E(xj) = mj for j ~ 1 and var(xj ) = 0'2, a constant,
the autocorrelations between the Xj are defined as
1
Pk = P-k = 2 E[(Xj - mj)(Xj+k - mj+k)] for k ~ 1.
0'
U=_l_ ,2
e-! dt
..n; J V = -i -
, e -112dt
2
..n;
J
are independent random variables, each having the rectangular distribution
in the (0, 1) interval.
If U and V are independent rectangular variables defined in the interval
(0,1), show that the random variables lW and Z defined by the equations
W = (-2 loge U)~ cos 2'ITV, Z = (-2Io~ U)lsin 2'ITV
SUPPLEMENT 357
are independent unit normal variables. Also, show how these results may be
used to construct from k + 2 (k ;;:.1) independent realizations of a rectangu-
lar random variable defined in the (0, 1) interval a function which has the X2
distribution with 2k + 1 d.f.
Hence prove that the marginal distribution of v has the probability density
function
nv"- 2
---:----::-[(n -1)(20. - v)"+1 - (n + 1)v 2(2a - V)"-1 + 2V"+1] for 0 ~ v ~ a;
2(n+ 1)0. 211
and that of u
for O~u~a
(n + 1)a 2n '
and
123 Two related components in an electronic unit are such that the failure
of one tends to affect the life-time of the other. If X and Yare continuous
random variables denoting the individual life-times of these components,
then a possible model for representing their joint distribution has the
probability density function
f(x, y) = atf3 2 exp[-132y-(al +13l-13~x], for O~X~ Y
= 1310.2 exp[-a2x - (0.1 + 131 - (2)y, for 0 ~ Y ~ X,
where the unrestricted range of both X and Y is from 0 to 00, and
al> 0.2, 131> 132 are positive parameters such that
Prove that the probability that X ~ Y is 0.1/(0.1 + 131). Hence derive the
joint distribution of X and Y subject to the condition X ~ Y, and then show
that in this case
(i) the marginal distribution of X has the probability density function
(0.1 + 131) e-<<x,+f3,)x, for 0 ~ X < 00; and
(ii) the conditional distribution of Y given X = X has the probability
density function
358 EXERCISES IN PROBABILITY AND STATISTICS
IZ7 A random variable X has a uniform distribution in the interval (0, a).
Ux\ and X2 are independent realizations of X, determine the joint probabil-
~y distribution of
~ence derive the marginal distribution of v and then verify that for any
positive constant (3"; O! 1[ (3 2
P(v;;;'(3)=Z 1-;].
Also, prove that 3lvl is an unbiased estimate of a and that the variance of
• • 1 2
Ihis estimate IS 2'0! •
128 (i) The probability that an event X will not be realized in the first n
trials is u". Prove that the expected number of trials required for
the realization of X is
00
L u.. where Uo = 1.
r=O
(ii) In a sequence of n Bernoulli trials, the probability of a success (S)
is p and that of a failure (F) q (p + q = 1). If Yn is the probability
that the pattern FSSS will not be realized in the n trials, prove that
Yn =
r=n/4
t (n -r r) (_ qp 3)(n-rl/3,
3
where, for any integer n;;;' 4, the summation over r is for r = nand
those values of n - r which are divisible by 3. Also, show that the
expected number of trials required for the realization of the
pattern FSSS is 1/qp 3.
Verify that for n = 15, the probability that the pattern FSSS
will be realized at least once in the sequence of trials is
4 qp 3(3 - 9qp 3+ 5q2 p 6).
129 Explain briefly the concept of the statistical independence of two
continuous random variables, and illustrate your answer by indicating (with-
out proof) the behaviour of the joint sampling distribution of the mean and
variance of a sample of N (> 1) random observations from a normal
population.
A random variable X is normally distributed with mean I-'- and variance
0'2. If Xl> X2 are the sample means and sf, s~ the sample variances obtained
respectively from two independent sets of n observations of X, prove that
the function
( ) (n + 1)0'2
var w = 2n 2
360 EXERCISES IN PROBABILITY AND STATISTICS
130 Explain very briefly, without proof, the relation between the X2 d'
distributions. an I·
If X~ and X~ are independent random variables each having th 2
distribution with n dJ., determine the joint distribution of the ran~ X
variables Rand e, where om
X~= R cos2 e and X~= R sin2 e.
Use this joint distribution to verify that the random variable
X2_X2
W=_I __2=cot2e
2XIX2
has the probability distribution with probability density function
1
(~ 1:)' (1 + w1
2)<n+1)/2' for -00< W <00.
B 2'2
Hence show that if F has the F distribution with n, n d.f., then
t= In [JF-~]
2 JF
has Student's distribution with n dJ. and, for F> 1,
F=1+-+-2t 2 2t [ 1+-t2]~ .
n In n
[Note: For p>O, r(2p)r(!) = 2 2p - 1r(p)r(p +!).]
131 State briefly (without proof) the relation between the unit normal and
X2 distributions.
(i) If a continuous random variable X has the Laplace distribution with
probability density function
f(x) =~ e- 1xl , for -oo<X <00,
prove that the characteristic function of X is
E(e itX ) = (1 + t2)-I.
(ii) If Y1 and Y2 are independent unit normal variables, determine the
joint probability distribution of the random variables
Ul=~(Yl+Y2) and U2=~(YI-Y2)'
(iii) By using the above results, prove that, if Z1> Z2, Z3 and Z4 are
independent unit normal variables, then the random variable
W=ZI Z 2+ Z 3 Z 4
may be expressed as half the difference of two independent X2 ,s
each having 2 dJ. Also, show that the probability distribution of W
is the same as that of X.
132 Cards are dealt out one by one from a standard pack of 52 cards of
which four are aces and the rest non-aces. If X is the random variable
denoting the number of cards which precede the appearance of the third
SUPPLEMENT
361
(.t, and Y another random variable ~enoting the number of cards which
plCn up after the appearance of the thIrd ace and before that of the f h
(.t, prove that ourt
E(Y)= ~8.
Also, if Yh Y2, ... 'YII are the deviations of the Xi from x (Yi == Xi - x),
show that, for any given i,
Hence prove that a necessary and sufficient condition for the x. (._
~, 2, ... , n) to have equal variance is that the Yi should be uncorrelated W'ith
x.
135 Three discrete random variables Xl> X 2 and X3 have the joint
probability-generating function
E«(JX, (JX2(JX3) = ~
1 2 3 211
[1 +!!.211 {I + (J2(1211+ (J3)1I}1I]1I
where n is a positive integer. Use this probability-generating function to
prove that
. n 2(211 + 1)11-1
(I) P(X2 = 1) = 2 11 (11+1) ;
211(11-1)
(ii) P(X1 = 11 X 2 = 1) = (211 + 1)11-1 ; and
... [211 (11+1) + (211 + 1)11 ]11
(m) P(X3 = 0) = 2 11 (11 +11+1)
2
E(R)= ~.
137 Three players A, Band C agree to playa series of games with a biased
coin which has a probability p (0 < P < 1) for a showing a head. In each
game, each of the three players tosses the coin once, a success being
SUPPLEMENT 363
recorded if the toss leads to a head. The series is won by the first player to
obtain a success in a game in which no other player obtains a success. If just
1\\,0 players obtain successes in the same game they continue to play without
lite other player who is deemed to have lost the series. If all three players
obtain successes in the same game they all continue to play.
Determine the probabilities of the possible outcomes of the first game
played, from the point of view of a particular player (say A). Hence deduce
lit at Un' the probability that A wins the series on the nth game, satisfies the
difference equation
Un = (1-3pq)un_l+ 2p3q2(1-2pq)n-2, forn~2, q=l-p.
Finally, if cf>(z) = Lr= 1 urz r is the generating function for u'" verify that
cf>(z) = pq2z[1-(1-2p)z]
[1-(1-2pq)z][1-(1-3pq)z] ,
SO that cf>(1) =~. Explain very briefly the significance of this value of cf>(1).
138 Two gamblers A and B agree to play a game of chance with initial
capital of £a and £b respectively, the stakes at each trial being £1 on the
occurrence or non-occurrence of an event E. If E occurs at a trial then B
pays A £1, whereas if E fails to occur at the trial, then A pays B £1. The
lrials of the game are independent, and the constant probabilities of the
occurrence or non-occurrence of E at a trial are p and q respectively, where
p+q = 1. A player wins when the capital of his opponent is exhausted.
If Un (0 < n < a + b) is the expected number of further trials required for
A to win when his capital is in, prove that Un satisfies the difference
equation
Un = 1 +pUn+1 +qu,.-I'
Hence show that the expected duration of play for A to win is
a (a+b)(qa_pa)pb
q_p (q _ p )(qa+b _ pa+b) if pi= q,
and ab if p = q.
139 A finite population consists of the N elements Xl> X 2 , ••. ,XN , and
Ihe mean and variance of the population are defined as
_ 1 N 1 N
X=-
Nj=l
LX; and S2=_-
N -1
L (Xj-Xf
j=1
where Z b Z2, ..• ,ZN are indicator random variables associated with the
individual elements of the population sampled such that Zj is either 1 or 0
according as X j is or is not included in the selected sample. Use this
364 EXERCISES IN PROBABILITY AND STATISTICS
NI
(r-l)! (s-r~I)! (N-s)! u r- 1 (v-u)·-r-l(l-v)N-., O:s;;;u:S;;;l; v;:ou.
P(X;::;'Y)=~.
nl +n2
Determine the joint probability distribution of X and Y given that X;:o Y.
Use this probability distribution to derive the joint probability distribution
of U = X - Y and Y, given that X;::;. Y. Hence deduce the probability
distribution of U given that X;::;. Y.
Also, show that the unconditional distribution of IUI has the probability
density function
E(X) = n( 1-~r
and
SUPPLEMENT 365
Also, obtain approximate expressions for E(X) and var(X) when sand
~rJJ in such a way that sIn = A, a constant.
II
1.3 A continuous random variable X has a uniform distribution in the
'olerval -a";;; X,,;;; a. If X,t+l is the median of a sample of 2n + 1 independent
~bservations of X, determine the probability distribution of X n +!. Use this
probability distribution to prove that, for any non-negative integer r,
r aT(2n+2)f{!(r+ 1)}
E(Ix,t +1 I ) = 22n + 1r(n + l)r{n +!(r+ 3)} .
Hence, in particular, verify that
E(lxn+li) = a/~.
[Note: It may be assumed that for large m > 0
r(m + 1) = .J21Tm(mle)"'.]
144 Suppose that Xl> X2, ••• ,Xn are serially correlated observations each
with zero mean and variance u 2 , and corr(Xj, Xi+k) = P for all i arid k = 1,2
such that 1,,;;; i ,,;;; i + k ,,;;; n and zero otherwise.
If x is the average of the Xi> prove that
. u2 [ 2(2n - 3)P]
(1) var(x) = -;; 1 + n ; and
Comment very briefly how this result may be used to test a specific null
hypothesis about the parameter (3.
366 EXERCISES IN PROBABILITY AND STATISTICS
U = L Xi and v= L Yi·
i=1 i=1
147 Suppose Xl is the smallest and XII the largest of n random observations
of a random variable X having the probability density function I(x) defined
in the range -oo<X <00. Find the joint probability distribution of Xl and x .
If another random sample of r observations of X is obtained, determine t1~~
conditional probability, given XI and XII' that these observations all lie
outside the range (XI' x,,). Hence show that the total probability of the
second sample observations lying outside the range of the first sample is
r- 2
[
X
1- f/(X) dx T/(x,,) dX II
P=
f(r+2) [
nr
r(r+ 1)]
l-~ approximately.
.
148 Prove that the probability-generating function of a Poisson variable X
with mean A is
E(OX) = e A (9-1).
If Xl and X 2 are two independent Poisson variables with means A1 and
A2 respectively, show that the joint probability-generating function of XI
and Y=X1 +X2 is
E(O;'!.Oi)=exp[Al(OI02- 1)+A2(02- 1)].
Hence prove that the marginal distribution of Y is Poisson with mean
Al + A2 but that the conditional distribution of Xl given Y = j is
when x is small.
151 A continuous random variable X has the rectangular distribution in
the interval (0, a). Prove that
P(X";;!a)=!.
If Y is another continuous random variable such that the joint distribution
of X and Y has the probability density function
1
f(x, y) = ( )' for 0,,;; Y,,;; a - X; 0,,;; X ,,;; a,
a a-x
prove that
P(Y";;!a) = pea - X - Y ";;!a) =!(1 + log., 2).
In an experiment a particle Xo with energy Eo is subjected to radiation.
There is a probability p that Xo will not split, and a probability 1- p that it
will split randomly into two particles Xl and X2 with energies EI and
Eo- EI respectively such that EI is uniformly distributed in (0, Eo). In a
second experiment the unsplit particle Xo or the particles Xl and X 2 are
again subjected to radiation under the same conditions as in the first
368 EXERCISES IN PROBABILITY AND STATISTICS
experiment. Prove that the probability that all the particles result in f
the two experiments are less than tEo in energy is g rOIl}
t(l- p )[1 +!(1- p )(1 + loge 2)2].
152 If X is a continuous random variable having the Laplace distrib t'
. h pro b a b'l'
Wit I Ity d enslty
. f unction
. U IOn
f(x)=texp-ix-8i, for-oo<X<oo,
prove that the moment-generating function of X about the origin is
E(e tX ) = e t8 (1- t 2 )-1.
Hence show that, if x is the sample mean of n random observations of X
then the standardized random variable '
x-8
z=--
mn
has the moment-generating function (1- t 2 /2n)-n.
Further, verify that the coefficients of skewness and kurtosis of the
probability distribution of z are 1'1 = 0 and 1'2 = 3/n respectively, and thai
for large n, z is approximately a unit normal variable. Alternatively
assuming that an approximating non-normal distribution representing th~
behaviour of z has the probability density function proportional to
e-!z2(1 + az 2+ (3z4),
determine the normalizing factor of this probability distribution. Also, if this
approximating distribution has the same first four cumulants as the exact
distribution of z, prove that
6
a = -6{3 = - 8n + 3 .
a -
_ (2n+1)
2n x,.
is an unbiased estimate of a with variance a 2 /4n(n + 1).
Also, derive the sampling distribution of the standardized random vari-
able
z=---
a-a
s.d. (a)
and then prove that, as n ~ 00, this distribution has the limiting form with
probability density function e- l for -00< z:s:;; 1.
156 A savings bank permits a client to withdraw money not more than
once a week. It is observed from the past experience of the bank that the
probabilities of a client withdrawing money in the week preceding and the
one following Christmas are PI and P2 respectively, and the probability of a
client using the facility in both weeks is P12, where O<PI2:S:;;min(pt> P2). The
bank has a total of n clients, who may be assumed to be operating their
accounts independently, and Xl, X 2 are discrete random variables denoting
the total number of withdrawals made in the week before and after
Christmas respectively.
Prove that the joint probability-generating function of Xl and X 2 is
E(Oi". 0~2) = [1 + Pl(OI -1)+ P2(02-1)+ pdOI -1)(02 -1)]".
Further, assuming that
al2
PI2=-,
n
where at> a2, and al2 are finite constants, determine the limiting form of the
above probability-generating function as n ~ 00. Show that for the limiting
joint distribution
al2
corr(Xt> X 2 ) = ~.
vala2
Hence verify that in this limiting case if Xl and X 2 are uncorrelated then
they are also independently distributed random variables.
370 EXERCISES IN PROBABILITY AND STATISTICS
157 Two continuous random variables X and Y have the joint distribur
with the probability density function IOn
0: - 2
(3 - yO< (y )
no: + 1) . (1 + x)13 . exp - 1 + x '
defined for O:s;; X, Y < 00, where 0: (> -1) and (3 (>0) are parameters sUch
that (3 > 0: + 2. Prove that
(i) the marginal distribution of Y has the probability density function
«(3 -0: -2)g(y)
o:s;; Y<oo,
no: + l)y13-o<-l '
where
Y
g(y) == Je- w 13 -
w 2 dw;
o
(ii) the conditional distribution of X given Y = y has the probability
density function
y13- 1 (Y )
exp - - -
g(y)(l+x)13 l+x '
and
... e- Yy13- 1 y
(Ill) E(X I Y = y) = «(3 -2)g(y) + (3 _2-1.
Also, if two new random variables U and V are defined by the relations
Y 1
U= l+X and V=l+X'
where 0: and (3 are positive parameters. Suppose that Xl> X2, •.• ,XII are /I
independent observations of X. Given that Xl is the smallest of the /I
observations, prove that the distribution of Xl has the probability density
function
no:- l exp -n(xI - (3)/0:, for Xl ~ (3.
Hence prove that
Also show that, for given Xl> the conditional distribution of Xj (if 1) has
the probability density function
o:-lexp-(Xj-XI)/O:, forxj~xl'
SUPPLEMENT 371
Ose this probability distribution to show that
t=-s-
x~ and
" / Vi~Xf,
W=i~Xi ,-"
where x and s are the sample mean and standard deviation respectively.
Prove that
(i) -JIi ~ w ~ JIi; and
(ii) t 2 = (n -1)w 2
n-w 2
Hence deduce that t is a monotonic function of w.
Starting with the joint distribution of x and s, use the polar transforma-
tion
x~= r sin 6, sJn -1 = r cos 6
to prove that the probability density function of the distribution of u = wlJli
is
1 . 2 !(,,-3)
BH, ~(n -1)} (1- u ) ,
Hence show that the probability distribution of v = Il.Jn -1 has the proba-
bility density function
1 2 _!"
BH, ~(n-l)}" (1 +v) , -oo<v<oo.
[Note: It may be assumed that the probability density function of the joint
distribution of x and s is
(n -1)!(,,-J)~ ]
2!("-2)rmf{~(n -1)} exp[-Mni 2+(n -1)s2}]s,,-2.
YI = Xl sin 6 + X 2 cos 6,
where
6 =It _1(-2 Pcr l cr2 )
2 an 2 2'
crl-cr2
then the random variables YI and Y2 are independently distributed with
zero means and
var(YI ) == cr~, = t(cr~ + cr~) +t[(cr~- cr~)2+4(pcrlcr2)2p;
var( Y2) == cr~2 = ~(cr~ + cr~) - mcr~ - cr~)2 + 4(pcr1cr2)2J!.
Hence, or otherwise, show that the moment-generating function about
the origin of R 2 is
E(e'R2) = [(1- 2tcr~,)(1- 2tcr~2)]-~
and then verify that the jth cumulant of R2 is
Kj(R 2 ) = 2j-l(cr~~ + cr~~ . (j -i)!
ANSWERS AND HINTS ON SOLUTIONS
Supplement
---------------------------------------------------
Finite differences and summation of series
= L" r(r+4)[l(r+1)(3)-2r]
r=1
n 11
=t L (r+4)(5)-2 L [(r+3)<4)-2(r+1)(2)]
n
-2 L [(r+2)(3)+(r+1)<2)-3r]
r=\
= h(n + 5)<6) - 2[k(n + 4)(5) -~(n + 2)(3) +!(n + 3)(4)
+Hn +2)(3) -~(n + 1)(2)]
=-k(n + 1)(2)[5n 4 + 34n 3 -14n 2 -331n -144].
FOrll=1, sum=-10.
II r
= L r(r+3) .1(r+1)(3)
r=\
=l L" [(r+3)(5)-2(r+2)(4)-2(r+1)<3)]
r=\
= HHn +4)<6) -hn + 3)(5)-!(n +2)(4)]
= -k(n + 2)(4l[5n 2 + 23n + 9].
For 11 = 2, sum = 20.
n-l r
(iii) Sum= L r(r+2)(r+4) L (s-1)s(s+2)
r=\ .=2
'1-1 r-l
= L r(r+2)(r+4) L [(t-2)<3)+(t+1)(2)]
r=1 1=\
11-1
=L r(r+2)(r+4)[!(r+2)<4)+Hr+ 1)(3)]
r=\
373
374 EXERCISES IN PROBABILITY AND STATISTICS
11-1
=! L [(r + 5)(7) - 6(r + 4)(6) + 3(r + 3)(5) + 3(r + 2)(4)]
r=1
11-1
+i L [(r+4)(6)-3(r+3)(5)-3(r+2)(4)]
r=1
= Uk(n + 5)(8)-~(n + 4)(7) +Hn + 3)(6)+~(n +2)(5)]
+Ht(n + 4)(7) -!(n + 3)(6) -~(n + 2)(5)]
= 4Ao(n + 2)(5)[15n 3+ 100n 2+ 125n -144].
= L (r+2)(r+6)[i(r+ 1)(3)-~r(2)+8r]
r=1
,.+1 ,.+1
=i L [(r+3)(5)+3(r+2)(4)]_~ L [(r+2)(4)+5(r+1)<3)+5r(2)]
r=1 r=1
n+l
+8 L [(r+2)(3)+5(r+ 1)(2)+5r]
r=1
= Hi(n + 5)(6)+~(n +4)(5)]-UHn +4)(5)+Hn + 3)(4)+i(n +2)(J)]
+ 8[Hn + 4)(4) +i(n +3)(3)+HI1 +2)(2)]
1=0
(11-2)
t
x' + 11
,,-I
r=O
L (n -1)
r
1" (n) x'
xr - - I
x t
1=1
1
= n(I1-1)x(1 + X),,-2+ 11(1 + X),,-I_- [(1 + x)" -1].
x
e;
Ur =
For 11 = 1, sum=3.
5 (i) The rth term of the series can be written as arx r where
ar = 2r2 and x =!, for r~O.
Observe that Llkao = 0 for k ~ 3. Hence the sum of the series
=--+
o 2x 4x 2
+---::-
I-x (l-x)2 (l-x)3
= 12.
(ii) Here the rth term of the series is arx r where ar = !r3 and x = t. Also
Llkao=O for k~4. Hence the sum is 13.
6 UL (1961).
(i) Since Llu" = U,,+1 - Un we have
U n +l = (1 + Ll)u"
= (1 + Ll)2U"-1 = ...
= (I+Ll)"ul.
(ii) Assume that the result holds for m -1 so that
Hence
Ll"'UI = L
",-I
(-1)'
(m -1)Llu",_ro where Llu,,,-r = U",+I-r - Um - r •
r=O r
The result for m is seen to hold by combination of binomial
coefficients
n
~ A-I A-I
(iii) '- Ur = L.1 U,,+1 - L.1 Ul
r=1
= Ll- 1[(1 + Ll)n U1 ]- Ll- t Ut
= [(1 + Ll)n -1]Ll- t u 1
7 UL (1961).
Assuming the form of f(x), show that
Ll2f(x) = 2h2a2+6h3a3+6h3a3x.
If this is == a + (3x, then
(iii) aAl3x = (A13 h -1)A13 and, in general, a rA I3x = (A I3h -1)'A I3x.
X
Hence, on summation,
II-I (A 13h -1)"-1
L a Al3x
r~O
r =-'------:-:---
Al3h -2 .
Also,
9 UL (1963).
(i) alog(1 + x) = 10g[1 +_h_]
l+x
so that
11.21 (1 )=1 [(1+X)(I+X+2h)]
L.l og +x og (l+x+h?
=IOg[I-(I~Xr{l+ I~J-2]
- log [1 - (1 ~ x r]
__ (_h
l+x
)2 +0(h 3 ).
The result stated now follows by noting that cos(x + It) = (1 + A) cos x
1-cos h = 2 sin2 h/2- h2/2 for small h. and
(iii) Ae 2= eX'(e2hx+h2 -1), and
A2ex2 = A[e'(e2hx+h2 -1)]
= e'(e4hx+4h2_2e2hx+h2+ 1)_eX2 . 2h 2 . (1 +2X2)+ O(h 3 ).
10 UL (1965).
(i) The general result follows on addition and for the particular series
note that
1
Ur = f(r) - f(r + 1), where f(r) = 6(3 2)( .
r- 3r+1)
(ij) For r = 2 and 3 the recurrence relation gives
28 - 8a + (3 = 0; -80+28a -8(3 = 0
whence a = (3 = 4.
The sum of the series is (1-4x)/(1+2x)2 for Ixl<!.
11 UL (1977).
(i) Here log x = (1 +!n)-l . log 0'
- (1-!n) log 0'
so that log(x/O') - -!n log a or
(ii) Here Ur = f(r) - f(r+ 1) where f(r) = (2r+ 1)/2r(r+ 1). Hence
L~'~ 1 Ur = f(1) - f(n + 1) ~ i as n ~ 00.
g(a,b)=2xa [1+(1-~rl]+IOg(1-~)
=:: [(!-~)+(!-t)~+(!-t)(~r +··1
Hence
x3
g(a, b» 6a 3 and
1 1 -5)
= --;- 3x 3 + O(x .
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT
379
Hence
1 1 1
-1--+--- on expansion.
X 2X2 2x 3 '
(-xY 1 (-xy+ 1
(iii) Note that Ur =--+-. 1 whence
r x r+
~ 1
L Ur = -log(l + x) - - [log(l + x) - x].
r=1 X
13 UL (1973).
(i) Note that Ur = (2 r -l)/(r+ 1)! whence
~ ~ ~
= 2 e -2X'[1 + 2 x 2 +3+'"
2x4 ]
(ii) L
"'-(3 (
a-
k) = ",-(3
L coefficient of x(3 in (1 + X)",-k
k=O (3 k=O
",-(3
( -1 )r-l r=1
x 1 +l1a '
= 0 on summation.
(ii) Here u r =f(r)-f(r+1), where f(r) = 4Ir 2(r+ 1f. Hence
L" U = f(l) -
r f(n + 1).
r=1
(iii) Here f(r+ 1) - f(r) = 14r6 -2r2 whence the expression for r6. Hence
f1 fI
L r6=9 L r2+1~[f(n+1)-f(1)],
and the answer follows on reduction.
17 UL (1969).
(i) Note that (r+s+1)(s+2)-(r+s)(s+2)=(s+2).(r+s)(s+1), whence the
result on summation. In particular, for s = 0,
Sr = !r(r+ 1) and S,,-r = !(n - r)(n - r + 1).
It now follows that
S.5,,-r =;\[(n + 3)(2)(r+ 1)<2) -2(n + 3)(r+ 2)(3) +(r+ 3)<4)].
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT 381
Hence
II
L SrSII-r =;t[(11 +3)(2)xi(n +2)(3)-2(n+3)Xi(n+3)(4)+!(n+4)(S)]
r=1
and the required result follows on reduction.
II (11)
(ii) Note that r~o r x r+1 =(1+x)II+1_(1+x)lI.
Ltl Jr r = n 2(J2
=!n 2 (n + 1)- n 2 var(.Jr)
=n 2(n+1)
2
[1- l1(n+1)
2V(n) J.
19 UL (1967).
(i) Note that ur+l-ur=24r7-8r3, whence
11 n n
L r =1 L r3+2~ L (u r+
r=1
7
r=1 r=1
1 - U r)
(iii) If x takes the values 2r -1 for 1 ~ r ~ 11 and s> r, then the required
sum is
II
L (2r-1) L (2s-1)
(2r-1)2 J
r=1 s>r
=2"1 [{r~1II (2r-1) }2 - r~1
II
20 UL (1966).
(i) We have Llf(r) =sin[O+(r+ l)ebJ-sin(O+rcp)
= 2 cos[ 0 + eb/2 + reb J sin( eb/2)
= 2 sin(eb/2) . sin[O +!( 7r + eb) + reb J.
Hence
Ll 2f(r) = 2 sin( eb/2)Ll sinE 0 +!( 7r + eb) + reb J
= [2sin(ebf2)f. sin[O+(7r+eb)+ rebJ.
The general result now follows by induction.
(ii) By definition U r = (1 + Ll)Ur-1 = (1 + Ll)'uo. Therefore
t
r=O
(n)(_l),,-r ur =
r
t
r=O
(n)(-l),,-r(1 + Ll)'uo
r
= Lto (;)(-l),,-r(l+LlY]. Uo
For ur =(-lY/(r+1),
Ll"uo=.to (;)<-1)"/(r+1)
=(-1)". t
(n+1)
n+1 r=O r+1
(-1)"(2,,+1-1)
n+1
( ...) H _ (r+2)(2)-2(r+2)+ 1
111 ere Ur - (r+2)!
r~1
" r [
u = 1
1 ] [1 1]
(n+1)! - 2-(n+2)! .
P(iii) = st (;)(~)"(~r-s.
zZ UL (1966).
Let A be the event that the player has a flush, and B the event that he
has at least one spade.
- 99
(iv) P(AB) = P(A) - P(AB) = 66640 = 0·0015.
31416
(v) P(A I B) = P(AB)/P(B) = 49 386 904 0·0006.
(vi) P(B I A) = P(AB)/P(A) = 1.
Z3 UL (1967).
(a) Let X be the event that there is at least one bad fuse in the selection
of P, and Y the event that all the v fuses selected are bad. Then the required
probability is
P(Y I X) = p(YnX)/p(X) = P(Y)/P(X),
whence the result.
(b) Let Z be the event that all the· fuses in the second selection of v are
good, and Tr the event that there are exactly r bad fuses in the first sample
of P, where l:so; r:SO; v. The required probability is
.
P(Z I X) = P(Z n X)/P(X) where X = LT r•
r=1
Therefore
p(znx) = f P(Z n Tr) f P(Z I Tr)P(T.}.
r=1
=
r=1
24 UL (1968).
(i) With replacement sampling the total number of sample points ar" .
from the numbers on the three tickets is (611)3. The numbelSlllg
sample points leading to the realization of S = 611 is the coefficie~t or
0 611 in the expansion of or
H(O) = (1- ( 611 )3/(1- 8?
The required coefficient is 18n 2 +9n-2=(3n+2)(6n-1) whe
the answer. ' nee
(ii) The probability-generating function of S is H(0)/216n 3 and E(S) 0
H'(1)/21611 3 •
When sampling is without replacement then the total number of sampl .
points is 611(611 -1)(6n -2), and the number of sample points leading to th~'
realization of S=611 is also less than 18n 2 +9n-2. From this we have te
exclude the sample points corresponding to the integral solutions of x + y :'
z = 611 whe~e (a) x = y = z and (b) y = z1= x so that?' +2y = 6n. The solution
correspondmg to (a) IS x = Y = z = 2n and the solutions corresponding to (h)
are obtained for 1:so; y:so; 3n. The solutions for (b) are 3n -1 in number and
hence the total number of sample points are 3(3n -1). Therefore the
excluded sample points corresponding to (a) and (b) are 9n -2, whence the
answer.
25 UL (1969).
(i) Suppose Xi is the event that the ith ball drawn is white, i == 1,2.
Then
. n(n+r)
P(I) = P(X1n X 2 ) = P(X2 1 X\)P(X1) = ( ) ( .
m +11 m+l1+r)
(ii) Similarly or by symmetry
P(ii)= m(m+r)
(m+n)(m+l1+r)
(iii) Suppose Yi is the event that the ith ball drawn is black, i == 1,2.
Then
P(iii) = P(Y2 1X1)P(X,)+ P(X2 1 Y1)P(Y,)
2mn
(m + n)(m + n + r) .
(iv) P(iv) = P(X2 ) = P(X2 1X,)P(X1) + P(X2 1 Y1)P( Y1)
n
=--
m +11
(v) P(v) = P(Y1 I X 2 ) = P(Y1n X 2 )/P(X2 )
= P(X2 1 Y1)P( Y 1)/P(X2 )
m
m+n+r
Expectation of white balls drawn is 2nl(m + n).
26 UL (1970).
Suppose X is ·the event that all r balls are of the same colour, Y tl!l'
event that all are black and Z the event that all are white. Then X == Y+Z.
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT 385
"here in the first experiment
P(Y) = P(Z) = (J / (2:).
/'he required probability is
P(Z I X) = P(Z n X)/P(X) = P(Z)/P(X) =!.
In the second experiment
(J 1
=~[1_k(~~1)][1 k(k;~-1)r1
-~ (1+2k~).
17 UL (1979).
P(Holmes dies) = xy + (1- x )(1- y) = a (1)
P(Holmes escapes) = (1- x)y = a (2)
P(Draw) = x (1- y) = 1 - 2a. (3)
From (2) and (3), x = y + 1- 3a whence on elimination of x from (1) we
obtain the stated quadratic in y.
The condition that the quadratic has real roots gives for a> 0, a ;;a;. §. The
inequality now follows since from (3) 1- 2a;;a;. o.
28 UL (1971).
(i) By the hypergeometric distribution
and
where
P(Xk ) = P(k) as in (i)
and
If it is known that A has all court cards, then the probability of B hav·
no court cards is Ing
Ek = f 21_kr(~)k(r-1)[1_ (~)k]
r=1 6 6
1 1-(i)k 1
= 2 k- 1 • 1-U2)k <2 k- 1 •
Hence, on the average, the player can never be on the winning side if the
non-returnable stake to play the game is £2 1 - k •
We have
Ek 1-(i)k 1-(fi)k+1
E k+1 =2 ·1-(i)k+1· 1-(f2)k
1-(i)k
>2· 1 _(i)k+1
_12 [ (i){1-(i)k-1}]
-11 1 + 1_(~)k+1
12
>11> 1, whence Ek > E k + 1 •
30 UL (1973).
Suppose Xr (0 ~ r ~ n) is the event that there are r white and n - r black
balls in the urn, and Y the event that in the sample of k balls drawn, there is
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT 387
oply one white ball. Then
Il-k+l
P(Y) = L P(Y I Xr)P(Xr),
P(Xr) = (;)prqll-r
apd
P(YI Xr)= G)(:~~)/(;) for 1~r~n-k+1
=0 for r=O.
Ip the second case
P(Xo) =0;
31 UL (1975).
The probability that the player gets far at the rth trial is (~~y-l. ~~.
Hence the expected gain of the player is
(25)r-l 11 11a
r~l
00
36 . 36' a r = 36-25a .
The player, on the average, will be on the winning side if the expected gain
>! which gives a >~.
32 UL (1977).
Irrespective of the suit of the cards, the following ten realizations of five
cards lead to a run: (A, K, Q, J, 10), «K, Q, J, 10,9), ... , (5,4,3,2, A).
Since there are four suits and in all 52 cards, the probability of a run is
whence the probability that the player has a run but not a running flush is
40. 255Wn = 81> say.
If it is known that the player has the jack of hearts, then the three
corresponding probabilities are
When the player has the jack of hearts, these probabilities are chan d
~~PI and ~P2 respectively. Hence the percentage change is ge to
35 UL (1969).
Suppose Hr is the event that the rth urn was selected and X the e
that the two balls drawn are white. Then vent
k
P(X) = L P(X I Hr)P(Hr) where P(Hr) = 11k.
r=l
n-r)2
(i) Here P(X I Hr) = (- - for 1 ~r~k-l
m+n
=1 for r = k.
Hence
G(Ol> O2 ) = L L O;:O~P(XI = r, X 2 = s)
r=Os=O
It nr
38 UL (1969).
Let X", Y,,, Z.. be the events that after n moves the system is in an
A-type, B-type or C-type cell. Then for n ~ 1
X,,-l + Y,,-l + Z,,-l = n, the sample space.
Hence
P(X,,) = a,,_IP(X.. I X.. - I) + ~"-IP(X" I Y,,-l) + 'Y.. -IP(X" I Z.. _I)
and similar equations for P(Y,.) and P(Z..).
Symmetry considerations now show that
P(X" I X,,-l) = 0, P(Y,.I X,,-l) =i, P(Z" I X .. - I ) =t
P(X" I Y,.-I) =!, P(Y,.I Y,,-l) =t, P(Z,. I Y,.-l) =t
P(X,. I Z .. -l) =i, P(Y,. I Z,,-l) =!, P(Z" I Z,,-l) =i.
Hence the difference equations
a" = !~"-I +i'Y"-1
~Il = iall-l +t~Il-1 +!'YIl-I
'Y" =~a"-l +t~"-l +i'Y,,-1
with the conditions a,,+~,,+'YIl=a"-I+~"-I+'YIl-l=1. Elimination of an
and 'Yn gives the following non-homogeneous equation
70= 120~" +27~"_1.
The transformation ~" = x,. +¥t reduces this to
40x,. +9x,,_1 =0.
Hence
10 (9)11
~Il = 21 + k - 40 for n ~ 1
where k is arbitrary.
For n = 0, ao = 'Yo = t ~o =! and so ~l = flo. Hence k = -li9. Therefore,
using earlier equations,
all =!-*~.. and 'Yll =!-a~".
By symmetry, the probabilities for specific A-type, B-type and C-type
cells are aa", i~" and h". As n ~ 00, these probabilities tend to i4, i.., /4.
392 EXERCISES IN PROBABILITY AND STATISTICS
39 UL (1967).
(i) Let Xo be the initial probability of A winning the series; this i' . I.
the probability of A to win whenever the scores of the two PIS d So
'h
are Ieve,I If XII IS t e pro b a b'l'
I Ity 0 f A ' 'h
winning t e series when ayers
he ..
n games ahead of B (-4~n~4) then the forward diffe" IS
, , lenee
equation IS
XII=PXII+I+qxlI_l+rx", for-4<n<4, x4=1, X-4==O,
whence
PX,t+2-(P+q)XII +1 +qxlI = O.
The roots of the auxiliary equation are 1 and q/p and so
XII = IX + {3(q/p )11,
where the arbitrary constants IX and (3 are determined by using the
conditions X4 = 1 and X-4 = 0, Hence
p8_ p4-lI q ll+4
XII = 8 8
P -q
and the stated result for Xo follows by putting n = 0,
(ii) This probability is obtained by putting -n for n and interchanging II
and q in the expression for XII'
40 UL (1967),
The probability that all the k bottles of the first customer are fresh is
Let Xr be the event that r fresh and v - r a day old bottles are sold oul
before the selection of the second customer, and Y be the event that she
selects k fresh bottles, then
v
where
41 UL (1966).
The difference equation for Un is
Un = p2 . U,.-3 + pq . U,.-2 + q . u,.-\> for n ;a. 3
with Uo = 1, Ul = q, U2 = pq + q2 = q. The roots of the auxiliary equation are
1, (-!p)(1 ± i.J3) whence the general solution
xt (=:~)= xt C~~)
= Nf' (t+n+k)
1=0 t
N-n
= L coefficient of ZO in z-l(1 + z)'+n+k
1=0
N-n
= coefficient of ZO in (1 + z)n+k L {(1 + Z)/Z}I
1=0
= coefficient of ZO in (1+z)n+k[(1+z)N-n+l. z-(N-n)_z]
= coefficient of zN-n in (1 + Z)N+k+l
= (N+k+l)= (N+k+l).
N-n n+k+l
For k = -1, we obtain the sum of the probabilities P(X = x). The expression
for E[(X + r -1)<')] is also obtained by the use of the equality whence the
required values of E(T) and var(T).
43 UL (1966).
The difference equation for Sn is
Sn+l =(1 +p)Sn +nR +(I-p)R +p. 2R, and So= R.
To solve this difference equation, make the transformation Sn = Xn + a + (3n,
394 EXERCISES IN PROBABILITY AND STATISTICS
(2N + 1- n)(1)2N+l-"
=1- L
N-"
-
r=O r 2
-ct>(- n )
\.J2N+1-n '
by using the normal approximation for the binomial distribution.
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT 395
For fixed N this probability is an increasing function of n and if
we assume that <1>(3) -1, then P(r ~ N + 1 - n) - 1 if
n2~9(2N +1- n),
. N-n+1 '
47 UL (1965).
The probability that the player loses for the first time at the (r+ 1)th trial
is prq and his net gain is rx - a. Hence
QO
E(S) = L (rx-a)prq
r=O
QO
= -a + pqx L rpr-1
r=1
d
= -a +pqx. dp [(1-p)-1], whence result.
E(S2)=X2q L {r(r-1)+r}pr-2axp/q+a 2
r=O
Hence var(S).
The inequality E(S) ~ a leads to x> a/p since 2q < l.
48 UL (1975).
and so
x=1
50 UL (1971).
~ ~
(i) L r(k)p(X = r) = f.L k L e-lLf.Lr-k/(r- k)!
r=a
~-k
=a e- IL e- 2/3/-L2
[
Le /-L +2e-/3/-L L e /-L ,
r=2 (r-2)! r=1 (r-1)!
whence the answer on summation.
Since 1- /-L < 0 and /-L e-/3 < 1, we have
Expected loss < 3a e 1 - 1L < 3a.
S2 UL (1969).
Assume a Poisson model for the distribution of errors with mean /-L.
Then the expected cost of correcting the errors made on a stencil is
=I e-IL/-Lr 2r(3r+2)
r=O r! . r+1
=2e- I /-L'[3r(r+1)-(r+1)-1]
1L
r=O (r+1)!
r r 1 r+l]
3 L _/-L__ L ~+- L _/-L_
<X> <X> <X>
=2e- 1L [
a -2[3/-L -1 +; (l-e- IL )]
E(Y2) = a f
+ N '\x e-A(x-l) dx
1
00
f
=a+N '\(u+1)e- AU du,
o
whence the result. Therefore the expected population at the start of the
r
third year is
=
a
[(,\ + 1)k _,\
,\ k-l
k] + N (~)k
,\ .
55 UL (1965).
00
E(T)= ~f x(1+{3x)exp{-~-{3(a-X)}dx
v27T{3 -00 2{3
=
-0<11+111 3 fOO
a ~2
{1
x(1 + (3x) exp - - (x - {32f dx.
}
27T{3 2{3
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT 399
The integration is completed by using the substitution z = x - (32 a d th
. 0 f th
the propertIes i 'Integra.I
e norma n en
Differentiate E(T) logarithmically with respect to a to maximize E(T).
S6 UL (1966).
For small a, (1 - 1~O) 100 ~ e -a. Hence the probability that out of ten
consignments exactly w will be rejected is
To solve this put z = 1-e-a , where z is small. The terms containing z and
Z2 vanish and we have approximately
Therefore
58 UL (1967).
The probability of finding < n defectives in the sample of N is
This is also the probability of accepting the batch after inspection, which
now contains only M - N components. For each component sold, the
expected net income is 1 - a6. Therefore
whence
1
6(M-N) ..
= (N+4)(4) Z~1 z(N-z+2)(N+4-a-az),
a + {3p, - ~
(TV27T
J(X 2_(T2) exp[-(x -p,)2/2(T2_'Y(X-p,)/(T] dx
_00
The stated result now follows by putting u = z + l' and then integrating over
u by using .the properties of the normal integral.
The expected cost after devaluation is
a' + {3' p,' - e!"Y'( p,' - 1'17')2
whence, under the given conditions,
, a+{3p,-a' , (a -a')-({3 -{3')p, +'Y{3'(T
p, = {3' ; 17=
'Y{3'
Under the modified system the proportion of underweight packets is
ct>[(l-p,')/(T'] whence the stated result.
60 UL (1972).
The expected payment per page is (in £)
ix e- IL + 1 x p, e- IL +ix [1-(1 + p,) e- IL ],
whence the answer.
If £0 is the new payment to the proof-reader for every page on which he
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT 401
detects no errors, then
9 x e - JL + 1 x,.., e- JL +lX[l-(l +,..,) e- JL ] = 1
whence
9 =i(S+,..,-e- JL ).
61 UL (1973). co
f
k e- Ax2 x 3 dx = 1.
o
The integral is readily evaluated in terms of the gamma function by using
the transformation z = x.Ji...
The average consumption of gas per day is
co co
f
2,\ 2 e-,\x x 4 dx = ~
o
2
JA. 0
fe-"u~ du, where u = '\x 2 ,
Note that
I e-~z2 I
00 00
65 UL (1977).
If k is the proportionality factor, then
I
00
X",+,-1
E(X') = k (1 + {3x",)2n dx.
o
The integral is reduced to a beta function of the second kind by using the
substitution z = {3x"'. Hence for r = 0, k = (2n -1)a{3.
Logarithmic differentiation of the probability density function gives the
equation
(a -1)(1 + (3x"')-2na{3x'" =0
for the stationary values of X. For the mode the second logarithmic
derivative of the probability density function of X is <0.
66 UL (1975).
'"
P(X~a)= B(m,
1
2)
IX Ill - 1 (I-X)dX,
o
whence the result on integration. P(X ~ (3) is obtained by symmetry. Hence
P(X~{3 X~a) P(X"':::{3)
P(X~{31 X~a) = ' ~
P(X~a) P(X~a)'
67 UL (1973).
1
whence the stated result. The median is obtained as a root of the equation
F(x) = o.
68 UL (1973). 00
1
= <1>(6) + r.::- [1(4, 00) - 61(2,00) + 31(0,00)]
nv27T
1
+ r.::- [1(4,6)-61(2,6)+31(0,6)],
nv27T
whence the result on reduction.
71 UL (1970).
Since the probability density function of X is a function of Ixl and the
distribution is symmetrical, all odd central moments of X and E(X) vanish.
If k denotes the proportionality factor, then
00
= 2k[a Je- x 2r dx + J
13x e-l3xx2r+1 dx J.
o 0
72 UL (1969).
The probability density function of X is an even function of x and .
E(X) = O. Hence 00 So
iJ.2r(X) = E(X2r ) = k J (a + X 4 )X 2r e- x2 dx
00
J
= 2k e- (a + X 4 )X 2r dx,
X2
o
where k is the proportionality factor. The integral is evaluated by the use or
the substitution x 2 = z.
For r = 0, E(X2r ) = 1 gives k = 4/(4a + 3)~.
Logarithmic differentiation shows that for a> 1, the mode of X is at the
origin. For a =~, var(X) = 1, 1'2(X) = -~ and maximum ordinate is 3/4/;.
For a unit normal variable Z, var(Z) = 1, 1'2(Z) = 0 and maximum ordinate
is 1/~ which is less than the maximum ordinate of X. But since
1'2(X) < 1'2(Z) , the frequency curve of X is to be "defined" as platykurtic.
73 UL (1968).
The probability density function of X is e- x/(1 +e- X )2.
= Jyt(l- y)-t dy
o
=B(I+t, I-f), Itl<1.
The moments are obtained by expansion of E(e tX )
as a power series in /,
whence 1'1 and 1'2'
(i) P(X:::; x) = ~ and P(X:::; x) = i give the first and third quartiles as
-log., 3 and log., 3. Hence the S.I.Q.R.
J dX]=1
00 00 00
k[J~-
l+x2
J~+
l+x4
x
2
l+x4 .
0 0 0
The second and third integrals cancel out as can be seen by putting x == 1/11
in the third integral. Hence k = 2/7r.
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT 405
Logarithmic differentiation leads to the stated equation for the turning
~alues.
The equation can be written as
(v2+~v+i)(v-~)=k where v=x 2.
rhus v -~. The modal values are approximately x = ±~, since the second
derivative is <0.
The improved approximation is obtained by setting v = 8 + ~ where 8 is
small. This gives
(8 +~)3_(8 +~)-2 = 0,
whence neglecting 1;2 and 83 we have 8 =;k,. Hence the improved approxi-
mation for the modal values.
7S UL (1967).
00
2a 2
E(e IX )=13 J[exp[-2ax/f3+tx]+ (1-a)2 exp[-2(1-a)x/f3+tx] ] dx
~
o
00
= 2a 2 J[exP[-Z(2a - f3t)] +
o
e :ayeXP[-z{2(1-a)-f3t}]] dz
where z = x/f3 and 2a - f3t > 0, 2(1- a) - f3t > 0. The stated result follows on
integration. To obtain the moments, note that
2[1-2a(1-a)] t 2
=f3t+f3 2a(1-a) 2+"" on expansion,
whence the E(X) and var(X). The lower limit for var(X) is obtained by
noting that a(1-a)~~.
76 UL (1966).
The moment-generating function of X is E(etX ) = ell" whence the mo-
ments.
The proportionality constant k for the probability density function of Z
is obtained from the equation
00
k Je-!z,[1+az(z+1)]dz=1,
whence
k = 1
.Jh(1+a) .
The vth moment of Z about the origin is
J
00
= IL~(Z).
406 EXERCISES IN PROBABILITY AND STATISTICS
For v = 2r + 1 and v = 2r, the required moments are obtained by using the
integral expressions for the moments of the unit normal variable X.
In particular, E(Z) = a/(1 +a) and var(Z) = 2-1/(1 +af.
77 UL (1965).
00
J
E(e'X) = A e-(A-')x dx = (1- t/A)-I.
o
The cumulants of X are obtained by the expansion of -log(1- t/A).
The point-probabilities of Yare c e- Ay where
00
c L e-A'=1 or c=1-e-A.
• =0
Hence
00
(e A - 1f - A2 e A
K2(X) - KiY) = A2(eA _1)2
~2wO 20 j =t
SO that
n 1
log L = constant-llog 0 - 20 [ns 2 + n(x - Of].
logL=constant-3nlogO+2
j=t 0
whence the stated value of 0. Var(O) = var(X)/9n.
E(L t ) = ~
4n
r
j=t
E(xD = !E(X2)
and
E(L2) = 1E(x2) = 1[var(x) + E 2 (x)],
whence the stated results.
82 UL (1971).
(i) P(X = n) = Probability that there are exactly r - 1 white balls in the
first n - 1 balls sampled
x Probability that the nth ball is white
= (nr-1
-l)pr-l(l_ p )II-r X p.
= (E.)r r
(n ~ l)(qO)1I
q lI=r r 1
=
~ (r+s-1)
(pO)' s~o s (qO)S = (pO)'(l- qo)-r.
408 EXERCISES IN PROBABILITY AND STATISTICS
(iv) E(r-1)=pr
n -1 n=r
I(n-2)ql1_r
r-2
~
= pr l.J (r+S-2) qS = pr(l_q)-r+l = p.
s=O S
83 UL (1969).
The point-probabilities of the truncated Poisson distribution are
so that
E(nr ) = NP(X = r)
and
1 <X>
E(IA- *) = - L rE(nr ) = IA-, on summation.
N r=2
A simple argument for the derivation of var(1A- *) is as follows.
Let Y 1 , Y 2 , ••• , YN be random variables denoting respectively the
number of prosecutions each motorist has. Then Yj are independent random
variables, each having the same distribution as that of X. Next, define new
random variables Zj such that Zj = Yj if Yj;;.2 and Zj = 0 otherwise. Then
N 1
NIA-* = j~ Zj and var(1A-*) = N var(Zj),
where
1 rlA- r
E(Zj) =-IL- L -, = IA-
<X>
e -1 r=2 r.
and
var(,l) =
var(1A- *)
(elLelL-1)[1_ (_1A-_)2]
elL - 1 ~ 1 as IA- ~ 00.
Note that IA- * is more easily calculated than ,1 and for large IA-, IA- * may be
preferred to ,1 as a quick estimate of IA-.
84 UL (1966).
The log likelihood of the sample observations is
11
n
2'
the integral being readily evaluated by the use of the transformation
x -IL = tan 0, -7T/2 ~ 0 ~ 7T/2.
8S UL (1977).
The log likelihood of the sample observations is
n n 1 n
log L = --log(27Tk) --log 0 - -
2 2 2kO i =1
(~- 0)2 L
whence the equation for 6 on differentiation. Since 6> 0, the positive root
of the quadratic equation gives the value of 6.
For the variance, observe that
d 2 logL n n 2 -2
d02 202- k0 3 (s +x)
n 1"
=20 2- k0 3 i~ x'f.
For the limiting variance put k = ILIO, whence
21L2 1 IL
k -+ o.
A
var(O) = - . 2 k 2 -+ - as
n IL+ n
86 UL (1973).
The least-squares estimate of IL is obtained by minimizing
whence
2 "
IL*= 1)
(
n n + v=I
L
(n-v+1)x".
Clearly,
4 "
var(1L *) = 2( 1)2 L (n - v + 1)2 var(Xv)
n n+ v=1
and
36 "
var(T) = 2(
n n+
1)2(2
n+
1f
v=l
L
(n - v + 1)4 var(Xv),
410 EXERCISES IN PROBABILITY AND STATISTICS
whence
2a 2 9a 2
var(IL*)= n(n+l); var(T) = (211 + If .
87 UL (1972).
1 /I l' /I
and
88 UL (1971).
Observe that
rv = Yv -a*- (3*(xv -x)
_ [ 1 (xv - X)2]
- Yv 1--- -
" [1
~ Yj -+
(xv - x)(Xj - X)]
.
n X j"ov n X
Hence
E(rJ = E(yJ - E(a*) - (xv - x)E«(3*) = 0
and
() a 2[1 - 1n (xv X- X)2]2 + j"ov,,[1-n + (Xv - x)(Xj
var rv =
X
~
- xf]2 2
a
= a
2[(n-l)2 n -1 (xv - X)4
- - +--+-'---'------,----'----
2(n -1)(Xv - xf
n n X2 nX
2(x -x) {
~X j~! (Xj-x)-(Xv-x)
/I }]
+
Hence
cov(O*, (3*) = E[(O* - 0)«(3* - (3)]
= E[(a* - a)«(3* - (3)]- xE«(3* - (3)2
XU 2
X·
When the Xi are the 11 natural numbers, then
n
whence
* * - [ 3(n + 1)
corr (e ,(3 ) - - 2(2n + 1) .
J!
90 UL (1967).
E(T- /.Lf = C 2 x E[Jl (Yi - /.L)2+ 2(I1C; 1)/.L Jl (Yi -IL)+ (nC~!)21L 2J
= nC 2u 2+(nC-1)21L2.
Hence the stationary value is C = l/(n + v 2 ).
91 UL (1965).
Clearly, E(x) = E(y) = IL;
u2 u2
var(x) =-;. var(y) = - [1 + (n -l)p].
n n
Now T is an unbiased estimate of IL if a + (3 = 1. Also
u2
var(T) = - [(1- (3)2+ (32{1 +(n -l)p}].
n
Hence for minimum var(T)
1+(n-1)p 1
u= . (3 = -,------
2+(n-1)p' 2+(n -l)p .
Since the Yi are equi-correlated, -[l/(n -1)] os:; p os:; 1, so that
(n-1fp2
-,,-'----,--'--'--,- ~ 0
4{1+(n-1)p} .
Hence for Pt=O. varG(x+y)]>minvar(T).
Note that x and y have unequal variances and so the best estimate of IL
is T.
412 EXERCISES IN PROBABILITY AND STATISTICS
92 UL (1968).
If x is the unknown (n + 1)th independent observation of X, then x - i .
normally distributed with E(x - x) = 0 and var(x - x) = (n + 1)0-2In. Henc~s
irrespective of the magnitude of fJ. '
x-x
2+6 2-6
1-6 1+36
Also, var(6) is readily obtained from d 2 10g L/d6 2.
With misclassification, the log-likelihood is
log L = constant + nllog(2+ 6) + (n2 + n3) log(3 -26) + n410g(1 + 36)
whence the equation for 6* is
nl 2(n2+ n 3) + 3n4 =0
2+6* 3-26* 1+36* .
For large samples var( 6*) > var( 6) if
4(3 - 26)(5 + 26 - 4( 2) > (2 - 36+ ( 2)(29 + 36)
which reduces to (6+2)(36+1»0.
98 UL (1969).
The log-likelihood of the observed sample is
log L = constant+ al log(2- 6) + a210g(1 + 6) + a310g 6 + a410g(1- 8)
whence, on differentiation, the equation for 6 is
_~+~+ a3_~_0
2-6 1+6 8 1-8- .
1 + 26(1- 6)
E(X) = N(1-26) and var(6*) = 2N '
99 UL (1970).
Since I~=l mi = N, a constant, therefore n=l dmJd6 = O. The probability
of obtaining the observed sample is
L = kN! . Ii (mi) a,
IT .1 i=l N
a,.
i=l
and the equation for 6 is obtained by logarithmic differentiation.
414 EXERCISES IN PROBABILITY AND STATISTICS
=
n
~ 1 [.f.
1=1
E(x7) - nE(x 2)]
=
n 1=1
[.f.
~ 1 var(Xj) - n var(x)], since E(Xi) = E(x)
whence the result.
(iii) This follows by noting that var(x);;:. 0 for all nand E(S2);;:. 0 for all
n ;;:. 2. Note that
var(xw ) = 2( 4 )2
n n+1 i=1
[f.
i 2 var(Xj) + 2 L ij cov(Xj, ~)]
i<i
40"2 [ 11 11-1 ]
= 2(
n n+1
)2 L f+2p i=1
i=1
L i(i+1) ,
( Y - ( 2) + 62 ]
62- 6 1 • E [ (X - ( 1)+ 61
=
X - 6 1 Y - 62 (X - ( 1)(Y - ( 2 )]
- 62-62 . E [1 --6-1-+~- 6 162 '
whence the result on taking expectations.
102 UL (1968).
Assume that for two random variables X and Y,
var(X) = O"f; var(Y) = O"~; cov(X, Y) = pO"10"2.
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT 415
Then var(X ± Y) = (0"1-0"2f+2(1 ±p)0"10"2;;:'0 for all 0"1 and 0"2' Hence for
0',==0"2,1±p;;:'0.
Expand and take expectation over the z's whence the answer.
103 UL (1969).
Here S" = Ill=2 Zi where
P(Z, = 1) =i; P(Z, =0) =~ for 2~r~ n.
Therefore
E(Zr) = t var(Z,) = /i;;
cov(Z" Z,+l) = --16; cov(Z;, Zj) = 0 for Ii - il ;;:,2.
Hence E(S,,) and var(S,,).
104 UL (1970).
(i) If E(X) = 81 and E( Y) = 82, then
cov(X + Y, X - Y) =E[{(X - 8,)+(Y -( 2 )}{(X -( 1)-(Y -82 m
= E[(X - 8 1f- (Y - ( 2)2]
= var(X) - yare Y), whence the result.
(ii) By definition,
lOS UL (1971).
(i) Let E(X;) = 8i, i = 1, 2, 3. Then, by definition,
cov(X1+ X 2 , X 2 + X 3 ) = E[{(X1- 81) + (X2 - 82 )}{(X2 - 82 ) + (X3 - 83 )}]
= E[(X2 - 82)2], since the X; are uncorrelated
=var(X2 ).
(ii) Here E(Si) = 0, i = 1, 2, 3. Hence
r "
cov(Sl> S2) = L L E(x.Xt)
.=11=r+1
r "
= L L cov(x., Xt)
.=11=r+1
=u 2 L L
r "
pl-.,
.=11=r+1
whence the result on summation.
Cov(Sl> S3) is obtained in a similar manner.
106 UL (1973).
Since E(Yi) = 0, we have
1 i+,,-1 i+r+n-1
cov(Yi> Yi+r) =2 L L E(x.Xt).
n s=i t=i+r
107 UL (1966).
For the Poisson distribution P(X = x) = e-"',A,x/X! whence
QO
Hence
E(e 'Z ) = e-',/i;; . E(e'x Jn7j;.)
= e-',/i;; . exp[nlL(e'/,/i;; -1)].
Hence
G (81) ( 2) = E[(0 1 (2)X,] . E[ 0~2]
= (ql + PI 81( 2)"'(q2 + P2 ( 2)"2.
The probability-generating function of Y is G(l, ( 2 ) whence P(Y = r), and
the probability-generating function of Xl given Y = r is
G(O) = L Pror,
r=O
r=O
= f t r(s) a;s.
r=O
Pr
s=O
= f a; f Prr(s)
s.
s=O r=s
418 EXERCISES IN PROBABILITY AND STATISTICS
i=1
and
E(e 'Z ) = e-P1/<To nk
i=1
(1 + P el/n,kuo)".
-~ [1 + V~~i)J.
112 UL (1966).
The sample of n unrestricted values can be selected in (~) ways. Next, if
x and yare the largest and smallest integers in the sample, then the choice
of the remaining n - 2 is restricted to x - y -1 numbers in all. Hence
P(X=x)= L
x-I ( z- 1)j(~
z=,.-1 n-2 n
where
Xf
z=,.-1
(Z-1)=
n -2
xf'
,,=0
(U+n-2)
U
x-n
= L coefficient of 0" in (1 + 0),,+,.-2
,,=0
=
x-,.
L coefficient of 00 {1 O}"
in (1 + 0),,-2 -+-
,,=0 0
= coefficient of 00
{1 + O}"
in (1 + 0),.-2 L - -
X-,.
,,=0 0
= coefficient of 00 in (1 + O)X-l • O-(x-,.l
=\x-n
(X-1)= (X-1).
n-1
Again, for fixed Z = z, we have z + 1 :s;;; X :s;;; N, whence the marginal distribu-
tion of Z is
=(N-Z)(:=~)j(~,
with the stated limits for Z.
113 UL (1966).
The probability that a prosecuted motorist has x prosecutions is
e-.... '.c.x/(1-e-.... ). x!, 1:s;;;x<oo
E[x(rl ] = e
1-e .... x=1 x!
f.Lr
- --
1-e-.....
Hence
E(x) = _1£_ .
1-e-.... '
Therefore
var(x) -1 1£ e-....
E(x) - -1-e-.... '
114 UL (1966).
Consider first the joint distribution of Xl> x,,+l and X Zv + 1 and then the
conditional probability of the 2r observations. It then follows that
Xv + 1
J J J[ Jf(x)dx ]v+r-t
00 00 00
= (2v+1)! . (2r)
p {(v-1)!F r
-00 Xl X v +l Xl
X,,+l
where O<u<v<w<l.
Integration over w is straightforward. Next transform from v to t by the
substitution 1- v = (1- u)t. The resulting double integral in u and t is
readily evaluated in terms of the beta function and the stated answer is
obtained by a suitable recombining of the gamma functions.
115 UL (1966).
By definition,
«I>(x) = k X
f e-!u 2
du
1
=-+-- L.e "u du
1 f _1 2
2 J2; 0
1
=-+-1 f L'" (-!u
x
--du
z)'
2 J2; r=O r!
o
whence the result on integration after changing the order of summation and
integration
E[etIX1 ]=_1_
J2;
f '"
etlxl-!x2 dx
= (2)! f'"
1T
t
eX-"X
1 2
dx
o
ff
I 00
{3
f{a+1)
f
(1- X)~-l e - Y(I-X){y(1- x)}"'(1- x) dy = (3(1- X)~-l
,
o for O::s;;;X::s;;;1.
f(:+
1
E(Y) = 24
a4 f'"x 2(a-x)(a 2-x 2)idx.
o
The integral is evaluated simply by using the substitution x = a sin O.
'"
P(Y::S;;;a) =~ f[2a(4a2_y2)i-(4a2_y2)]y dy.
4a
o
Evaluate the integral by using the substitution y = 2a sin O.
422 EXERCISES IN PROBABILITY AND STATISTICS
119 UL (1967).
co
[J J
00 00
= ~ e-(1+ila)" du + e-(1-ila)" du ]
o 0
1
1 + t 20'2'
E(e i1z ) = E[eil(1-.... )y']E[eit....y2]
1 1
1 + t 2(1_1J.)2 . 1 + t 21J. 2
1 [(1-1J.)2 IJ. 2 ]
=1- 2 1J. 1+t 2(1-1J.)2 1+t21J.2·
The sampling distribution of z now follows directly by the use of the
Inversion Theorem.
120 UL (1967).
(i) From the definition we have
cOV(Xj_k, Xj+2) = a cov(Xj+l> Xj-k) + (3 cov(Xj, Xj-k)
so that Pk+2 = apk+I + (3Pk'
(ij) Similarly,
var(Xj+2) = a 2 var(Xj+l) + (32 var(Xj) + 2a(3 cov(Xj, Xj+I) +var(ej+2)
so that
var(e) = 0'2(1- a 2- (32)-2a(30'2pl>
where from (i) for k = -1, PI = apo + (3P-l> Po = 1, PI = P-I' Hence
PI = a/(1- (3) and var(e) follows.
(iii) The general solution of the difference equation in (i) is
Pk = AA~ + BAt with Po= 1, PI = a/(1-a)
Use of the initial conditions now gives
A = -A2+ a /(1-(3).
Al - A2 '
whence Pk'
121 UL (1967).
Note that
1 1 z
v=-tan- -
21T W
It then follows that
a(u, v)
--=
1
--exp[I( 2 2)]
-2" W +z
a(w, z) 21T
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT 423
whence Wand Z are independent N(O, 1) variables. Therefore W+Z2=
- 2 loge U is distributed as x2 with 2 d.f.
If Ul> U2, ... ,Uk+2 are independent realizations of U, then
n=1-2Ioge Uj is distributed as x 2 with 2k d.f.
Also (-2 loge Uk+l)~cos(21TUk+2) is an independent N(O,1) variable.
Hence w2+ L~= I - 2 loge u; is distributed as X2 with 2k + 1 d.f.
122 UL (1967).
The probability density function of the distribution of XI and x,. is
4n(n-l) 2 211-2 ,,::: ""'"- >-
211 (X I1 -Xl) XIX,,, O-"Xl-.,a;x,,"'-XI·
a
Hence the probability density function of the distribution of U and v is
n(n-1)
----'-.........,,--'- (UV)"-2(U 2 - v 2) 0 ~ V ~ a' v ~ U ~ 2a - v
2a 2n " •
P(X~Y)= J Jf(x,Y)dXdY=~
al+~l
o 0
124 UL (1968).
Consider first the joint distribution of Xl> Xv+l and X2v+l and then .t~e
conditional probability of the 4r observations. Then the required probabIlity
424 EXERCISES IN PROBABILITY AND STATISTICS
is
=Px.y.u[y~:O;:;~J.
But since (U - X)/(l + ( 2 )! is N(O, 1) and independently distributed of Y, it
follows that Z and (U-X)/(1+u 2 )! have the same distribution.
126 UL (1968).
The cumulant-generating function of X is
00
n
E(e'X) = " E[e'x/,,]
i=1
L r(Ur-1 - ur ) =L
,=0
Ut·
r=1
(ii) Here Y,,-1 - y" = probability that the pattern is realized for the first
time at the nth trial
= y,,_ 4 X qp 3
whence the difference equation for y", n;;' 4, with the initial condi-
tions YO=Yl=Y2=Y3=1.
The generating function for this difference equation is
L Yr = ep(1) = l/qp 3.
r=O
129 UL (1968).
Here Xl> X2, sf and s~ are all independently distributed and F=sVs~ has
the F distribution with n -1, n -1 d.f. Now
Xl + FX 2
w= 1+F and E(wIF)=p.,
where
{ F} 1 J
00
p"-1)/2 dF
EF (1+F)2 = (n-1 n-1) (1+F)"+1
B -2-'-2- 0
n-1
4n .
130 UL (1968).
The distribution of W is obtained directly by using the transformation
suggested. It then follows that w = t/.Jii, where t has Student's distribution
with n d.f. Also,
whence
np2-(4t2+2n)F+n =0.
The positive root of this quadratic gives the required answer since for F> 1,
t>O.
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT 427
131 UL (1968).
f
00
00
= fe -x cos tx dx
o
= 1- t 2 E(e itX ).
(ii) U j and U 2 are independent N(O,!) variables.
(iii) We can write
W = a[(Zl + Z2)2 - (Zl - Z2f + (Z3 + z4f - (Z3 - Z4f]
where !(Zl ± Z2) and !(Z3 ± Z4) are independent N(0,1) variables.
Hence W = !(xi - x~) where xi and x~ are independent X2 variables
each with 2 d.f.
Finally,
132 UL (1968).
P(X = r) = Probability that the first r cards include any two aces and r - 2
non-aces
x Probability that the (r + 1)th card is an ace
(~)(,~82) 2
(5 2) . 52-r' for 2,,;;; X,,;;; 50.
r
P(¥ = s ! X = r) = Probability that s non-aces turn up after the third ace
x Probability that the (r + s + 2)th card is the fourth ace
Hence
P(X=r, ¥=s)=P(X=r)P(¥=s!X=r).
12r(r-1)
52(4)
428 EXERCISES IN PROBABILITY AND STATISTICS
Therefore
50-. 12r(r-l)
P(¥=S)= r~ 52(4)
--'-------;-:-;--- • f or 0 :0;;;; ¥
4(51-s)(3)
:0;;;;
48 •
52(4)
133 UL (1968).
I 00
I 00
whence var(Z).
134 UL (1968).
Here
var(i) = \
n
[ti=1 O'~ + i""jL PO'iO'j]
=2 1 [ (l-p) L
n
" O'~+P ("
i=1
L O'i
i=1
)2] .
The inequality now follows since var(i);;;: O.
Set ~ - IL = Zi and let z be the mean of the Zi' Then
cov(i, Yi) = E[Z(Zi - z)]
1"
=E [ Zi'- L Zj ] -E(Z2)
n i=1
= AO'~ + BO'i - C,
where
I-p P n
A=->O, B=- L O'j' C=var(i»O.
n n j=1
Hence if 0'1 = 0'2 = ... = 0'", then cov(i, Yi) = O. Conversely, if
cov(i, Yi)=O, then AO'~+BO'i-C=O. Since A>O and C>O, there is one
positive root O'i = 0' say. But A, B and C are independent of i, so that the
conditions cov(i, Yi) = 0 give the equations
AO'~+ BO'i - C = 0 for i = 1, 2, ... , n,
which have the solution 0'1 = 0'2 = ... = 0'".
135 UL (1969).
Let G(Oh O2 , ( 3 ) denote the joint probability-generating function of Xl>
X 2 and X 3 •
(i) P(X2 = 1) = coefficient of O2 in G(I, O2 , 1).
J
00
1 J
00
u 2/4cr2 2u
=u-/; e- u du = -/;.
o
If Xl> X 2 , X3 are three independent observations of X, then
IX2 - XII + IX3- X 2 1+ IX3- XII = 2R,
where for i l' j E(IXi - Xj D= 2u/.Jrr.
137 UL (1969).
Let X, Y and Z be the events that A, Band C respectively have
success in a game. Then for A to win, the eight possible outcomes of the fin
game may be classified as follows.
E I : XYZ, XYZ-game reverts to initial state
P(E I ) = p3+ q3= 1-3pq
E 2 : XYZ, XYZ, XYZ-A loses
P(E2) = pq(1 +q)
E3: XYZ-A wins; P(E3) = pq2
E 4 : XYZ-game undecided between A and C; P(E4 ) = p2q
Es: XYZ-game undecided between A and B; P(Es) = p2q.
Next, if S.. denotes the event that A wins the series on the nth game, the
for n~2
5
P(S..) = L P(S.. I Ei )P(E;)
i=1
where, by definition,
P(S.. I E 1) = Un-I; P(S.. I E 2 ) = P(S.. I E 3 ) = 0;
P(S.. I E 4 ) = P(S.. I Es) = (1- 2pq) .. -2pq.
Hence the stated difference equation for u.,. It then follows that
00 00 00
or
1 2 p 3q 2 z 2
-</I(z)-ul=(1-3pq)</I(z)+1 (1 2 ) , Ul=pq2,
Z - -pqz
whence </I(z).
138 UL (1969).
When A has £n another trial of E must be made. If this trial results in
the occurrence of E, then the expected number of further trials needed for
A's win is u,.+l, whereas if this trial results in the non-occurrence of E, then
the expected number of further trials needed for A's win is Un-l. Hence
Un = 1 + Pu,.+l + qU,.-l, where Uo = 0 and Ua+b = O.
The difference equation is reduced to the homogeneous form
Vn = pVn+l + qVn-l
by the substitution u,. = Vn +an where a = 1!(q - p). Hence
Un =_n_+ kl(q!p)n + k2
q-p
where, by using the boundary conditions the constants
(a+b)pa+b
k l =-k2=-( q-p )( q a+b -p a+b)' (P=l=q)·
dF(x ) dF(x )
x -r- - -
s
dXr dxs
For the uniform distribution F(x) = x/a, whence the distribution of u and v.
For fixed u, u:s;; v:s;; 1, so that the probability density function of the
marginal distribution of u is
1
N!
------:------:----:--,-
J r-l(
U V - U
u r- 1(1- u)N-r
)s-r-l(1 - v ) d v = --:-----'-----...:...--
(r-l)! (s - r-l)! (N -s)! B(r, N - r+ 1)
"
by using the substitution v - u = z(l- u). The conditional distribution of v
given u has the probability density function
o 0
n2 . .
- - - on mtegratIOn.
nl +n2
Hence for X ~ Y, the probability density function of the conditional joint
distribution of X and Y is
nl(nl + n2) exp- (nix + n2Y) = nl(nl + n2) exp-[nl(x - y)+ (nl + n2)Y],
X~ Y; O:s;; Y <00.
lf we now set u = x - Y and v = y, then it readily follows that U and V are
independently distributed. Hence, for U ~ 0, the conditional distribution of
U has the density function
00
J
nl(nl + n2) e- n," e-(n,+n2)u dv = nl e-n,u.
o
Again, by symmetry the conditional distribution of Y - X given Y - X ~ 0
has the density function n2 e- n2".
Also, P(X ~ Y) = n2/(nl + n2) and P(X:s;; y) = nl/(nl + n2)' Hence the
unconditional distribution of IUI has the density function
432 EXERCISES IN PROBABILITY AND STATISTICS
142 UL (1969).
Let Xi be a random variable such that
Xi = 1 if the ith quadrat is empty
= 0 if the ith quadrat is not empty.
Then
and
11
X= L X;.
i=1
It is now easily shown from first principles that
143 UL (1969).
Observe that whether X +1 >0 or I1 X,,+1 <0,
P(X~x..+l)P(X.;;;x..+l)=4
X~+I) .
1 ( 1-7
Therefore
= (2n + 1)! a
(n !f2211 + 1
r J
1
i~l var(x;)+ i~
1 ["
(i) var(x) = n2 II
j~i cov(X;, Xj) ]
1 ["-1
= n2 nu 2+2 i~ cov(X;, X;+1)+2 i~ cov(X;, Xi+2) ,
n-2 ]
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT
433
whence the result on reduction.
145 UL (1969).
The Jacobian of the x ~ u transformation is n! Also, observe that
n n
UI = n(xl-m and L Uj = L (Xi -Xl)'
i=2 i=2
It then follows that the Uj are independent and identically distributed
random variables such that for any j"2u/a has the X2 distribution with 2 d.f.
The ratio n(n -l)(XI - ml'f.r=2 (Xj - Xl) has the F distribution with 2,
2(n-1) d.f.
146 UL (1966).
Clearly, both 20u and 2v/0 are independently distributed as X2'S each
with 2n d.f., whence the joint distribution of u and v. Next, the Jacobian of
the (u, v) ~ (w, z) transformation is 2w/z. Hence the probability density
function of the joint distribution is
2
{f(nWexp[-(Ow/z+wz/0)].w2n-l.z-t, O:os;w<oo; O:os;z<oo.
,
1! 1! ~~-2)!
[fxf(x) dx ]n-2f(XI)f(Xn), -00< Xl <00; XI:OS;Xn <00.
Also, given Xl and x'" the probability that the r observations of the second
sample all lie outside the range (xt> Xn) is
Hence the integral expression for P. To evaluate the double integral use the
434 EXERCISES IN PROBABILITY AND STATISTICS
transformation
u=
",
J f(x) dx, v= r
f(x) dx.
Then
1 1
JJ
P = n(n -1) du (v - u)"-2[1- (v - u)]r dv.
o lC
149 UL (1966).
Unconditionally, 0 ~ X ~ a/2 and 0 ~ Y ~ a. For fixed X, X ~ Y ~ a - X.
Hence on integrating the joint distribution with respect to Y, the probability
density function of the marginal distribution of X is
n(n -1)a-"2"-lx"-2(a -2x), O~X ~a/2.
(a - xy+l- x r + 1
E(yrIX=x)= (a-2x)(r+1) , r;::O
whence E(Y I X = x) and var(Y I X = x).
150 UL (1966).
<?learly, <I>~x)=t+"'.(x) a~d <I>(x)[1-<I>(x)]=!-",2(x). On expansion of
the mtegrand m the detinmg mtegral for "'(x) and term-wise integration .
obtain ' we
and
Hence
2X2 2X4 7x 6
4<1>(x)[1-<I>(x)]-1-- + - - - + ...
7T 37T 457T
-2X2hr[1 + 2( 7T - 3)
-e 37T 2 x 4 -'" ]
.
151 UL (1967).
Here
P(Y ::o.;;ta) = 1- P(Y~ta)
=t(1-log 2)
=P(a - X - y::o.;;ta), by symmetry.
After the first experiment there are two possibilities:
(i) probability p that Xo does not split;
(ij) probability 1- p that Xo splits into two particles XI and X 2 with
energies EI and Eo-EI respectively.
Now P(E I ::0.;; tEo) =t. Also, if EI ::0.;; tEo, then any further splitting or not
of XI must lead to particles or particle with energy 1!Eo. Therefore the
probability of Xl giving particles with energy ::o.;;tEo after the second experi-
ment is t(1- p).
Next, the probability that X 2 has energy ~tEo is ! and the probability
that in the second experiment X 2 splits into two further particles X3 and X 4
is 1- p. If the energies of these particles are E2 and Eo - EI - E 2 , then it is
required that E 2 ::o.;;tEo and E o -E I -E2 ::o.;;tEo simultaneously.
436 EXERCISES IN PROBABILITY AND STATISTICS
E(eIX)=~ J exp[tx-\x-O\]dx
00
=~ J exp[t(u+O)-\u\]du
00 00
Mx(t)=E(eIX)=e'8 1- n 2 ( t2)-" ,
whence
t 2 )-" .
-_ ( 1- 2n
E(z) = 0; ( ) _1+3a+1513.
varz- 1 +a+ 313 '
( ) _ 3+15a + 10513
fL4 z - 1+a+313 .
For the exact distribution of z,
E(z) = 0; var(z) = 1; Kiz) =3/n.
Hence
6
a =-613=---.
8n+3
153 UL (1967).
If y denotes the number of points obtained uppermost on anyone die,
then
6 6
E(y) = L i.s=!; E(y2) = L i. S2 =9j;
0=1
var(y) =~~.
0=1
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT 437
Hence
7nl 7
E(XI)=T; E(X2) = E(X3) = 2" n2.
35nl 35n2
var ( Xl) =1:2 ; var(x2) = var(x3) = 1:2 .
It then follows that
var(XI + X2) = var(xi + X3) = H(nl + n2);
COV(XI + X2, Xl + X3) = E[(XI + X2)(X2 + X3)]- E(XI + x2)E(XI + X3)
= E(x~) + E(XI)E(X2) + E(XI)E(X3) + E(X2)E(X3) - ?(nl + n2)2
=~~nl'
Hence
corr(XI + X2, Xl + X3) = ~ .
nl+n2
!J
~ ~
155 UL (1967).
The distribution of x,. has the probability density function
n I (X2)"-l 2x
(11-1)! I! 0; o~'
whence for any r> 0
E(xr) =2n
11 0211
'"
J
- X211 + r- l dx
11 11
= - (
2n-) or.
2n + r
o
It then follows that
n0 2
var(xl1 ) = (11 + 1)(211 + 1)2·
Hence & and var(&) as stated.
Note that the limits for z are
=(1 - -1 -)211[ 1+ Z 2
• 11-1 ( 1+- J
1 )-~ --i>e'-1 as n --i> 00,
2n + 1 2{n(n + I)}> n
for -00< z < 1.
156 UL (1967).
Probability of a withdrawal only in the week before Christmas is PI - P12;
Probability of a withdrawal only in the week after Christmas is P2 - P12;
Probability of no withdrawals in these two weeks is 1- PI - P2 + P12. Hence
the probability-generating function of the withdrawals of a single client in
the two weeks is
(1- PI - P2+ pd+(Pl- P12)81 +(P2- pd 02+ PIP2 0 1 02
= 1 + Pl(OI -1) + P2(02-1) + PliOI -1)(02 -1),
whence E(O~'. 0~2)= G(Ol, O2 ).
ANSWERS AND HINTS ON SOLUTIONS: SUPPLEMENT
439
Under the stated conditions the limiting form of G(O 1> 0)'
2 IS
exp[al(OI -1) + a2(02 -1)+ aliOI -1)(02 -1)].
The moment-generating function about the origin of Xl and X is G(' ,
whence the cumulants on logarithmic expansion. For corr(X 2X ) == eO" eh,)
.
moment-generatIng f unction
. 0
fX I and X 2'IS exp[a l (e"-l)1>+a2(e"-1)]
2 ,tc
whence independence of Xl and X 2 • '
157 UL (1969).
(i) The probability-density function of the marginal distribution of Y is
=
(3-a-2 cxI
r( a + 1) . y
1
-(l-+-x-)-=-/3 exp - -l-+-x dx.
(y )
o
To integrate over x for fixed y use the substitution w = y/(l + x),
whence the stated result.
(ii) This follows readily by using P(X I Y = y) = P(X = x, Y = y)/
p(Y=y).
•.•
(111) E(XI Y=y)= g(y)
y/3-1 I
=
x e- y /(1+x)
(l+x)/3 dx.
o
P(X ~ Xl) = ~ I
=
e-(x-/3)/cx dx = e-(x,-/3)/CX.
O=Itan [-2<TI-<T2
2
P<T <T2]
-I
2
1
2·
For this value of 0, Y1 and Y2 are independent normal variables with zero
means and variances <T~1 and <T~2. '
Since R 2 = Xt + X~ = Yt + Y~, the moment-generating function of R 2 is,
readily obtained from the joint distribution of Y1 and Y2 • Logarithmic
expansion leads to the cumulants of R2.