Scribd
Upload
100 views608 pages

Probability and Random Processes

Uploaded by

Edith
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
100 views608 pages

Probability and Random Processes

Uploaded by

Edith
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
Probability and Random Processes GEOFFREY GRIMMETT and DAVID STIRZAKER Third Edition Probability and Random Processes GEOFFREY R. GRIMMETT Statistical Laboratory, University of Cambridge and DAVID R. STIRZAKER Mathematical Institute, University of Oxford OXFORD UNIVERSITY PRESS OXFORD UNIVERSITY PRESS Great Clarendon Street, Oxford 0x2 6DP Oxford University Press is a department of the University of Oxford. It furthers the University’s objective of excellence in research, scholarship, and education by publishing worldwide in Oxford New York Athens Auckland Bangkok Bogoté Buenos Aires Cape Town Chennai Dar es Salaam Delhi Florence Hong Kong Istanbul Karachi Kolkata Kuala Lumpur Madrid Melbourne Mexico City Mumbai Nairobi Paris Sao Paulo Shanghai Singapore Taipei Tokyo Toronto Warsaw with associated companies in Berlin Ibadan Oxford is a registered trade mark of Oxford University Press in the UK and in certain other countries Published in the United States by Oxford University Press Inc., New York © Geoffrey R. Grimmett and David R. Stirzaker 1982, 1992, 2001 ‘The moral rights of the author have been asserted Database right Oxford University Press (maker) First edition 1982 Second edition 1992 ‘Third edition 2001 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without the prior permission in writing of Oxford University Press, or as expressly permitted by law, or under terms agreed with the appropriate reprographics rights organization. Enquiries concerning reproduction outside the scope of the above should be sent to the Rights Department, Oxford University Press, at the address above You must not circulate this book in any other binding or cover and you must impose this same condition on any acquirer A catalogue record for this title is available from the British Library Library of Congress Cataloging in Publication Data Data available ISBN 0 19 857223 9 [hardback] ISBN 0 19 857222 0 [paperback] 10987654321 ‘Typeset by the authors Printed in Great Britain on acid-free paper by Biddles Ltd, Guildford & King's Lynn Lastly, numbers are applicable even to such things as seem to be governed by no rule, I ‘mean such as depend on chance: the quantity of probability and proportion of it in any two proposed cases being subject to calculation as much as anything else. Upon this depend the principles of game. We find sharpers know enough of this to cheat some ‘men that would take it very ill to be thought bubbles; and one gamester exceeds another, as he has a greater sagacity and readiness in calculating his probability to win or lose in any particular case. To understand the theory of chance thoroughly, requires a great knowledge of numbers, and a pretty competent one of Algebra. John Arbuthnot An essay on the usefulness of mathematical learning 25 November 1700 To this may be added, that some of the problems about chance having a great appearance of simplicity, the mind is easily drawn into a belief, that their solution may be attained by the mere strength of natural good sense; which generally proving otherwise, and the mistakes occasioned thereby being not infrequent, it is presumed that a book of this kind, which teaches to distinguish truth from what seems so nearly to resemble it, will be looked on as a help to good reasoning. Abraham de Moivre The Doctrine of Chances 1717 Preface to the Third Edition This book provides an extensive introduction to probability and random processes. It is intended for those working in the many and varied applications of the subject as well as for those studying more theoretical aspects. We hope it will be found suitable for mathematics undergraduates at all levels, as well as for graduate students and others with interests in these fields, In particular, we aim: # to give a rigorous introduction to probability theory while li theory in the early chapters; 1g the amount of measure @ to discuss the most important random processes in some depth, with many examples; @ to include various topics which are suitable for undergraduate courses, but are not routinely taught; # to impart to the beginner the flavour of more advanced work, thereby whetting the appetite for more. ‘The ordering and numbering of material in this third edition has for the most part been preserved from the second. However, a good many minor alterations and additions have been made in the pursuit of clearer exposition. Furthermore, we have included new sec sampling and Markov cl in Monte Carlo, coupling and its applications, geometrical prob- ability, spatial Poisson processes, stochastic calculus and the It6 integral, 1t6’s formula and applications, including the Black-Scholes formula, networks of queues, and renewal-reward theorems and applications. In a mild manifestation of millennial mania, the number of exer- cises and problems has been increased to exceed 1000. These are not merely drill exercises, but complement and illustrate the text, or are entertaining, or (usually, we hope) both. In a companion volume One Thousand Exercises in Probability (Oxford University Press, 2001), we give worked solutions to almost all exercises and problems The basic layout of the book remains unchanged. Chapters 1-5 begin with the foundations of probability theory, move through the elementary properties of random variables, and finish with the weak law of large numbers and the central limit theorem; on route, the reader meets random walks, branching processes, and characteristic functions. This material is suitable for about two lecture courses at a moderately elementary level. The rest of the book is largely concerned with random processes. Chapter 6 deals with Markov chains, treating discrete- time chains in some detail (and including an easy proof of the ergodic theorem for chains with countably infinite state spaces) and treating continuous-time chains largely by example Chapter 7 contains a general discussion of convergence, together with simple but rigorous viii Preface to the Third Edition accounts of the strong law of large numbers, and martingale convergence. Each of these two chapters could be used as a basis for a lecture course. Chapters 8-13 are more fragmented and provide suitable material for about five shorter lecture courses on: stationary processes and ergodic theory; renewal processes; queues; martingales; diffusions and stochastic integration with applications to finance. We thank those who have read and commented upon sections of this and earlier editions, and we make special mention of Dominic Welsh, Brian Davies, Tim Brown, Sean Collins, Stephen Suen, Geoff Eagleson, Harry Reuter, David Green, and Bernard Silverman for their contributions to the first edition. Of great value in the preparation of the second and third editions were the detailed criticisms of Michel Dekking, Frank den Hollander, Torgny Lindvall, and the suggestions of Alan Bain, Erwin Bolthausen, Peter Clifford, Frank Kelly, Doug Kennedy, Colin McDiarmid, and Volker Priebe. Richard Buxton has helped us with classical matters, and Andy Burbanks with the design of the front cover, which depicts a favourite confluence of the authors. This edition having been reset in its entirety, we would welcome help in thinning the errors. should any remain after the excellent TEX-ing of Sarah Shea-Simonds and Julia Blackwell, Cambridge and Oxford GRG April 2001 s. 1 Events and their probabil 1d 12 13 14 15 1.6 17 18 Contents Introduction 1 Events as sets 1 Probability 4 Conditional probability 8 Independence 13 Completeness and product spaces 14 Worked examples 16 Problems 21 2 Random variables and their distributions 21 22 23 24 25 26 27 Random variables 26 The law of averages 30 Discrete and continuous variables 33 Worked examples 35 Random vectors 38 Monte Carlo simulation 41 Problems 43 3 Discrete random variables 31 32 33 3.4 35 3.6 37 38 39 Probability mass functions 46 Independence 48 Expectation 50 Indicators and matching 56 Examples of discrete variables 60 Dependence 62 Conditional distributions and conditional expectation 67 Sums of random variables 70 Simple random walk 71 3.10 Random walk: counting sample paths 75 3.1 Problems 83 x Contents 4 Continuous random variables 41 42 43 44 45 4.6 47 48 49 4.10 al 4.12 4.13 414 Probability density functions 89 Independence 91 Expectation 93 Examples of continuous variables 95 Dependence 98 Conditional distributions and conditional expectation 104 Functions of random variables 107 Sums of random variables 113 Multivariate normal distribution 115 Distributions arising from the normal distribution 119 ‘Sampling from a distribution 122 Coupling and Poisson approximation 127 Geometrical probability 133 Problems 140 5 Generating functions and their applications 5.1 5.2 53 54 55 5.6 5.7 58 59 5.10 5.1 5.12 Generating functions 148 Some applications 156 Random walk 162 Branching processes 171 Age-dependent branching processes. 175 Expectation revisited 178 Characteristic functions 181 Examples of characteristic functions 186 Inversion and continuity theorems 189 ‘Two limit theorems 193 Large deviations 201 Problems 206 6 Markov chains 6.1 62 63 64 65 6.6 67 68 6.9 6.10 OAL 6.12 6.13 6.14 6.15 Markov processes 213 Classification of states 220 Classification of chains 223 Stationary distributions and the limit theorem 227 Reversibility 237 Chains with finitely many states 240 Branching processes revisited 243 Birth processes and the Poisson process 246 Continuous-time Markov chains 256 Uniform semigroups 266 Birthdeath processes and imbedding 268 Special processes 274 Spatial Poisson processes 281 Markov chain Monte Carlo 291 Problems 296 Contents xi 7 Convergence of random variables 10 i 71 72 73 14 75 16 17 78 19 7.10 Tl Introduction 305 Modes of convergence 308 Some ancillary results 318 Laws of large numbers 325 The strong law 329 ‘The law of the iterated logarithm 332 Martingales 333 Martingale convergence theorem 338 Prediction and conditional expectation 343 Uniform integrability 350 Problems 354 Random processes 8.1 8.2 83 84 8.5 8.6 8.7 Introduction 360 Stationary processes 361 Renewal processes 365 Queues 367 The Wiener process 370 Existence of processes 371 Problems 373 Stationary processes ol 9.2 93 94 95 9.6 97 Introduction 375 Linear prediction 377 Autocovariances and spectra 380 Stochastic integration and the spectral representation 387 The ergodic theorem 393 Gaussian processes 405 Problems 409 Renewals 10.1 10.2 10.3 10.4 10.5 10.6 ‘The renewal equation 412 Limit theorems 417 Excess life 421 Applications 423 Renewal-reward processes 431 Problems 437 Queues 1d 11.2 11.3 114 115 Single-server queues 440 MIM/1 442 MIG/I 445 GIM/1 451 GIGN 455 xii Contents 11.6 Heavy traffic 462 11.7 Networks of queues 462 11.8 Problems 468 12 Martingales 12.1 Introduction 471 12.2 Martingale differences and Hoeffding’s inequality 476 12.3 Crossings and convergence 481 12.4 Stopping times 487 12.5 Optional stopping 491 12.6 The maximal inequality 496 12.7 Backward martingales and continuous-time martingales 499 12.8 Some examples 503 12.9 Problems 508 13 Diffusion processes 13.1 Introduction 513 13.2 Brownian motion 514 13.3 Diffusion processes 516 13.4 First passage times 525 13.5. Barriers 530 13.6 Excursions and the Brownian bridge 534 13.7. Stochastic calculus 537 13.8 The It6 integral 539 13.9. It6's formula 544 13.10 Option pricing 547 13.11 Passage probabilities and potentials 554 13.12 Problems 561 Appendix I. Foundations and notation 564 Appendix IL. Further reading 569 Appendix III. History and varieties of probability 571 Appendix IV. John Arbuthnot’s Preface to Of the laws of chance (1692) 573 Appendix V. Table of distributions 576 Appendix VI. Chronology 578 Bibliography 580 Notation 583 Index 585 1 Events and their probabilities Summary. Any experiment involving randomness can be modelled as a prob- ability space. Such a space comprises a set & of possible outcomes of the experiment, a set ¥ of events, and a probability measure P. The definition and basic properties of a probability space are explored, and the concepts of condi- tional probability and independence are introduced. Many examples involving modelling and calculation are included. 1.1 Introduction Much of our life is based on the belief that the future is largely unpredictable. For example, games of chance such as dice or roulette would have few adherents if their outcomes were known in advance. We express this belief in chance behaviour by the use of words such as ‘random’ or ‘probability’, and we seek, by way of gaming and other experience, to assign quantitative as well as qualitative meanings to such usages. Our main acquaintance with statements about probability relies on a wealth of concepts, some more reasonable than others. ‘A mathematical theory of probability will incorporate those concepts of chance which are expressed and implicit in common rational understanding. Such a theory will formalize these [Link].a collection of axioms, which should lead directly to conclusions in agreement with practical experimentation. This chapter contains the essential ingredients of this construction. 1.2 Events as sets Many everyday statements take the form ‘the chance (or probability) of A is p’, where A is some event (such as ‘the sun shining tomorrow’, ‘Cambridge winning the Boat Race’, ...) and p is a number or adjective describing quantity (such as ‘one-eighth’, ‘low’, ...). The occurrence or non-occurrence of A depends upon the chain of circumstances involved. This chain is called an experiment or trial; the result of an experiment is called its outcome. In general, we cannot predict with certainty the outcome of an experiment in advance of its completion; we can only list the collection of possible outcomes. (1) Definition. ‘The set of all possible outcomes of an experiment is called the sample space and is denoted by &. 2 1.2 Events and their probabilities (2) Example. A coin is tossed. There are two possible outcomes, heads (denoted by H) and tails (denoted by T), so that @ = {H, T}. We may be interested in the possible occurrences of the following events: (a) the outcome is a head; (b) the outcome is either a head or a tail; (©) the outcome is both a head and a tail (this seems very unlikely to occur); (@) the outcome is not a head. e (3) Example. A die is thrown once. There are six possible outcomes depending on which of the numbers 1,2, 3, 4, 5, or 6 is uppermost, Thus 2 = (1, 2, 3, 4, 5, 6). We may be interested in the following events: (a) the outcome is the number 1; (b) the outcome is an even number; (©) the outcome is even but does not exceed 3; (d) the outcome is not even e We see immediately that each of the events of these examples can be specified as a subset A of the appropriate sample space @. In the first example they can be rewritten as (a) A={H), (&) A={H) UIT), © A=(H}N{T), @ A=(H, whilst those of the second example become @) A=(l), (b) A=(2,4,6}, (c) A= (2,4,6}9{1, 2, 3}, (d) A= {2,4,6)}°. The complement of a subset A of Q is denoted here and subsequently by A®; from now on, subsets of & containing a single member, such as (H}, will usually be written without the containing braces. Henceforth we think of events as subsets of the sample space 2. Whenever A and B are events in which we are interested, then we can reasonably concern ourselves also with the events AUB, ANB, and A®, representing ‘A or B’,*A and B’,and ‘not A’ respectively. Events, A and B are called disjoint if their intersection is the empty set 2; @ is called the impossible event. The set Q is called the certain event, since some member of & will certainly occur. Thus events are subsets of 2, but need all the subsets of & be events? The answer is no, but some of the reasons for this are too difficult to be discussed here. It suffices for us to think of the collection of events as a subcollection F of the set of all subsets of Q. This subcollection should have certain properties in accordance with the earlier discussion: (a) if A,B € F then AUB Ee F and ANBeEF; (b) if A € F then Ao F; (©) the empty set 2 belongs to F. Any collection ¥ of subsets of 2 which satisfies these three conditions is called a field. It follows from the properties of a field F that if Ai,A2,...,An€F then (Jape F; izl 1.2 Bvents as sets 3 ‘Typical notation _ Set jargon Probability jargon 2 Collection of objects Sample space o Member of Elementary event, outcome A Subset of 2 Event that some outcome in A occurs a Complement of A Event that no outcome in A occurs ANB Intersection Both A and B AUB Union Either A or B or both A\B Difference A, but not B AOB ‘Symmetric difference Either A or B, but not both ACB Inclusion If A, then B 2 Empty set Impossible event 2 Whole space Certain event ‘Table 1.1. The jargon of set theory and probability theory, that is to say, F is closed under finite unions and hence under finite intersections also (see Problem (1.8.3). This is fine when @ is a finite set, but we require slightly more to deal with the common situation when is infinite, as the following example indicates. (4) Example. A coin is tossed repeatedly until the first head turns up; we are concerned with the number of tosses before this happens. The set of all possible outcomes is the set Q = (01, 02, 03, ...}, where «; denotes the outcome when the first i — 1 tosses are tails and the ith toss is a head. We may seek to assign a probability to the event A, that the first head occurs after an even number of tosses, that is, A = {«02, «04, 06, ... }. This is an infinite countable union of members of @ and we require that such a set belong to F in order that we can discuss its probability. e Thus we also require that the collection of events be closed under the operation of taking countable unions. Any collection of subsets of & with these properties is called a o-field. (5) Definition. A collection ¥ of subsets of @ is called a o-field if it satisfies the following conditions: (a) BEF; (b) if Ay, Aa... € F then UPS, Ai € F (© ifA € F then Ao e F. It follows from Problem (1.8.3) that o-fields are closed under the operation of taking countable intersections. Here are some examples of o-fields. (6) Example. The smallest o-field associated with Q is the collection F = {@, 2}. e (7) Example. If A is any subset of 2 then F = {@, A, A°, 2} is ac-field. e (8) Example. The power set of 2, which is written (0, 1} and contains all subsets of @, is obviously a o-field. For reasons beyond the scope of this book, when & is infinite, its power set is too large a collection for probabilities to be assigned reasonably to all its members. @ 4 1.3. Events and their probabilities To recapitulate, with any experiment we may associate a pair (2, F), where @ is the set of all possible outcomes or elementary events and F is a o-field of subsets of & which contains all the events in whose occurrences we may be interested; henceforth, to call a set A an event is equivalent to asserting that A belongs to the o-field in question. We usually translate statements about combinations of events into set-theoretic jargon; for example, the event that both A and B occur is written as A B. Table 1.1 is a translation chart, Exercises for Section 1.2 1. Let {A; :i € 1} be a collection of sets. Prove ‘De Morgan’s Laws": (Ua) =A (Ma) =Uai 2. Let A and B belong to some o-field F. Show that F contains the sets AB, A\ B, and AA B. 3. A conventional knock-out tournament (such as that at Wimbledon) begins with 2” competitors and has n rounds. There are no play-offs for the positions 2, 3, ..., 2” — 1, and the initial table of draws is specified. Give a concise description of the sample space of all possible outcomes. 4, Let F bea o-field of subsets of © and suppose that B ¢ F. Show that $= (ANB: Ae Flisa o-field of subsets of B. 5. Which of the following are identically true? For those that are not, say when they are true. (@ AU(BNC)=(AUB)N(AUC); (6) AN(BNC)=(ANB)NC; (© (AUB)NC=AU(BNC), @ A\ (BNC) =(A\ B)U(A\C). 1.3 Probability We wish to be able to discuss the likelihoods of the occurrences of events. Suppose that we repeat an experiment a large number N of times, keeping the initial conditions as equal as possible, and suppose that A is some event which may or may not occur on each repetition. Our experience of most scientific experimentation is that the proportion of times that A occurs settles down to some value as N becomes larger and larger; that is to say, writing N(A) for the number of occurrences of A in the N trials, the ratio N(A)/N appears to converge to a constant limit as NV increases. We can think of the ultimate value of this ratio as being the probability P(A) that A occurs on any particular trial; it may happen that the empirical ratio does not behave in a coherent manner and our intuition fails us at this level, but we shall not discuss this here. In practice, N’ may be taken to be large but finite, and the ratio N(A)/N may be taken as an approximation to P(A). Clearly, the ratio is a number between zero and one; if A = © then N(@) = 0 and the ratio is 0, whilst if A = @ then N(2) = N and the + Augustus De Morgan is well known forhaving given the first clear statement of the principle of mathematical induction, He applauded probability theory with the words: “The tendency of our study is to substitute the satisfaction of mental exercise for the pernicious enjoyment of an immoral stimulus”. +This superficial discussion of probabilities is inadequate in many ways; questioning readers may care to discuss the philosophical and empirical aspects of the subject amongst themselves (see Appendix IID. 1.3 Probability 5 ratio is 1. Furthermore, suppose that A and B are two disjoint events, each of which may or may not occur at each trial. Then N(AUB) = N(A) + N(B) and so the ratio N(A U B)/N is the sum of the two ratios N(A)/N and N(B)/N. We now think of these ratios as representing the probabilities of the appropriate events. The above relations become PAUB)=P(A)+P(B), =P) =0, (QV = 1 This discussion suggests that the probability function P should be finitely additive, which is to say that if Ay, Az, ..., An ate disjoint events, then P Us) Pan; = ist aglance at Example (1.2.4) suggests the more extensive property that P be countably additive, in that the corresponding property should hold for countable collections Ay, Ag, ... of disjoint events. These relations are sufficient to specify the desirable properties of a probability function P applied to the set of events. Any such assignment of likelihoods to the members of ¥ is called a probability measure. Some individuals refer informally to P as a ‘probability distribution’, especially when the sample space is finite or countably infinite; this practice is best avoided since the term ‘probability distribution’ is reserved for another purpose to be encountered in Chapter 2 (1) Definition, A probability measure P on (2, F) is a function P : ¥ > [0, 1] satisfying (@) P@)=0, F(Q)=1; (b) if Ai, Ap, ... isa collection of disjoint members of F, in that Aj; = @ for all pairs i, j satisfying i # j, then »(Ua) = > P(A). ia The triple (@, FP), comprising a set Q, a o-field F of subsets of 2, and a probability measure P on (&, F), is called a probability space. A probability measure is a special example of what is called a measure on the pair (2, F). A measure is a function j. : F — [0, 00) satisfying (2) = 0 together with (b) above, A measure j1 is a probability measure if px(Q) = 1. We can associate a probability space (@, FP) with any experiment, and all questions associated with the experiment can be reformulated in terms of this space. It may seem natural to ask for the numerical value of the probability P(A) of some event A. The answer to such a question must be contained in the description of the experiment in question. For example, the assertion that a fair coin is tossed once is equivalent to saying that heads and tails have an equal probability of occurring; actually, this is the definition of fairness. 6 1.3. Events and their probabilities (2) Example. A coin, possibly biased, is tossed once. We can take 2 = (H,T) and F = {2, H, T, Q}, and a possible probability measure P : F — [0, 1] is given by P(2)=0, PCH) Pp, PT)=1—p, P)=1, where p is a fixed real number in the interval [0, I]. If p = 4, then we say that the coin is fair, or unbiased. e (3) Example. A die is thrown once. We can take @ = (1,2,3,4,5,6}, F = (0, 1)®, and the probability measure P given by P(A) =o pi forany ACQ, ica where pi, p2, ..-. Po are specified numbers from the interval (0, 1] having unit sum. The probability that i turns up is p;. The die is fair if p; = } for each i, in which case P(A) = HIA|_ for any AC, where || denotes the cardinality of A. e ‘The triple (, F, P) denotes a typical probability space. We now give some of its simple but important properties. (4) Lemma. (a) P(A) = 1- P(A), (b) if BD A then P(B) = P(A) + P(B\ A) = P(A), (©) P(AU B) = P(A) + P(B) — P(AN B), (4) more generally, if Ay, A2,..-, An are events, then P Us) Pad) = POPN AY) + YO PAIN ANA, =» =I igi icjek + (-1)"*'P(AL AQ N+ An) where, for example, 3; ; sums over all unordered pairs (i, j) with i # j. Proof. (a) AUA® = Qand AN AS = 2, 0 P(A U A®) = P(A) + P(A®) = 1 (b) B = AU (B\ A). This is the union of disjoint sets and therefore P(B) = P(A) + P(B \ A). (©) AUB = AU(B\ A), which is a disjoint union. Therefore, by (b), P(AU B) = P(A) + P(B \ A) = P(A) + P(B \ (AN B)) = P(A) + P(B) — (ANB). (d) The proof is by induction on n, and is left as an exercise (see Exercise (1.3.4)). 1.3 Probability 7 In Lemma (4b), B \ A denotes the set of members of B which are not in A. In order to write down the quantity P(B \ A), we require that B \ A belongs to F , the domain of P; this is always true when A and B belong to F, and to prove this was part of Exercise (1.2.2), Notice that each proof proceeded by expressing an event in terms of disjoint unions and then applying P. Itis sometimes easier to calculate the probabilities of intersections of events rather than their unions; part (d) of the lemma is useful then, as we shall discover soon. The next property of P is more technical, and says that P is a continuous set function; this property is essentially equivalent to the condition that P is countably additive rather than just finitely additive (see Problem (1.8.16) also). (5) Lemma. Let Ay, Az, ... be an increasing sequence of events, so that Ay © Ay © A3 S and write A for their limit: > 4 Cea Ai = tim Aj. Mae Then P(A) = lim; P(Ai). Similarly, if B,, Bo, ... is a decreasing sequence of events, so that By > By > By 2+», then B=()Bi = lim B satisfies P(B) = limj-so0 P(B)). Proof. A = Ay U (Ap \ Ai) U(A3 \ Az) U--> is the union of a disjoint family of events. Thus, by Definition (1), P(A) = P(A1) + DS P(Ai+1 \ AD) " nel = P(A) + Jim, Lire —P(A)] = lim POA, - To show the result for decreasing families of events, take complements and use the first part (exercise), . To recapitulate, statements concerning chance are implicitly related to experiments or trials, the outcomes of which are not entirely predictable. With any such experiment we can associate a probability space (Q, F , P) the properties of which are consistent with our shared and reasonable conceptions of the notion of chance. Here is some final jargon. An event A is called null if P(A) = 0. If P(A) = 1, we say that A occurs almost surely. Null events should not be confused with the impossible event 2. Null events are happening all around us, even though they have zero probability; after all, what is the chance that a dart strikes any given point of the target at which it is thrown? That is, the impossible event is null, but null events need not be impossible. 8 1.4 Bvents and their probabilities Exercises for Section 1.3 1, Let Aand B be events with probabilities P(A) = j andP(B) = 4. Show that fs < P(ANB) < 4, and give examples to show that both extremes are possible. Find corresponding bounds for P(AU B). 2. A fair coin is tossed repeatedly. Show that, with probability one, a head turns up sooner or later. Show similarly that any given finite sequence of heads and tails occurs eventually with probability one. Explain the connection with Murphy’s Law. 3. Six cups and saucers come in pairs: there are two cups and saucers which are red, two white, and two with stars on. If the cups are placed randomly onto the saucers (one each), find the probability that no cup is upon a saucer of the same pattern. . 4, Let Ay, Ap, ..., An be events where n > 2, and prove that e(Uas) = Dra Drang + SS PN 4; 1A) iI 7 ig i 1, be events such that P(A,) = 1 for all r. Show that P((Y22, Ar) = 1 6. You are given that at least one of the events A,, 1 1/n and q < 2/n. 7. You are given that at least one, but no more than three, of the events A;, 1 3, The probability of at least two occurring is 5. If P(Ar) = p, P(Ar As) = gr #8, and P(Ar As Ay) =x,7 <8 3/Qn), and q <4/n. 1.4 Conditional probability Many statements about chance take the form ‘if B occurs, then the probability of A isp’, where B and A are events (such as ‘it rains tomorrow’ and ‘the bus being on time’ respectively) and p isa likelihood as before. To include this in our theory, we return briefly to the discussion about proportions at the beginning of the previous section. An experiment is repeated N times, and on each occasion we observe the occurrences or non-occurrences of two events A and B. Now, suppose we only take an interest in those outcomes for which B occurs; all other experiments are disregarded. In this smaller collection of trials the proportion of times that A occurs is N(A 9 B)/N(B), since B occurs at each of them. However, N(ANB) — N(ANB)/N N(B) —N(B)/N If we now think of these ratios as probabilities, we see that the probability that A occurs, given that B occurs, should be reasonably defined as P(A 9 B)/P(B). Probabilistic intuition leads to the same conclusion. Given that an event B occurs, itis the case that A occurs if and only if A 0 B occurs. Thus the conditional probability of A given B 1.4 Conditional probability 9 should be proportional to P(AN, B) which isto say that it equals aP(AN B) for some constant @ he conditional probability of 2 given B must equal 1, and thus @P(2N B) = 1, yielding « = 1/P(B). We formalize these notions as follows. (1) Definition, If P(B) > 0 then the conditional probability that A occurs given that B occurs is defined to be P(ANB) PUL B)= We denote this conditional probability by P(A | B), pronounced ‘the probability of A given B’, or sometimes ‘the probability of A conditioned (or conditional) on B’. (2) Example. Two fair dice are thrown. Given that the first shows 3, what is the probability that the total exceeds 6? The answer is obviously }, since the second must show 4, 5, or 6. However, let us labour the point. Clearly 2 = {1,2,3, 4,5, 6}, the sett of all ordered pairs (i, j) for i, j € {1,2,...,6), and we can take to be the set of all subsets of 2, with P(A) = |A|/36 for any A CQ. Let B be the event that the first die shows 3, and A be the event that the total exceeds 6. Then B=({G,b):16}, ANB =({(,4), (3,5).G.6)}, and PANB) _|ANB|_ 3 PB) iB. 6 U P(A| B)= (3) Example. A family has two children, What is the probability that both are boys, given that at least one is a boy? The older and younger child may each be male or female, so there are four possible combinations of sexes, which we assume to be equally likely. Hence we can represent the sample space in the obvious way as & = (GG, GB, BG, BB) where P(GG) = P(BB) = P(GB) = P(BG) = }. From the definition of conditional probability, P(BB | one boy at least) = P(BB | GB U BG U BB) _ P(BBN (GBUBGUBB)) ~— P(GBUBGUBB) P(BB) 1 ™ B(GBUBGUBB) 3 A popular but incorrect answer to the question is 3. This is the correct answer to another question: for a family with two children, what is the probability that both are boys given that the younger is a boy? In this case, P(BB | younger is a boy) = P(BB | GBU BB) _ P(BBA GB UBB)) PBB) it ~ P(GBUBB) PGGBUBB) 2° Remember that A x B = (a,b) :a € A, b € B) and that A x A = A? 10 1.4 Events and their probabilities ‘The usual dangerous argument contains the assertion P(BB | one child is a boy) = P(other child is a boy). Why is this meaningless? [Hint: Consider the sample space.] e The next lemma is crucially important in probability theory. A family By, Bz, ..., By of events is called a partition of the set Q if Bi B)=o when i#j, and (JB; Each elementary event « € & belongs to exactly one set in a partition of & (4) Lemma. For any events A and B such that 0 0 for alli. Then P(A) =)“ P(A | B)PB))- Proof. A = (AN B)U (AN B®). This is a disjoint union and so P(A) = P(AN B) + P(AN BS) = P(A | B)P(B) + P(A | B°)P(B®). The second part is similar (see Problem (1.8.10)). . (5) Example. We are given two urns, each containing a collection of coloured balls. Urn I contains two white and three blue balls, whilst urn II contains three white and four blue balls. A ball is drawn at random from urn J and put into urn Il, and then a ball is picked at random from urn I and examined. What is the probability that its blue? We assume unless otherwise specified that a ball picked randomly from any urn is equally likely to be any of those present. ‘The reader will be relieved to know that we no longer need to describe (2, F ,P) in detail; we are confident that we could do so if necessary. Clearly, the colour of the final ball depends on the colour of the ball picked from urn I. So let us ‘condition’ on this. Let A be the event that the final ball is blue, and let B be the event that the first one picked was blue. Then, by Lemma (4), P(A) = P(A | B)P(B) + P(A | BY)P(BS). We can easily find all these probabilities: P(A | B) = P(A | urn II contains three white and five blue balls) = 3, P(A | B°) = P(A | urn II contains four white and four blue balls) = b P(B)=3, PCB) = 3. 1.4 Conditional probability i Hence P(A) e g34+3°5 Unprepared readers may have been surprised by the sudden appearance of urns in this book. In the seventeenth and eighteenth centuries, lotteries often involved the drawing of slips from urns, and voting was often a matter of putting slips or balls into urns. In France today, aller aux umes is synonymous with voting. It was therefore not unnatural for the numerous Bernoulli and others to model births, marriages, deaths, fluids, gases, and so on, using urns containing balls of varied hue. (6) Example. Only two factories manufacture zoggles. 20 per cent of the zoggles from factory Land 5 per cent from factory Il are defective. Factory I produces twice as many zoggles as factory IT each week. What is the probability that a zoggle, randomly chosen from a week’s production, is satisfactory? Clearly this satisfaction depends on the factory of origin. Let A be the event that the chosen zoggle is satisfactory, and let B be the event that it was made in factory I. Arguing as before, P(A) = P(A | B)P(B) + P(A | B°)P(BS) 4.2419 1251 3°3+20°3 = If the chosen zoggle is defective, what is the probability that it came from factory I? In our notation this is just P(B | A®). However, PIBOA®) _ P(AS| BYP(B) _ PBI A = Fa = aay This section is terminated with a cautionary example. It is not untraditional to perpetuate errors of logic in calculating conditional probabilities. Lack of unambiguous definitions and notation has led astray many probabilists, including even Boole, who was credited by Russell with the discovery of pure mathematics and by others for some of the logical foundations of computing. The well-known ‘prisoners’ paradox’ also illustrates some of the dangers here. (7) Example. Prisoners’ paradox. Ina dark country, three prisoners have been incarcerated without trial. Their warder tells them that the country’s dictator has decided arbitrarily to free one of them and to shoot the other two, but he is not permitted to reveal to any prisoner the fate of that prisoner. Prisoner A knows therefore that his chance of survival is }. In order to gain information, he asks the warder to tell him in secret the name of some prisoner (but not himself) who will be killed, and the warder names prisoner B. What now is prisoner A’s assessment of the chance that he will survive? Could it be }: after all, he knows now that the survivor will be either A or C, and he has no information about which? Could it be 4 after all, according to the rules, at least one of B and C has to be killed, and thus the extra information cannot reasonably affect A’s earlier calculation of the odds? What does the reader think about this? The resolution of the paradox lies in the situation when either response (B or C) is possible. An alternative formulation of this paradox has become known as the Monty Hall problem, the controversy associated with which has been provoked by Marilyn vos Savant (and many others) in Parade magazine in 1990; see Exercise (1.4.5). e 12 1.4. Events and their probabilities Exercises for Section 1.4 1, Prove that P(A | B) = P(B | A)P(A)/P(B) whenever P(A)P(B) # 0. Show that, if P(A | B) > P(A), then P(B | A) > P(B). 2. Forevents Aj, A,..., An satisfying P(Ay 1. Az +++ Ay—1) > 0, prove that P(A, 1A A+++ An) = P(A1)P(A2 | Ar)P(A3 | Ay Ag) <= P(An | AL Ad 1+ An). 3. Aman possesses five coins, two of which are double-headed, one is double- normal. He shuts his eyes, picks a coin at random, and tosses it. What is the probability that the lower face of the coin is a head? He opens his eyes and sees that the coin is showing heads; what is the probability that the lower face is a head? He shuts his eyes again, and tosses the coin again. What is the probability that the lower face is ahead? He opens his eyes and sees that the coin is showing heads; what is the probability that the lower face is a head? He discards this coin, picks another at random, and tosses it. What is the probability that it shows heads? 4. What do you think of the following ‘proof” by Lewis Carroll that an um cannot contain two balls of the same colour? Suppose that the umn contains two balls, each of which is either black or white; thus, in the obvious notation, (BB) = F(BW) = P(WB) = P(WW) = }. We adda black ball, so that P(BBB) = P(BBW) = P(BWB) = F(BWW) = §. Next we pick a ball at random; the chance that the ball is black is (using conditional probabilities) 1-}+3-4+3-4+4-4 = 3. However, if there is probability 3 that a ball, chosen randomly from three, is black, then there must be two black and one white, which is to say that originally there was one black and one white ballin the urn. 5. The Monty Hall problem: goats and cars. (a) Cruel fate has made you a contestant in a game show; you have to choose one of three doors. One conceals a new car, two conceal old goats. You choose, but your chosen door is not opened immediately. Instead, the presenter opens another door to reveal a goat, and he offers you the opportunity to change your choice to the third door (unopened and so far unchosen). Let p be the (conditional) probability that the third door conceals the car. The value of p depends on the presenter’s protocol. Devise protocols to yield the values p = }, p = 3, Show that, for a € (4, 3], there exists a protocol such that p = a. Are you well advised to change ‘your choice to the third door? (b) Ina variant of this question, the presenter is permitted to open the first door chosen, and to reward you with whatever lies behind. If he chooses to open another door, then this door invariably conceals 4a goat. Let p be the probability that the unopened door conceals the car, conditional on the presenter having chosen to open a second door. Devise protocols to yield the values p = 0, p = 1, and deduce that, for any a € [0, 1], there exists a protocol with p = a. 6. The prosecutor’s fallacy}. Let G be the event that an accused is guilty, and 7° the event that some testimony is true. Some lawyers have argued on the assumption that P(G | T) = P(T | G). Show that this holds if and only if (G) = P(T). 7. Urns, There are n ums of which the rth contains r ~ 1 red balls and m —r magenta balls. You pick an urn at random and remove two balls at random without replacement. Find the probability that: (@) the second ball is magenta; (b) the second ball is magenta, given that the first is magenta. +The prosecution made this error in the famous Dreyfus case of 1894. 1.5 Independence 1B 1.5 Independence In general, the occurrence of some event B changes the probability that another event A occurs, the original probability P(A) being replaced by P(A | B). If the probability remains unchanged, that is to seca |B) Py ihen we call A and B ‘independent’. This is well defined only if P(B) > 0. Definition 1.4.1) of conditional probability leads us to the following. (1) Definition. Events A and B are called independent if (res B) = P(A)P(B). € 1} is called independent if Ie More generally, a family {Ay : for all finite subsets J of J. Remark. A common student error is to make the fallacious statement that A and B are independent if AN B = 2. If the family (A; : i € 7} has the property that P(A;NAj) =P(ADP(A;) — for alli # j then it is called pairwise independent. Pairwise-independent families are not necessarily independent, as the following example shows. (2) Example. Suppose 2 = (abe, acb, cab, cba, bea, bac, aaa, bbb, ccc}, and each of the nine elementary events in @ occurs with equal probability 4. Let A, be the event that the Ath letter is a. Itis left as an exercise to show that the family (A, A2, A3} is pairwise independent but not independent. e (3) Example (1.4.6) revisited. The events A and B of this example are clearly dependent because P(A | B) = $ and P(A) = 3. e (4) Example. Choose a card at random from a pack of 52 playing cards, each being picked with equal probability 3. We claim that the suit of the chosen card is independent of its rank. For example, Pking) = 4, P(king | spade) = 74. Alternatively, P(spade king) = 35 = j - 7 = P(spade)P(king). e Let C be an event with P(C) > 0. To the conditional probability measure P(- | C) corresponds the idea of conditional independence. Two events A and B are called conditionally independent given C if (5) P(ANB|C)=P(A| C)P(B | C); there is a natural extension to families of events. [However, note Exercise (1.5.5).] 14 1.6 Events and their probabilities Exercises for Section 1.5 1, Let A and B be independent events; show that A°, B are independent, and deduce that A°, B° are independent. 2. We roll adie n times. Let Aj; be the event that the ith and jth rolls produce the same number. Show that the events {Ajj : 1 (0, 1] be given by: QB) Pi2(Aq Xx A2) = Pi(A1)P2(A2) for Ay € Fi, Az © Fe It can be shown that the domain of P12 can be extended from #1 x F) to the whole of g = o(Fi x F2). The ensuing probability space (Q1 x Q2, 9, Piz) is called the product space of (21, Fi, P1) and (Q2, F2, P2). Products of larger numbers of spaces are constructed similarly. The measure Pj2 is sometimes called the ‘product measure” since its defining equation (3) assumed that two experiments are independent. There are of course many other measures that can be applied to (21 x 22, 9). In many simple cases this technical discussion is unnecessary. Suppose that 1 and 22 are finite, and that their o-fields contain all their subsets; this is the case in Examples (1.2.4) and (1.4.2). Then J contains all subsets of 21 x 22 16 1.7 Events and their probabilities 1.7 Worked examples Here are some more examplesto illustrate the ideas of this chapter. The reader is now equipped to try his or her hand at a substantial number of those problems which exercised the pioneers in probability. These frequently involved experiments having equally likely outcomes, such as dealing whist hands, putting balls of various colours into urns and taking them out again, throwing dice, and so on. In many such instances, the reader will be pleasantly surprised to find that it is not necessary to write down (Q, F , P) explicitly, but only to think of Q as being acollection {«1, a2, ..., @v} of possibilities, each of which may occur with probability 1/N. Thus, P(A) = |A|/N for any A C Q. The basic tools used in such problems are as follows. (a) Combinatorics: remember that the number of permutations of n objects isn! and that the number of ways of choosing r objects from n is ("). (b) Set theory: to obtain P(A) we can compute P(A‘) = 1 — P(A) or we can partition A by conditioning on events B;, and then use Lemma (1.4.4). (©) Use of independence. (1) Example. Consider a series of hands dealt at bridge. Let A be the event that in a given deal each player has one ace. Show that the probability that A occurs at least once in seven deals is approximately } Solution. ‘The number of ways of dealing 52 cards into four equal hands is 52!/(13!)4. There are 4! ways of distributing the aces so that each hand holds one, and there are 48!/(12!)* ways of dealing the remaining cards. Thus 4148/24 1 521/034 — 10° Now let B; be the event that A occurs for the first time on the ith deal. Clearly B; Bj = 2, i 4 j. Thus P(A) 1 P(A occurs in seven deals) = P(B) U---U Bz) =) P(B)) using Definition (1.3.1). 7 Since successive deals are independent, we have P(B;) = P(A® occurs on deal 1, A° occurs on deal 2, ; A® occurs on deal i — 1, A occurs on deal ’) = P(A‘)! P(A) using Definition (1.5.1) =(i- 4) b Thus 1 1 P(A occurs in seven deals) =)“ P(Bi) ~ > ( 7 7 Can you see an easier way of obtaining this answer? ° (2) Example. There are two roads from A to B and two roads from B to C. Each of the four roads has probability p of being blocked by snow, independently of all the others. What is the probability that there is an open road from A to C? 1.7 Worked examples 7 Solution. P(open road) = P((open road from A to B) 9 (open road from B to C)) = P(open road from A to B)P(open road from B to C) using the independence. However, p is the same for all roads; thus, using Lemma (1.3.4), (open road) = (1 — P(no road from A to B))” = {1 — P((first road blocked) n (second road blocked) y = {1 — P(first road blocked)? (second road blocked) |” using the independence. Thus @) (open road) = (1 — p?)?. Further suppose that there is also a direct road from A to C, which is independently blocked with probability p. Then, by Lemma (1.4.4) and equation (3), P(open road) = P(open road | direct road blocked) - p + P(open road | direct road open) - (1 — p) = (1 =p’): p+1-(1 =p). e (4) Example. Symmetric random walk (or ‘Gambler’s ruin’). A man is saving up to buy anew Jaguar at a cost of N units of money. He starts with k units where 0 < k < N, and tries to win the remainder by the following gamble with his bank manager. He tosses a fair coin repeatedly; if it comes up heads then the manager pays him one unit, but if it comes up tails then he pays the manager one unit. He plays this game repeatedly until one of two events occurs: either he runs out of money and is bankrupted or he wins enough to buy the Jaguar. What is the probability that he is ultimately bankrupted? Solution. This is one of many problems the solution to which proceeds by the construction of a linear difference equation subject to certain boundary conditions. Let A denote the event that he is eventually bankrupted, and let B be the event that the first toss of the coin shows heads. By Lemma (1.4.4), 6) Pe(A) = P(A | BYP(B) + Px(A | BYP(BS), where F; denotes probabilities calculated relative to the starting point k. We want to find P(A). Consider P;(A | B). If the first toss is a head then his capital increases to k + 1 units and the game starts afresh from a different starting point. Thus P(A | B) = Px41(A) and similarly Py(A | B®) = Pg—1(A). So, writing py = Px(A), (5) becomes: © Pk = 5(Peri + Pkt) if O P(A| BSAC) and P(A| BNC) > P(A| BNC) but (12) P(A | B) < P(A| B®). named after Simpson (1951), was remarked by Yule in 1903. ‘The nomer instance of Stigler's law of eponymy: “No law, theorem, or discovery is named after its originator’ applies to many eponymous statements in this book, including the law itself. As remarked by A. N, Whitehead, “Everything of importance has been said before, by somebody who did not discover it” 20 1.7 Events and their probabilities h D3 & Ds Figure 1.1. Two unions of rectangles illustrating Simpson's paradox. ‘We may think of A as the event that treatment is successful, B as the event that drug Tis given toa randomly chosen individual, and C as the event that this individual is female. The above inequalities imply that B is preferred to B® when C occurs and when C® occurs, but BS is preferred to B overall. Setting a=P(ANBNC), b=P(ASNBNC), c=P(ANBNC), d=P(ASN BNC), e=P(ANBNC), f=P(ASNBNC), g=PIANB'NC), h=PASN BNC), and expanding (11)-(12), we arrive at the (equivalent) inequalities (13) ad >be, eh> fg, (at+e\(dt+h) <(b+ fi(c+s), subject to the conditions a, b,c,...,h > Oanda+b+e+-+-+h = 1. Inequalities (13) are equivalent to the existence of two rectangles Ry and Ro, as in Figure 1.1, satisfying area(D}) > area(D), area(D3) > area(D4), area(R}) < area(Rp). Many such rectangles may be found, by inspection, as for example those witha = 3,b = +, He=hfahsaha Similar conclusions are valid for finer € 1) of the sample space, though the corresponding pictures are hardér to c= $d partitions (C; draw. Simpson’s paradox has arisen many times in practical situations. There are many well- known cases, including the admission of graduate students to the University of California at Berkeley and a clinical trial comparing treatments for kidney stones. e (14) Example. False positives. A rare disease affects one person in 10°. A test for the disease shows positive with probability ;%; when applied to an ill person, and with probability 7p When applied to a healthy person. What is the probability that you have the disease given that the test shows positive? Solution. In the obvious notation, PC | ill) Gill) (I)PGl) + PG | healthy)P(healthy) _ 7-105 _ 9 2 ~ 10-3 + hl — 10-5) 99+ 105-1 ~ 1011 is rather small. Indeed it is more likely that the test was incorrect. @ PAIL) = Bey The chance of bein; 1.8 Problems 21 Exercises for Section 1.7 1. There are two roads from A to B and two roads from B to C. Each of the four roads is blocked by snow with probability p, independently of the others. Find the probability that there is an open road from A to B given that there is no open route from A to C. If, in addition, there is a direct road from A to C, this road being blocked with probability p independently of the others, find the required conditional probability. 2. Calculate the probability that a hand of 13 cards dealt from a normal shuffled pack of 52 contains exactly two kings and one ace. What is the probability that it contains exactly one ace given that it contains exactly two kings? 3. Asymmetric random walk takes place on the integers 0, 1, 2, ..., /V with absorbing barriers at 0 and N, starting at k. Show that the probability that the walk is never absorbed is zero. 4. The so-called ‘sure thing principle’ asserts that if you prefer x to y given C, and also prefer x to y given C®, then you surely prefer x to y. Agreed? 5. A pack contains m cards, labelled 1,2,...,m. The cards are dealt out in a random order, one by one. Given that the label of the kth card dealt is the largest of the first k cards dealt, what is the probability that itis also the largest in the pack? 1.8 Problems 1. A traditional fair die is thrown twice. What is the probability that: (a) a six tums up exactly once? (b) both numbers are odd? (©) the sum of the scores is 4? (@) the sum of the scores is divisible by 3? 2. A fair coin is thrown repeatedly. What is the probability that on the nth throw: (a) ahead appears for the first time? (b) the numbers of heads and tails to date are equal? (©) exactly two heads have appeared altogether to date? (@) at least two heads have appeared to date? 3. Let Fand 9 be o-fields of subsets of 2. (a) Use elementary set operations to show that F is closed under countable intersections; that is, if Ay, Ap,... ate in F, then so is 1; Ay (b) Let = FG be the collection of subsets of @ lying in both Fand g. Show that J is a.o-field. (©) Show that FUG, the collection of subsets of @ lying in either For @, is not necessarily a o-field 4, Describe the underlying probability spaces for the following experiments: (a) a biased coin is tossed three times; (b) two balls are drawn without replacement from an urn which originally contained two ultramarine and two vermilion balls; (c) a biased coin is tossed repeatedly until a head turns up. 5. Show that the probability that exactly one of the events A and B occurs is P(A) + F(B) — 2F(A 1 B). 6. Prove that P(A U BUC) = 1 — P(A® | BSN C®)P(BE | C)B(C*). 2 1.8 Events and their probabilities 7. (a) If Ais independent of itself, show that P(A) is 0 or 1 (b) IF P(A) is 0 or 1, show that A is independent of all events B. 8& Let Fbe ao-field of subsets of , and suppose P : F > {0, 1] satisfies: (i) P() = 1, and (i) P is additive, in that P(A U B) = P(A) + P(B) whenever AN B = @. Show that P(2) = 0. 9. Suppose (2, F. P) is a probability space and B € F satisfies P(B) > 0. Let Q: F > [0, 1] be defined by Q(A) = P(A | B). Show that (2, F, Q) is a probability space. If C € Fand Q(C) > 0, show that Q(4 | C) = F(A | BC); discuss. 10. Let By, Bp, ... be a partition of the sample space @, each B; having positive probability, and show that P(A) = P(A | B)P(B)) i LL. Prove Boole’s inequalities: o(U ) < SO P(A), (A 4) zl - Sup. 12, Prove that »(Aa) = OP) = OP U A) + SO PAU A; U AR) 1 7 ig icjck moe (CIMP(A] U Ag UU An), 13. Let Aj, Az,..-, An be events, and let Ny be the event that exactly k of the A; occur. Prove the result sometimes referred to as Waring’s theorem: = Yanan N Ay) iy che P(A). 17. In Problem (1.8.16) above, show that B and C are independent whenever By and Cy are inde- pendent for all n. Deduce that if this holds and furthermore Aj, —> A, then P(A) equals either zero or one. 18, Show that the assumption that P is countably additive is equivalent to the assumption that P is continuous. That is to say, show that if a function P : ¥ —~ [0, 1] satisfies P(@) = 0, P(Q) = 1, and P(AU B) = P(A) + P(B) whenever A, B € Fand AM B = @, then P is countably additive (in the sense of satisfying Definition (1.3.1b)) if and only if P is continuous (in the sense of Lemma (1.3.5)). 19. Anne, Betty, Chloé, and Daisy were all friends at school. Subsequently each of the (3) = 6 subpairs meet up; at each of the six meetings the pair involved quarrel with some fixed probability p. or become firm friends with probability 1 — p. Quarrels take place independently of each other. In future, if any of the four hears a rumour, then she tells it to her firm friends only. If Anne hears a rumour, what is the probability that: (a) Daisy hears it? (b) Daisy hears it if Anne and Betty have quarrelled? (©) Daisy hears it if Betty and Chloé have quarrelled? (@) Daisy hears it if she has quarrelled with Anne? 20. A biased coin is tossed repeatedly. Each time there is a probability p of a head turning up. Let pa be the probability that an even number of heads has occurred after m tosses (zero is an even number). Show that pp = 1 and that py = p(1— pp—1)+(1—p) Py ifn > 1. Solve this difference equation 21. A biased coin is tossed repeatedly. Find the probability that there is a run of r heads in a row before there is a run of s tails, where r and s are positive integers. 22. A bowl contains twenty cherries, exactly fifteen of which have had their stones removed. A greedy pig eats five whole cherries, picked at random, without remarking on the presence or absence of stones. Subsequently, a cherry is picked randomily from the remaining fifteen. (a) What is the probability that this cherry contains a stone? (b) Given that this cherry contains a stone, what is the probability that the pig consumed at least one stone? 23. The ‘ménages’ problem poses the following question. Some consider it to be desirable that men and women alternate when seated at a circular table. If n couples are seated randomly according to this rule, show that the probability that nobody sits next to his or her partner is, ee 4 2n (2n-k ae a( A Jew ‘You may find it useful to show first that the number of ways of selecting k non-overlapping pairs of adjacent seats is (""-*)2n(2n — k)~! 24 1.8 Events and their probabilities 24, Anum contains b blue balls and r red balls. They are removed at random and not replaced. Show that the probability that the first red ball drawn is the (k + 1)th ball drawn equals ("0") / (ey) Find the probability that the last ball drawn is red, 25, An um contains a azure balls and c carmine balls, where ac 4 0. Balls are removed at random and discarded until the first time that a ball (B, say) is removed having a different colour from its predecessor. The ball B is now replaced and the procedure restarted. This process continues until the Jast ball is drawn from the urn, Show that this last ball is equally likely to be azure or carmine. 26. Protocols. A pack of four cards contains one spade, one club, and the two red aces. You deal two cards faces downwards at random in front of a truthful friend. She inspects them and tells you that one of them is the ace of hearts. What is the chance that the other card is the ace of diamonds? Perhaps 5? Suppose that your friend’s protocol was: (a) with no red ace, say “no red ace”, (b) with the ace of hearts, say “ace of hearts”, (c) with the ace of diamonds but not the ace of hearts, say “ace of diamonds”. Show that the probability in question is 4. Devise a possible protocol for your friend such that the probability in question is zero. 27. Eddington’s controversy. Four witnesses, A, B, C, and D, at a trial each speak the truth with probability + independently of each other. In their testimonies, A claimed that B denied that C declared that D lied. What is the (conditional) probability that D told the truth? [This problem seems to have appeared first as a parody in a university magazine of the ‘typical’ Cambridge Philosophy Tripos question.) 28. The probabilistic method. 10 per cent of the surface of a sphere is coloured blue, the rest is red. Show that, irrespective of the manner in which the colours are distributed, it is possible to inscribe a cube in S with all its vertices red. 29, Repulsion. The event A is said to be repelled by the event B if P(A | B) < P(A), and to be attracted by B if P(A | B) > P(A). Show that if B attracts A, then A attracts B, and B° repels A. If A attracts B, and B attracts C, does A attract C? 30. Birthdays. If m students born on independent days in 1991 are attending a lecture, show that the probability that at least two of them share a birthday is p = 1 — (365)!/{(365 — m)!365""}. Show that p > } when m = 23 31, Lottery. You choose r of the first positive integers, and a lottery chooses a random subset L of the same size. What is the probability that: (a) L includes no consecutive integers? (b) L includes exactly one pair of consecutive integers? (c) the numbers in L are drawn in increasing order? (d) your choice of numbers is the same as L? (e) there are exactly k of your numbers matching members of L? 32. Bridge. During a game of bridge, you are dealt at random a hand of thirteen cards. With an obvious notation, show that P(4S, 3H, 3D, 3C) ~ 0.026 and F(4S, 4H, 3D, 2C) ~ 0.018. However if suits are not specified, so numbers denote the shape of your hand, show that (4, 3, 3, 3) ~ 0.11 and P(4, 4, 3,2) ~ 0.22. 33. Poker. During a game of poker, you are dealt a five-card hand at random. With the convention that aces may count high or low, show that: PA pair) = 0.423, P(2 pairs) ~ 0.0475, P(3 of akind) ~ 0.021, P(straight) ~ 0.0039, P(flush) ~ 0.0020, (full house) ~: 0.0014, PG of akind) ~ 0.00024, (straight flush) ~ 0.000015. 18 Problems 25 34, Poker dice. There are five dice each displaying 9, 10, J, Q, K, A. Show that, when rolled: FC pair) ~ 0.46, P(2 pairs) ~ 0.23, P(3 of akind) ~ 0.15, (no 2 alike) ~ 0.093, (full house) ~ 0.039, P(4 of akind) ~ 0.019, PG of akind) ~ 0.0008. 35. You are lost in the National Park of Bandrika+. Tourists comprise two-thirds of the visitors to the park, and give a correct answer to requests for directions with probability }. (Answers to repeated questions are independent, even if the question and the person are the same.) If you ask a Bandrikan for directions, the answer is always false. (a) You ask a passer-by whether the exit from the Park is East or West. The answer is East. What is the probability this is correct? (b) You ask the same person again, and receive the same reply. Show the probability that it is correct, is } (© You ask the same person again, and receive the same reply. What is the probability that itis correct? (@) You ask forthe fourth time, and receive the answer East. Show thatthe probability its corect is 3 (©) Show that, had the fourth answer been West instead, the probability that East is nevertheless correct is 7. 36. Mr Bayes goes to Bandrika. Tom is in the same position as you were in the previous problem, but he has reason to believe that, with probability €, East is the correct answer. Show that: (a) whatever answer first received, Tom continues to believe that East is correct with probability «, (b) if the first two replies are the same (that is, either WW or EE), Tom continues to believe that East is correct with probability €, (©) after three like answers, Tom will calculate as follows, in the obvious notation: 9 lle [rage Past comect | WWW) = 55 (East correct | EEE) = Evaluate these when 37. Bonferroni’s inequality. Show that °(U 4) = OP) — DPA Ay. yan ah rk 38. Kounias’s inequality. Show that o(U 4) < min {Sea = Arn vo}. rire 39. Then passengers for a Bell-Air flight in an airplane with n seats have been told their seat numbers. ‘They get on the plane one by one. The first person sits in the wrong seat. Subsequent passengers sit in their assigned seats whenever they find them available, or otherwise in a randomly chosen empty seat. What is the probability that the last passenger finds his seat free? 1A fictional country made famous in the Hitchcock film ‘The Lady Vanishes’ 2 Random variables and their distributions Summary. Quantities governed by randomness correspond to functions on the probability space called random variables. The value taken by a random vari- able is subject to chance, and the associated likelihoods are described by a function called the distribution function. ‘Two important classes of random variables are discussed, namely discrete variables and continuous variables. ‘The law of averages, known also as the law of large numbers, states that the proportion of successes in a long run of independent trials converges to the probability of success in any one trial. This result provides a mathematical basis for a philosophical view of probability based on repeated experimenta- tion. Worked examples involving random variables and their distributions are included, and the chapter terminates with sections on random vectors and on Monte Carlo simulation. 2.1 Random variables We shall not always be interested in an experiment itself, but rather in some consequence of its random outcome. For example, many gamblers are more concerned with their losses than with the games which give rise to them. Such consequences, when real valued, may be thought of as functions which map @ into the real line IR, and these functions are called ‘random? variables’. (1) Example. A fair coin is tossed twice: 2 = (HH, HT, TH, TT}. For @ € Q, let X(w) be the number of heads, so that X(HH) 1, X(T) =0. X(HT) = X(TH) Now suppose that a gambler wagers his fortune of £1 on the result of this experiment. He gambles cumulatively so that his fortune is doubled each time ahead appears, and is annihilated on the appearance of a tail. His subsequent fortune W is a random variable given by W(HH) = W(HT) = W(TH) = W(TT) = 0 e {Derived from the Old French word randon meaning ‘haste’ 2.1 Random variables 27 Figure 2.1. The distribution function Fy of the random variable X of Examples (1) and (5).. After the experiment is done and the outcome € Q is known, arandom variable X : 2 — R takes some value. In general this numerical value is more likely to lie in certain subsets of R than in certain others, depending on the probability space (, ¥ ,P) and the function X itself, We wish to be able to describe the distribution of the likelihoods of possible values of X. Example (1) above suggests that we might do this through the function f : R — [0,1] defined by f(x) = probability that X is equal to x, but this turns out to be inappropriate in general. Rather, we use the distribution function F :R— R defined by F(x) = probability that X does not exceed x. More rigorously, this is (2) F(x) =P(A(x)) where A(x) C is given by A(x) = {o € @ : X(w) < x}. However, P is a function on the collection ¥ of events; we cannot discuss P(A (x)) unless A(x) belongs to ¥, and so we are led to the following definition. (3) Definition, A random variable is a function X : 2 —» R with the property that (o € Q: X(@) (0, 1] given by F(x) = P(X = x). This is the obvious abbreviation of equation (2). Events written as (w € @ : X(w) < x} are commonly abbreviated to {e : X(w) < x} or {X < x}. We write Fy where it is necessary to emphasize the role of X. 28 2.1 Random variables and their distributions Figure 2.2. The distribution function Fy of the random variable W of Examples (1) and (5). (5) Example (1) revisited. ‘The distribution function Fy of X is given by 0 ifx <0, 4 if02, and is sketched in Figure 2.1. The distribution function Fy of W is given by 0 ifx <0, Fwa)=4 3 if04, and is sketched in Figure 2.2. This illustrates the important point that the distribution function. of a random variable X tells us about the values taken by X and their relative likelihoods, rather than about the sample space and the collection of events. e (6) Lemma. A distribution function F has the following properties: (@) lim FG) =0, tim F(@) = 1, (b) ifx < y then F(x) < F(y), (©) F is right-continuous, that is, F(x +h) > F(x) ash | 0. Proof. (a) Let By = {w € 2: X(w) < —n) = (X <—n). The sequence By, Bo, ... isdecreasing with the empty set as limit. Thus, by Lemma (1.3.5), P(B,) —> P(@) = 0. The other part is similar. (b) Let A(x) = {X < x}, AG, y) = (x < X < y}. Then A(y) = AQ) U AG, y) isa disjoint union, and so by Definition (1.3.1), P(A(y)) = P(A(x)) + P(AQ, y)) giving F(y) = F(X) +PQ@ < X < y) > FQ). (©) This is an exercise. Use Lemma (1.3.5). . 2.1 Random variables 29 Actually, this lemma characterizes distribution functions. Thats to say, F isthe distribution function of some random variable if and only if it satisfies (6a), (6b), and (6c). For the time being we can forget all about probability spaces and concentrate on random variables and their distribution functions. The distribution function F of X contains a great deal of information about X. (7) Example. Constant variables. The simplest random variable takes a constant value on the whole domain Q. Let c € R and define X : @ > R by X()=c forall weQ ‘The distribution function F(x) = P(X < x) is the step function 0 x R be given by X(H) 1, X(T) = Then X is the simplest non-trivial random variable, having two possible values, 0 and 1. Its distribution function F(x) = P(X < x) is 0 x <0, FQ@)=)1-p Osx<1, 1 xel X is said to have the Bernoulli distribution sometimes denoted Bern(p). e (9) Example. Indicator functions. A particular class of Bernoulli variables is very useful in probability theory. Let A be an event and let /4 : 2 —> R be the indicator function of A; that is, 1 ifweAd, USO { 0 ifwe a. Then [4 is a Bernoulli random variable taking the values 1 and 0 with probabilities P(A) and P(A) respectively. Suppose {B; : i € 1} is a family of disjoint events with A © Uje, Bi Then (10) In =D lane, an identity which is often useful. e 30 2.2 Random variables and their distributions (11) Lemma. Let F be the distribution function of X. Then (a) P(X > x) =1- FQ), (b) Pa x)= F(x), T2(x) = P(X < x) = F(-x), where x is large and positive, We shall see later that the rates at which the T; decay to zero as x —> oo have a substantial effect on the existence or non-existence of certain associated quantities called the ‘moments’ of the distribution. Exercises for Section 2.1 1. Let X be arandom variable on a given probability space, and let a € R. Show that (i) aX is a random variable, (ii) X — X =0, the random variable taking the value 0 always, and X + X = 2X. 2. Arandom variable X has distribution function F. Whats the distribution function of Y = aX+b, where a and b are real constants? 3. A fair coin is tossed n times. Show that, under reasonable assumptions, the probability of exactly heads is ('2)(4)". What is the corresponding quantity when heads appears with probability p on each toss? 4, Show thatif F and G are distribution functions and < A < 1 then AF +(1—A)G isa distribution function. Is the product FG a distribution function? 5. Let F be a distribution function and r a positive integer. Show that the following are distribution functions: (@) Foy’, (b) 1- (1 F@®)}", (©) F(x) + (1 — F(x)}log{l — F(x)}, (@) (F(x) = Ne +expll = FO). 2.2. The law of averages We may recall the discussion in Section 1.3 of repeated experimentation. In each of N repetitions of an experiment, we observe whether or not a given event A occurs, and we write N(A) for the total number of occurrences of A. One possible philosophical underpinning of probability theory requires that the proportion N(A)/N settles down as N — oo to some limit interpretable as the ‘probability of A’. Is our theory to date consistent with such a requirement? With this question in mind, let us suppose that Ay, A2,... is a sequence of independent events having equal probability P(A;) = p, where 0 < p < 1; such an assumption requires of 2.2 The law of averages 31 course the existence of a corresponding probability space (@, F , P), but we do not plan to get bogged down in such matters here. We think of 4; as being the event ‘that A occurs on the ith experiment’, We write 5, = "7, [4,, the sum of the indicator functions of Ay, A2,-.. Ans S; isarandom variable which counts the number of occurrences of A; for 1 0, P(p-e1 as n> 0, There are certain technicalities involved in the study of the convergence of random variables (see Chapter 7), and this is the reason for the careful statement of the theorem. For the time being, we encourage the reader to interpret the theorem as asserting simply that the proportion n~'S,, of times that the events Ay, A2,..., An Occur converges as n — 00 to their common probability p. We shall see later how important itis to be careful when making such statements. Interpreted in terms of tosses of a fair coin, the theorem implies that the proportion of heads is (with large probability) near to 4. As a caveat regarding the difficulties inherent in studying the convergence of random variables, we remark that itis nor true that, in a ‘typical’ sequence of tosses of a fair coin, heads outnumber tails about one-half of the time. Proof. Suppose that we toss a coin repeatedly, and heads occurs on each toss with probability p. The random variable S, has the same probability distribution as the number H, of heads which occur during the first n tosses, which is to say that P(S, = k) = P(Hy = k) for all k. It follows that, for small positive values of , P(1s.= pte) = > P(A, =k). 7 ken(p+e) We have from Exercise (2.1.3) that n P(Hy =k) (; Jota = py * for O» (Dota or where m = [n(p + )], the least integer not less than n(p + €). The following argument is standard in probability theory. Let 2 > 0 and note that e > (P+ if k > m. Writing q = 1— p, we have that P( Se =p +) z Se ett-nwo(t) ptr 0, an inequality that is known as ‘Bernstein’s inequality’. It follows immediately that P(a~!S, > p+e)— 0asn > oo. Anexactly analogous argument shows that P(n7!S, < p—€) > 0 as n —> 00, and thus the theorem is proved. . Bernstein’s inequality (4) is rather powerful, asserting that the chance that S, exceeds its mean by a quantity of order n tends to zero exponentially fast as n > oo; such an inequality is known as a ‘large-deviation estimate’. We may use the inequality to prove rather more than the conclusion of the theorem. Instead of estimating the chance that, for a specific value of 1n, Sy lies between n(p — €) and n(p + €), let us estimate the chance that this occurs for all large n. Writing An = {p — € ei" 50 as m— oo, giving that, as required, 1 (6) P(p-e< 18, < pte foraltn > m) +1 asm — 00. n Exercises for Section 2.2 1. You wish to ask each of a large number of people a question to which the answer “yes” is ‘embarrassing. The following procedure is proposed in order to determine the embarrassed fraction of the population. As the question is asked, a coin is tossed out of sight of the questioner. If the answer would have been “no” and the coin shows heads, then the answer “yes” is given. Otherwise people respond truthfully, What do you think of this procedure? 2. Accoin is tossed repeatedly and heads turns up on each toss with probability p. Let Hy and Ty be the numbers of heads and tails in n tosses. Show that, for ¢ > 0, 1 P(2p-1-«< (ty =p) <2p-1 46) = 1 asin 00. n 3, Let {X; : r > 1} be observations which are independent and identically distributed with unknown distribution function F. Describe and justify a method for estimating F (x). 2.3. Discrete and continuous variables 33 2.3 Discrete and continuous variables Much of the study of random variables is devoted to distribution functions, characterized by Lemma (2.1.6). The general theory of distribution functions and their applications is quite difficult and abstract and is best omitted at this stage. It relies on a rigorous treatment of the construction of the Lebesgue-Stieltjes integral; this is sketched in Section 5.6. However, things become much easier if we are prepared to restrict our attention to certain subclasses of random variables specified by properties which make them tractable. We shall consider in depth the collection of ‘discrete’ random variables and the collection of ‘continuous’ random variables. (1) Definition. The random variable X is called discrete if it takes values in some countable subset {x1,2,...}, only, of R. ‘The discrete random variable X has (probability) mass function f : R > [0,1] given by f(x) = P(X =x). We shall see that the distribution function of a discrete variable has jump discontinuities at the values x;,x2,... and is constant in between; such a distribution is called atomic. this contrasts sharply with the other important class of distribution functions considered here. (2) Definition. The random variable X is called continuous if its distribution function can be expressed as Foy= f° frau rer, for some integrable function f : R +> [0, 00) called the (probability) density function of X. The distribution function of a continuous random variable is certainly continuous (actually itis ‘absolutely continuous’). For the moment we are concerned only with discrete variables and continuous variables. There is another sort of random variable, called ‘singular’, for a discussion of which the reader should look elsewhere. A common example of this phenomenon is based upon the Cantor ternary set (see Grimmett and Welsh 1986, or Billingsley 1995). Other variables are ‘mixtures’ of discrete, continuous, and singular variables. Note that the word ‘continuous’ is a misnomer when used in this regard: in describing X as continuous, we are referring to a property of its distribution function rather than of the random variable (function) X itself. (3) Example. Discrete variables. The variables X and W of Example (2.1.1) take values in the sets (0, 1,2} and (0, 4) respectively; they are both discrete. e (4) Example. Continuous variables. A straight rod is flung down at random ontoa horizontal plane and the angle w between the rod and true north is measured. The result is a number in Q = [0,2r). Never mind about ¥ for the moment; we can suppose that F contains all nice subsets of , including the collection of open subintervals such as (a,b), where 0 4m To see this, let 0 R be given by X(T) . X((H, x)) = ‘The random variable X takes values in (—1} U [0, 2s) (see Figure 2.3 for a sketch of its distribution function). We say that X is continuous with the exception of a ‘point mass (or atom) at -1. e 2.4 Worked examples 35 Exercises for Section 2.3 1. Let X be a random variable with distribution function F, and let a = (dm : -00 < m < 00) be a strictly increasing sequence of real numbers satisfying a—y, > —co and dy, —> 00 as m —> 00. Define G(x) = F(X < dm) when d,—1 ER be continuous and strictly increasing. Show that ¥ = g(X) isa random variable. 3. Let X be a random variable with distribution function 0 ifx 1. Let F be a distribution function which is continuous and strictly increasing. Show that ¥ = F~'(X) is arandom variable having distribution function F. Is it necessary that F be continuous and/or strictly increasing? . 4. Show that, if f and g are density functions, and 0 < A <1, then Af + (1 —A)g isa density. Is the product fg a density function? 5. Which of the following are density functions? Find c and the corresponding distribution function F for those that are. ex? x>1, a) f(x) = @ fo) { 0 otherwise. (b) f(x) =ce*(1+e*)?, xER, 2.4 Worked examples (1) Example. Darts. A dart is flung at a circular target of radius 3. We can think of the hitting point as the outcome of a random experiment; we shall suppose for simplicity that the player is guaranteed to hit the target somewhere. Setting the centre of the target at the origin of IR?, we see that the sample space of this experiment is Qa (ix, yx? +y? <9. Never mind about the collection ¥ of events. Let us suppose that, roughly speaking, the probability that the dart lands in some region A is proportional to its area |A|. Thus 2) P(A) = |Al/(9z). The scoring system is as follows. The target is partitioned by three concentric circles C1, C2, and C3, centered at the origin with radii 1, 2, and 3. These circles divide the target into three annuli Ay, Ap, and A3, where A= {(x,y):k-1< yx? + y? b, Indicate how these distribution functions behave as a —* —00, b > 00, 2.5 Random vectors Suppose that X and Y are random variables on the probability space (2, FP). Their dis- tribution functions, Fy and Fy, contain information about their associated probabilities. But how may we encapsulate information about their properties relative to each other? The key is to think of X and Y as being the components of a ‘random vector’ (X, ¥) taking values in iR?, rather than being unrelated random variables each taking values in R. (1) Example. Tontine is a scheme wherein subscribers to a common fund each receive an annuity from the fund during his or her lifetime, this annuity increasing as the other subscribers die. When all the subscribers are dead, the fund passes to the French government (this was the case in the first such scheme designed by Lorenzo Tonti around 1653). The performance of the fund depends on the lifetimes L, L2,..., Ln of the subscribers (as well as on their wealths), and we may record these as a vector (L.1, L2,..., Ln) of random variables. @ (2) Example. Darts, A dart is flung at a conventional dartboard. The point of striking determines a distance R from the centre, an angle © with the upward vertical (measured clockwise, say), and a score S. With this experiment we may associate the random vector (R, ®, S), and we note that S is a function of the pair (R, ). e (3) Example. Coin tossing. Suppose that we toss a coin n times, and set X; equal to 0 or 1 depending on whether the ith toss results in a tail or a head. We think of the vector X = (Xi, X2,..., Xp) as describing the result of this composite experiment. The total number of heads is the sum of the entries in X. e An individual random variable X has a distribution function Fy defined by Fy (x) = P(X < x) for.x € R. The corresponding ‘joint’ distribution function of a random vector (X1, Xo,..-, Xn) is the quantity P(X, < x1, Xz < x2,...,Xn [0, 1] given by Fx(x) = P(X < x) forx eR", As before, the expression {X_< x} is an abbreviation for the event {o € @ : X(w) < x}. Joint distribution functions have properties similar to those of ordinary distribution functions. For example, Lemma (2.1.6) becomes the following. (3) Lemma. The joint distribution function Fy of the random vector (X, Y) has the follow- ing properties: (@) limy,y co Fx.v(x, y) = 0, limy,y-+00 Fx, (ty) = 1, (b) if (41, 91) S (x2, Y2) then Fy.y (x1, 1) < Fx.y (x2, 92), (©) x,y is continuous from above, in that Fry(xtuyy+v) > Fxy@y) as u,v} 0. We state this lemma for a random vector with only two components X and Y, but the corresponding result for n components is valid also. The proof of the lemma is left as an exercise. Rather more is true. It may be seen without great difficulty that (6) yim, Fxy(, y) = Fx(x) (= P(X = x) and similarly a dim, Fx.v@, 9) = FrQ) (= PY sy). This more refined version of part (a) of the lemma tells us that we may recapture the individual distribution functions of X and Y from a knowledge of their joint distribution function. The converse is false: it is not generally possible to calculate Fx,y from a knowledge of Fy and Fy alone. The functions Fy and Fy are called the ‘marginal’ distribution functions of Fx,y (8) Example. A schoolteacher asks each member of his or her class to flip a fair coin twice and to record the outcomes. The diligent pupil D does this and records a pair (Xp, Yp) of outcomes. The lazy pupil L flips the coin only once and writes down the result twice, recording thusa pair (Xz, Yi) where X, = Yz. Clearly Xp, Yp, X, and Y,, are random variables with the same distribution functions, However, the pairs (Xp, ¥p) and (Xz, Yz.) have different joins distibution fonctions. In particular, POXn = Yo = heads) = sinee only one ofthe four possible pairs of outeomes contains heads only, whereas P(X, = Y; = heads) = }. @ Once again there are two classes of random vectors which are particularly interesting: the ‘discrete’ and the ‘continuous’. (9) Definition. The random variables X and ¥ on the probability space (@, ¥ , P) are called (jointly) discrete if the vector (X, Y) takes values in some countable subset of R? only. The jointly discrete random variables X, Y have joint (probability) mass function f : R? — 10, 1) given by f(x, y) = P(X =x, Y=y).

You might also like