GIFT OP

JULIUS WAHOENHEIM 7
 TREATISE
ON

PROBABILITY:
FORMING

THE ARTICLE UNDER THAT HEAD IN THE SEVENTH EDITION OF

THE ENCYCLOPEDIA BRITANNICA.

BY

THOMAS GALLOWAY, M.A. F.R.S.

SECRETARY OF THE ROYAL ASTRONOMICAL SOCIETY.

EDINBURGH:
ADAM AND CHARLES BLACK, NORTH BRIDGE,
BOOKSELLERS TO HER MAJESTY FOR SCOTLAND.

1839.
 EDINBURGH:
BALFOIR JACK, Printers, Niddry Street.
 CONTENTS.

INTRODUCTION Page 1.

SECTION I.

GENERAL PRINCIPLES OF THE THEORY OF PROBABILITY. P. 16.

ARTICLES. 1-4. Definition of the terms Chance and Proba

bility. 5, 6. Measure of mathematical probability. 7. Probability

of compound events. 8. Example in numbers. 9. Probability of
an event which may happen in different ways. 10. Identity of the
formulae for simultaneous and successive events.

SECTION II.

OF THE PROBABILITY OF EVENTS DEPENDING ON A REPETITION OF

TRIALS, OR COMPOUNDED OF ANY NUMBER OF SIMPLE EVENTS, THE
CHANCES IN RESPECT OF WHICH ARE KNOWN A PRIORI AND CON
STANT P. 27.

ARTICLES. 11, 12, Probabilities of the different combinations that

may happen in a series of trials. 13. Application of the binomial

theorem. 14.Example in a particular case. 15. Extension to the
case of any number of simple events. 1621. Examples. 22, 23.
Artifice for abbreviating the calculation. 24, 25. Determination of

G00735
 IV CONTENTS.

the number of trials required to render the probability of an event

equal to an assigned fraction.

SECTION III.

OF THE PHOB ABILITY OF EVENTS DEPENDING ON A REPETITION OF

TRIALS, OR COMPOUNDED OF ANY NUMBER OF SIMPLE EVENTS, THE

CHANCES IN RESPECT OF WHICH ARE KNOWN A PRIORI, AND VARY

IN THE DIFFERENT TRIALS P. 51.

ARTICLES. 26. Expression for the probability of the different pos

sible results in particular cases. 27, 28. Extension to the general

case. 29. Examples. 30. Solution of a question proposed by
Huygens.

SECTION IV.

OF MATHEMATICAL AND MORAL EXPECTATION P. 59.

ARTICLES. 31. Definition of the term

expectation. 32. Example
of mathematical expectation. 33. Mathematical expectation not
ap
plicable in particular cases. 34. Hypothesis of Bernoulli. 35, 36.
Formula to express the value of a moral
3739. Con
expectation.
sequences of the formulae. 40. Application of the theory of Ber
noulli to the subject of insurances. 41.
Petersburg problem.

SECTION V.

OF THE PROBABILITY OF FUTURE

EVENTS DEDUCED FROM EXPERI-
ENCE
P. 75.

ARTICLES. 42. Hypotheses

respecting the causes of an event. 43.
ermination of the probabilities of the
different hypotheses. 44.
bility of a future event deduced from the
probabilities of the
>eses. 45. Extension of the formula to
finite
any number of
causes or hypotheses. 46. Sense in which the term cause is used in
 CONTENTS.

this theory. 47, 48. Examples of the formulae. 49. Case in which

the different causes are not equally probable. 50. Application of

the formulae to physical or moral events. 51, 52. Extension of the
formulae to the case of an infinite number of different causes.

SECTION VI.

OF BENEFITS DEPENDING ON THE PROBABLE DURATION OF HUMAN

LIFE P. 90.

ARTICLES. 53. Principles on which the probability of life is com

puted. 54, 55. Method of computing the value of an annuity on a

single life. 56. Of a life annuity for terms of years. 57. Of an

nuities on joint lives. 58. On the survivor of any number of lires.

59 63. Methods of computing the values of assurances on lives.

SECTION VII.

OF THE APPLICATION OF THE THEORY OF PROBABILITY TO TESTIMONY,

AND TO THE DECISIONS OF JURIES AND TRIBUNALS P. 100.

ARTICLES.64, 65. Expression for the probability of an event at

by a single witness on an assumed hypothesis. 66. Case in

tested

which the event attested is extremely improbable. 67. Case in

which the character of the witness is altogether unknown. 68. Ex.

pression of the probability of an attested event, regard being had
to the a priori probability of the event. 69, 70. Probability of
events attested by several witnesses. 71. Formulae for the case of

conflicting testimony. 72 75. Successive testimony, or tradition.

76. Application to the verdicts of juries. 77 81. Probability of

acquittal and condemnation under different hypotheses. 82, 83.

Probability of a verdict being correct when pronounced by a giren

majority. 84. Numerical expression for the error of a verdict when

arbitrary values are given to the constants. 85,86. Values of the con

stants, deduced from the records of the criminal courts in France.

VI CONTENTS.

SECTION VIII.

OF THE SOLUTIONS OF QUESTIONS INVOLVING LARGE NUMBERS. P. 128.

Form of the of
ARTICLES 87-89. expression for the probability
the most probable event in a large number of trials. 90. Stirling s

theorem for approximating to its numerical value. 91. Example of

a large
the formula. 92. Probability of any assigned result of
series of trials. 93. Example of the rapid diminution of the terms

of the binomial towards the beginning arid end of the series when
the exponent is large. 94. Probability of the number of occur

rences of an event falling within given limits. 95. Determination

of the limits for an assigned probability. 96. Method of computing

the value of the integral which expresses the probability. 97. Ge
neral consequences of the formulas. 98. Method of using the table.
99. Application to numerical examples. 100 102. Extension to
the case of probabilities deduced from experience. 103. Applica
tion to a question of mortality. 104. To the determination of a

physical cause. 105. To determine the limits of loss or gain in a

multitude of pecuniary risks. 106. General deductions relative to
the ultimate results of play. 108, 109. Digression relative to a
question respecting the duration of play.

SECTION IX.

OF THE MOST PROBABLE MEAN RESULTS OF NUMEROUS DISCORDANT

OBSERVATIONS, AND THE LIMITS OF PROBABLE ERROR P. 164.

ARTICLES. 110. Object of the

application of the theory. 111.
Statement of the problem. 112123. Solution of the problem,
and expression for the probability of the of a
average large number
of observations falling within 124.
assigned limits. Connection of
the special constant to which the
average continually approaches,
with the centre of
gravity of a curve line. 125126. Cases in
which the number of possible values of the
phenomenon is limited.
 CONTENTS. Vll

127130. Examples when the law of the probability is supposed to

be known. 131 132. Form of the function which expresses the
law of probability, on the hypotheses of positive and negative errors

being equally probable. 133. Illustration by means of a curve.

134. Determination of the constants on this hypothesis. 135142.
Investigation of an expression for the constant of the law of facility,
in terms of the results of observations already made. 143. Differ

ent expressions for the limits. 144. Use of the formulae in deter

mining the relative values of tables. 145 148. Expressions for the

precision, weight, probable error, and mean error of the results of ob

servations. 149. Determination of the most probable value of a

function of several quantities, separately determined by observation,

with the precision and weight of the determination.

SECTION X.

OF THE METHOD OF LEAST SQUARES P. 200.

ARTICLES. 151-154. Method of forming equations of condition.

155-157. Determination of the most probable values of the ele

ments from the equations of condition, and the probable amount of

the remaining errors. 158. Agreement of this solution with the

method of least squares. 159. Example in the case of a single ele

ment. 160. Example in the case of sereral elements.

 substitute the following List of
The Reader requested to
is

Errata that at the end of the volume.

for

ERRATA.
40, note, for Essai sur les Probabilites, read Theorie Analytique
Page
des Probabilitcs.
read adapt.
43, line 7 from foot, for adopt
with. In the note, read Traite Ele-
49, line 13, read beginning
mentaire.

54, line 7, for a TO, read a m .

81, first line, for (1+2+3.. . + ), read (1+2+3.. .+n).

89, line 9, for TO m, read m+m .

135, last line, for read (+*)-"-*- .

(n+ar)-+^-J,
_ 188 line 8, for e~\ read -**.

193, line 8, delete products of the.

198, line 9, for e e", read e e".
,

203, line 19, for kc t

read kc.
line 3 from foot, for (1), read (2).

209, last line, for les read des.

210, line 2 from foot, for 1833, 1834, and 1835, read 1834,
1835, and 1836.
211, last line, for Abhandhmgen, read Nachrichten.
 PREFACE.

IN drawing up the following treatise for the Ency

clopaedia Britannica, my design has been, to pre

sent a general view of the principles, applications,
and more important results of the mathematical
theory of Probability, as laid down in the best
and most recent works on the subject particu
;

larly those of Laplace and Poisson and, without

:

entering into the details of mathematical difficul

ties, to explain the methods of applying analysis to

the solution of the principal questions, as fully as

the limits of the space which could be appropriated
to the article would permit.
In the prosecution of this design, questions con
nected with lotteries and games of chance, the sub

jects to which the earlier writers on Probability

chiefly confined themselves, and which are frequent

ly supposed to form a principal part, if not the whole
of the science, occupy but a small portion of the work ;
indeed they are only introduced as furnishing exam-
 PREFACE.

pies to illustrate the principles or the formulae. The

theory of moral expectation has been developed
at some length, because it admits of important

applications in the affairs of life. The subject of

mortality, and the computation of annuities and
other pecuniary interests depending on life contin

gencies, (the most common application of the

theory,) having been treated in detail under other

heads in the Encyclopedia, scarcely came within the

scope of the article ; but the method of applying the

principles of the science to the data furnished by the
mortality tables, has been explained, and the formulae
of most frequent practical use have been deduced.
The section on the probability of testimony and the
verdicts of juries, is more expanded, and contains
formulae which resolve the principal questions the

subject presents, having regard to the different

causes of uncertainty. In the following section,
formulae are given for obtaining approximate values
of expressions which involve very large numbers.
Such expressions occur in all the higher and
more interesting applications of the doctrine of
chances, and give rise to the greater
part of the
mathematical which attend the subject.
difficulties

The probability of the result of a large number

of experiments or observations being contained
within given limits, is
expressed by a definite inte
gral,which occurs in various other applications of
in order to facilitate the
analysis ;
computation of
such probabilities, a table of the values of this inte-
 PREFACE. XI

gral is
given at the end, by which, and by a simple
arithmetical operation, we are enabled to solve a mul
titude of questions of the most interesting and other
wise In the investigation of the most
difficult kind.

probable mean value of a quantity, to be deter

mined in magnitude or position, from a series of ob
servations liable to error, and the determination of
the limits of probable uncertainty, I have followed the

very general and elegant analysis of Poisson. The last

section contains the application of the general theory

to the determination of the most advantageous

method of combining equations of condition, and an
explanation of the celebrated rule of least squares,
which is shewn to coincide with that
pointed out by
the theory as giving the most probable values of the
elements involved.
The principal application which has hitherto been
made of the theory explained in the last two sec
tions,has been to questions of astronomy ; but
itmay be applied, with equal advantage, to every
branch of physical inquiry, or statistical research,
where the object is to deduce precise and accu
rate mean results from multiplied observations, or
to detect the existence of general laws, or the opera
tion of constant causes, among
large masses of phe
nomena.
Although considerable pains have been taken to
treat the subject in an elementary manner, it was im

possible, in some places, particularly the three last

sections, to avoid discussions which belong to the
 PREFACE.

more abstruse parts of the mathematics. The

higher and more useful parts of the theory, do
not admit of being treated in any other way,
unless the reader is prepared to abandon the ac
curacy of exact science, and to accept of asser
tion for proof. have endeavoured, however, to
I

place the demonstrations in their clearest light, and

to give the expressions in their simplest forms ; and
it is nothing in the present work be
hoped there is

yond the reach of those who have a moderate know

ledge of the differential and integral calculus, and
are acquainted with the first principles of the theory
of finite differences.
It may be proper, perhaps, to give a reason for de
viating from the usual notation, by writing fractional
expressions in the form of ratios or quotients. This

practice was adopted for the purpose of economising

space in printing, when it was perceived that the
article was likely to exceed the prescribed limits.
It may be considered objectionable on the ground of
its being less familiar to the reader ; but as it is here

used it can hardly occasion any real embarrassment.

LONDON,
February 12, 1639.
 PROBABILITY.

THE is an extensive and very im

doctrine of probability

portant branch of mathematical science, the object of which

is to reduce to calculation the reasons which we have for

believing or expecting any contingent event, or for assent

ing to any conclusion which is not necessarily true. When
it is considered that the whole edifice of human science, with
the exception of a few self-evident truths, such as the axioms
of geometry, is nothing more than an assemblage of propo

sitionswhich can only be pronounced to be more or less pro

bable, the importance of a calculus which enables us to ap

preciate exactly the degree of probability existing in each

case, willbe readily understood.
Our reasons for judging an event to be probable or im
probable, are derived from two distinct sources ; first, an

a priori knowledge of the causes or circumstances which

determine its occurrence ; and, secondly, when the causes
are unknown, experience of what has already happened in
the same circumstances, or in circumstances apparently si

milar. Suppose, for example, a hundred white balls to be

 ? i

placed in an urn along with fifty black balls, and that a per

son, blindfold, proceeds to draw a ball, there is to us, who

are acquainted with the contents of the urn, a determinate

probability that the ball which is drawn will be white. The

balls being supposed to be all in precisely the same circum
stances with respect to facility of drawing, we assume that

there is the same chance of drawing any one ball as of draw

ing any other ; and, consequently, since there are two white
balls for each black ball, and therefore two chances of draw

ing one of the first colour for each chance of drawing one
of the second, we conclude the event which consists in the

drawing of a white ball to be twice as probable as the op

posite event, or the drawing of a black ball. In this case
our knowledge of the contents of the urn enables us to

judge of the probable result of the drawing. Suppose, how

ever, that antecedently to the drawing, we were entirely ig
norant of the contents of the urn, but that after a great
number of trials have been made, (the ball drawn being al

ways replaced in the urn after each trial, in order that the
circumstances may be the same in all the trials) it has been
observed that a white ball has been drawn twice as often
as a black ball, we presume that the urn contains twice as

many white as black balls, and consequently affords twice

as many chances of drawing a white ball as of drawing a

black ; presumption becomes stronger in propor
and this

tion to the number of instances included in the observation.

In this case experience makes up for the want of a priori

knowledge, and affords a measure of the probability of the
result of a future trial.

It is only in a comparatively small number of cases that

all the possible ways in which an event may happen are

known a priori, and in which, consequently, the ratio of the

 HISTORY OF THE SCIENCE. 3

number of chances favouring the event to the whole num

ber of existing chances is determinate. In fact, most of the

questions of this class to which the calculus can be applied,

are connected with lotteries and games of hazard. The re

sults obtained from the analysis of such questions cannot be

considered as being of any great value in themselves, but

they frequently throw light on subjects of far higher impor

tance which present analogous combinations. It is true that

the mathematical theory comes in aid of moral considerations,

and demonstrates the ruinous tendency of gambling even
when the conditions of the play are equal, mathematically

speaking ; but, unfortunately, those who indulge a passion

seldom capable of appreciating the force of
for this vice are

such arguments. The principal advantage which has resulted

from the application of analysis to games of chance is the ex
tension and improvement of the calculus to which it has led.

The calculation of the probabilities of events, the chances

of which are not known a priori, but inferred from experi

ence, is founded on the presumed constancy of the laws of

nature, in obedience to which events depending on constant

though unknown causes, are always reproduced in the same

orderwhen considered in large numbers. Among the vari
ous phenomena of the physical and moral world, nothing is
more remarkable than the constancy which is observed to
prevail in the recurrence of events of the same kind. The
ratio of male to female births furnishes a noted instance.
If we consider only a small number of births, nothing can be
more uncertain than the result but taking a very large num
;

ber, as those of a whole kingdom in the course of a year,

the proportion of males to females is found to be almost inva

riable, and nearly as 21 to 20. The mean duration of hu

man life affords another familiar example. Notwithstand-

4 PROBABILITY.

of con
ing the proverbial uncertainty of life, the differences
stitutions, and the various accidents to which mankind are

exposed, the average duration of the lives of a large

num
ber of individuals living in the same country is always found
tobe very nearly the same, insomuch that pecuniary risks
depending on it, if undertaken in sufficiently large numbers,
are among the least uncertain of all commercial speculations.
A similar constancy is remarked in the results of statistical in
quiries of every kind. The number
of crimes of the same spe

cies committed in a year, the ratio of the number of acquit

tals to the number of trials, the number of conflagrations, of

ships lost in a particular trade, of letters which pass through the

post-office, of patients admitted into the public hospitals ;

in ev

ery case the numbers in a given time are observed to fluctuate

between very narrow limits, and to approach nearer and near
er, as the observation more extended, to fixed mean values.
is

This constant approximation to fixed ratios, which is

proved by all experience, in the recurrence of events of the

same kind, enables us to apply the calculus of probabilities
to many of the most interesting questions connected with our
social and political institutions ; and to determine the aver

age result of a series of coming events with as much precision

as if their chances were determinate, and known a priori, like

that of obtaining a given point with the throw of a die. What

ever be the nature of the phenomenon under consideration,
whether it
belong to the physical or moral order of things,
the calculus is
equally applicable when the requisite data
have been determined from experience.
The foundations of the mathematical theory of probabili
ties were laid by Pascal and Fermat about the middle of

the 17th century. Among some other questions relating

to chances, the following was proposed to Pascal. " Two
 HISTORY OF THE SCIENCE. 5

persons sit down to play on the condition that the one who
first gains three games shall be the winner of the stakes.

The having gained two games, and the second one,

first

they agree to leave off and divide the stakes in proportion

to their respective probabilities of winning what share is :

each entitled to take ?" Pascal solved the question, but by

a method which was applicable only to the particular case.

Fermat, to whom it was communicated by Pascal, employ

ed the direct and general method of combinations, and gave
a solution which could be applied to the case of any num
ber of players. His reasoning, however, did not at first

appear to Pascal to be satisfactory, and a correspondence

on the subject took place between these two illustrious

geometers, which is
preserved in their respective works, and
throws some light on the history of mathematics in that age. 1

About the same period Huygens composed his tract De

Ratiociniis in Ludis Ale<z, which was first
published in the
Exercitationes Geometricce of Schooten in 1658, This was
the first systematic treatise which appeared on the doctrine
of chances. It contained an analysis of the various ques
tions which had been solved by Pascal and Fermat, and
at the end five new questions were proposed, the solutions
of which, simple as they may now appear, were then attend
ed with considerable difficulty. The analysis of two of them
was in fact given for the first time by Montmort, half a cen

tury after their publication. Huygen s tract was translated

into English and published in 1 692, with some additional re
marks relative to the advantage of the banker in the game
of Pharaon, in an Essay on the Laws of Chance, edited and

1
OBuvres de JBlaise Pascal, tome iv. Paris, 1819; Opera Petride
Fermat, Toloste, 1679.
 b PROBABILITY.

of
supposed to have been written by Motte, then Secretary
the Royal Society.
James Bernoulli appears to have been the first who per
ceived that the theory of probability may be applied to much
more important purposes than to regulate the stakes and ex
pectations of gamesters, and that the phenomena,
both of the
moral and physical world, anomalous and irregular as they

appear when viewed in detail, exhibit, when considered

in

a
large numbers, constancy of succession which renders their

occurrence capable of being submitted to numerical esti

mation. The Ars Conjectandi, published in 1713, seven

years after the death of the author, contains

a number of in
teresting questions relative to combinations
and infinite

series ; but the most remarkable result which it contains is a

theorem respecting the indefinite repetition of events, which

may be said to form the basis of all the higher applications

of the theory. It consists in this, that if a series of trials be

instituted respecting an event which must either happen or
fail in each trial, the probability becomes greater and greater,
as the number of trials is increased, that the ratio of the
number of times it
happens, to the whole number of trials,
will be equal to a priori probability in a single trial ;
its

and that the number of trials may be made so great as to

give a probability, approaching as nearly to certainty as we
please, that the difference between the
ratio of its occurren

ces to the number of trials, and the fraction which measures

its a priori probability, will be-less than any assigned quan
tity. Bernoulli informs us, that the solution of this impor
tant theorem had engaged his attention during a period of
twenty years.
In the interval between the death of Bernoulli and the

appearance of the Ars Conjectandi, Montmort published

 HISTORY OF THE SCIENCE. 7

his Essai cT Analyse sur les Jcux de Hazard. The first

edition was in 1708 ; the second, which is

considerably ex
tended, and enriched by several letters of John and Nicolas
Bernoulli, appeared in 1713. The work possesses consi
derable merit ; but being chiefly confined to the examina
tion of the conditions of games of chance, many of whicli
are now forgotten, it has lost much of its
original interest.
About the same Demoivre began to turn his atten
time,
tion to the subject of probability, and his labours, which
were continued during a long life, contributed greatly to the
advancement of the general theory, as well as the extension
of some of its most interesting applications. Demoivre s

first publication on the subject was a Latin memoir De Men-

sura /Sortis, in the Transactions of the Royal Society for

1711. His Essay on the Doctrine of Chances first

appeared
in 1716 ; a second edition in 1738 ; but the third and most

valuable, including also his Treatise on Annuities on Lives,

is dated 1756. This work contains a great variety of ques
tions relating to chances, solved with much clearness and

elegance ; but it is chiefly remarkable for the theory of re

curring series, there given for the first time, which is of im

portant use in investigations of this kind, and is in fact

equivalent to the methods employed in the modern calculus

for the integration of equations of finite differences having

constant co-efficients. Of the particular results obtained by
Demoivre, one of the most important in reference to theory,
is an extension of the theorem of James Bernoulli, above

mentioned. It follows from Bernoulli s theorem, that if we

have a given probability that the ratio of the number of
occurrences of an event to the whole number of trials, will

approach to the a priori probability of the event within

certain given limits, those limits will become narrower and
8 PROBABILITY.

narrower, as the number of trials is

multiplied ; but in order
to complete the theorem, it is
necessary to assign the nu
merical value of the probability that in a large number of
future trials, the number of occurrences will fall within as

signed limits. For this purpose we must find the product

of the natural numbers 1, 2, 3, 4, &c., up to the number of

trials ; an operation which, if attempted by direct multipli

cation, becomes very laborious, even when the number of

trials is inconsiderable, and when the number is
great, as
10,000 for example, altogether beyond the reach of human
is

industry. A formula was however discovered by Stirling.

by means of which an approximate value of the product is
found by the summation of a few of the first terms of a series
which converges the more rapidly as the number of trials is
greater. With the aid of this formula, Demoivre was en
abled to assign the probability in question, and thus give a

practical value to the theorem of Bernoulli.

The objects and important applications of the theory of

probabilities having been made known by the works now

mentioned, the subject has ever since been regarded as one
of the most curious and interesting branches of mathemati
cal speculation, and accordingly has received more or less

attention from almost every mathematician of eminence. A

great variety of questions connected with and especially
it

relating to lotteries, are interspersed in the volumes of the

Paris and Berlin Memoirs, (particularly the latter,) by John
and Nicolas Bernoulli, Euler, Lambert, Beguelin, and others.
D Alembert has likewise treated of the theory in several of
the volumes of his Opuscula; and it is not a little remarkable,
that in some instances its first
principles should have been
misunderstood by so ingenious and profound a writer. In
the St. Petersburg Memoirs, (vol. an interesting
v.) there is
 HISTORY OF THE SCIENCE. 9

paper by Daniel Bernoulli on the relative values of the ex

pectations of individuals who engage in play, or stake sums
on contingent benefits, when regard is had to the difference
of their fortunes ; a consideration which, in many cases, it

is
necessary to take into account ; for it is obvious, that the
value of a sum of money to an individual, depends not mere
ly on its absolute amount, but also on his previous wealth.
On this principle Bernoulli has founded a theory of moral

expectation, which admits of numerous and important appli

cations to the ordinary affairs of life. The Transactions of
the Royal Society for the years 1763 and 1764, contain two
papers by the Rev. Mr. Bayes, with additions to the latter
by Dr. Price, which deserve to be noticed, inasmuch as the
principles on which the probability of an event is determin
ed, when the event depends on causes of which the exis
tence and influence are only presumed from experience,
are there for the first time correctly laid down. The ques
tion proposedand solved by Bayes was this a series of ex :

periments having been made relative to an event, to deter

mine the presumption there is, that the fraction which mea
sures its
probability falls within given limits.
One of the earliest applications of the theory of probabi

lity was to determine, from observations of mortality, the

average duration of human life, and the value of pecuniary

interests depending on its continuance or failure. This

particular application appears to have been

thought of, first

be carried into practical effect, in

or at least attempted to

Holland, by Hudde and the celebrated pensionary De Witt ;

but the first tables of mortality, with the corresponding va

lues of annuitieson single lives, were constructed by our

illustriouscountryman Dr. Halley, and published in the
Philosophical Transactions for 1693. For the history of
10 PROBABILITY.

this branch of the subject, we refer to the two articles, AN

NUITIES and MORTALITY in the seventh edition of the En
cyclopaedia Britannica. We may remark, however, that

although the English writers, who have expressly treated

of it, have almost without exception confined themselves to
the explanation of the methods of computing annuity tables,
and of determining from them the values of sums depend
ing on life contingencies, the aid which this branch of eco

nomy derives from the general theory of probabilities, is by

no means confined to the consideration of such elementary

questions. The number of observations necessary to in

spire confidence in the tables, the extent to which

risks

may be safely undertaken, the comparative weights of dif

ferent sets of observations, and the probable limits of de

parture from the average results of previous observations in a

given number of future instances, are all
questions of the ut
most importance, which come within the scope of the calculus,
and cannot, in fact, be justly appreciated by any other means.
The application of the theory of probability to the subject
of jurisprudence, and the verdicts of juries and decisions of

tribunals, has been discussed by the Marquis Condorcet in va

rious articles in the Encyclopedic Methodique ; but more
Essai sur F application de fAnalyse a la
especially in his
Probabilite des Decisions renducs a la Pluralite des Voix,
Paris 1785; a work of great ingenuity, and abounding with
interesting remarks on subjects of the highest importance to

humanity. James Bernoulli, it

appears, had intended to
treat jurisprudence as a branch of probability in the Ars
Conjectandi, but his premature death prevented that work
from being completed. There is a memoir on the subject

by his nephew Nicolas, in the Leipsic Acts for 171 1. The

most important questions to be determined, are the number
 HISTORY OF THE SCIENCE. 1 1

of jurors of which a jury ought to consist, and the majority

which should be required to agree in a verdict in order to
afford, on the one hand, the greatest probability that an ac

cused person will not be wrongly condemned ; and, on the

other, to give to society the greatest security that its inter
ests will not be compromised, by allowing too great facili
ties for the guilty to escape. This important subject has
been treated more profoundly, and with numerical elements
derived from much better data than existed in the time of

Condorcet, in a recent work by Poisson, to w hich

r
we shall

presently allude.
Another of the moral subjects to which the theory of pro
bability has been applied, and connected with the preceding,
is the appreciation of the evidence of testimony. In mat
ters of this kind, it is
easy to see that the calculus must be
founded almost entirely upon hypothetical data. The vera

city of a witness can scarcely be made the subject of direct

experiment ; and by reason of the complicated circumstan

ces with which the facts forming the subject of testimony
are usually accompanied, and the numberless ways in which
mankind are influenced by their passions, credulity, or ig

norance, it is
perhaps equally impossible to deduce an aver
age value from the comparison of a great number of state
ments which have been ascertained to be true or false. Nu
merical results can therefore only be obtained by having re

course to hypotheses, and consequently must be considered

as only probable approximations. The knowledge, however,
which is thus obtained of the various combinations of the

quantities concerned, affords important aid in guiding our

judgments in complicated cases, and when we have to de

cide upon conflicting testimony. Approximations deduced
from a train of accurate and systematic reasoning, are al -
12 PROBABILITY.

ways to be preferred to the most specious arguments drawn

from any other source.
The analysis of probability has been applied with signal
advantage in many researches of Natural Philosophy, but

especially in appreciating the mean errors of observations.

Owing to the imperfections of sense and of instruments, phy

sicalmagnitudes are only susceptible of being measured with
in certain limits of accuracy and where the last degree of
;

precision is indispensable, as in practical astronomy, it is on

lyby means of a very great number of measures, compared

with one another, and combined according to the methods
which this calculus points out, that we can obtain the near
est approximation to the true values which the observations
are capable of giving. The mean errors of observations were
treated as questions of probability by Lagrange in the Turin
Memoirs for 1773 ; but it is to Laplace that the theory owes
its The me
principal extension and most important results.
thod of combining numerous equations of condition now

universally followed, known as the method of minimum

squares, and which Laplace has demonstrated to be that

which leaves the least probable amount of error in the final

equations, was made known by Legendre in an Appendix to

his Nouvelles Mcthodes pour la Determination des Orbits

des Comctes, published in 1806. A similar method, however,

or rather the same, (for they are identical in principle,) had
been discovered by Gauss, and employed by him for several
years before the work of Legendre made its appearance.
1

Laplace s great work, the Theorie Analytique des Proba-

bilites, first published in 1812, is one of the most remarkable
productions that has ever appeared in abstract science. The

Theoria Motus Corpcrum Ccelestium) p. 221.

1
 HISTORY OF THE SCIENCE. 13

principles of the calculus, as well as the peculiar methods

of analysis which it
requires, and the most interesting and
difficult questions which it
presents, are here discussed in a
far more general manner than had been attempted by any
former writer on the subject and it may be said, accord
;

ingly, to have placed the theory under an entirely new aspect.

It is much to be regretted that so little pains have been taken
by the illustrious author to render the work intelligible to the

generality of mathematical readers. Consisting for the greater

part of separate memoirs presented at different times to the

Academy of Sciences, arranged without regard to symmetry

or order, it abounds with repetitions which only serve to em
barrass the student; while the deficiency of explanation com
bined with the subtlety of the analysis, and the inherent intri

cacy of the subject, render it often a painfully difficult task to

seize the force of the demonstrations. Notwithstanding these

defects, however, it forms one of the most splendid creations

of mathematical genius ; and is alike admirable, whether we

regard the extension which has been given to the calculus,

or the results which have been arrived at, or the tone of

lofty philosophy in which subjects bearing on some of the

most important concerns of mankind are treated.
Next to the Theorie Analytique of Laplace, the most im
portant work which has hitherto appeared on the subject of
probability is the recent one of Poisson, entitled, Recherches
sur la Probabilite des Jugements, (Paris 1
837.) Although it
might be inferred from the title that this work relates only to
a single though very interesting application of the theory, the

greater part of it is devoted to the development and demon

stration of the general principles, and the discussion of the
principal questions which present themselves in the differ
ent applications and it is only in the last of the five books
;
1 4 PROBABILITY.

of which it consists that the special subject to which the title

refers is taken into consideration. In applying the theory

to the decisions of tribunals, Condorcet and Laplace had been
unable to obtain positive results from the want of authentic
data ; but the recent publication by the French government
of the Comptes Generaux de VAdministration de la Justice

Crimmellc, in France, having furnished an immense collec

tion of facts from which the requisite data could be obtained,
Poisson was led to consider the subject anew, and the results

of his investigations, which are of singular interest, are given

in the work now mentioned. Poisson had already given a

theory of the mean errors of observations in the Additions

to the Connaissance des Terns for 1827 and 1832.

It is in these two works of Laplace and Poisson that the

higher and more abstruse parts of the theory of probabilities

must be studied. A
very clear exposition of the principles,
accompanied with many interesting remarks on the uses and

applications of the theory,

is
given by Lacroix in his valu
able little work, Traite Elementaire du Calcul dcs Proba-
bilites, Paris 1822.

Since the time of Demoivre, the English treatises on the

general theory of probability have neither been numerous,

nor, with one or two exceptions, very important. Simpson s
Laws of Chance (1740) contains a considerable number of
examples, in the solution of which the author displays his

usual acuteness and originality, but as they belong entirely

to that class in which the chances are known a priori, they

give no idea of the most interesting applications of the theory.

Dodson s Mathematical Repository contains a large selection
of the same kind. The Essay in the Library of Useful
Knowledge, by Mr. Lubbock, gives a more comprehensive
and philosophical, though an elementary view of the sub-
 HISTORY OF THE SCIENCE. 15

ject ; but by far the most valuable work in the language is

the Treatise in the Encyclopedia Metropolitana, by Pro

De Morgan, 1837- In this very able production, Mr.
fessor

De Morgan has treated the subject in its utmost generality,

and embodied, within a moderate compass, the substance of
the great work of Laplace.
Within the limits to which the present article must be
confined, would be hopeless to attempt giving a complete
it

view of a branch of science which embraces so many com

plicated and intricate subjects of research, and which re

quires the aid of some of the most abstruse and recondite

theories of the modern mathematics.
In the higher appli
cations of the theory, the analysis of many of the questions

which arise, in order to be made intelligible, would require

an extent of development and a parade of mathematical for

mulae altogether incompatible with the plan and scope of

this work. All that we can propose to ourselves, therefore,

is to explain as briefly as may appear consistent with perspi

cuity, the general principles of the theory, and to give an
outline of the manner
which these are applied to some cf
in

the more important questions which have been investigated

by Laplace and Poisson. The examples will be selected

with a view to shew the nature of the principal results of the
mathematical theory, as well as the peculiar methods of ana

lysiswhich are of most general application.

16 PROBABILITY.

SECTION I.

GENERAL PRINCIPLES OF THE THEORY OF PROBABILITY.

1. The term probable., in its popular acceptation, is used

in reference to any unknown or future event, to denote that

in our judgment the event is more likely to be true than

not, or more likely to happen than not to happen. With
out attempting to make an accurate enumeration of the va
rious circumstances which are favourable or unfavourable
to its occurrence, or to balance their respective influences,

we suppose there is a preponderance on one side, and ac

cordingly pronounce it to be probable that the event lias

occurred, or will occur, or the contrary.

2. If we can see no reason why an event is more likely
to happen than not to happen, we say it is a chance whe
ther the event will happen or not or if it may happen in ;

more ways than one, and we have no reason for suppos

ing it will happen in any one of these ways rather than in

another, we say it is a chance whether it will happen in

any
way or in any other. Suppose, for
assigned example, an
unknown number of balls of different colours to be placed

in an urn, from which a ball isabout to be extracted by a

person blindfold. Here we have no reason for supposing

that the ball about to be drawn will be of one colour rather
than another, that it will be white rather than black,
 GENERAL PRINCIPLES. 17

or red ; and accordingly say it is a chance whether the ball

will come out of a particular colour, or a different. In this

instance, then, the term chance denotes, simply, the absence

of a known cause. If, however, we are made acquainted
with the number of balls in the urn, and the number there
are of each of the different colours, the term is used in a

definite sense. For instance, suppose the urn to contain

ten balls, of which nine are white, and the remaining one

black, we say there are nine chances in favour of drawing a

white ball, and one chance only in favour of drawing the
black ball. Chance, in this sense, denotes a way of hap
pening, or a particular case or combination that may arise
out of a number of other possible cases or combinations ; and
an event becomes probable or improbable according as the
number of chances in its favour is greater or less than the
number against it. Chance and presumption are also fre

quently used synonymously with probability.

3. The mathematical probability of any event is the ratio
of the number of ways in which that event may happen
to the whole number of ways in which it
may either hap
pen or Thus, recurring to the previous example, the
fail.

event, namely, the drawing of a ball from an urn con

taining 9 white balls 1 black, and

may happen in 10 dif
ferent ways, inasmuch asany one of the 10 balls may be
drawn but in one only of those ways will the event be a
;

black ball and therefore the probability of drawing the

;

1
black ball is
-y ^.
In like manner, as there are 9 differ

ent ways in which a white ball may be drawn, or 9

chances of drawing a white ball, and ten chances in all, the
9
probability of drawing a white ball at the first trial is T S .

It follows immediately from this definition, that the proba

bility of drawing a ball of either colour will remain the same,

 18 PROBABILITY.

however the number of balls in the urn may be increased,

provided those of each colour are increased in the same pro

portion. For instance, suppose the number of white balls
to be 45, and the number of black balls to be 5 ; the num

ber of chances in favour of drawing a black ball is 5,

while there are 50 chances in all, consequently the pro

bability of a black ball being drawn is
/y=r^ . In the same
9
manner, the probability of drawing a white ball is
|^=r-j g ;

the same as before. Generally, let E and F be two con

trary events, that is to say, such that the one or the other
of them must necessarily happen, and both cannot happen

together ; and let a be the number of chances or combina

tionswhich produce the event E, and b be the number of
combinations which produce the event F, or cause the fail

ure of E ; then the probability that E will happen is

j
j

and the probability that F will happen, or that E will not

happen is . In future, the term probability will be us

ed only to signify mathematical probability.

4. It is to be carefully remarked, that the different

chances or combinations which form the elements of pro

bability are supposed to be perfectly equal. If this equa

lity does not hold, and there is any circumstance respect

ing the event under consideration which renders one com
bination or set of combinations more likely to occur than

another, the different combinations must be multiplied by

numbers proportional to their respective facilities, after

which the units each multiplier may be regarded as so

in

many distinct chances, from which the probability of the

event will be found by the above formula. This is
equiva
lent to saying that a combination or chance which is twice
 GENERAL PRINCIPLES. 19

as likely to happen as another, must be regarded as two

equal and similar combinations in comparison of that other ;

a proposition which is
sufficiently obvious.
5. It follows from the above definition, that the probabi

lity of any contingent eventmeasured by a fraction less

is

than unity, and may have any value between and 1. It

follows, also, that the sum of the two fractions which mea
sure the probabilities of two contrary events is
equal to
unit, which is the measure of certainty, inasmuch as either
the one or the other necessarily occurs. Thus, in the last

example, the probability of the event E is -, and that of

the contrary event F is , and -4 7= 1 . Hence

a +b ab*a+b
if denote the probability of any event E, and q the pro
p
bability of the contrary event F, we have q~ 1 p. This

consequence of the definition is of great importance in the

calculation of probabilities.

6. We have here supposed the result of a trial to be ne

cessarily one or other of two events E and F ; but it is easy
to imagine the trial to be of such a kind that it
may give
rise toany one of a number of events E, F, G, H, &c. each
having a given number of chances in its favour. This case
is
represented by supposing an urn to contain balls of as
many different colours or sorts as there are different events.

Let the urn be conceived to contain a balls of the sort

which produces the event E, b of the sort which produces

F, c of the sort which produces G, and so on ; and let a-f-

b-\-c + d, &c. k, so that k is the whole number of balls in

the urn. The probabilities of the different

events E, F, G,
H, &c. are then, respectively, by the definition,
20 PROBABILITY.

abed
T T T T &c<

the sum of which =1. In fact, if a ball be drawn at all,

it must be of one or other of the different sorts contained in

the urn ; and consequently the sum of all the probabilities

amounts to unit or certainty.

7. When an event is
compounded of two or more simple
events independent of each other, the probability of the

compound event is equal to the product of the probabilities

of the several simple events of which compounded. it is

Let us imagine two urns, A and B, of which A contains a

white balls and b black, and B contains a white and b
black. Make a-\-b= c, and a +b =:c , and let the com
pound event whose probability is to be determined be the

drawing of a white ball from both urns. Now, as each of

the c balls in A may be drawn with any one of the c balls

in B, the whole number of ways in which the balls in A

may be differently combined by pairs with the balls in B,

or the whole number of possible cases is cc But the num .

ber of cases favourable to the compound event is evidently

the number of different ways in which a white ball may be
drawn from A with a white ball from B, and therefore equal
to aa . Hence by the definition (4), the probability that a
ft ft?
white ball will be drawn from both urns is .
Now, if
cc

p denote the probability of drawing a white ball from A,

and p that of drawing a white ball from B, we have by the
a a aa
definition p ,and whence -pp. :

c c cc

In general, let p denote the probability of an event E,

p that of another event E , p" that of a third E", and so

on ; then the probability of the concourse of the events E,
 GENERAL PRINCIPLES. 21

E , E", &c., OY the probability that they will all happen, is

p Xp Xp"> &c. ; that is to say, the probability of an event

compounded of any number of simple and independent

events, is the product of the respective probabilities of
the several simple events.
The probabilities that the several simple events E, E ,

E," &c., will not all happen, or that some of them will hap
pen and others fail, are easily determined in the same man
ner ; it will be sufficient to indicate their several expres
sions. Suppose there are only three simple events, of which
r
the probabilities are respectively p, p and p" and let
, ;

q=l p, q Ip, q"=l p". The product pq q" ex

presses the probability of the compound event which con
sists in E happening and E and E" both failing qp q" is ;

the probability that E will happen, and that E and E" will

both fail ; pp p" is the probability they will all three happen ;

1
pp p" is the probability they will not all three happen, or
that one of them at least will fail ; qq q" is the probability

they will all fail; and qq q" is the probability they will
1

not all three fail, or that one at least of them will happen.

8. As an example of the application of this rule, suppose

itwere required to assign the probability of throwing aces,
at one throw, with two common dice. As a common die
has six symmetrical faces, there are in respect of each die
six ways equally possible, in which the simple event may
happen. The probability therefore of throwing ace with
one die p-=\. In respect of the second die,
is
J, that is,

we have also p =J
hence the probability of the compound
;

event, or that aces will be thrown is j0p J X 5V The = =

probability that aces will not be thrown at any assigned trial

is therefore (5) 1
J
gs
=f ; and the odds against throwing
aces at any given trial are 35 to 1 .
 22 PROBABILITY.

Again, suppose two numbers, each consisting of 7 di

gits, to be taken
at random, (for instance from a table of

logarithms), and let be proposed to assign the proba

it

bility that the substraction of the one from the other will

be performed without its being necessary, in any case,

to increase the upper figure. Here, as each digit may
have any one of the ten values from to 9 both inclusive,

and as each of those values in the upper line may be com

bined with any one of them in the lower line, there are 100
different combinations or equally possible cases for each par

tial Now, if the upper figure be 0, there is

substraction.

only one of those cases favourable to the event, or which

will admit of the substraction being performed, namely,

when the figure below is also 0. If the upper figure be 1,

there are two cases favourable, namely, those in which

the under figure is or 1 . If the upper figure be 2, there
are three favourable cases, namely, when the under figure
is 0, 1, or 2. Proceeding in this way through all the di

gits, the whole number of favourable cases is found to be

Hence, for each partial substraction there are 55 favourable

cases out of 100 possible cases ; therefore (4) the probabi

lity that any one of the figures in the upper line is not less
than the corresponding figure in the under line is
-f^ ; and
/
we have j9=//=// =c.=yg5 for the probability of each
of the seven simple events or partial substractions, whence,

by (7), the probability of the compound event is

=:(-55) =0152243,

which is less than ^\, and greater than ^.

9- When an event may happen in several different ways,
 GENERAL PRINCIPLES. 23

each independent of the others, the probability of the event

isthe sum of all the partial probabilities taken in respect
of each of the different ways.

Suppose there are n different urns A t A 2 A 3 ...... A w , , ,

each containing balls of two colours, white and black, and

let the whole number of balls in each urn respectively, be

1? 2 , 3 , ...... n,

and the number of white balls in each be

and let the event E be the extraction of a white ball in

drawing a ball from any urn at random. In this case there

are n different ways, all equally probable, in which the event

may happen, for it may be drawn with equal facility from

any one of the urns. The probability that the ball will be

drawn from any given urn, A l , is therefore ; and if it

be drawn from this urn, the probability of its

being white is

; therefore, by (7), the probability of a white ball being

ci

drawn from A . is -. In like manner the probabi-

n Cj

lity of a white ball being drawn from A2 is shewn to be

1 a \ a~
.
-
; trom A. to be ,
and so on. Denoting
n c2 n cs

therefore by p the whole probability of the event E, the

proposition affirms that

To prove this, let the fractions > &c. be reduced

to a common denominator, and suppose the equivalent frac

tions to be
>

&**.# j** -
rr*^/-^
a
..-.
24 PROBABILITY.

We may now
777
conceive the urns
7
A 1? A 2 A 3 ...A n to
, be re
placed by others, each containing the same number, y, of
balls, and of which the first contains a l white balls, the se

cond a 2 ,and so on; and it is evident that the chance of a white

ball being drawn from this new system of urns will be pre
cisely the same as it was for a white ball being drawn from
the first system. Now the probability of drawing a white
ballfrom the new system will not be altered by placing the
whole of the ny balls in a single urn, for they may still be
conceived as arranged in groups, disposed in any manner

whatever, each group containing the same number of balls,

and the same proportion of white to black as were in the

separate urns and as each group contains the same

; num
ber of balls, the chance of laying the hand on any one group
is the same as that of laying it on any other. The probabi
lity of drawing a white ball from the single urn, is therefore

the same as for drawing it from the

group of separate urns
which contain each the same number of balls. But the pro

bability of drawing it from the single urn is the ratio of the

number of white balls contained in the urn to the number of
both colours, therefore (this probability being p) we have

p= (
a i+ 2 +3 ...... +o);

whence, substituting for , , c., their respective

,7 7
O, O.Q n
values, ,
>
&c., we have

As a particular case suppose three urns A, B, C to be placed

 GENERAL PRINCIPLES. 25

together, of which A contains 2 white balls and 1 black ;

B 3 white balls and 2 black, and C 4 white and 3 black,

and let it be required to determine the probability p of a
white ball being drawn from the group by a person who is

ignorant of the contents of the different urns. As there is no

reason for selecting one urn in preference to another, the pro

bability that he will put his hand into the urn A is ^ ; and if
he draw from this urn the probability that a white ball will
be drawn is , there being 2 cases favourable to that event,
and 3 cases in all. The probability of both events is there
fore \ X f = In like manner, the probability of the ball

being drawn from B is \ ; and if drawn from B the proba

bility of its being white is f ; therefore, the probability of
this compound event is
\ X f = } Lastly, the probability
of the ball being drawn from C is
J ; and if drawn from C
the probability of its being white is i ; therefore, the pro

bability of this compound event is

J X f = ^\ Hence,
by the proposition now demonstrated, the complete pro

bability of the event E is

/>= +i+A= *ff-

If all the balls had-been placed in a single urn, the proba

bilityof drawing a white ball would have been T9j, for there
are 3 5 + + =
7 15 balls in all, of which 2+34-4=9
are white. But T9^ n
J-J| a fraction which differs sensibly
;

from Uf, the measure of the probability of the same event

when the balls are distributed in the manner above supposed
amongst the different urns. The distinction between the two
cases is
important.
10. The rule laid down in (7) for finding the probability

of a compound event applies alike whether the simple events

are determined simultaneously or in succession. In fact,

when the simple events are entirely independent of each

26 PROBABILITY.

other, the chances which determine the compound event are

not influenced in any way by the intervention of time. Sup
pose, for compound event to be the throwing
example, the
of a certain number of points with a given number of m
dice ; the chances for and against the event are obviously
the same whether the m dice are thrown at once, or a single
die is thrown m times successively. But as the determin

ation of the probability of a compound event is in general

facilitated by supposing the simple events to_be decided one

after the other, it will be convenient to view the subject in

this light in explaining themethod of forming the differ

ent combinations of the chances by which the probabilities
of compound events are determined.
 EVENTS DEPENDING ON REPETITION. 27

SECTION II.

OF THE PROBABILITY OF EVENTS DEPENDING ON A REPE

TITION OF TRIALS, OR COMPOUNDED OF ANY NUMBER
OF SIMPLE EVENTS, THE CHANCES IN RESPECT OF WHICH
ARE KNOWN A PRIORI, AND CONSTANT.

Suppose an urn to contain a -f- b balls, a white and

1 1 .

b black, and let a ball be successively drawn, and replaced

in the urn after each drawing, in order that the chances in
favour of drawing a ball of either colour may be the same
in every trial, and let it be required to find the respective

probabilities of the different possible results of any number

of drawings.
Let us first
suppose the number of trials to be two. The
event may happen in any of these four different ways : first

white, second white; first white, second black; first black, se

cond white ; first black, second black. Assuming W to re

present the simple event which consists in the drawing of a

white ball, and B that of a black ball, and supposing the or
der of the arrangement of the two letters to correspond with
the order of succession of the simple events, the four pos
sible cases or combinations will be represented thus :

WW, WB, BW, BB.

Now let the probability of drawing a white ball in any
trial be />, and that of drawing a black ball be q, (whence,
23 EVENTS DEPENDING ON REPETITION.

p=.
-
a
-> q
-- 1>
-
)
the probabilities of the four possible

compound events are by (7) respectively as under :

probability of WW p y> p -=.

p-
of WB = p= pq x q
ofBW = q Xp=pq
of BB = ? X =? </
2

If we disregard the order of succession, and consider the two

arrangements WB and BW, which are equally probable, as

forming the same compound event, namely, a ball of both
colours in the two trials, the probability of this event, by

(9), becomes 2 pq. The sum of the probabilities of all the

possible arrangements is therefore

whence it
appears that the probabilities of the different ar
rangements in two trials are respectively the terms of the

development of the binomial, (p qY- +

Let us next suppose the number of trials to be three.
The different arrangements that may be formed of the
pimple events in three trials, with the probability of each

respectively, are as follows

:

WWW, probability of which ~ppp~p~

WWB, .......................... =ppq=p
2
q
WBW, ..........................

BWW, ..........................

WBB, ...........................

BWB, ........................... -qpq-pq^

BBW, .......................... ^qqp=pf
BBB, .......................... =q(/q=f/-<
It thus appears that the probability of obtaining two events
of one kind, and one of the other, is the same? in whatever
order they succeed each other, and, in fact, is
independent
 CHANCES KNOWN A PRIORI, AND CONSTANT. 29

of the order. Disregarding, then, the order of succession,

and considering the combination of two white balls with
one black, in whatever order they may be arranged, as the
same compound event, the probability of its occurrence in

any order whatever, being the sum of its

probabilities in
each particular order (9), is fypq. In like manner, regard

ing the combination of two black balls with one white, in any
order of arrangement, as the same compound event, its pro
2
bability is
3pq . The compound event resulting from three
trials must then happen
in one of four different ways,
namely,
3 white balls; 2 white, combined withl black, in any order;
2 black, combined with one white, in any order ; or, lastly,

3 black; and the sum of the probabilities of these different

cases is

Hence the probabilities of all the different possible combi

nations in three trials are respectively given by the deve

5
lopment of the binomial (p-{-q) -
1 2. In
general, let^? denote the probability of any simple
event E, then the probability of E happening twice in two
2
trials is j of happening thrice in 3 trials jo 3 , and of hap
,

pening m times in m successive trials, //". In like manner,

the probability of the contrary event F being q (jo-f^r 1),

the probability of F happening n times in n successive trials

is
q Hence (7) the probability of E happening m times,
n
.

and then F happening n times in succession, in m-\-n trials,

m n But the probability of these events happening in
is p q .

any assigned order is the same as that of their happening in

m n
any other assigned order ; therefore p q is the measure of
the probability that E will occur m times, and F will occur
n times in a determinate order. Now, let m + n=h, and let
U be the number of different ways in which m events E, and
 30 EVENTS DEPENDING ON REPETITION.

n events F, can be combined in h trials, and P be the proba

bility of any one of these combinations whatever, or the proba

bility of E occurring m times, and F occurring n times in h

trials, without regard to the order in which they succeed
each other, we have then
P=U/y.
In order to determine the value of U, we may suppose
the events in question to be so many different things repre

sented by the letters A, B, C, D, E, &c. of which there are

m of one kind, and n of another, and make m--n=h then ;

by the algebraic theory of combinations, we have

1 2 3 - *
u= -

1 . 2 .3 mxl .2 .3 n
This value of U is symmetrical in respect of m and n, and
may be otherwise written in either of the two following

forms,

h(h\)(h 2) h m-f-1
1.2.3 m
jr _k(hl)(h 2) h n+l
1.2.3 n
which shew that the probability P, or the product Upm q"
is the
(m-j-l)th term of the development of the bino
mial (p+q)* arranged according to the
increasing powers
of^?, or the (w-fl)th term of the same development ar
ranged according to the increasing powers of q. Hence
we conclude that when p and q remain constant, the pro
babilities of all the different compound events which can
be formed by the combination of the simple events E and
F in h trials, are expressed by the different terms of the
formula (p-)-q) h expanded by the binomial theorem.
The whole number of possible cases is evidently A-f-1,
for in h
experiments, E may occur h times, h 1 times,
 CHANCES KNOWN A PRIORI, AND CONSTANT. 31

h 2 times h h times; this last being the case in

which the contrary event F occurs in all the trials. The
different cases are unequally probable, both by reason of the

greater or smaller number of combinations by which they

may be produced, and which in reference to each case is

represented by U, and by reason of the inequality between

p and q. It will be shewn afterwards, that when pz=q, and
h is a whole number, the most probable case is that in which

the occurrences of E and F are equal ; and if h is an odd

number, the two most probable cases are those in which
the difference in the number of occurrences of E and the
number of occurrences of F is unity.
13. In order to place the proposition now demonstrated
in a clearer light, let us consider separately the different
terms of the development of (p + q) \ namely,

_- 1.2.3
h-n+ 1
n f *

The first term p h expresses the probability that the event

E will in h
every one of the h trials. The second term hp ~^q
expresses the probability that E will occur h 1 times, and
F once, without distinction of order ; that is to say F may
happen at the first or last or any intermediate trial. If a

determinate succession is
proposed, for example, that of
h 1 times the event E in succession, and F in the next trial,

the probability of the event in the assigned order is found

and h ~ l
by suppressing the coefficient h, is consequently p q.

The
~
third term
[.2
^ V*~V expresses the probability
32 EVENTS DEPENDING UPON REPETITION.

that the result of h trials will be h 2 times the event E,

and twice the event F, without distinction of order. Jf a

particular order be assigned, it is

necessary to suppress the
coefficient, and the probability of the simple events occur
-2 2
ring in that particular order is
ph q .

The general term

h(hl)(h2)
~- h n +~ph~n qn ex-
\

1 2 o it

presses the probability that the result of h trials will be

(h n) times the event E, and n times the event F in any
order. The probability of (h n) times E and n times F
an assigned order h ~n n.
in is p q
14. If we suppose the event E to be such that the chances
in favour of its happening or failing are equal, that is, if

pq-=.^, the different terms of the binomial Q? + #) >

on sup

pressing the coefficients, become all

equal ; so that a parti

cular order being assigned in each of the possible cases or

combinations, all the cases become equally probable. Thus,

suppose a shilling to be tossed 100 times in succession, the
probability of head turning up in every trial is (^)
l
. The
probability of 50 heads and 50 tails in any assigned order
50 50 100 if the probability
is(i) x(i) =(i) ; wi+w=100,
ofm heads and n tails is also (l)(l) =(i) +-(r)ioo.
n w

Hence the probability of any compound event formed by

the combination of two simple contrary events succeeding
each other in an assigned order, and each having the same

probability, independent of the ratio of the simple events,

is

and depends only on the number of trials. Before the trials,

it is an even
wager that head will be turned up in succession
100 times, and that the result of 100 trials will be 50 heads
and 50 a given order of succession, or any proportion
tails in

of heads to tails in an order arbitrarily chosen. This con

sideration is
frequently lost sight of in reasoning about those
 CHANCES KNOWN A PRIORI, AND CONSTANT. 33

events of the natural world, which are termed extraordinary

and miraculous. If in tossing a shilling 100 times into the
air, thenumber of heads turned up is found nearly equal to
the number of tails, the event excites no surprise some ;

thing like was expected. On the contrary, if the diffe

it

rence between the number of heads and the number of tails

is considerable, the event is termed extraordinary ; and if

head turned up in every trial without exception, we should

scarcely be persuaded that such an event was entirely the re

sult of chance, and independent of a special cause. Never
theless, the a priori probability that every trial will give
head, is same as the probability of throwing
precisely the

any given number of heads and tails in an assigned order of

succession. It will, however, be proved afterwards, that if

such an event as throwing head 100 times in succession were

actually observed, the probability of a special cause having

intervened, would approach very nearly to certainty.
15. Hitherto we have supposed the compound event to

be formed by the combination of two simple events only,

E and F, one of which necessarily excludes the other. Let
us now suppose there are any number of simple events, E 15
Ea E5 , , &c. of which the respective probabilities a.rep 19 p,^

p3 ,
&c. and such that one or other of them necessarily hap

pens in each trial, so that p^ -\-p 2 -{-p 5 +> &c.= 1, and de

termine the probability of any assigned combination of them
in a given number of trials. This case may be represented

by supposing an urn to contain a number of balls of as many

different colours as there are distinct events ; the event E,
will be the drawing of a ball of the colour and its proba ,

bility p will be the fraction whose

t
numerator is
equal to the
number of balls of the colour i, and denominator the whole
number of balls in the urn. Now the probability of the
34 EVENTS DEPENDING ON REPETITION.

event Ej happening m times in succession is//? by

that of E2 happening n times in succession is
p\ ; that of

E3 happening r times in succession p^ ; and so on. There

fore (7) the probability of the compound event which is

formed by the occurrence of m times E 1? n times E2 ,

r times

E5 , and so on, these events succeeding each other in or

r
der, is the product p p p5 , &c. But the probability of
the simple events succeeding each other in any particular
order is the same as that of their succeeding in any other

assigned order (12) ; consequently, if U denote the num

ber of different ways in which events m E 1? n events E2 , r

events E3 , &c. can be combined, or succeed each other,

and P be the probability of the compound event in any order
whatever, we have,

Assuming h=m+n-\-r+, &c. we have also by the theory

of combinations,

the factor
TT,- 1-2.3 .................. h
~1 .2.3...mxl .2.3...nxl.2.3...rx
U
__
being the coefficient of the term which has for
&c.

its multiplier p p p*, &c. in the expansion of the mul

(Pi+Pv+Ps + & C *) whence
A
tinomial >

We shall now proceed to give some examples of the ap

plications of the preceding formulae.
16. Let it be proposed to assign the probability P, of

throwing ace once, and not oftener, in four successive throws

of the same die. Simpson, p. 15.
Here, the chance of throwing ace in a single trial being
we have jo=}, and consequently 5 =^, and also 7<t=4.
,1,

Now the compound event being the occurrence of the sim-

 CHANCES KNOWN A PRIORI, AND CONSTANT. 35

pie event E, whose probability is jo, once, and of the con

trary event F three times, the probability of the compound

event is that term of the development of (p-\-q)* which is

5
multiplied by pq .
If, therefore, in the formula,

M 1.2. 3 h m fJ n
"~
,

T)
1 2 ^ w? y 1 2 ^ ii *

we make p^ (? B> 7^=4, m=l, ?i=3, we shall have

1 .2.3.4 /5\ _12_5 5

X J_ X
""1X1.2.3 6 \6/ "324
which is the probability required, and the same as that of

throwing one ace, and not more than one, at a single throw
with 4 dice.
The probability of the contrary event, that is to say, the

probability of either not throwing an ace at all, or of throw

ing more aces than one is 1

1|| ^|| ; and therefore the =
odds against throwing one ace and no more in 4 throws of a
common die are 199 to 125, or 8 to 5 very nearly.
17. If in this example it had been proposed to assign the

probability of throwing ace once at least, instead of once

it would have been
and not more, necessary to have includ
ed those cases in which the ace occurs twice, or three times,
or in each of the four trials. The binomial (jo-f-^) 4 gives

the first term of which expresses the probability of throw

ing ace four times in succession the second that of throw;

ing ace three times, and another number once ; the third
that of throwing ace twice, and a different face twice ; the
fourth that of throwing ace once, and a different face three

times ; and the fifth that of throwing a different face in each

of the four trials. But as every one of these compound
events, excepting the last, satisfies the condition of ace be

ing thrown once at least, the whole probability of that event

 36 EVENTS DEPENDING ON REPETITION.

must be the sum of the probabilities of the different events

by which it
may be produced (9) and is
consequently

U;^W
f
}
W V^YW V V- 1

+T6") T+TS-J Uv +4 U;UJ

\*
5 ] 1 5 671
1298

In general, the sum of the first n-\- 1 terms of (p + 9 )" ex

r "

presses the probability of obtaining not less than h n events,

the probability of each of which is/?,
or not more than n con
trary events, the probability of each of which is
q.

Since p +<?= 1, the sum of all the terms of the series pro
duced by the expansion of (p+q) h is equal to unit, and
therefore the sum of any number of the terms is equal to
unit diminished by the sum of the remaining terms. This
consideration frequently gives the means of abridging the
calculations. preceding example, instead of
Thus, in the

. >v expanding the binomial (^-fu) 4 order to find the proba m

bilities of throwing 4 aces, 3 aces, 2 aces, and 1 ace only,

in a series of 4 trials, we might have sought the probability

of not throwing ace at all. The probability of not throwing
ace in a single trial is
|j,
and therefore (7) that of not throw
4 Hence
ing it in 4 trials is () =T 6/^5- the probability of
the contrary event, namely, that ace will be thrown once
or oftener, is 1
-f^fy^-fifo ; the same as before.
18. Let a shilling be tossed ; what is the probability that
more than 3 heads will turn up in the first 10 trials? In

case,/>=i, g=, A=10

h
this ; therefore (p+q) =(i+i) 10
=:()io(l -f l)i. Now the term of this developmentlast

expresses the probability that head will not turn up in any

one of the ten trials ; the last but one, the probability that
it will turn up once ; the last but two, the probability that
it will turn up twice ; and the last but three, the probabil

ity that it will turn up three times ; therefore the four last
 CHANCES KNOWN A PRIORI, AND CONSTANT. 37

terms include all the different ways in which the ten trials

give not more than three heads ; and their sum consequently
expresses the probability that not more than 3 heads will
be thrown. Now the last four (or first four) terms of the
T
expan:ion of (
1
-f-
1
) are

1Q 9 10 9 8
i in
1.2 1.2.3
and their sum is 176, which multiplied by (i) 10= To35^

gives yVy^j for the probability that not more than 3 heads
will turn up ; whence the probability of the contrary event
or that more than 3 heads will be thrown, is 1
JQ"^
n34 = M
8
i ; and the odds in favour of throwing heads more
than three times in 10 trials are 53 to 11.

19- A
engage in play; the probability of A s
and B
a
winning game is
p, and the probability of B s winning a
game is
q ; A s winning m
required the probability P, of
games before B
wins n games, the play being supposed to
terminate when either of those events has occurred.
It is evident that the question must be decided at the

by the (m + n l)th game for supposing m-\-n 2

latest, ;

games to have been played, there is only one combination

according to which the match can remain undecided, name
ly, that in which A has won m 1, and Bn 1
games ; and
in this case the next game necessarily decides the match.

Suppose m+ x games to have been played. The proba

bility that of these games m have been won by A, and x by
m +*
B, is
represented by the term of the binomial (p-\-q)
m
in which the factor p q* occurs (13) which term ;
is

1.2. 3 ......... m + x
1 .2 .3
But A cannot win m games out of m-\-x exactly unless he
wins the last game, for otherwise he must have won m games
38 EVENTS DEPENDING ON REPETITION 7
.

out of m+x 1, if not out of a smaller number. In order

therefore that A may win m games out of m+x exactly, it

is necessary in the first place that he wins m 1 out of

m-\-x 1 in any order, and then that he wins also the next

game. Now the probability of his winning m 1

games out
of m-Jf-x 1 in any order (13) is

1 . 2 .3 ......... m+x 1
_j x
.2.3...x pm q 5
1 . 2.3...m 1X1
and the probability of his winning the following game is p,
whence the probability of both events is (7)

1.2.3 ...... m x 1
1.2.3 ...... m 1X1 .2.3 ......
which, therefore, expresses the probability of As winning
m games out of m + x exactly.
If we suppose a?=0, this formula becomes m which is
jo ,

the probability of As winning m games in succession. If

.*:=: 1, it becomes mp mq, the probability that A wins m games
If x= 2, becomes - m 2
outofw-j-1. it
p q , the pro

bability that A wins m games out of m + 2. If x=3, it be-

comes --123
m(m+\)(m4-2)
P *"^ ^G Probability that A wins

m games out of w-f-3 ; and so on. Continuing this process

we arrive at the term multiplied m x
till
by p q the sum of the
,

probabilities of all the different compound events is

which expresses the probability of As winning m games

out of a number not greater than m-\-x.
Now it has been shewn, that the match is
necessarily
decided by (m + n 1) games; consequently the solution
 CHANCES KNOWN A PRIORI, AND CONSTANT. 39

of the question is obtained by substituting n 1 for x in

the last formula, which will then express the probability

of As winning m games in any order, out of a number not
greater than m-\-n 1. On making this substitution, we
obtain

v ^
?=/>" -j
l+mq+ f
m(m -f- 1 ).~.m .
-f- n
1 .2 n 1
,.-}.
The probability Q, that the match will be decided in fa

vour of B, or that B will win n games out of a number not

greater than m-\-n 1, is found by changing m into n, and

p into q, and is therefore

n(n+l) m + n-2
1.2 m\ P
2 i
As an example, let us suppose p -, q-=. -, m=4, and
o o
n=2. The probability of As winning the match, or the
value of P, becomes

and the probability of Bs winning the match, or the value

ofQ,

/l\f 4 2.3/2v 2.8.4/2v) _ 131

W I ^3^1.2MU h
l-2.3W j""243
In this example the skill A is
supposed to be twice as
of

great as that of B, and the number of games that must be

won by him in order to gain the match is also twice as great

number
as the required to be won by B in order that B may
40 EVENTS DEPENDING ON REPETITION.

gain ; one might therefore suppose, that when they begin

to play the chances in favour of each are equal. But the
result shews that the chances in favour of A are fewer than
those in favour of B in the proportion of 1 12 to 131 ; whence
it
appears that it would be unsafe to wager that a player
who has two chances in his favour while his adversary has

only one, will gain four games before his adversary shall
have gained two.

Suppose A and B, engaged in play, agree to leave off be

fore the match is decided, it is evident that the stakes ought
to be shared between them in proportion to their respective

probabilities of winning, and consequently the share of each

is found from either of the above expressions for P and Q.
This was one of the questions proposed by the Chevalier
de Mere to the celebrated Pascal, to which allusion has al

ready been made.

20 An urn contains n-\-l balls, marked with the num
bers 0, 1, 2, 3
n\ a ball is successively drawn and re
placed in the urn, so that the chance of drawing any given
number remains the same in each trial, what is the probability
that in h trials the sum of the numbers drawn will be equal to*? 1

The solution of this problem depends on the number of

ways in which the number s can be formed by the addition of
h different numbers, each of which may have any value from
to n. If we suppose the numbers marked on the balls to

be indexes of a certain quantity #, and develope the expres

2 n h
sion (x-\-x l -\-x -\-x ) the coefficient of any term of
,

the development will indicate the number of different ways

in which the balls may be drawn, sum of the num
so that the

bers drawn in h trials shall be equal to the sum of the in-

1
Demoivre, Miscellanea Analytica, p. 196 ; Laplace, Essai sur les

Probability, p. 253, et seq.

 CHANCES KNOWN A PRIORI, AND CONSTANT. 4

dexes of x in that term. If, therefore, we denote by N the

coefficient of that term of the development in which the sum
of the indexes is s, then N
will be the number of cases fa

vourable to the event. But the whole number of possible

cases is (/z-J-l)*; therefore the probability of the event is

On account of the particular form of the polynomial in

question, the value of is found withoutNdifficulty.

Because x-}-x 2 n
1
-
_x n + l

-{-x ...... -\>x =:

1
, therefore
1 x
-}-x
2
...... +xn ) h =(l x n + l )\l x)~
h.
Now, ex
pressing these two factors in series, we have (1 xn + l )h

1.2

and the coefficients of the several terms of the product of

these two series in which the sum of the indexes is s will
be found as follows :

(1.) Multiply the first term of the first series by that term
of the second series of which the argument is x 1 the coeffi-
;

cient of the product will be

/i(/*-fl)(/i
v + 2) ...... h+s ~1
1.2.OQ
,
...... S

(2.) Multiply the second term of the first

by that
series
s~ n ~ l
term of the second series which has for its argument x ;

the coefficient of the product will be

,_ k(h + \)(h + 2) ...... k + sn 2

1.2.3 ...... s n 1

(3.) Multiply the third term of the first series by that

term of the second series which has for its argument
*_2(n+i). t h e coefficient of the product will be
42 EVENTS DEPENDING ON REPETITION.

h(h\) X A(A+ !)(/* + 2) ...... h+s2n3

~T72~ 1.2.3 ......

(4.) Proceed in the same manner with the fourth term of

the first series, and so on with the others, advancing at each
new multiplication one term to the right in the first series,

and n-{- 1 terms to the left in the second series, until a term
is reached in the first series, the exponent of x in which is

equal to, or greater than s. The sum of the several products

thus obtained will be the value of N. We have therefore

...... h+s-l
1.2.3
h
1 .2.3 ...... s n 1

h(h + 1
)(h + 2) ......
1.2 1.2.3 ...... s2n 2
_&c.
The series now found for N may be changed into another,

having a more elegant form, by reducing all the terms to

others having the common denominator 1.2. 3 ...... h 1.

This will be accomplished by leaving out of the numerator

and denominator of the first term all the numbers after h 1
to 5, (including s), when s is greater than h 1, or by in

serting the numbers between s and h 1

(the last included),
when s^Ji 1 ; by leaving out of the numerator and deno
minator of the second term all the numbers from h 1 to

s n, or by inserting those numbers ; and so on with the

other terms. If we then make the common denominator
1.2. 3.../ l=/e, we shall have

_ (
s n )(s n + 1
) ...... (s n -f // 2)
 CHANCES KNOWN A PRIORI, AND CONSTANT. 43

_&c.
to be continued till the last factor of one of the terms be
comes or negative. If we also make

+ h2=f(n+ !)=/
s 2n 4.h3=f2(n + )=/" 1

s3n+h4=f3(n+ !)=/"
&c.
and write the factors in each of the terms in the reverse
order, the above value of N will become

+ &C.
As an example of the application of this formula, let
21.
itbe required to assign the probability of throwing the point
16 with 4 common dice. (Simpson, p. 53.)
A die having no face marked 0, it is necessary, in order
to adopt the formula to this case, to suppose the number of

points on each face to be diminished by unit, which is equi

valent to supposing s h to be substituted for s. The num
bers are then 0, 1,2, 3, 4, 5, and we have rc=5, ^=4, and
)=12. Hence
44 EVENTS DEPENDING ON REPETITION.

f =f-2(n+})=3
f"=f 3(n+l )=3,
and A 1 . 2 . 3> Substituting these values in the formula,
we find

336)
9.8.7Xg(=

or N= 125. Now the probability of the event is N-*-(w + 1)*;

and in the
present case (ra-j- l)
A
=6
=il296 ; consequently 4

the probability required, namely that of throwing the point

16 with 4 dice, is
TO^T-*

22. In the numerical solution of questions of this sort, it

sometimes happens that the labour may be abridged by

computing the probability of throwing a different point from
that which is proposed, but which has the same number of

chances in its favour. For example, let it be proposed to

determine the probability that in throwing 10 dice the sum
of the points will be 50. In this case, the smallest number
of points that can possibly be thrown and the greatest
is 10,
60; and the chances in favour of
throwing 10 and of throw
ing 60 are obviously equal. The probability of throwing
any given number of points above 10 is also evidently the
same as that of throwing the number which is as much under
60 ; and consequently the probability of throwing 50 is the
same as the probability of throwing 20, these numbers being
at equal distances from the extremes. Now to find the pro

bability of throwing 20 with 10 dice, or, which is the same,

the probability that in 10 successive drawings from an urn
 CHANCES KNOWN A PRIORI, AND CONSTANT. 45

containing 6 balls, marked with the numbers 0, 1, 2, 3, 4, 5,

the sum of the numbers drawn will be 10, we have ^10,

7*=5, 5=10; whence /=19,/=13,/"=7,/ " negative,
and =1 . 2... 9 Substituting these numbers in the series
for N, and observing that since f. 7i-f 2= 1, the third

term becomes negative, we have

19.18.17.16.15.14.13.12. 11 .

*=-1.2.3.4.S.6.7.8.9 (=923 8)

13.12.11 .10.9.8.7.6.5.
1.2.3 .4.5.6.7.8.9 (=-7150
-
)

and consequently N= 85228. Dividing this by (n + =6 1

)
A 10

= 60466176, the probability of throwing 20, or of throwing

QOOQ i

50, with 10 dice is found = , or between - and

23. The probability that the whole number of points

drawn in h trials will not exceed s is found by substituting
for s the different values 0,
1, 2 ...... s in the series for N,

and taking the sum of the results. This labour, however,

may be avoided by means of a property of the figurate num
bers. It is well known that the sum of the series of num
bers obtained by giving n every value from n= 1 to n=v (v
being any number whatever) in the formula

*(*+!)(* + 2) ...... (n + u)
1-2 3 ...... w+1
is expressed by this other formula

(p + 2) ......

or, which is the same thing, that the sum of the series ob
tained by giving x successively every value from x=l-{-u
to &= 1
-f u -{-
v in the formula
46 EVENTS DEPENDING OX REPETITION-

gfo l)(g2) ...... (xu)

1.2.3 w+1
is
expressed by this other formula

(x + l)x(x\) ...... xu
1 .2.3 .......... w + 2
in which #=1 -j-w-}-i\
Comparing the N (20) with
different terms of the series for

these last formulae, be evident, that on giving s every

it will

value successively, from to s in the value off, and denot

ing by N
the sum of all the results, we shall have

y, ,(
1.2.3.4 ... h

l)(/-2)...(/-A+2)
1.2.3.4 ... h
// //
l)(/ -2)-(/ -^+2) A(A-l)
1.2.3.4 ... h 1.2
&c.

for the probability that the number of points thrown will

not be greater than s.

As an example, let/*=lO, s=5, rc=5; we have then

f=s + h 1 = 14, / =/ O+l)=8, whence/ + 2=0, A

and consequently the second term vanishes. Hence

18. 14. 13. IS. 11. 10. 9. 8. 7. 6

1.2.3.4.5.6 .7.8.9.10-
The probability, therefore, of the sum of the numbers
drawn in 10 drawings not exceeding 5, is 3003 -f-(6) 10 or
,

3003
fifi >.{.>
And this is also the probability that in throw-
OU"xOO I/O

ing 10 common dice the sum of the points does not exceed
16. (Simpson, p. 60.)
24. In the preceding questions the number of trials, de-
 CHANCES KNOWN A PRIORI, AND CONSTANT. 47

noted by h, has been supposed to be given ; and the ob

ject, in every case, has been to determine the value of a

given term, or of a given number of terms of the series pro

duced by the development of the binomial (p-j-^y. But
there is a numerous class of questions in respect of which
the exponent h is unknown, and is
required to be deter
mined from the condition that an assigned term, or the sum
of a certain number of assigned terms of the development,
must have a given value. For example, let it be proposed
3d
to determine how often a common die must be thrown in

order to give the probability of ace turning up once at least

equal to a given fraction u. Here the probability of throw

ing ace in any throw being , we havep=r, and q=$>
Now, as every term of the development of (p q} h except + ,

ing the last, gives a combination in which ace occurs once

or oftener, the question requires a value to be found for h,

such, that the sum of the first h terms of that development,

shall be equal to u. This may be done, in general, by the
common methods of trial and error; but in the present case,

the last term being the only one not included among those
which contain a chance of throwing ace, it is evident that
it is
only necessary to find the last term alone in order to
have the probability of not throwing ace in h trials, which by
the question is 1 u. The last term of the development
is
(f ; therefore we must have the equation qh ~l u;
whence h log q=\og (1 u,) and ^=log (1 u) -i-log q.

Let 1 M=- and <7=-, we shall then have log (1 u)-=.

log log y, and log #=log b log c; whence

A _log log
y^
log b log c

Substituting in this general formula the particular numbers

 48 EVENTS DEPENDING ON REPETITION.

given in the question, namely b=5, c=6 ; and supposing

w=J, and consequently /3= 1, yr=2, we have /*=. 7,- ;

logb log 5
whence, by computing from the logarithmic tables, h=;3 8.

From this it follows, that in four trials the probability of

throwing ace once at least, is greater than the probability

of not throwing it at all.

If the question had been to determine in how many throws

with two dice one may undertake, on an equality of chance,
to throw aces at least once, we should have had p=^, q=
f f and consequently 635, and c=36. Substituting these
,

numbers in the general formula, and observing, that in this

case also 0=1, y=2, we get A= -=-

--
=24-6.
log 36 log 35
The probability of not throwing aces once is therefore

greater than the opposite probability or that of throwing

aces once or oftener, when the number of throws is 24, but
less when the number is 25.

These two questions are celebrated in the early history of

the theory of Probability, from the circumstance that the

Chevalier de Mere, by whom they were proposed to Pascal,
declared the two results above stated to be inconsistent with
each other, and thence took occasion to question the accu

racy of the theory of combinations by means of which they

had been obtained. He reasoned thus : Since the proba

bility of throwing ace with one die is

J, and that of throw

ing aces with two dice ^ of 5 = 35 ; therefore, if there be a

given probability in favour of throwing ace in fourthrows with

one die, there must likewise be the same probability of throw

ing aces with two dice in 6 X

4 24 throws in other words, ;

the chances in favour of an event E in a single trial,

being
six times more numerous than those in favour of F, there
 CHANCES KNOWN A PRIORI, AND CONSTANT. 49

will be as many chances in favour of Fin six trials as there are

in favour of E in one. The error consists in supposing that

the number of trials must increase or diminish exactly in the

inverse ratio of the probability ofobtaining the proposed point.

25. The general question may be enunciated as follows :

Let p the probability an event E will happen, q the proba

bility it will fail ; how many are required to give a
trials

probability =u that E will happen k times.

Let r=the number required. Taking the sum of all the

x
terms of the development of (p-\-q) in which the exponent
of p is less than k, we shall have the probability that the
event does not happen k times in x trials. This sum must

consequently be made equal to 1 u ; therefore, beginning

the last term, and writing the terms in the reverse order,
we have the equation

Let p-=. e q, and this equation becomes

from which the value of x may be found by the ordinary

methods of converging series.

Ifp=q^ then e=l and; if we also suppose M=^, and

consequently 1
u=^, the equation will become
x( _... _
1
Pascal, (Euvres, torn. iv. p. 367 ; Lacroix, Elemental, p. 36.
D
50 EVENTS DEPENDING ON REPETITION.
"-J-
I

But the first side of tins equation is the expansion of (1 X 1

)*

continued to k terms ; therefore, since the sum of the first k

terms is
equal to one-half of the whole series, and the terms
of the first half of the series are the same as those of the

last, it follows that the whole number of terms must be 2.

But the whole number of terms in the expansion of (1 -f- 1)*

is#+l; therefore 2k=zx+l, and.x=2k 1. Suppose k =

10, then #=19 ; hence in tossing a shilling it is an even bet

that head will turn up 10 times in 19 throws.

 CHANCES KNOWN A PRIORI, AND VARYING. 5 1

SECTION III.

OF THE PROBABILITY OF EVENTS DEPENDING ON A RE

PETITION OF TRIALS, OR COMPOUNDED OF ANY NUMBER
OF SIMPLE EVENTS, THE CHANCES IN RESPECT OF WHICH
ARE KNOWN A PRIORI, AND VARY IN THE DIFFERENT
TRIALS.

26. Let us suppose the trials to consist in drawing balls

from an urn containing a white balls, and b black balls, and

that when a ball is extracted it is not returned to the urn.

Make a + b=c, and let the extraction of a white ball be the

event W, and that of a black ball the event B. At the first

trial the probability of W is

(4), and that of B, -. But at

the second trial, the number of balls in the urn is diminish

ed by 1 and the probability of drawing a white ball at the

;

second trial is therefore not the same as it was in the first,

but is influenced by the event which has already taken
place. If W happened at the first trial, the number of
white balls remaining in the urn is then a 1 ; the num
ber of black is b, and the number of both colours c 1.

The probability of W at the next trial is therefore

-^-,and that of B is -. In like manner, if B happened

c 1 c 1

at the first trial, the probability of W at the second is

-
52 EVENTS DEPENDING ON REPETITION.

and that of B is . Hence (7) the different combinations

which can arise from two trials are the following :

WW, WB, BW, BB,

the probabilities of which are respectively,

a(a 1) ab ba b(bl)
c(c 1) c(c 1) c(c 1) c(c 1)

Now if we neglect the order of succession in the two cases

in which W and B are combined, the probability of the

compound event which consists of the extraction of a ball

of each colour in the two trials, is r, and the probabi-

c(c 1)

lities of the three possible combinations are respectively :

a(a 1) 2ab b(bl)

c(clY c(c\y c(cl)
Comparing these with the probabilities of the same com
binations when the chances are constant, or the ball is re

turned to the urn after each drawing, namely

J??
"? "?" "?*

the analogy of the two cases. is obvious-

After two balls have been drawn, the whole number re

maining in the urn is c 2 ; but the number of each colour

depends on the two events that have already occurred. If

two white balls have been drawn, the probability of draw-
Q 2
ing a white ball at the next trial will be - ; but we have
c

just seen that the probability of WW is -

c(c
^; therefore
1)

WWW q ) (a
~2A
(7) the probability of is
^~\
C(C i)(C
,

4)
Thepro-
 CHANCES KNOWN A PRIORI, AND VARYING. 53

bability of drawing a black ball after two white have been

drawn is - (for there are

now c 2 balls in the urn, of

WWB
^
/
c(c-\)(c
-
which b are white)

2)
. On
; therefore the probability of

forming in this manner all the differ-

is

ent possible combinations which can result from three trials,

we find

c(c l)(c 2)
_
-- _ ab(a
^ 1)
J.

If we disregard the order of succession in those combina

tions into whichW and B both enter, and consider the oc
currence of W twice and B once as the same compound
event ; and also the occurrence of W once and B twice as
the same compound event, in whatever order they occur,

the probability of the former will be , . ., and of

the latter -
- 3ab(b
^
c(c\)(c
1)^

2)
.
54 EVENTS DEPENDING ON REPETITION.

27. In general, if m! -\-n f balls have been drawn, of which

in have been found to be white and n black, the number
of white balls in the urn will now be a m ,
the number of
black b n and the whole number of both colours
9
c m nf.

Hence the probability of drawing a white ball in the next

trial will be -, , ; and that of drawing a black ball

Now if in these two fractions we substitute

c m! n

successively for m and n all the different numbers from

to m 1 and n 1
respectively, the product of the m-\-n
numbers thusobtained will (7)be the probability of drawings
white balls and 72 black in an assigned order, mm-\-n trials.

Let this probability be denoted by K, and we shall have

ft(g-l)(a-2) (-+ l)x 6(6-1 )(6-2) (6-rc+l)

c(c 1
(c2)
) (cmn + 1
)

whatever the given order may be. Hence, if we denote by

P the probability of m white balls and n black being drawn
in any order whatever, in h trials, we shall have P= UK,

1 .2.3 h
where, as in (12), U .

1 .2.3 mxl -2.3 n

the co-efficient of that term of the binomial (p-\-q) h which
has for its argument pm q n ; this co-efficient
expressing all
the different arrangements which can be formed of
things m
of one kind, and n things of another.
28. When the urn is
supposed to contain balls of more
than two different colours, the probability of any
proposed
number of each colour being drawn in a given number of
trials is found with the same facility. Suppose it to contain
!
of the first colour, a 2
of the second, a 5 of the third, and
so on, and let al +a 2 -\-a. &c.=c then ; the probability that
 CHANCES KNOWN A PRIORI, AND VARYING. 55

in m-^-n + rJf&LC.h trials, there will be drawn m of the

first colour, n of the second, r of the third, &c. is

U x^Oi !)(! 2) ......... (!

X 2( 2 1)(^ 2) ......... (tf 2
__
Xa 3 (a 5 1)(3 2) ......... (a 3

_ _ K
X 4 1)K 2) ......... ( 4

X &c.

where, as in (15),
1.2.3 ............ h
rj,_
1 .2.3 ...... m*l .2.3 ...... wxl .2-3 ....... X&c. /

29. The following examples will shew the use of the pre
ceding formulae.
Suppose a bag to contain 1 6 balls, of which 8 are white
and 8 black, what is the probability that in drawing 8 balls
from the bag the whole of them will be white ?

Applying the formula (27) to the solution of this ques

tion, we have a=8, 6=8, cnl6, i=8, n=0, and as the

probability required is that of drawing white balls only, b

cannot enter into any of the factors of the numerator ;

hence
8. 7. 6. 5. 4. 3. 2. 1_ 1
;
""16.15.14.13.12.11.10. 9~~12870
and since m=h, U=l, the probability sought, is therefore

T270-
Let there be a heap of 20 cards, wherein are 7 diamonds,
6 hearts, 4 spades, and 3 clubs required the probability ;

that in drawing 8 of them at a venture there shall come out

3 diamonds and 2 hearts? (Simpson, p. 21.)
The probability required in this case being that of draw

ing 3 diamonds, 2 hearts, and 3 other cards which are nei-

56 EVENTS DEPENDING ON REPETITION.

ther diamonds nor hearts, the spades and clubs may be con
sidered as forming one parcel, containing 7 cards. We have
then in the formula (28) a l =7, 2 =6,
a 3 =7, c=20; m=3,
/* 2, r=3, 7/ =8 ; therefore

1.2.3.4.5.6.7.8
U
=l. 2. 3X1. 2 X 1. 2. 3-
and the probability required becomes,
7.6. 5x6 . 5x7 . 6 . 5 1225
X 18 16 14
20 . 19 . 17 . . 15 . .
13~3978*
The odds against the event are therefore 2753 to 1225, or

nearly 9 to 4.
Let 4 cards be drawn from a pack of 52 ; what is the

probability of drawing one of each sort

?

In this case we have a 1

= 13, a 2 :=13, = 13,
3 =rl3, 4

c=52 ; also m=l, ra=l, r=l, s=l, k4, whence U =

1 .2.3.4
-
=r 24, and the probability required becomes, on

substituting these numbers in the formula (28),

13.13.13.13_
X 52. _ 21.97 1
n<
51 .50. 20825"" 49"" 9

The odds against this event are nearly 8 to 1.

30. The following question, proposed by Huygens, and

solved by Demoivre and Bernoulli (Ars Conjectandi, p. 59)>

belongs to the class of problems now under consideration.

An urn contains 12 balls, of which 4 are white and 8
black. Three gamesters A, B, and C agree that the first
who, blindfold, shall draw a white ball shall be the winner of

the stakes. They also agree that A shall draw first. B second,
C third, A fourth, and so on ; and the balls drawn are not
replaced in the urn. It is proposed to find their respective

probabilities of winning.
 CHANCES KNOWN A PRIORI, AND VARYING. 57

Here the play terminates as soon as a white ball is drawn,

and it must therefore terminate with the 9th trial, if not

sooner, inasmuch as, after 8 black balls have been drawn,

the urn will contain only white balls, and the probability of

drawing a white ball at the next trial will become certainty.

The question will therefore be solved, if we determine the

probabilities of the play ending with the 1st, 2d, 3d, 4th, &c.

games respectively, and take the sum of the probabilities

of its ending with the 1st, 4th, and 7th, for the probability of
A s winning; the sum of the probabilities of its ending with
the 2d, 5th, and 8th, for the probability of B s winning ; and
the sum of the probabilities of its ending with the 3d, 6th,
and 9th, for the probability of Cs winning.
For the sake of rendering the solution more general, let
a be the number of white balls in the urn, b the number of
black, and let a -f- b=c. The probability of drawing a white
ball at the first trial, or of the play ending with the first

a
trial, is then .

The probability of the play ending with the second trial

is
compounded of the probability of a black ball being drawn
at the first trial, and a white at the second ; and the proba
a
bility of both events (26) is
-
..

The probability of the play ending with the third trial is

compounded of three separate probabilities, namely, that a

black ball will be drawn at the first trial ; that a black ball
will be drawn at the second ; that a white ball will be drawn
at the third ; and the probability of the concourse of these

In general the probability of a black ball being drawn in

58 EVENTS DEPENDING ON REPETITION.

x 1 trials successively, and a white ball at the xih is

I,(b
,
!)(&
r-r?
2) (br + 2)a .
substituting tor #. b. and
c(c l)(c 2) (c a?+l
c in this formula the numbers proposed by Huygens, we
obtain in respect of the 1st, 4th, and 7th trials, or the pro

bability in favour of A,
8-7.6 4 8.7. 6.5.4.3 4 _ 77
12
+ 12. 11 .10* 9
"
"12.11 .10. 9- 8. 7 "6" ""165
;

in respect of the 2d, 5th, and 8th trials, or the probability

and in respect of the 3d, 6th, and 9th trials, or the proba

bility in favour of C,

8 . 7 8.7.6 .5.4
4_
8.7.6 .5.4.3.2.1
"*"
12.11*10"*" 12. 11. 10.9.8 7 12.11.10.9.8.7.6.5
4 __ 35
*T~~ 165*

The chances in favour of A, B, and C, are therefore pro

portional to the numbers 77, 53, 35, respectively.
If the condition of the play had been that the ball was to

be returned to the urn after each trial, the chances in fa

vour of the three gamesters would have been easily found

by the formula (12) to be respectively as the numbers 9>

6, and 4.
 MATHEMATICAL AND MORAL EXPECTATION.

SECTION IV.

OF MATHEMATICAL AND MORAL EXPECTATION.

v 31. In the theory of probability, the term expectation is

used to denote the product found by multiplying the value

of a casual benefit into the probability of the event on
which it is
contingent taking place. But the value of a be
nefit may be estimated either with respect to iis absolute

amount, or to the amount of relative advantage it affords the

individual who receives it. This consideration has led to a

distinction between mathematical and moral expectation.

When we place the circumstances of the individual entirely
out of consideration, and have regard merely to the abstract
or absolute value of the benefit, the product of its amount
by the probability of obtaining it is the mathematical expec
tation of the individual ; but when a relative value is as

signed to the benefit, the product of this relative value by

the probability of obtaining it is called the moral expecta
tion^ because it is estimated by certain moral considerations

respecting the circumstances or fortune of the individual in

whose favour the expectation exists, on the principle that a
sum of money which may be relatively of very little import
ance to a man in possession of a large fortune may be of

great importance to another who favourably circum

is less

stanced. We shall first consider the mathematical expecta

tion.
CO MATHEMATICAL AND MORAL EXPECTATION.

32. Suppose A and B to engage in play let/> be the pro ;

bability of A s winning a game, q the probability of B s win

ning it, and s a sum of money staked on the issue of the

game. By the definition, the mathematical expectation of

A is ps, and that of B is qs. Now if we suppose these expec
tations to be purchased by A and B, the sums they ought
respectively to pay for them, or in other words to stake on
the issue of the game, must be proportional to their respec
tive expectations, in order that they may play on equal terms.
Let therefore a be the sum staked by A, and b the sum
staked by B, we have then ps qs a b, and consequently
: : : :

pbqa. Now suppose a~\-b=:s, or that the sum played for

is the amount of the stakes ; then, since b is the sum A ex

pects to gain, and p is the probability of his gaining it, pb is
the mathematical value of As expectation of gain. In like

manner qa is the mathematical value of B s expectation of

gain. Hence it follows, that when the sum staked by each is

proportional to his probability of winning, the mathematical

expectations of the two players are equal ; so that after the

stakes have been placed, and before the event is decided,

they might exchange places without advantage or disadvan

tage to either. It follows likewise, that since the sum which
the one must gain is
just that which the other must lose, the

product qa, which is B s expectation of gain, may be regard

ed as A s expectation of loss or (if taken with a negative
;

sign) as part of A s whole expectation, which then becomes

pb qa. Butj?> ^#=0; whence the condition of A before
the event is decided is not altered by the circumstance of
his having staked on the issue of the play.
33. This conclusion at first sight appears paradoxical ; for

it is certain, that after the stakes are placed, A must either

gain the sum b or lose , and therefore his fortune will of
 MATHEMATICAL AND MORAL EXPECTATION. 61

necessity either be increased by the gain of his adversary s

stake, or diminished by the loss of his own. The explana

tion depends on theorems which will afterwards be demon

strated relative to the repetition of trials, from which it re

sults, that though in a single trial the player must either

lose or gain, yet on multiplying sufficiently the number of

games, a probability will at length be obtained, approaching
as nearly to certainty as we please, that the sum gained or

lost in the long run will not exceed a certain given fraction

(which may be as small as we please) of the whole sum

staked, provided the play is undertaken on terms of mathe
matical equality. But this indefinite repetition of the ha
zard is
practically impossible and innumerable cases may
;

easily be imagined, in which an individual will be guided

by other considerations than the mere mathematical value
of the expectation in undertaking or declining a risk. A
person of moderate fortune would scarcely be persuaded to
risk L.500 for the expectation of gaining L.5, though the

chances might be 100 to 1 in favour of the event which

would produce that sum ; but numbers would be found wil

ling enough pay L.5 for the expectation of gaining L.500,

to

the chances being 100 to 1 against them. In both cases,

however, the expectation would be purchased at its real

abstract value. According to the formula of mathematical

expectation, the man whose sole fortune consists of a lot
tery ticket which has an equal chance of turning up a prize
of L.20,000 or a blank, an equally advantageous posi
is in

tion as he who is in possession of L.I 0,000 ; yet no man

of ordinary prudence, if offered his choice of the two states,

would hesitate as to which he ought to give the preference.
Common sense will prevent a man from risking a sum, the
loss of which would be attended with great privations, even
 2 MATHEMATICAL A^ D MORAL EXPECTATION.
7

when, mathematically speaking, the chances are consider

ably in his favour. It is also obvious that two individuals

whose fortunes are very unequal cannot engage in play with

the same advantage, although the chances in favour of each,
in respect of a single game, are precisely the same. The
one who has a large fortune can repeat the hazard so often
as to obtain a probability almost equal to certainty that his

loss will not amount to any given sum ; whereas the other,
who cannot continue the play in case of loss, runs the risk
of being ruined. It is thus evident, that in a multitude of

cases the abstract theory of probability is not alone sufficient

an expectation, and that in dealing with
to give the value of

contingent events, an individual must be guided to a cer

tain extent by considerations of relative advantage.
34. Various hypotheses have been imagined for the pur

pose of reducing such relative or moral considerations to ac

curate calculation but that which appears the most natu
;

ral, and applicable to the greatest number of cases, consists

in supposing the relative value of any infinitely small sum

to be directly proportional to its absolute value, and inversely

as the fortune of the individual who has an expectation of

receiving it. This principle was first proposed by Daniel

Bernoulli in the Petersburg Commentaries (vol. v.), and is

there applied by him to the solution of a number of ques

tions of great practical interest.

Let x be the absolute value of the capital, or, as it is de

nominated by Laplace, the physical fortune, of an individu
al ; then, according to the hypothesis of Bernoulli, the mo
ral advantage which he derives from an infinitely small incre

ment of fortune dx> is measured by the expression c ,

c
x
being a constant to be determined by the nature of the ques-
 MATHEMATICAL AND MORAL EXPECTATION. 63

tion. No\v, if we suppose the physical fortune to arise from

the accumulation of the elements dx, and denote by y the

relative or moral value of the fortune, of which the absolute
or physical value is a?, we shall have
r ax
dx
yl c =c log. x -f- constant.

To determine the constant, we may suppose yo^ when x

has a given value = ; this gives c log. a-f-constant,
/v

whence yc (log. x log. a), or y=c log. ; and it is to

be observed, that those values of x and yean never become

negative, for as Bernoulli has remarked, it is
only the per
son who is dying of hunger that can be said to possess ab

solutely nothing. In every other circumstance the mere pos

session of existence may be accounted a moral advantage, to

which, however, it would be absurd to attempt to assign a nu

merical value.
35. From the above formula, it is
easy to deduce a nu
merical expression for the value of a moral expectation. Let
a be the original fortune of the individual, and a, /3, y, &c.
sums to be received on the occurrence of certain contin

gent events, E, F, G, &c. This being supposed, if the event

E happens, the absolute fortune of the individual becomes
a-j-a, and its relative value, therefore, according to the for

mula, is c log. - . If F happens, his absolute fortune be

comes -f/3, to which the corresponding relative value is

c log. - ; and so on. Now, let the probabilities of the

events E, F, G, &c. be respectively p, q, r, &c. (assuming

p _j. q jf r 4. &c. = 1 , so that one or other of the events will ne

cessarily happen), and let Y represent the relative fortune of

64 MATHEMATICAL AtfD MORAL EXPECTATION.

the individual arising from his expectation, then, since the

value of a benefit in expectation is equal to the amount of
the benefit multiplied by the probability of obtaining it, we
have

Y=c log.
--a +fJ log. -*+r log.
*
+ &c. }
{?
Let also X denote the absolute value of Y ; then, by the
X
formula, we have Y=c log. . On comparing these two

values of Y, we get

log.
^log. ^ -K "og. +r log. + &c. ;

and on passing to numbers,

X g a

therefore, since + + r-f- &c. =1,

/? </
r
X=(a + a)(a + /3)(a + y) ,
&c.
In this expression X denotes the absolute value of the

original fortune and of the expectation added together if, ;

therefore, we deduct a from X, the difference will be the

value of the expectation, or the sum which, if it were to be
received certainly, would procure the individual the same
relative advantage as his expectation.

36. If the sums a, & y, &c. are supposed to be very small

in comparison of a, so that quantities of the order J J

may

be neglected, the preceding equation becomes

-i f &c.

whence, since p-\-q-}-r-\- &c.z=l,

 MATHEMATICAL AND MORAL EXPECTATION. 65

Deducting from this the original fortune a, the remainder

pa~\-q[B-\-ry--8zc. is the value of the expectation, or the sum

equivalent to themoral advantage. But the value of the ma
thematical expectation of the benefits a, )3, -y, &c. of which the

probabilities are respectively p, q, r, &c. is also pa-{-qj3-{-ry

-f-&c. (31), therefore, when the contingent benefits are very

small in comparison of the original fortune, the moral ad

vantage and the mathematical expectation are sensibly the

same.
37. From the formula X= (a -f a)p (a + $y*( a -\-y} r &c. Ber
noullideduces the consequence that gambling or betting is
attended with a moral disadvantage, even when the chances
of gain or loss, mathematically speaking, are perfectly equal.
To shew this, he proposes the following question. A, whose
fortune is 100 crowns, bets 50 crowns with B, on the issue
of an event of which the probability is ^, on these terms :

if the event
happens, A is to receive from B 50 crowns ; if
it fails, he is to pay B 50 crowns ; what is the relative va
lue of As fortune, after undertaking the bet, and before the
event is decided? In this case, we have a=:100, a 50,
50,7=0; also/?=J, <?=J, r=0; and the formula (35)
becomes

x (100 50)

whence X=^/150 x50=87 ; and, consequently, the con

dition of A is
worse by 13 crowns than it was before he ha
zarded the bet. The moral disadvantage is therefore equi
valent to this sum, though the terms of the play, according
to the mathematical theory, are equal.

38. The conclusion arrived at in this particular case is

easily shewn to be universally true. Let a be the capital

of the player, p his probability of winning, q his probability
G6 MATHEMATICAL AND MORAL EXPECTATION.

of losing, and s the sum at stake. In order that he may

play on terms of mathematical equality, the part of the
stakes contributed by himself, or the sum which he can lose,

must be ps (32), and the part contributed by his adversary,

or thatwhich he may gain, must be qs. The equation in

(35) therefore becomes

X= (a -j- qs)P X (aps)i ,

and if it can be shewn that this value of X is less than ,

it

will follow that his condition is rendered worse in conse

quence of having staked on the game. Now, dividing by

,
and taking the logarithm of both sides of the equation, we

get log. =jo log. f 1

-}-
1
-f- q log. ( 1 the diffe
J,

rential of which (making s variable) is

X
d\og. =pqds
^ a a
But the second side of this equation is
evidently negative ;

therefore d log. X-j-a is

negative ;consequently the loga
rithm of X-f-a is
negative, and X must be less than a. In
all cases, therefore, the bet, if on even terms, produces a
moral disadvantage.
39- Another consequence deduced by Bernoulli from this

theory of moral expectation, is, that when property of any

it is more advantageous
kind is
exposed to a risk or hazard,

to expose it independent of each

in parts to several risks

other, than to expose the whole at once to a single risk, al

though the probability of loss be in both cases precisely the

same. To prove this, he takes the following example. mer A

chant has a capital of L.4000, besides goods of the value of
L.8000, which must be transported by sea. The probabi
lity of the loss of a vessel in the voyage being -j^,
let it be
 MATHEMATICAL AXD MORAL EXPECTATION. 67

proposed to find the value of the moral expectation of the

merchant in the case of the goods being embarked in a single

vessel, and also in the case of one half being embarked in one
vessel and the other half in another. Supposing the mer
chandise embarked in one ship, the absolute fortune of the
merchant will be increased to L.I 2,000 in the event of the

safe arrival of the ship, \md will be reduced to L.4000 in the

event of its being lost. The probability of the first of these
events is
T%, and of the second
J5 ; therefore his absolute
fortune becomes, in virtue of his expectation,

X
whence X= 10751. Deducting his other capital, L.4000,
there remains L.675 1 for the value of the moral expectation
in respect of the venture.
Let us next suppose the merchandise embarked in equal

parts in two ships. In this case there are three compound

events to be considered, 1st, Both vessels may arrive in safety ;

the probability of which is X y^

= 7
y
2d, One may

arrive in safety and the other be lost ; the probability of

9 1
which, as it
may happen in two ways,
(ll)is2Xy^X
1 8
==
TOO
3C * ^ Ot ^ ma^ ^e ^ St ^e P r k akility f wm ch is

To
* To
= Too*
^ ^ie ^ rst ^ ^ese events happen, the

capital of the merchant will become L.4000 + L.8000=

L.12,000; if the second happen it will be L.4000 -f L.4000
= L.8000 ; and if the third happen it will be only L.4000.
With these numbers the formula becomes

X=(12 ; 000)
T ^ x (8000)^ X (4000)T K
 KT1CAL AM> MORAL KXTECTATION.

w hence X= 11 0:U. Deducting his other capital, which was

<1 to no risk, there remains 1*7033 for the value of the
moral expectation. This sum exceeds the former by I-.2S2 ;

and it is ea>il\ found b\ following the same process e.f rea

soning, that in pro|>ortion M

.
numb* .
:he mor;,.

ed, and approaches its limit, which f! the value of the ma

thematical expectation, or ^ of LjBOOOsLJTVXK
40. The theory of moral expectation enables us likewise to
assign the circumstances in which it is
advantageous or
otherwise to insure property against particular hazards.
There an* three principal questions to be considered ii .

rence to The amount of premium the insur

tins subject ; 1.

ed may pay without disadvantage 2. The ratio of his for ;

tune to the value of the stun exposed to risk, in order that it

may be advantageous to insure at a given premium ; and 3.

The capital which the insurer or underwriter ought to pos

se**, in order that he may insure a given risk with probable
advantage to himself, and safety to the insured.
1 ,ei * be the value of a
cargo which a merchant emb.,
A ship, p the probability of the safe arrival of the vessel, and
n his capital independently of s. The mathematical \ alne of
the premium for insurance is $; for, if we denote the pre
mium by y* then .vis the sum the insurer will gain if the vend
reaches its destination in safety, and * */ is the sum ]

lose if it does not ; and by the theorem for the mat hen

expectation ;>y=r q(* y) ; whence, since ;> + </=!. y =.$.*.

\ f, therefore, the merchant insures i : . .
^sol ute for-
tune becomes a + * qx .,nd if he does not insure,
it is tlie value of X in the equation X=r. (^-|-.<)
r c
. Hence
it will be advantageous or oil -
insure according as
is
greater or h |
-^->^ ./.
/

\on the logarithm

 M \ I III M \ I d \ I \ M. M..K \ll.ri.|\| ,,.

"i tin in i "i d" <

i
id", (/ ) ,

quh
i

pi | /

li lil Id ill. mli "i.il / MM) lli, ld".n illim ,,| flu ,

"
( i, i id. di /. id i,/
|
,

)-f-y log//,
i

equi
.di i.Mn /
,
/ "-|--v

UK . i ... i " i"",. a i,

.
M,.,,, a
/ propel | /,..
,

ther( fbri the in infc ifc than ii nd

gnl
i . .
<
ton
in ", n, "i, .id
<|ui nll\ i.il. in
( / i :. i lli. ,/
j //A (
|

lli. in in in. . i .ill, nd, d \\ if li .idv.inl.i", .

Let Ul i"

nine i ,,
\ /> (,, j
s
)
,/ . md i will In- II,, mm || M m, i

li.ii l cdiild .illdid I,.

p. ls
id, n, m, i .,!... , lli, mill,, m.ili

.In, dl lli, ii I \\ illioiil nidi.il di .i.K.ml.i", . ||

ih. in
l

VA j
, i., i, i.,!,v, fortuni i in. ir.i-.rd
by ii

..id If he -.-IN .
in"" In- .1 loner. In pruefice tbe
|
pre
mium m;i\ In- I. nil M leHM than (/v-f..r, but greater
Hun V dlh.il ; ulnlr lli, m MM.I pfiyH HIOI C llliUl tllC IHll-

IK MI. in, ..I \.diK dl id, rink, he gaiim a moral

advantage by
th< I I -HI .1. I Kill.

To Hdl\ ih, mud .

i|in
i
ion, let c be the prc^rniuni de-
m.md, d idi m mm" ,,
.unoutit*; then, tbe other capi-
.1 MI id, m, ii ii.mi in- fortune being imured
t

being O| ftfttr
i " A / ; ululr il d, i.d ..--. (I.,- rink on value
j
hiniMelf, it.M

beCORK ("| M / . II, ill. n I,,., ,

ili, value of // be deter-
m:iK d hdin HK (
.jiLilKiii ,r -. ,
( we nlluM have
| j s)"<i\
il ;midiiiii nl i In (Hi-dii in in order that
:i|ni.il pdHHeHH il

m..\ l, mui-iillif :\ m..ti, i ..I md il, M .,,< to him whether he

m m, , or noi. A ,m ,

tuunple, let tbe value of the tncr-

(!i;oi.l, .. 01 .
be I
/
=L.HOO, and />=};;. Tin-
<
ijii.it
KHI I IK n In < dim

" nd iu :5043. It follow

appro .
 rO MATHEMATICAL AM> MOK.M. t: X I F.CTATIOX.

therefore, that unless his other capital amount to L.5043, it

would t)e disadvantageous to neglect insuring, although the

premium demanded exeeed the mathematical value of the

risk (which is
^ X L.UU)00=L.500) by L.300.
The third question, the amount of capital the underwriter

ought to |
determined precisely in the same way.

Let b be his capital. After accepting the risk of the sum >

tor the premium t\ his capital will become h-\-e in the case

of the vessel arriving in safety, and b s-{-e in the case of

its being lost. The formula of the moral expectation there

fore becomes X=(&+ e)p (b s-f-f)
7
; and in order that there

may be neither advantage nor disadvantage in undertaking

the risk, this value of X must be equal to his original capi
tal, l>.
Supposing, therefore, ,v, e, p, q> to have the same sig

nifications as above, the equation from which b is to be de

termined is
b=(b+8W)$*(b 9200)
25
, whence b= 14243.
Unless, therefore, the capital of the insurer amounts to

L.I 4,243, there would be a moral disadvantage in undertak

a cargo worth L.10,000 tor a pre
ing the risk of insuring
mium of L.800 ; and it is easy to see, that if a smaller pre
mium were demanded, the capital ought to be still
greater.
On making e=600, (which still exceeds the mathematical
value of the risk), the value of b becomes L. J; 1^878. I lence

it follows, that a company possessing a large capital may not

in speculations which might prove

only with safety engage
ruinous to another whose resources are more limited, but
1
even derive from them a sure profit.
41. The theory of moral expectation which we have now

been considering had its origin in a problem proposed by

1
See the Commcntarii Acad. AfropofitaMB, torn. v. ; Laplace,

Thhrie dcs Prob. p. 4. 3 2 ; Lacroix, Trait J Ekmcntaire, p. 132.

 MATHEMATICAL AXD MORAL EXPECTATI 71

Nicolas Bernoulli to Montmort^ which, from its

having been
discussed at great length by Daniel Bernoulli in the Peters

burg Memoirs* has been usually called the Petersburg prob

lem. It is this A and B play at heads and tails. A agrees
:

to pay B 2 crowns if head turn up at the first throv,

crowns if it turn up at the second, and not before 8 ; if

it turn up at the third, and not before ; and, in general.

2 n crowns if it turn up at the nth throw, and not before :

required the value of B s expectation ? Here the proba

bility of head turning up at the first throw is ^ the proba ;

bility of its turning up at the second, and not at the first, is

i Xi i ; the probability of its not turning up either at the

or second, and of l
-: its
turning up at the third, X-^xi
= J, and so on. Hence the probabilities of B receiving 2,
16 2" crowns

are respectively
-^ -p
,
^ ,
consequently (31)

the mathematical value of Bs expectation is

Y X2+-5-X4+ -i X-+^Xl6... + _Lx2 crowils .

Now. as no limit can be assigned to n, inasmuch as it is

possible thathead may not turn up till after a very great,

or any assignable number, of throws, this series, of which
each term is unity, may go on for ever, and consequently
the value of Bs expectation becomes infinite. Yet it is ob
vious that no one would pay any considerable sum for the
expectation. This disagreement between the dictates of
common sense and the results of the mathematical theory,

appeared to Montrnort to involve a great paradox ; although

the question differs in this respect from no other question
of chances in which the contingent benefit is
very great, and
72 MATHEMATICAL AND MORAL EXPECTATION.

the probability of receivingit

very small. If the play could
be repeated an infinite number of times, B might undertake
to pay without disadvantage any sum, however large, for his

expectation. A result, however, more in accordance with or

dinary notions, is obtained from the principle of Bernoulli.
Let a be the amount of Bs fortune before the play begins,
x the value of his expectation, or the sum he pays A in con
sideration of the agreement, and make z=a x. If head

turn up at the first throw, Bs fortune becomes z~\-2; if at

2
the second, and not before, 2-J-2 ; if a t the third, and not
3
before, ;z-J-2 j and so on. But the probabilities of these

events being respectively ,

-^-
........ ,
the formula
2 4 o 2i

for the moral expectation becomes (35)

Now the sum which B ought to pay will be determined by

making the value of his moral expectation, after the bet,

and before the play begins, equal to his previous fortune ;

we have therefore a=X, that is,

n 2n
The general term of this series being (z-J-2 )
/ z \ 2n
22( 1+ J
, the equation may be put under the form

and since the logarithm of the first factor of this expression

 MATHEMATICAL AND MORAL EXPECTATION. 73

2 log we have =2 log 2+

2, log
log(l +y)+ -j-
log

from which a value of z may be found by trial and error

for any given value of a. Suppose z=lQO ; on computing
the first 10 terms of the series there results a= 107-89,
whence (since x=a z) #=7 89 ; that is to say, if B pos
sessed only 100 crowns before beginning the play, it would
be morally disadvantageous for him to risk 8 crowns for the
expectation, although its mathematical value be infinitely
If we suppose z 1 000, the sum of 1 1 terms gives
great.

a=:1011, nearly; so that if B possessed a fortune of 1011

crowns, the value of the moral expectation would, to him,
be about 1 1 crowns.
It is scarcely necessary to remark, that the results de
duced from the principle of Bernoulli are of a character

widely different from those which are calculated according

to the mathematical expectation. The latter gives the pre

cise value of a contingent benefit, without any assumption or

hypothesis respecting the personal circumstances of the indi
vidual who may gain or lose it ; whereas the considerations
of relative advantage, of which it is the object of Bernoulli s

theory to take account, are entirely arbitrary, and by their

very nature incapable of being made the subject of accurate

computation. It is evidently impossible to have regard to,

or appreciate, all the circumstances which may render the

same sum of money a more important benefit to one man than
to another; and consequently every rule that can be given for

the purpose must be liable to numerous exceptions. The

principle, however, is thus far valuable, that it
gives in the
most common cases a plausible and judicious estimate of
74 MATHEMATICAL AND MORAL EXPECTATION.

the value of things which are not susceptible of exact ap

preciation ; and it has the advantage of being readily sub

mitted to analysis. A different principle, proposed by the
celebrated naturalist Buffon, consists in making the value
itself of a casual benefit, instead of its infinitely small ele

ments, inversely proportional to the fortune of the expec

tant ; but as this hypothesis has seldom been adopted, it is

unnecessary to discuss it in this place.

 FUTURE EVENTS DEDUCED FROM EXPERIENCE. 75

SECTION V.

OF THE PROBABILITY OF FUTURE EVENTS DEDUCED FROM

EXPERIENCE*

42. In the preceding part of this article it has been as

sumed, in every case, that the number of chances favour

able and unfavourable to the occurrence of a contingent
event is known a priori, and consequently, that the proba

bility of the event, or the ratio of the number of favourable

cases to the whole number of cases possible, can be abso

lutely determined. But in numerous applications of the

theory of probabilities, and these, generally speaking, by far
the most important, the ratio of the chances in favour of an
event to those which oppose it is altogether unknown \ and
we can form no idea of the probability of the event except

ing from a comparison of the number of instances in which

it has been observed to happen, with the whole number of
instances in which it has been observed to happen and fail.

In order to assign the probability of a contingent event in

such cases, it is
necessary to consider all the different causes
or combinations of circumstances by which the event could
possibly be produced, and to determine its probabilities suc

cessively on the hypotheses that each of these causes exists

to the exclusion of all the others. The comparative facili

tieswhich these hypotheses give to the occurrence of the

event which has actually arrived, will then enable us to de-
76 PROBABILITY OF FUTURE EVENTS

termine the relative probabilities of the different hypothe

ses, and consequently their absolute probabilities, since their

sum is
necessarily equal to unity ; and when the probabili
ties of the different hypotheses, and of the occurrence of the

event on each hypothesis, have been determined, the pro

bability of the event occurring in a future trial will be found

by the methods already explained.

43.Taking a simple case, let us suppose an urn to contain
4 counters, which are either white or black ; that the num
ber of each colour is unknown, but in four successive draw

ings (the counter drawn being replaced in the urn after

each trial) a white counter has been drawn three times, and
a black one once ; and let it be proposed to assign the pro

bability of drawing a counter of either colour at the next

trial.

In the present case three hypotheses may be formed re

lative to the number of white and black counters in the urn.

1st, The urn may contain 3 white counters and 1 black

2d, It may contain 2 white and 2 black 3d, It may con ;

tain 1 white and 3 black ; for a counter of each colour hav

ing been drawn, the other two possible cases, namely, that
they are all white or all black, are excluded
by the observa
tion. Now, p lt p2 pv be the probabilities respectively of
let ,

drawing a white counter on each hypothesis, and q^ q.2) q3) the

probabilities of drawing a black. Supposing the first hypo
thesis to be true, or that the compound event which has been
observed was produced by the cause indicated by that hypo

thesis, we havejpjrrj, <7i=i ; and the probability of the ob

served event, or that 3 white counters and 1 black would be

drawn, (12) is
4pf gr^fj. The second hypothesis gives
pt\, q3 ^ whence 4p23 q a =^- The third hypothesis
3
gives >
3 =J, 23=|> whence 4/? 3 3=^. The probabilities
 DEDUCED FROM EXPERIENCE. 77

of the observed compound event, on each of the three hy

potheses, are therefore, respectively, |, J, fa

and the ;

question now arises, how are the probabilities of the differ

ent hypotheses to be estimated ? As we have no data, a

priori, for determining thi> question, we must assume the

probabilities of the different hypotheses to be respective

ly proportional to the probabilities they severally give of

the observed compound event ; in other words, we must
assume the probability of any hypothesis to be greater or less

according as it affords a greater or smaller number of com

binations favourable to the event which has been observed

to take place. Thus, if C and Qbe two independent causes

from which an observed event E may be supposed to arise,
and C furnishes 20 different combinations out of a given num
ber, favourable to the occurrence of E, while C\ furnishes

only 10 such combinations out of the same number, we na

turally infer that the probability of the cause C having ope
rated to produce E, is twice as great as the probability that
the event was produced by the operation of the cause C^
Applying this principle to the present example, the probabi
lities of the three hypotheses are respectively proportional
to the three fractions |, , fa, or to the numbers 27, 16,
3 ; no other hypotheses are admissible, the sum of
and as

their probabilities must be unity ; therefore, making sr the l

probability of the first

hypothesis, w2 that of the second, and
TT3 that of the third, we have
27 16 3
*- 2
i=46 =46> ^=46-
44. Having found the probabilities of the different hy
potheses, that of drawing a white counter at the next trial
is obtained without difficulty ; for according to what was
78 PROBABILITY OF FUTURE EVENTS

shewn in (9), the probability of this

simple event must be

equal to the sum of its probabilities relative to the different

hypotheses, each multiplied into the probability of the hypo

thesis itself. Now it has been seen that, on the first hypo
of drawing a white ball is | on the
thesis, the probability ;

second f and on the third ; and that the probabilities of

,

the hypotheses are respectively J j, f -fa ; therefore the ,

probability of a white counter being drawn at the next

trial is

5x
1 46
T4 ^+-*^+l
X 46 T 4 x-5- -^
46"184
1

In like manner, the probability of a black counter being

drawn at the next trial is

+A
27 68
4_JL X 1?
T X JL-
IX +T ;
4 46 46 46~~184

and the sum of these two fractions is

unity, as it
ought to be,
*
drawn must necessarily be white or black.
since the counter

45. The reasoning which has been employed in this par

ticular case is of general application. Let E be an observed

event, simple or compound, of which the particular cause is

unknown, but which may be ascribed to any one of the

causes, C,, C2 C3
, , CM , which, before the event has

happened, are all equally probable, and such that the opera
tion of any one of them excludes that of the others, so that
the event E is produced by one of them alone, and not by the

joint agency of several of them. Let the probabilities of the

observed event E on the hypothesis that it has proceeded
from each of those causes be respectively P,, P 2 P 3 ,..,P W , ,

so that if the cause, for instance, C were the true one, the t

probability of the event E, previous to the observation,

would be P f ; and let the probabilities (as determined by the
 DEDUCED FROM EXPERIENCE. 79

event) of the existence of the different causes be respective

ly v lt w 2 ,
-

3,
......... wn . From the principle laid down in

the preceding paragraph, namely, that the probabilities of

the different causes or hypotheses are proportional to the

probabilities they respectively give of the observed event,

we have

whence, making P i -f P 2 , +P s
...... -f- P n = 2P,., and observ
ing that w 1 -j-z3- 2 -j-
r
3 ......-f-wa =1 (since it is assumed
that there are no other causes than those specified from
which the event could arise), we have

whence it
appears that the probability of each hypothesis re
specting the cause of the observed event is found by divid

ing the probability of the event on the supposition that that

particular cause alone existed, by the sum of its probabili
ties in respect of all the causes. Let us now assume the
probabilities of a future event E r
(which may be the same
with E or different, but depending an the same causes) in

respect of the several hypotheses, to\}e,p 19 99 p s j ..... ./?* ; p

so that if the particular cause C, be the true one, the proba

bility of E is
pi and let ; U be the probability of E in respect
of all the causes, then by (9)? n will be equal to the sum of
the probabilities p^p 2 p 3 ......pn, relative to the different

hypotheses, each multiplied by the probability of the hypo

thesis ; that is to say we shall have

or n=2p,w,, the symbol 2 indicating the sum of all the dif

ferent values of jo and v in respect of the different causes

Cp C 2 C 3 , , ...... Cn .
 80 PROBABILITY OF FUTURE EVENTS

46. It may be worth while to remark that the word cause is

not here used in its

ordinary acceptation to denote the com
bination of circumstances, physical or moral, of which the
event is a necessary consequence. In the sense we have
used the term, the cause C is that which gives rise to the
determinate probability P, that the event E will happen ;

but so long as this probability falls short of certainty, its

existence also implies that of another probability, 1 P,

that the contrary event F will happen. If we make P=l,
the existence of the cause C would necessarily involve the
occurrence of E ; and it is in this particular sense that the

word cause is
ordinarily used. In the theory of probabi
lities the causes of events are considered only in reference
to the number of chances they afford for the occurrence of

those events which they may possibly, but do not neces

sarily, produce.
47. The following example may serve to illustrate the
method of applying the preceding formulae. An urn
contains n balls, which are known to be either white or
black. A ball is drawn at random and found to be white ;

required the probability of drawing a white ball at the

next trial ?

In this case, the number of hypotheses that may be made

respecting the contents of the urn, is n ; for we may sup
pose that it contained one, or two, or any number of white
ballsfrom 1 to n, and each of these cases may be consider
ed as a distinct cause of the observed event E. Let these
causes or hypotheses be C 15 C 2 C 3 , , C n and let us
,
sup
pose the true cause was C,, or that the urn contained i white
balls. On this hypothesis the probability of the observed

event E is -
, whence P =-
f ; and therefore, making i succes-
n n
 DEDUCED FROM EXPERIENCE. 81

sively equal to 1, 2, 3,....w, we have 2P< 2-[- 3.. +);*.)

n-(!-{-

But the sum of this arithmetical series is ^- - , therefore

2

2,P=J(ra -[-!), and consequently,

Wf
_P _ 2?

~2P,~W!~
which is the probability of the assumption that the event

proceeded from the cause C,, or that the urn contained i

2
white balls. If we suppose i n we have sr B =r - for the

probability that all the balls are white ; and if we also sup

pose n 3, this becomes ^ ; whence if an urn contain 3 balls

which must be either black*or white, and a white ball be
drawn at the first trial, it is an even wager, after the trial,

that all the balls are white.

48. Having found, from the observed event E, the pro
babilities of the different hypotheses, we have now to deter

mine the probability n of the event E (the drawing of a

white ball) at the next trial. Here two cases present them
selves ; according as the ball is
replaced in the urn, or is not ;
or in general, according as the law of the chances remains

constant during the series of trials or varies.

1st, Let us suppose that the ball has been replaced in the
urn. In this case the probability of the event E , on the hy

pothesis that the urn contains i white balls, is ; that is to

say pf==. But the probability zrt of this hypothesis, as

found above, is ; therefore whence

/?^,-^
82 PROBABILITY OF FUTURE EVENTS

the general formula (45) n= 2/?, w,.becomes n=E -%- =

2
2-
Ze 2 . Now Si 2 =2z(*+ 1
)
2z. But by the pro

perty of the figurate numbers referred to in (23), the sum

of the series of numbers obtained by giving i every value

from l to izzn in the formula -

1 .
~
2
is
expressed by

; therefore S?

We have also as above

consequently
.
a _n(
3

and therefore

2.3
2d, Suppose the ball which has been extracted is not re

placed in the urn. In this case, on the hypothesis that the

urn at first contained i white balls, the probability of draw

----;
ing a white ball at the next trial is
ft I
that is,/^^^-
n 1
;

and the probability of the hypothesis is the same as in the for-

2i(t 1)

2 ^ 2 "~ 1 ^ Now
and consequently 11=27^=
_ \\ n (n+l}

the value of 2i(i be found by writing

1) will evidently
1 for n in the above expression for 2i(+ 1 ) whence ;
 DEDUCED FROM EXPERIENCE. 3

~v
2z(z
i\
1
)=
(
v
H l) w x w
( - +l)
- ,
l ^
r *u-
and, therefore, in this case
o

n_
2 2
y (*-l )*(*+!)_
A
-(rc_l)ra(ra+l) 3 -3
When n is a very large number, the ratio of 2n -j- 1 to
3, the value of n in the former case, does not sensibly
differ from f and therefore in both cases ri f
, Hence it
follows, that if an event, depending on unknown causes, can

happen only in one of two ways, and it has been observed

to happen once, the odds are two to one in favour of its

happening in the same way at the next occurrence.

49. The expression for & in (45) was determined on the
supposition that previously to the experiments being made,
we are entirely ignorant of the relative numbers of the two
sorts of balls in the urn, and have no reason to suppose one
hypothesis more probable than another. If, however, we
happen to know, previously to the experiment, that the dif
ferent causes C 1? C 2 C 3 , , &c. have not all the same num
ber of chances in their favour, or that the probabilities ojf
the different hypotheses have relative values, it becomes ne

cessary to introduce those relative values, in consequence

of which Z3-
1?
w2 , &c., will receive a modification. Let us
conceive a number of urns, each containing balls of two

colours, black and white, to be distributed in n groups,

Aj, Ag A ,
3
An , in such a manner that the ratio of
the number of white balls to the number of black balls

is the same in respect of each urn belonging to the same

group, and consequently that the probability of drawing

a ball of either colour is the same from whichever urn
in the group it
may happen to be drawn, but different in

respect of the different groups ; and let the probabilities of

drawing a white ball from each of the different groups be re-

84 PROBABILITY OF FUTURE EVENTS

spectively P 1? P 2 P 3 , , ...... Pn .
Now, let us suppose there
are a 1 urns in the group A T, 2 in the group A2 ,
and so on,
and let s = the whole number of urns, so that s^^a l + 2

-}-a 3 ...... -\-a n ; then, if we make rzXj, =X 2 ,

and so
s s

on, A! willbe the a priori probability that a ball drawn from

any urn at random, will be drawn from the group A x X 2 ;

the probability it will be drawn from the group A 2 and, ;

it will be drawn from the

in general, X f the probability group
A,. This being premised, suppose a trial to be made, and
that the event E is a white ball ; the probability w of the t

hypothesis that the ball was drawn from the group A, is

found as follows. The a priori probability of the ball be

ing drawn from the group A (

is X4 ; and if the ball is actu

ally drawn from that group, the probability of its

being
white is P, ; therefore the probability of both events is

X P
>,P ;
and consequently (45), w,r=- =5-,
the symbol ofsum-

mation 2 extending to all the values of i from z= 1 to z=:w.

50. In the applications of the theory to physical or moral

events, the different groups of urns here imagined may be

regarded as so many independent causes C 15 C 2 C 3 &c.
, ,

by any one of which the event E might have been produced ;

W| is the probability that the event was produced by the par

ticular cause C, ; P, is the probability that the cause C, if it

had alone existed, would have produced the observed event

E ; and X,- is the probability, previously to the experiment,
that Q would be the efficient cause. The formula *r t -=:

XP
, therefore, shews that the probability of any one of
2X f
i i

the possible causes (C t )

of an observed event is
equal to the
product of the probability (P,) of the event taking place if
 DEDUCED FROM EXPERIENCE. 85

that cause acted alone multiplied into the probability A, that

the cause C.is the true one, and divided by the sum (2A,P,) of
all the similar products formed relatively to each ofthe causes
from which the event can be supposed to arise.
51. The formulae now obtained can only be used when
the number of hypotheses is finite ; but in the applications
of the theory most frequently happens that an infinite
it

number of hypotheses may be made respecting the causes of

an observed event, as would be the case in the above ex

ample if the number of balls in the urn had been unknown.

In such cases, in order to find the values of -sr and n, it be
comes necessary to transform the sums 2 into definite in
tegrals, which is accomplished by means of the theorem

SXi^y^
X dx, where X is a function of x. Suppose a ball
1

to have been drawn a great number of times in succession

from an urn (the number in which is unknown) and re

placed in the urn after each drawing, and that the result
has been a white ball m times and a black ball n times, the

probable constitution of the urn, and thence the probability

of drawing a w hite
r
ball at a future trial will be found as
follows. Assume the hypothesis that the ratio of the num
ber of white balls to the whole number in the urn is x :
I,

and let & be the probability of the hypothesis. On this

hypothesis the probability of drawing a white ball in any

trial is x, and that of drawing a black ball 1 x, and
consequently, the probability of drawing m white and n
black in m -f n trials is Vx m ( 1 x) by (12). We have there
n

fore for the probability of the observed compound event

P=Uxm (l x) ;
n
whence
consequence of the above
in

formula for transforming a sum into a definite integral

m n
2P=U/^a? (l x) dx{\3 being independent of x) and
therefore
86 PROBABILITY OF FUTURE EVENTS

p ar
m
(l *)"
" m
SP f Q
l
x
The value of the integral in the denominator of this frac
tion is obtained by the usual method of integrating by parts.
Since

therefore

In like manner we getjx m + l n ~ l dx

(l x)

Continuing this operation n times, or till the exponent of

x] becomes n n=.0, the last integral will be

(1

//>+&:=
m-f/z+l
t
-
m + n +i
-
;

therefore, collecting the several terms into one sum, we have

1-1
/-m/i y.,7 _
y?
1)(^_ 2) ......... 2.1 .

When #=0, all the terms of this series vanish, and when
x~\ they all vanish excepting the last therefore between ;

the limits #.=0 and ar=l, the value of the integral is the
last term of the series when x in that term = 1 ; that is to

say,

/ ~(i+l)(f
For the sake of brevity, let the symbol [#] be adopted to
 DEDUCED FROM EXPERIENCE* 87

represent the continued product 1 . 2 . 3...X of the natural

numbers from whence by analogy [#+?/] will repre
1 to x, 1

sent the continued product of the same series from 1 to the

number denoted by x+y* Multiplying, then, the numera
tor and denominator of the above expression by 1 .2.3...

...?wi=[m], we get

whence the probability of the hypothesis, in consequence of

the equation above found, becomes

From this value of zr we are enabled to deduce that of n,

the probability of drawing a white ball at the next trial.

By
(45) n=2zrp. Now, since by hypothesis the number of
white balls in the urn is to the whole number of both co
lours in the ratio of x to 1, the probability of drawing a

white ball is#; consequentlyj0=:#, and therefore n=2zrx

^ But the
_ _
1
xm + 1
(l
n
x) dx will evidently be obtained by substi
ofy^

tuting w-f-1 for m in the expression found for C xm

l

n
(1 x) dx. This substitution gives

whence, observing that [m-\- 1] -i- [m~\ = m-^ 1, and

=m-f n-}-2, we have

m+n + 2
1
This convenient notation lias been adopted by Mr. De Morgan.
 PROBABILITY OF FUTURE EVENTS

The probability of the contrary event, or of drawing a

black ball, is 1 n= -- n-le- 1

m-)--t-2
. As the numbers m and n

become larger, these two fractions approach nearer and

nearer to their limits -- and -

7?fc Tfc

,
which are the apri or i

probabilities p and q of the respective events when the ratio

of the number of white balls in the urn is to that of the black

balls as m to n.

52. The probability of drawing m white balls and n

black balls in m -\-
n future trials is found in a similar man
ner, and the problem may be thus stated. E and F are two

contrary events, depending on constant but unknown causes;

and it has been observed, that in m-\-n-=.h successive in
stances the event E has occurred m times and Fn times,

required the probability that in m -{-n =zh future instances,

E will occur m times and Fn times.

Assume, as in the last case, the facility of the occurrence

of E to that of F to be in the ratio of x to 1 x ; we have

then, as before, for the probability of the hypothesis, w=
x m (\ x)
n. Now on this hypothesis the probabi

lity of E in the next instance is a?, and that of F is 1 a?,

whence the probability of m times E and n times F in the

next h? trials being denoted by /?,

we have (12) p=U xm
1 2 .3...h
<!-*), making U = ! . 2 . 3 ...
.

m xl
,
. 2 . 3...,,,
=

We have therefore ^ u/
~WM "
+*

)+ for the probability of the compound event on

this hypothesis. To find its probability n on the infinite
 DEDUCED FROM EXPERIENCE. 89

number of hypotheses formed by supposing x to increase by

infinitely small increments from x to a?=l, we have

n=2zrp=: J"Q zrpdx. On substituting for Tsp the value just

m m
x) + dx, and
n==U n n
found, we get ft x +
JQ (\
O] \_n~\

it is manifest that the value of this integral will be obtained

by substituting m+m for m, and n-\-n for n in the value

J
# m (l n
x) dx found above. This substitution gives

whence we conclude
[^

The most probable hypothesis will be found by making

the value of w a
maximum, or its differential coefficient equal
~
xm (\
ww
*~
to zero. Differentiating the equation zr= x}\

and making
QX
=0, we get 7??(1 cc)
= nx, whence x =
. The most probable supposition, therefore, respect

ing the contents of the urn is, that the two sorts of balls are
in thesame proportions as have been shewn by the previous
drawings. We shall have further occasion for these for
mulae when we come to consider the cases in which m and
n are large numbers.
90 BENEFITS DEPENDING ON THE

SECTION VI.

OF BENEFITS DEPENDING ON THE PROBABLE DURATION OF

HUMAN LIFE.

53. In applying the principles of the theory of probabi

lity to the determination of the values of benefits depend

ing on life, the fundamental element which it is

necessary
to determine from observation is the probability that an in

dividual at every given age within the observed limits of

the duration of life, will live over a given portion of time,

for instance one year ; for when this has been determined
foreach year of age, the probability that an individual, or

any number of individuals, will live over any assigned num

ber of years, is easily deduced. Thus, if the probabilities
that an individual A, whose age isy, will live over 1, 2,

3...X years, be denoted respectively by/^, />, p z .*>px ; and

if </i> <? a> 9s <?*
denote the same probabilities in respect

of an individual whose age T/+ 1 years; r 1 ,r ,r 3 ...r^. the

is
<J ,

same in respectof an individual whose age is y-j-2 years, and

so on ; then, since the probability p2 which A has of living
over 2 years obviously compounded of the probability/*,
is

of his living over 1 year, and of the probability q that, hav {

ing attained the age y+l> he will live another year, we

have, by (7), /> 2 =/ )
1 91 Again, the probability ps that A
will live over three years, being compounded of the proba-
 PROBABLE DURATION OF HUMAN LIFE. 91

bility /? 2 that he will live over two years, and of the proba

bility r l that, having attained the age y-f-2 years, he will

survive another year, we have /> 3 =r/) 2 r l =p q r In like l l l
.

manner J 4 =/) i q^ r i
* lf
and so on ; so that the probabili
ties jo 2 , j 3 , /> 4 /?* are successively derived from j^, q lt
r 15 s lt &c. which are supposed to be the data of observation.
If a large number n of individuals, all born in the same

year, were selected, and if it were observed that the num

ber of them remaining alive at the end of the first year is
nv at the end of the second year ra
fl ,
at the end of the
third e
3 , and so on, then the probabilities p l p 2 p3
, , , c.

would be given directly by the observation, being respec

tively equal to the quotients , , , &c. But the most

n n n
accurate observations of mortality are furnished by the ex

perience of the annuity and assurance where they offices,

are not made on an isolated number, diminishing, and con

sequently giving a less valuable result every year, but on a
comparison of the numbers which, in a series of years, en
ter upoii and survive each year of age. This observation

gives p 19 q^ r 1? st , &c, whence p^ p 3 p 4 , , &c. are found,

1
as above, for every year of life.

54. The values of annuities on lives, and of reversionary

sums to be paid on the failure of lives, are found by com
bining the probabilities pv p l2 , p3 , &c. with the rate of in
terest of money. Let r= the rate of interest, that is to say,
the interest of L,l for a year, and v= the present value of

L.I to be received at the end of a year, we shall then have

r=l-t-(l-f-7*) Now an annuity, payable yearly, is always

understood in this sense, that the first payment becomes due

1
For further details on this subject, see MORTALITY, vol. xv. p. 550.
92 BENEFITS DEPENDING ON THE

at theend of a year after the annuity is created. Suppose

then the annuity to be L.I, the present value of the first

payment, if it were to be received certainly, is v ; but the

receipt of this sum

contingent on the annuitant being
is

alive at the end of the year, the probability of which we sup

pose to be p\ ; therefore (7) the present value of L.I sub

ject to the contingency, is

vp r In like manner, the present
value of L.I to be received certninly at the end of x years
is but the annuity will only be received at the end of
if ;

the #th year if the annuitant be then living, the probability

of which ispx ; therefore the present value of that particu

lar payment is v*px . Hence if A denote the present value

of the annuity, or the sum in hand which is
equivalent to
all the future payments, we shall have A=2v /? s

jr ; the sum
2 including all values of a? from x 1 to x-=. the number for

which p=0. If the annuity be a pounds, its value is ob

viously =. a"2v p x
sc
=aA .

55. The denoted by Ev*p x may be divided into two

series

parts, ^v np n -if- 2?ry; z where n is to be taken from 1 to n, and

,

z from n -j- 1 to the number for which p vanishes. The first gives
the value of the temporary annuity on the given life for n
years, and the second the value of the deferred annuity, that
is to say, of the annuity to commence n years hence if the

individual shall be then living, and to continue during the

remainder of his life. Let A be the value of the annuity
on the life of a person now aged y years for the whole of
life, A ( n) the value of a temporary annuity on the same life

for n years, and A^ the value of an annuity deferred n

years on the same life, we have then Ar=A ( n)

-f A (dM >.

To find A* dn> , let A n be the value of an annuity on a life

aged y-\- n years. If the person now aged y years lives over
n years, the value of an annuity on the remainder of his
 PROBABLE DURATION OP HUMAN LIFE. 93

life will then be A n. The present value of this sum, if it

were to be received certainly, # n A n, and the probability

is

n
of receiving it is
pn ; therefore its value is v p n k n Hence .

A( dn ^=v np n A n and , A>=A vnp n A n -,

temporary and deferred annuities are

so that the values of

readily computed from tables of A and p for all the diffe

rent ages.

56. The equation A=A<* W >

+ A< dn >
gives a formula by
which the values of A are readily deduced from one ano

ther. Let n=l ; we have then A=rA^ 1) -|-?^ i A 1

. But
A (rt)
, the value of an annuity for one year, is merely the
value of the first payment to be received in the event of the

given life surviving one year. Its value is therefore vp\ ; and
we have consequently A =vp l -}-vp l
A 1? orA=?y 1 (l-}-A 1 ).
This formula, which gives the value of an annuity at any

age in terms of the next higher age, and greatly facilitates

the computation of the annuity tables, is due to Euler.

57. The value of an annuity on the joint lives of any num

ber of individuals, that is, to continue only while they are all

living, is calculated precisely in the same manner as the an

nuity on a single life. Let there be any number of individuals,

A, B, C, D, &c. and let the probabilities of each living over
one year be respectively p l9 q^ r^ s 1} &c. and let PI be
the probability that they will all live over one year ; then
P = piX
l q-iX rl x s 19 &c.

P2 = p2 X q2 X rQ X ss, c.

PS =
p x X qx X r x x f# &c.
and the value of an annuity of L.I on the joint lives is
2t-*P a from #=1 to x= the number which renders any one
.,

of the probabilities^, q, r, 5, &c. nothing.

58. The value of an annuity on the survivor of any num-
94 BENEFITS DEPENDING ON THE

her of given lives, that is, to continue so long as any one of

them exists, is thus found. The probability that A will be
alive at the end of the #th year being /?,, the probabi

lity that he
will not be alive at the end of that time is
1 px . The probability that all the lives will be extinct
at the end of the #th year is therefore

and the probability that they will not all be extinct, or that
at least one of them will be in being, is

i-0-/>.)0-?.)(i-^)0-*.) &*
which becomes by multiplication

p xq*p*r*
+P* q* rx +p x qx sx ...... +qx rx -.s
x +&C.
P* q* r x sx &c<

4-&c.
Multiplying each of the terms by v*, and taking the sums of
the respective products from #=1, and observing that
x the value of the annuity on the joint lives of
^.v px q x is

A and B, *Zv x
pxqxrx that on the joint lives of A, B, and

C, and so on, we have this rule :

The value of an annuity on the survivor of any number

of lives is
equal to the sum of the annuities on each of the

lives, minus the sum of the annuities on each pair of joint

lives, plus the sum of the annuities on the joint lives taken

by threes, and so on. When there are only two lives, the

value of the annuity on the life of the survivor becomes

59* Let V denote the value of an assurance on the life

of A, or the present worth of L.I to be received at the end

of the year in which A shall die. In respect of any year,
the #th, after the present, the probability of A dying in the
 PROBABLE DURATION OF HUMAN LIFE. 95

course of that year is/V-i p* For let u be the probabi

lity that a life x 1
years older than A will live over one

year, then 1 u is the probability of a life of that age not liv

ing over one year ; therefore px i being the probability of A

living over x 1
years, Pxi(l u) is the chance of his liv
ing over x 1 and dying in the following year (7).
years,
But px-i (1 u)=px_i px-i u ; and by (53), px ^i u=pf ;
therefore px i px is the chance that A will survive x 1
years and not survive x years. Now v* is the value of L.I

to be received certainly at the end of the #th year ; there

fore in respect of the xth year the value of the expectation

is V X (PJI PX)> whence we have for the value of the as

surance

fromvr=l to #=r the number which makes j=0. Now$ if

l
we observe that /? ==1, and Z.v*p. -.i=.vZv*- p.e__i,
c it will be
x
obvious that Zv x px _\-=.v(\ -\-*Lv p ) ; whence, denoting
x as in (54) the value of the annuity
2v p x by A, (A being
on the given life), we have
V=r(l+A) A; or V=*> (1 v)A.
60; The values of assurances on joint lives, (that is, to be

paid at the end of the year in which any one of the lives shall

fail),
or on the survivor of any number of joint lives, are cal

culated from the corresponding annuities by means of the same

formula. Thus, let A be the value of an annuity of L.I
on any number of joint lives, and the value of an assur V
ance of L.I on the same joint lives, then V =w (1 v)A .

If A" be the annuity, and V" the assurance on the life of

the survivor of any number of given lives, we have still

V"=v (I )A".
61. Assurances on lives are usually paid not in single pay

ments, but by equal yearly payments, the first

being made
96 BENEFITS DEPENDING ON THE

at the time the contract is entered" into, and the succeeding

ones at the end of each future year during the life of the
assured. The present value of the sum which the assured

contracts to pay is therefore equal to the first payment add

ed to the value of an annuity of the same amount on his life ;

and if the assurance is made on terms of mathematical equa

lity, this sum must be precisely equal to the value of the
assurance in a single payment. Therefore, if y denote the
amount of the yearly payment, we have the equation
y(l+A) = V; whence ?/=V-T-(l-f. A).
62. The value of a temporary assurance for n years, that

is, of an assurance to be paid only in the event of the indi

vidual dying before the end of n years is thus found. Let
V be the present value of L.I, to be paid on the death of a

person now aged y years, and V n the present value of L.I,

tobe paid on the death of a person now aged y-\-n years.
At the end of n years from the present time, the value of
L.I assured on the life of a person now aged y years
will be Vn , if he be then But the present value of
living.
L.l to be received certainly at the end of n years is?;" and ;

the probability that the life will continue n years is

p n \ there

fore the present value of Vn , subject to the contingency of

the life continuing n years, is vnp n Vn .
If, therefore, we sub
tract this from V, we shall have the value of the temporary
assurance in a single payment, namely V v np n Vn .

The equivalent annual premium found by observing,

is

that as the first payment is made immediately, and n pay

ments are to be made in all, the value of all the premiums

after the first is that of a temporary annuity of the same

amount for n 1 years. Denoting therefore the annual
premium by u, and the value of a temporary annuity for

n 1
years by A <*"
>, the value of all the premiums is
 PROBABLE DURATION OF HUMAN LIFE. 97

M-fwA<
n > m >
and we have consequently
=rw(l -f)A< ;

=yv n
n V n1 whence

63. The following question is of frequent occurrence.

Required the present value of a sum of money to be receiv

ed at the end of the year in which A dies, provided he die
while B is
living.
Let the sum be L.I, W=r its
present value, p x -=. the

probability of A living over x and q x

years, the probabi ~
lity of B living over x years. The chance of receiving the
sum end of any given year, the #th, depends on two
at the

contingencies ; 1. A may die in the course of that year, and

B live over it ; 2. A and B may both die in that year, A
dying first. The probability of dying in the a?th year A
has been shewn (59) to be px_i px ; whence (7) the

probability of the first contingency is (p x p x )q x The \

probability that and B will A

both die in the .rth year is

(p *_! p x )(q x \ q x ) and for so short a period as one year,

;

itmay be considered an even chance whether A or B will

die first, whatever be the difference of their ages therefore ;

the probability in respect of the second contingency is

%(p x \ Px)(q x ##) \ Hence the whole probability of

the sum being received at the end of the #th year, is

(Px-l p + ^(px-ipx)(q*-iq )=l;(p*-ip

x )q x x x)

fe+i + $ *)> which being developed, and multiplied by vx,

becomes

and the sum of all the values of this expression from

#=1, gives the value of W.

It has been already shewn (59) that 1 v x (p x^i
l
/?*)=
v (1 v)A, where A= the annuity on the life of A. In like
98 BENEFITS DEPENDING ON THE

manner, if we denote by AB the value of an annuity on the

joint lives of Ajind B, we shall have *Zv x (p x-.iq x-.ip*q*)
=-v (1 #)AB, which is the value of an assurance to be

paid on the death of the first dying. Assume p such that

jt/^rrjt/j />.,,_!, then p x is evidently the probability that an

individual A one year younger than A, will live over x years

x _ x xx ; denoting by
Pi Pi
AB the value of an annuity on the joint lives of A and B.

Again, let q x ~q\q x -.\> then q x is the probability that

B who, is one year younger than B, will live over x years,
1 __
and I v xp x qx__i = *Lifpx q x AB ; denoting by AB the
9 i

value of an annuity on the joint lives of A and B . Collect

ing the different terms, we have therefore

W=ifv(l v)AB + 4"A B --7 AB whence W

7
}
P i 9 i

is
easily computed from tables of annuities on joint lives.
If A and B are both of the same age, the two last terms
destroy each other, and W is
equal to ^ the value of L.I, to
be paid on the failure of the joint lives, as it evidently ought
to be, since there is in this case the same chance of A dy
ing before B as of B dying before A.
The formula gives the value of L.I in a single payment ;

the equivalent yearly payment is W divided by 1-fAB, for

the contract ceases on the failure of the joint lives by the

death of either.
It would be easy to extend the formula to the case of an
assurance to be paid on the contingency of the failure of

any number of lives during the continuance of any number

of other lives, or of an assurance to continue only during a

stated time ; but as it is not our purpose to give solutions

 PROBABLE DURATION OF HUMAN LIFE. 99

of the various problems of this kind which may occur in

practice, but merely to shew the manner in which the ge

neral principles of the theory are applied to them, we shall

not pursue the subject farther, but refer the reader to the
article ANNUITIES, and to the standard works of Baily 1 and
2
Milne, in which it is treated in detail.

The Doctrine of Life Annuities and Assurances analytically inves

1

tigatedand practically explained, fyc. By Francis Baily. London,

1813. This werk is now out of print, but a French translation of
it has recently been published at Paris.
2
ATreatise on the Valuation of Annuities and Assurances on Lives
and Survivorships, Sfc. By Joshua Milne. London, 1815.
 100 APPLICATION TO THE

SECTION VII.

OF THE APPLICATION OF THE THEORY OF PROBABILITY

TO TESTIMONY, AND TO THE DECISIONS OF JURIES AND
TRIBUNALS.

64. The case of a witness making an assertion may be

represented by an urn containing balls of two colours, the
ratio of the number of one colour to that of the other
being
unknown, but presumed from the result of a number of ex
periments, which consist in drawing a ball at random, and
replacing it in the urn after each trial. A true assertion

being represented by a ball of one colour, and a false one

by a ball of the other, it follows from the theorem in (51),
that if a witness has made m-\-n assertions, of which m are
true and n false, the probability of a future assertion being
n ~*~
true is -, and that of its being false Let
m+n+2 m+n+2
the of these fractions be represented by v, and the se
first

cond by w, then v is the measure of the veracity of the in

dividual, or the probability of his speaking the truth, and
w the opposite probability, since v+wl. In general, the

existing data are insufficient to enable us to determine the

numerical values of v and w in this manner ; and therefore
in applying the formulae to particular cases, we must assign
arbitrary values to these quantities, foundedon previous
knowledge of the moral character of the individual, or on
 DECISIONS OF JURIES ANt>

some notions, more or less sanctioned by experience, of the

relative number of true and false statements made by men
in general, placed in similar circumstances.

65. Having assumed v and iv, let us suppose a witness to

testify that an event has taken place, the a priori probabi

lity of which is jo, and let it be proposed to determine the

probability of the event after the testimony. In this case

the event observed (E) is the assertion of the witness, and

two hypotheses only can be made respecting its cause 1st, ;

that the event testified really took place ; and 2d, that it

did not. On the first hypothesis the witness has spoken

the truth, the probability of which is v ; and an event has
occurred of which the probability is
p ; therefore (7) the

probability (Pj) of the coincidence is vp. On the second

hypothesis, the witness has testified falsely, the probability

of which is w ; and the event attested did not happen, the

probability of which is q\ therefore the probability (P 2 ) of
the coincidence is wq. Hence, by the formula (47) z^rr:
P,-r-sP )the probability (ar 1 of the first hypothesis becomes
( ,

,
and the probability (zr 2 ) of the second
vp-\-wq
The sum of these two probabilities is unit, a condition
which ought evidently to be fulfilled, since no other hypo
thesis can be made, and consequently one or other of the

two must be true. It is to be observed, that these values

of w-j and sr 2 are the respective probabilities, after the tes

timony has been given, that the event attested took place,
and that it did not.

Since -0-,=
vp
vp
-
+ wq
, we have -or. p = p(v -- vp
-
vp -f- wq
wq)

= p\v(\p)
L-2 -- vp
L-L

wq
wq}
L> *~-i--*/
= p(vqwq)
vp wq
;
.
but v w=v 1 + v=i
 A.MM 1. vTION TO THE

_Y 1, therefore *-./>= ^- ?. This traction being

positive or negative, according as 2r 1 is

greater or less
than unity, or as r is
groat or or loss than i, it follows that
it c^i, then w,^/); that is to say, the probability of the

event alter the testimony is

greater than its a priori proba

bility when the veracity of the witness is

greater than ^.
On the contrary, it the veracity of the witness is less than

J,
the effect of the testimony is to render the probability of

the event less than its a priori probability.

66. If the event asserted by the witness be of such a na
ture that its occurrence is a priori extremely improbable,
so that p is a very small fraction, and q consequently ap

proaches nearly to unity, although at the same time the ve

racity of the witness be great, and measured by a fraction

approaching to unity, the value of w, becomes nearly equal

\op-r-ic, (for on this supposition /j-j-wv/H-r is
nearly equal
to w). But it is obvious, that however great the improba
bility of a witness giving false testimony may be supposed,
the improbability of a physical event may be any number of
times greater ; in other words, however small a value may
be given to ir, the value of p may still be any number of
times smaller ; so that notwithstanding the veracity of the

witness, the probability of the event after the testimony,

namely arjSr/J-f-ir may be less than any assignable quan

tity. On this principle mankind do not easily give credence

to a witness asserting a very extraordinary or improbable

event. The odds against the occurrence of the event may

be so great, that the testimony of no single witness, how
ever respectable his character, would suffice to induce be

lief.

(37. In the case of the character of a witness being alto-

 DECISIONS OF JURIES AJCD TRIBUNAL*.

gether unknown, we may suppose v to have aQ possible values

within certain limits, and to find the value of w l by integrat
ing the fractionyW , </; between those limits. Since ,=:

we liave fa.dc-=. I which on substitut-

tp-f- /
,.

J vp-^wq ,

ing 1 v and 1
/>
for w and q respectively, become*

/, , the integral of which is

^/ I
p + (2p l)v
/ \
] p 1
{fj 2/* 1 ( 2y>
I
j
*
\ J J
C being a constant, the value of which will be determined
from the assumed limits. If v be fllj l iftfil to vary between
the limits r=0 and rz=l, then

and if we assume /?=J, we have /V l <r=r|(l Jlog.3),

which, since the logarithm is the Napierian logarithm, and

Nap. log. 3=1.0986, becomes f X .4507=.676, or nearly f.
Whence we see, that on this hypothesis the probability of
the event is diminished in consequence of the testimony.
68. The credit due to the ^ftiftMFffy of a witness depends

not merely on his good faith, but also on the probability that
he is not himself deceived with respect to the event he as
serts. The chances of a witness being deceived through

credulity or ignorance are much more numerous in general

than the chances of intentional fraud ; and this must be

the case more particularly when the event is of such a na
ture that it
may happen in various ways which may be mis
taken one for another : as for instance, in the case of a lot

tery ticket being drawn, and the witness asserting that it

bears a particular number, which might with equal proba
bility be any other number on the wheel. The following
 104 APPLICATION TO THE

question will illustrate the method of applying the calculus

when a distinction is made between these sources of error.

An urn contains s balls, of which a\ are marked A 1? 2

marked A2 a n marked An A . ball having been drawn

at random, a witness of the drawing affirms that the ball

drawn is marked Am ; required the probability of the testi

mony being true.
Here we have s=:a l +a a -{-a 3 +an , (n being the
number of the different indices or sorts of balls) ; so that if

we make p l =a --s,p 2 =za 2 -i-s l .pm =an -^s, then pi is

the a priori probability that the ball drawn is of the class

marked A pe
1? the probability that it
belongs to the class
whose index is A2 , and so on. It is evident that n diffe

rent hypotheses may be made respecting the index of the

ball which has been drawn, for it
may belong to any one of

the different classes A A 2 ...A n

l?
. Let the probabilities of
these hypotheses be respectively ar
p w 2 ... WB , (that is, in re

spect of any particular index , w t

is the probability after the
assertion that the ball drawn is marked A,) ; and let the

probabilities of the assertion on each of these hypotheses be

respectively P,, P 2 ...P n , (that is, if the ball drawn be mark

ed A,, then P, is the probability the witness will assert it

to be marked A m ). v be the veracity of the wit

Lastly, let
ness, and u the probability that he has not been deceived.
(1.) Let us first consider the hypothesis that the ball drawn
is marked A m and consequently that the assertion is true.
,

In order to find Pm , the probability of the assertion being

made, there are four cases to be considered. 1st, we may

suppose the witness is not deceived himself (w), and that he

speaks the truth (v). The probability of the assertion in

this case is uv. 2d, The witness knows the truth, but in
tends to deceive, or testifies falsely. In this case the proba-
 DECISIONS OF JURIES AND TRIBUNALS. 105

bility of the assertion being made, on the hypothesis under

consideration, is 0. 3d, The witness has been deceived him

self,but intends to speak the truth. In this case also the proba

bility of the assertion being made is 0. 4th, The witness has

been deceived himself, and intends to deceive. In this case
the assertion might be made ; and to find the probability of

its
being made we have to consider, that since the witness
has been deceived, he must have supposed some other index
than A^ to have been drawn ; and since he intends to de
ceive, he must assert some other index to be drawn than
that which he supposes to be drawn. Setting aside, there
fore, the index which he supposes to have been drawn, there

remain n 1 others, any one of which he is as likely to name

as any other. The probability, therefore, of his naming A m
when he intends to deceive is
l~(n 1). Hence the pro

bability of the assertion in this case is

compounded of the
probabilities of three simple events, as follows : 1 . Probabi

lity the witness is deceived =(1 u) ; 2. Probability he in

tends to deceive =r(l v) ; 3. Probability he names Am

= 1 .
:
(n 1
). The probability of the assertion is therefore in

this case =(l Adding this to the pro

w)(l v)-~-(n 1).

bability found in the first case, we have P m ,

the whole pro

bability of the assertion being made on the hypothesis that

the index of the ball drawn was Am , namely

(2.) Let us now consider one of the remaining hypotheses,

and suppose that the ball actually drawn was marked A,, and

not Am , as attested by the witness. As before, there are

four possible cases for consideration. 1st, The witness
knows the fact and speaks the truth. In this case the as
sertion could not be made, or its
probability is 0. 2d, The
 106 APPLICATION TO THE

witness knows the fact, and intends to deceive. In this case

the probability of his asserting A m to be drawn is compound
ed of the probability that he is not deceived (u), the proba

bility that he testifies falsely (1 v), and the probability that,

knowing the index A f to be drawn, he selects A m from among
then 1 which remain after rejecting A (l-=-(w 1)). The
probability of the assertion being made in this case is there
fore u (1 v]-^-(n 1). 3d, The witness is deceived, and
intends to speak the truth. By reasoning as in the last
case, it is easy to see that the probability of the assertion
being made in this case is v (1 u)-s-(n 1). 4th, The wit
ness is deceived, and intends to deceive. The probability
of the assertion being made in this case will be found by con

sidering, that as the witness is himself deceived, he must

suppose some particular index to be drawn different from A f,

(which drawn by hypothesis), for instance A the proba

is ,

bility of which is ]-r-(n 1); and intending to deceive, he

must fix on some index different from A e, which he sup

poses to be drawn ; and he announces A m the probability ,

of which selection is also l-?-(ra 1). The probability,

therefore, that the witness supposes A e to be drawn, and
annnounces Am , is \-s-(n I)
2
. But it is evident, that
whatever can be affirmed with respect to the particular in
dex A e may be affirmed with equal truth of every one of
,

the other n indexes, excepting A,, which is

actually drawn,

(since by hypothesis the witness is

deceived), and Am , which
he announces, (since by hypothesis he lies). There are
therefore n 2 different ways in which he may at the same
time be deceived, and intend to deceive, and announce Am ;

consequently the probability of this announcement in any of

2
these ways is (n 2)-f-(?z I) Multiplying this into the .

probability of his being deceived (1 u), and the probability

 DECISIONS OF JURIES AND TRIBUNALS. 107

-___-
- -
of his giving false testimony (1 v), the probability of the

assertion in this case becomes - - Hence

v\f
the whole probability of the assertion, in all the cases in
cluded in the hypothesis that the ball actually drawn was
marked A,, is

u(l-v)
n1 nl (nl)
2

As this expression will evidently be the probability of the

assertion on any other of the n 1 hypotheses that the ball
actually drawn was marked with an index different from A m ,

the sum of the probabilities of the assertion on all these hy

potheses is 2P Z,
where i is successively each of the num
bers 1, 2, 3,-..n, excepting m.
We have now to find wm , the probability of the first
hypo
thesis. Since the hypotheses, in the present question, are not
all equally probable a priori^ we must have recourse to the
formula (49)ar t
=X P -r-2X,. lp
z J l.,
and consequently in the present
case we have
> Pm
AmJ
_ -

~X m P m + SX P; f

the sign of summation S including every value of i from

to n,excepting i=m. Now the value of P, being the
same in respect of each of the hypotheses which suppose
the assertion untrue, SXj^rrPiZX, ; and the sum of all the
values of X f from i=Q to i=n being 1, on excluding
\ m ~a m -t-s, we have SX =(s t ct^-r-s. Substituting this,

together with the values of Pm and P,, as above found, and

r
making u 1 u, v 1 v, the formula becomes, after the

proper reduction,
a m {(n l)ut>+uV}
108 APPLICATION TO THE

which is the probability of the hypothesis that a ball mark

ed A m was drawn, or that the testimony is true.
When there are no two balls in the urn having the same

index, the numbers a lt a^ a s &c. become each =1, and

,

s~?i. In this case the formula gives

(n I
)uv -f. u v
\) (u v + uv ) + (n 2)u v

which, on observing that uv+u v + uv -\-u v =.\, becomes

by reduction

This is the probability of the truth of the testimony of a

witness, who affirms that the number m is drawn from an

urn which contains n balls, numbered I, 2, 3... n. It is ob
vious, that when u and v are fractions approaching to unity,

and n is a considerable number, the second term becomes

very small, and may be neglected. The probability then

becomes simply wm =M#.
69. We now proceed to consider the probability of an event
attested by several witnesses and first let us suppose the wit
;

nesses to agree in their testimony. The measures of the ve

racity of the several witnesses being respectively v lt v a , v s ,
&c., and the a priori probability of the event being p, we
have by (58) for its probability after the testimony of the
first witness,

In order to find the probability of the event after the se

cond witness gives his testimony, we may suppose the a

priori probability to be changed from p to by the ^ testi

mony of the first witness, and the same formula gives

 DECISIONS OF JURIES AND TRIBUNALS. 109

Let a third witness now come forward, and give testi

mony in favour of the same event. Its probability after his

testimony will become in like manner

In general, let -sr

x be the probability of an event after it

has been attested by x witnesses, and let vx be the veracity

of the last witness, then Trx \ being the probability of the
event after x 1
eyewitnesses have each testified in its fa

vour, we have

If we suppose the witnesses all

equally credible, or that
l
=v 2 =v 3 ...=:v*, this becomes
V*

v p
Now, if v=-|, then (1 v)-?-v=:l, and w,=/?; whence
it
appears that the probability of an event is not increased
by the testimony of any number of witnesses, when the vera
city of each is only ^; but when v is greater than |, the event

becomes more probable as the number ofwitnesses is greater,

and when v is a considerable fraction, its probability in
creases very rapidly with the number of witnesses.

70. When the values of v and p are given, that of x in

the last formula may be found so as to render -srx of any

given value. Hence we may find the number of witnesses

required to make it an even wager, whether an event ex
ceedingly improbable, and in favour of which they give una-
 110 APPLICATION TO THE

nimous testimony, has happened or not. For

example,
let the odds against the event be a million million to one,

tllat is let P= = Tot? and let " the

i,ooo,oo.Uo,ooi

veracity ofeach witness be. In order that &* may equal J,

/I v\* 1 p 1 v 1
we must have (
v
)
= 1. Now = -,and
\ J p v 9

Zf = 10 12 therefore 12
=l whence x log ^=r
P
,
(-Vxl0
\y/
;
y
1 12
log -p or x log 9= 12 and therefore x = =12-6

nearly, so that 13 independent witnesses would suffice to

render it more probable that the event really took place

than that it did not.

This example is given by Mr. Babbage, (Ninth Bridge-

water Treatise, Note E), with a view to shew the fallacy
of Hume s celebrated argument respecting miracles. What
the example proves is
simply this, that if we suppose an urn
to contain a million million of white balls, and only one black
ball, and that on a ball being drawn at random from the
urn, thirteen eyewitnesses of the drawing, each of whom
makes only one false statement in ten, without collusion,
and independently of each other, affirm to A, who was not
present at the drawing, that the ball drawn was black, then
A would have rather a stronger reason for believing than
for disbelieving the testimony. But it is
sufficiently obvious,
that the event attested in this case, though exceedingly im

probable a priori, cannot be regarded as in any way mira

culous. On the contrary, the black ball might be drawn
with the same facility, and was a priori as likely to be drawn,
 DECISIONS OF JURIES AND TRIBUNALS. 1 1 1

as any other specified ball in the urn. Let it be granted

that an event is within the range of fortuitous occurrence,

and that there exists a single chance in its favour out of

any number of millions of chances,

may then happen in it

any one trial nay, a number

; of trials may be assigned,
such that its non-occurrence would be many times more im

probable than the contrary.

71. Let us next consider the case of a number of wit
nesses contradicting each other. If the first witness an
nounces an event of which the probability is />, then the

probability, after the testimony, of its having happened is

ar,,
and the probability that it has not happened 1
a-,. Sup
pose a second witness now to appear, and testify that the
event has not happened, and let the probability of the truth
of his testimony be denoted by sr 2 ; then 1
sr, being the
probability before his testimony was given that what he as
serts is true, and v 2 being the measure of his veracity, we

have, as in (69), ^ =V2

2(
* - >
\7~/ii -- 7 whence, since

WT1 \
tnere results
X P)

for the measure of the probability that the event has not

happened. The probability that it has happened is there

fore 1 z?-
2,
and accordingly if w 2 be less than -J,
there
isa stronger reason for believing that the event happened
than that it did not. The method of forming the expression
for the probability of the event, after it has been attested
or denied by a third witness, or any number of successive wit
nesses, is obvious.
112 APPLICATION TO THE

If we suppose the values of v l and v 2 to be equal, the ex

pression becomes w 2
1
p, which is the a priori probabi
lity that the event did not happen. It is obvious that this
must be the inasmuch as two contradictory testimonies
case,
of equal weight neutralize each other. In general, the pro

bability of an event which is affirmed by m witnesses, and

denied by n witnesses, all
equally credible, is the same as
that of an event which is affirmed by m n witnesses who
agree in their testimony.
72. When a relation has been transmitted through a se
ries of narrators, of whom the first only has a direct know
ledge of the event, and each of the others derives his know
ledge from the relation of the preceding, the probability of
the event is diminished by every succeeding relation In
order to obtain a general expression for the probability of

traditionary testimony, we may take the event considered

in (68), namely the extraction of a ball marked A m from an

urn containing s balls, of which are marked A t ,* 2 mark t

ed A 2 ,... n marked A n there being in all n different in

,

dices. Now suppose the relation to have passed through a

chain of narrators, T, T T 2 ,...T.r
t, ,
in number .r-f-1, of
whom the first only was an eyewitness of the event, each
of the others receiving his knowledge of it from the one pre

ceding him, and communicating it in his turn to the suc

ceeding, the question is to determine the probability that a

ball marked Am was drawn, after this event has been nar
rated by Tx , the last witness of the series.

73. In order to apply the general formula of (68) to this

case, it is necessary to remark that the event observed is the

attestation of T, of his having been informed by T,_i; that
the ball drawn from the urn, the drawing of which was seen
by T, was marked Am . There are n different hypotheses
 DECISIONS OF JURIES AND TRIBUNALS. 113

respecting the index of the ball actually drawn, but it is

only necessary to consider two of them, namely, the hypo

drawn was marked A m and any
thesis that the ball actually ,

one of the other hypotheses which consist in supposing that

a ball with a different index from A m was drawn, for ex

ample A f
. Let the probability of the attestation, on the hy

pothesis that the index of the ball drawn was A m be denot ,

ed by y x and its probability on the hypothesis that the in

,

dex was A, by y ,, (yx and y x corresponding to P m and

P in (68), which express the same probabilities in respect
{

of the eyewitness T). then by (68), the probability of the

hypothesis that Am was drawn is

But since y x is the same for all the hypotheses that the in

dex drawn was different from Am , SA^rrf/^ZA, ; and by

(68) Z\-=(s a m )--s, and A m = m -j-*, therefore

+ (s 0/
We have now to find yx and y x in terms of x. Let v, v lt
v2 , &c. be the respective probabilities of T, T I} T 2 , &c.

speaking the truth, then the probability of T, speaking the

truth is ?;,,and the probability that he does not 1 v,, whether

because he is dishonest, and intends to deceive, or because

he has mistaken the statement of the preceding witness. Now

there are two ways in which it may happen that A m is announ
ced by T,. First, if he speaks the truth, and has been inform
ed by the preceding narrator Tj_i that A m was the index
drawn ; secondly, if he lies, and has been informed by T ,_i
that a different index fromA m was drawn. Assuming yx ^\
to have thesame signification with respect to T^_i that yf
has been assumed to have with respect to T,, (that is to say,
114 APPLICATION TO THE

the probability of the assertion being made by Tx i on the

hypothesis that the ball actually drawn was A m ), the proba
bility of the first of these combinations is vx yx_\* With re

spect to the second case, it is to be observed, that

T, an if

nounces a different index from that which has been announ

ced to him by T^_i, the chance of his announcing Am out
of n 1 indexes different from that announced by T^ i is

l-i-(72 1) ; and on multiplying this by the probability

1 v, that the testimony of Tx is false, and by the probability
1 yx _i that Tj_i has announced a different indexfrom A TO ,

we have, for the probability of the second combination

(l_vj(l_^_ )^_(/i_l). The whole probability of T, tes

1

tifying that A m was drawn, is therefore, on the first hypo

thesis, given by the equation,

yx =vxyx^ + (\ O(l-^_0 :
.(n 1).

This is an equation of finite differences of the first order,

the complete integral of which is

y*~n*
_1 C(nv, \)(nv, l)...(^-t !)(>, l)i

(n\)*
In order to determine the arbitrary constant C, it is to be ob
served, that, since T/!, y^-y^ as well as v v v 2 ...vx apply to
the narrators T T 2 ...T,
t, respectively, if we suppose
x=o
the resulting value of the integral will be the probiibility
that A m was announced by the eyewitness T, on the hy

pothesis that Am
was actually drawn. Let this probability
be Pm ; then the equation becomes P m =C-f-l-r-ra whence

C=(nP m 1
) -T-
n. If, therefore, we make.

This is easily verified ; for on changing x into x 1 in the in

1

tegral, and forming the expression vx y^i + (1 Uj)(l yx i)-r-

(rc 1), there results an identical equation.

 DECISIONS OF JURIES AND TRIBUNALS. 1 15

we obtain, on the first hypothesis, for any value of #,

In the same manner we find the probability y x of the

testimony given by T*, on the hypothesis that the ball ac

tually drawn was A t

. The probability of the event, after

being testified by the eyewitness, being on this hypothesis

P,, we have,

Substituting these values of yx and y x in the expression

above found for zrm , we obtain for the probability of the event
observed by T, and narrated by T,, the narration having

passed from one to another in the manner supposed,

74. Since =v x----f-, and since vx is always less

n 1 n 1

than unity, and n always greater than unity, each of the

terms of the series represented by X, whether positive or

negative, is a proper fraction, whence the value of X be

comes smaller and smaller as x increases. Supposes in

finite, then X 0, and vr m ~am -t-s, which is the a priori pro

bability of the event. Hence we see that the probability of

an event transmitted through a series of traditionary evi

dence becomes weaker at every step, and ultimately equal
to the simple probability of the event, independent of any
testimony.
75. When the urn is supposed to contain only n balls,
each having a different index, the expression for wm is great

ly simplified ; for, in this case, a m -=:l,sn ; therefore, (since

P m +(n 1)P ( =1) the denominator becomes n, and we have

consequently vrm ~l -\-(nP m l)X|-j-^, which coincides

116 APPLICATION TO THE

with the value of yx found above, that is to say, with the

probability of the event being testified by T, on the hypo

thesis that it
actually happened. Laplace, in solving this
particular case of the problem, (p. 456) assumes that the

probabilities here denoted by yx and wm are identical. They

are, however, as is evident from the above analysis, quite
distinct in their nature, and their values are only equal in the

particular case in which s am is to am in the ratio of m 1

to 1. (Poisson, p. 112.)
76. The question of determining the probability that the
verdict of a jury is correct, is
precisely analogous to that of

finding the probability of an event attested by one or more

witnesses. Let us first take the case of a single juror, and
assume w=r the probability that the juror gives a cor
rect verdict, (that is, correct in respect of the facts), and

/?= the probability that the accused

guilty before being is

put on his trial. Suppose the verdict guilty to be return

ed ; two hypotheses may be made respecting the cause of
the verdict, first, that the accused is guilty ;
secondly, that
he is innocent. On the first hypothesis, the accused will
be condemned if the juror gives a right verdict, the probabi
lity of which is u. On the second hypothesis, the accused
will be condemned if the juror gives a wrong verdict, the
probability of which is 1 u. But the a priori probabili
ties of these causes (the guilt or innocence of the accused)

being respectively p and 1 p we have by

9 (49)

w-j being the probability of the first hypothesis, or the pro

bability that the accused is
guilty after the verdict has been
given, and *r 2 the probability resulting from the verdict that

the accused is innocent.

 DECISIONS OF JURIES AND TRIBUNALS. 117

77. Suppose the verdict not guilty to be given, and let

a\ and w r
2 be the probabilities after the verdict of the two

hypotheses. On the first hypothesis, namely, that the ac

cused is
guilty, this verdict
will be given if the juror gives
a wrong verdict, of which the probability is 1 u ; and on
the second hypothesis, the verdict will be given if the juror

gives a right verdict, of which the probability is u ; and the

probabilities of these hypotheses before the verdict being

respectively p and 1
p as before, we have
-"
>
~ -
From the above value of ir,,
we obtain & l p =
,
~
1)
; a fraction which is positive or negative
*00 p)
according as u is greater or less than J. Hence it appears
that the guilt of the accused is only rendered more probable
by the verdict guilty being pronounced, when the probabi

lity that the juror gives a correct verdict is

greater than \.
In like manner it is shewn that ts \ (the presumption of the
guilt of the accused after a verdict of acquittal), is
greater
than p when u is less than ^.
78. The a priori probability of the condemnation of the
accused before he is
put on his
p) J trial is up-\-(\ w)0 :

for there are two ways in which condemnation may this

take place ; first, if the accused be guilty, and the juror give
a correct verdict, the probability of which concurrence is

up ; and, secondly, if the accused be innocent, and the ju

ror give a wrong verdict, the probability of which is (1 u)
(lp). Therefore, making c = the probability of a ver
dict of condemnation, we have c=up-\-(\ u)(l p) ; and
for a verdict of acquittal, 1 c=(l u)p + u(\ p).
79. Let us next suppose that after the verdict ef the first
118 APPLICATION TO THE

juror has been pronounced, the accused is

put onhis trial be
fore a second juror, u s be the probability the second
and let

juror gives a correct verdict, and c 2 be the probability the ac

cused will be pronounced guilty by him. After the verdict

guilty has been pronounced by the first

juror, the probabi

lity of the guilt of the accused is

zr^ and it is evident that
c 2 will be found by substituting u2 for w, and & l for p in

the above value of c, whence c a z=u a w l + (\ U Q )(\ -sr^).

The probability of a verdict of condemnation by both ju

rors is cc 2 ; therefore, (observing that w-jsrttp-r-c), we have
for this probability

The probability of the guilt of the accused after a verdict

of acquittal has been pronounced by the first juror
being w j,
the probability of a verdict of acquittal being given by the

second juror is 1 cQ (l u^-sr\ +u 2 (\ wJ ; therefore,

(observing that ^ = (1 u)p--(l c), wehavefor the pro

bability of a verdict of acquittal by both jurors

Adding the probability of a verdict of condemnation by both

jurors, to
that of acquittal by both, we have uu s ^(\ u)
(1 u2) for the probability of both giving the same verdict.
This result is
independent of/?, and is evidently true a priori,
inasmuch as there are two ways in which the same verdict

may be given, namely, when both jurors are right, and when
both are wrong.
The probability of acquittal by the second juror, after a
verdict of guilty by the first, is 1 c 2 rr (1 U 2 )^i -f-

u 2 (l wj; multiplying by c, and substituting for c and

23-j
their values, we have for the probability of a verdict of

guilty by the first, and not guilty by the second,

 DECISIONS OF JURIES AND TRIBUNALS. 119

In like manner, if the accused has been acquitted by the

first juror, the presumption
of his guilt becomes w A ,
and the

probability of a verdict guilty by the second is c^rrw^ j

.j_(l
w 2 )(l yr\) ; therefore the probability of a verdict

of not guilty by the first, and of guilty by the second is

(1 C)c fl
= (l U)u 2 p + U(l M 8 )(l-_p).
The sum of these two expressions gives for the probability
of a discordant verdict, u(\ ^2) + (l u)u 2 .

80. If we now suppose wrrw 2 and make , 1 u w, the

probability that the two jurors will agree in their verdict,
2
whether they are both right or both wrong, isu^ + iu ; and
the probability of a discordant verdict uw+uw2uw. The
2
u *-\-2mv-\-2v 2
<

sum of the two expressions is

(u-}-w) ; and
therefore the probabilities of the different cases are respec
2
tively given by the developement of the binomial (u + w) .

pursuing this reasoning, it is easy to see that if there

By
be any number h whatever of jurors, or voters on any ques
tionwhich admits only of simple affirmation or negation, all
being supposed to possess the same integrity and know
ledge, so that there is the same probability u of a correct
decision in respect of each, the probablities of the different
h
cases are found by the development of the binomial (u+w) .

The probability of a correct verdict being pronounced una

h
nimously is u ; of an erroneous one being pronounced una
nimously isand the probability that a correct verdict
up ;

will be given by m
of the jurors, and an erroneous one by
1 .2.3 h
m n
is Uu w , where U =
n> , ,
.

1 .2.3...wxl .2.3...T*
81. The probability that the accused will be pronounced
guilty by m jurors, and acquitted by n> on the supposition
that the value of u isthe same for each juror, is thus found.
There are two ways in which this event may take place ;
120 APPLICATION TO THE

1st, if the accused be guilty (the probability of which is

p),
and m jurors decide correctly, and n wrongly (the probabi
m
of which is Uu iv n ) ; the probability of the condemna
lity
therefore \Ju m w p.
n
tion taking place in this way is 2d, If
the accused be innocent (the probability of which is
q) and
n jurors decide rightly, and m wrongly
(the probability of
which is Uu"w m ) ; the probability of the event taking place
in this way is therefore Uun wm q. Let G therefore denote
the whole probability of the verdict, and we have
G= U(u mw p + u wmq).
n n

Hence the probability that the accused will be condemned

unanimously by a jury consisting of ^jurors is v^p ufq ; +
and the probability that he will be unanimously acquitted
uh q-\-whp.
82. Suppose the accused to have been pronounced guilty

by m jurors, and not guilty by n jurors, the probability of

the verdict of the majority being correct is found from the
formula in (49). Two hypotheses may be made 1st, the :

accused is
guilty ; 2d, he is innocent. The probability Pj
of the observed event (the condemnation by m, and acquit
m n
tal by n jurors) on the first hypothesis is Uu w ; and the

a priori probabilities of the two hypotheses (or the proba

bilities denoted by X x and X 2 in (49), being p and q ; there
fore if 73-
^
denote the probability of the verdict being cor

rect, that is, the probability of the first

hypothesis after the
verdict has been pronounced, and zr s the probability of its

being wrong, we shall have (49)

um w np u n wmq
~~ ~~
u m wnp -f un w m q
757
1 *

If the verdict has been pronounced unanimously, then

h and wzrO, and the formulae become

 DECISIONS OF JURIES AND TRIBUNALS. 121

If pq J, and m w=:z, we have then

~~
wn + i
u +w*
But this is the probability of a verdict being correct which
has been pronounced unanimously by i
jurors ; whence it

follows that the probability of a decision rendered by a given

majority being correct, is the same as that of a decision ren
dered unanimously by a jury equal in number to the differ
ence between the majority and minority, and is therefore

independent of the total number of jurors. This, however,

is only true on the supposition that the value of u is known
a priori ; for if u be not absolutely known, the weight of the
verdict depends on the ratio of the majority to the whole num
ber of jurors. This is in accordance with common notions,
for it will readily be admitted that a verdict given unani
mously by a jury of 10 will be entitled to much more weight
than one pronounced by a jury consisting of a large num

ber, as 100, in which 55 are of one opinion, and 45 of the

opposite. In this case, the opinion of the minority throws

great doubt on the correctness of the verdict. It is to be

observed, however, that the probability of a verdict being

given by a small majority becomes less and less as the num

ber of jurors is increased.

83. When the number who dissent from the opinion of

the majority is unknown, and we merely know that the ma

jority exceeds the minority by at least i jurors, the proba

bility of the verdict being correct is found as follows. Sup

pose the verdict to be guilty. On the hypothesis that it is

correct, the probability of the accused being found guilty

by h x and not guilty by # jurors,

} is
by the formula in (80),
122 APPLICATION TO THE
x
Uuh *iv .
Now, if we give x successively all the values

0, 1, 2,... w, where n=.^(h z), and assume U to denote the

value of when a?=0, U A its value when #=1, and so on

U ;

and also make W=the probability of the accused being

pronounced guilty by h n at least, we shall have

In like manner, if W denote the probability of a verdict

guilty by h n jurors at least, on the hypothesis that the

accused is not guilty, we shall have,

whence, p and q being as above the a priori probabilities

of the two hypotheses, the probability that the verdict guilty
is correct, when pronounced by h n jurors at least, becomes

84. It is evident that no application can be made of these

formulae without assigning arbitrary values to u and />,
un
less, indeed, we have data for determining their mean values
from experience. With respect to p, we may assume, for
the sake of shewing the general consequences of the formu

lae, its value to be J ; for it cannot well be supposed less than

J, or that a person brought before a jury is more likely to be

innocent than guilty ; and if it much exceeds J and ap

proaches to unity, a verdict of guilty may be expected from

any jury, however constituted. When a mean value of u
cannot be determined from experience, the only way of ob

taining numerical results, is to suppose u to have all

possible
values within given limits, and to integrate the equations

between those limits. As it seems unreasonable to suppose

that a juror is more likely to give a wrong verdict than a
right one, we may assume that u cannot be less than ^.
Suppose, then, that u increases by infinitely small increments
 DECISIONS OF JURIES AND TRIBUNALS. 123

from w=J to tt=l, and let it be proposed to determine the

probability that a decision is correct when the accused has

been pronounced guilty by m jurors,
and not guilty by n.
Here an infinite number of hypotheses may be made re
specting the value of u, and we must therefore have recourse
to the formulae in (51.) Let u=x be one of those hypo
theses, Px = the probability on that hypothesis of the event
observed (thatis, of the accused being pronounced guilty

by m, and not guilty by n jurors,) 33-*:= the probability of

the assumed hypothesis, and n =
the mean probability of
the correctness of the verdict from all the hypotheses. By
the formulae in (81) we have

and as all the hypotheses are supposed equally probable,

we have (45) ^rrP^-r-SP,. But between the proposed

limits tP x =pft x m (l
l
>x)
n dx
+ (1 jo)/jV (1
m
x) dx ;

if, therefore, we make j5=^j we shall have by reason of

and therefore

for the probability of the hypothesis.But (82) the proba

bility on this hypothesis

of the accused being guilty, is

; multiplying this by the probabil-

ity of the hypothesis, wx we

, obtain for the probability of

the verdict being correct

n
x m (\m
x) --J^x (l
n
x) dx ;
and, therefore, for the probability of the verdict being cor
rect on all the hypotheses from x=^ to #=],
 124 APPLICATION TO THE

Hence the probability that a verdict given by a majority

m out of m-)-n=h jurors is
wrong, is

m
f l
Jc
(lx} n dx
which, on effecting the integrations by the formula in (51)
becomes after reduction

"
I 1-2 ~--
~T-2-3
(h+l)h(hl) ...... (h n + 2) I
1-2-3 ................ n j

Assuming h (the number of jurors) =12, and making n

successively 0, 1, 2, 3, 4, 5, the series gives

1 14 92 378 1093 2380

8192 8l92 8192 8192
for the respective probabilities of the error of a verdict when
pronounced unanimously by 12 jurors, by a majority of 11
to 1, of 10 to 2, of 9 to 3, of 8 to 4, and of 7 to 5. In the

last case the probability of the error is nearly =f .

85. From these results it

appears that the chance of a
verdict being wrong which has been pronounced unanimous
ly by twelve jurors is very small ; but it is to be remarked,
that they have been deduced on the supposition that the

unanimity proceeds from agreement in the same opinion, and

that the jurors are unbiassed by each other. In this country,

where unanimity is compelled by law, the mean probability

of a correct verdict can scarcely be considered as greater than

that of a verdict pronounced by a simple majority; for,

though in most cases the verdict may be supposed to repre

sent the opinion of a larger majority than seven, it may

happen, not unfrequently, that a smaller

number than five,

possessing greater energy

or perseverance, may persuade the
 DECISIONS OF JURIES AND TRIBUNALS. 125

others into a surrender of their judgment. In fact, unless

the presumption of the guilt of the accused be very great,

it would scarcely be possible, without concert, to procure an
unanimous verdict in any case. It is also to be observed,
that the assumption of all values of u from Jto 1
being equal
ly probable, may lead to results widely different from the
truth. The mean value of u, which depends on the general
intelligence of the class of persons from amongst whom the
lists of jurors are made up, can only be rightly determined
from data furnished by experience. One of the elements,
however, which require to be known for this purpose, is the
number of jurors who concur in, and dissent from, the verdict.
The forced unanimity ofthe law renders it
impossible to obtain
this element from the records of the English courts ; but in
France and Belgium, where the majority and minority are
known and recorded, the same obstacle does not exist, and
the " Comptes Generaux de FAdministration de la Justice
Criminelle" published by the French Government, have
enabled Poisson to deduce mean values of u and p for that

country, and consequently to obtain the necessary data for

one of the most interesting applications of the theory of
Probabilities. The general results were as follows During :

the six years from 1825 to 1830 inclusive, the system of cri
minal legislation in France underwent no change ; the jury
consisted of 12, and a simple majority was only required to
concur, though when it
happened that the majority was the
least possible, the Court had power to overrule the verdict.
On comparing, according to the rules of the theory, the ver
dicts given in the cases tried before the criminal courts

during those six years, it was found that for the whole of

France, the probability (u) of a juror giving a correct

verdict was a little greater than with respect to crimes
126 APPLICATION TO THE

against the person, and nearly equal to \% with respect to

crimes against property ; without distinction of the species

of crime, it was found to be a very little below J. The other

element, the probability (/>) of the guilt of the accused be

fore the trial, was found not much to exceed ^ (being be

tween 0.53 and 0.54) with respect to crimes against the

person, while it a little exceeded in respect of crimes

against property. Without distinction of crime, its value

was very nearly 0.64.

86. On substituting these values of u and p (namely
w=j, jo=.64, whence w=^, q=.36) in the formula in (81),

and making m=7, /z=5, and consequently

X.36)=:.07 nearly. Hence it

may be ex
pected, that in a hundred trials it will happen only seven
times that the accused will be pronounced guilty by the
smallest possible majority. If w=12, and ra=0, we shall

have uh p+w q=:. 02027= ^

h
nearly, for the probability of
b
an unanimous verdict of guilty, and u q-\-w p=:.Oll4 for ft

the probability of an unanimous verdict of not guilty.

Making the same substitutions in the formula in (82), we

have for the probability of a verdict guilty being correct,

from which 5 jurors out of 12 dissent, zj^rr-j-f ; and ss-

<2 -=.^
for the probability of its being wrong.
Substituting the same values in the series represented by
W and W in (83), and supposing n to have all values
37
from w=0 to n=5, there results W= X7254, W=
- Wp 126915984 118
1
x 239122, whence
^--^
=
y^^ = TT9

nearly. This is the probability that a verdict guilty, pro-

 DECISIONS OF JURIES AND TRIBUNALS. 127

nounced byamajorityof sevenagainst fiveatleast, is correct.

The probability of the same verdict being wrong, is there
fore Yig ; so that out of 119 verdicts, respecting which we
know nothing else, than that seven at least of the jury con
curred in finding the accused guilty, we may expect one to
be wrong, or that one person out of 1 1 9 so condemned will
be innocent.
 128 SOLUTION OF QUESTIONS

SECTION VIII.

OF THE SOLUTION OF QUESTIONS INVOLVING LARGE

NUMBERS.

87. The probabilities of the different compound events

which can from the combination of any number of
result

simple events, E,, E 2 E 3 &c. being (13) measured respec

, ,

tivelyby the several terms of the developement of the mul

tinomial (p-j-^-f /-}-&.)*, the most probable of those com

pound events will be that which corresponds to the term

having the greatest numerical value. Let us consider the

case of two simple contrary events E and F, the proba
of which are respectively p and q, and suppose the
bilities

number of occurrences to be 4. Neglecting the order of

occurrence, the different combinations, with tlieir respec
tive probabilities, are the following :

EEEE, EEEF, EEFF, EFFF, FFFF,

Now it is evident that the numerical values of these proba

bilities depend on the ratio of p to <?, as well as on the co
efficient by which they are multiplied, and that values may
be given to p and </,
such that any one of the terms may be
made the greatest or the least in the series. If we suppose

p*= q, and consequently /?=J, <7=J (since p-\-q=l) the

probabilities of the different cases become respectively
 INVOLVING LARGE NUMBERS. 129

whence it
appears, that the most probable combination is
2 2
that which corresponds to 6/? <? , or in which each of the

simple events occurs twice, the probability of this combina

tion being while that of either of the simple events oc
-^
curring four times in succession is only -J^.

When the number of trials is 5, the probabilities of the

several cases are respectively
5 2 5
p , 5p*q, lOpSq ,
Wp*q*, 5pq*, q ,

which, when p=q, become

&> A JS & &, *V>
so that there are two different combinations equally pro

bable, namely, that in which E occurs three times and F

twice, and that in which E occurs twice and F three times ;
and of the six possible combinations these two are the most
probable, having in their favour a number of chances twice
as great as the two cases in which one of the events occurs
only once, and the other four times, and ten times greater
than the two cases in which either of the simple events oc
curs in each of the five trials.

From these two instances it

may be inferred in general,
that when h is an even number, the most probable com

pound event is that of which the probability is represented

h
by the middle term of the developement of (p + q) > and
that when h an odd number, there are tv/o compound
is

events equally probable, and more probable than any other,

namely, those corresponding to the terms which occupy the

middle of the series, supposing in both cases p=q. This
A A
supposition gives (p-J-^) =(l l)Xi) + therefore in the
case in which h is an even number, the general expression
for the greatest term is

h(h-l)(h-2) (h-^h+l)
~~
(a)
1.2.3
130 SOLUTION OF QUESTIONS

and when h is odd, the general expression for either of the

two equal terms, which are greater than any of the other
terms, is

A(A-I)(A-2) ...... {A-KA+1)+1]

1-2.3 ...... KA+I)
88. When p and q are unequal, the greatest term of the
h
expansion of (p-\-q) will not occupy the middle of the
series, but its place may be found by comparing two con

secutive terms. Let h=m+n. The general term of the

series then becomes

1.2.3 ......... h
1.2.3 ......
and the term immediately preceding is

1 .2.3 ......... h
1.2.3 ...... (m+l)xl.2.3 ......
Dividing the first "of these by the second, we get for the

quotient (m-\-\)q--np, which, therefore, is the ratio of two

consecutive terms taken at any part of the series. If this

ratio be greater than 1, the term which has been taken

as the dividend is
greater than the preceding one which
has been taken as the divisor ; and it is evident that the
terms must go on increasing, from the beginning of the se
ries, so long as the ratio in question is greater than 1 But .

if the ratio be less than 1, the preceding term is

greater than
the succeeding, and the terms will become less and less as

they are nearer the end of the series. Let (m -|- 1

) q-r-np 1 ;

then, since p-\-q=\, and m-\-n=h, we have n=(h-\-\)q,

and consequently the ratio of any term to the next preced

ing is
greater or less than 1 according as n is less or

greater than (A-f- l)q. Now n is necessarily a whole num

ber therefore if (h
; + 1
)q be a whole number, take ?i= (h + 1
)</,

and the two terms of the series given by the expansion

 INVOLVING LARGE NUMBERS. 131

of (p-i-qY, in which the exponents of q are n 1 and n, will

be equal to each other, and each greater than any other

term of the series. But if (A+ 1 )q be not a whole number,
let (h +1)7 x be the nearest whole number less than

(h +
)</,
1 and make
n=(h -f- )q ; 1 x then the greatest term
of the developement will be that in which the exponent of
is 71.
q
Since n=i(h -f. 1
)<?-#, we have q(n -j- x) -~(h -f 1) whence ,

n+x = x
p= 1 7- m-f-1
?" , ,
q
and therefore i
_
= -m n+x .

A-fl p
Now x is by hypothesis less than 1, therefore if m and n
are large numbers, we have, very nearly, q :pzzn : m; or,

since m+n^h, mhp, nzzhq. It follows therefore, that

the greatest term of the developement of the binomial

A
4- ?) *s tna * i n which the exponents
ofp and 5 are to each
other in the ratio ofp more nearly in that ratio to q, or

than are any other two numbers whose sum is h. In other

words, the most probable combination of two simple events,

E and F, in any number of trials, is that in which the num
ber of occurrences of E is to the number of occurrences of
F in the ratio of their respective probabilities.

89. In the same manner it

may be shewn, that when
there are more than two simple events, of which one must
occur in every trial, the most probable result of any number
of trials is that combination in which the number of repe
titions of each simple event is in proportion to its probabi

lity in a single trial. Thus, the probabilities of the simple

events being respectively^, q, r, &c. the most probable com
pound event is that whose probability is expressed by^that
A
term of the expansion of (p + g+/ +&c.) f

,
which has for
hp hg hr
its
argument p , q ,
r ,
&c.
90. Having determined the form of the greatest term of
132 SOLUTION OF QUESTIONS.

the series, we have next to find a method of approximating

to its numerical value ; for its coefficient containing the pro

duct of the natural numbers from 1 to h inclusive, its di

rect calculationbecomes impracticable even when h is only
a moderately large number. The theorem which gives the

approximate value of this product is known by the name of

Stirling s Theorem, having been discovered by that mathe

matician. As its investigation is a matter of pure analysis,
we shall not stop to give it here, but refer the reader to the
Treatise on Differences and Series, by John Herschel,
Sir

in the translation of Lacroix s Elementary Treatise on the

Differential and Integral Calculus, p. 568.
1
The theorem
is as follows : Let x be any number, then

1 2 3 *=*^V2^(l+ i + HI? + &C >

where e is the number of which the Napierean logarithm

is unit, or the number 271828, and TT the ratio of the cir
cumference of a circle to the diameter, or 3-14159.

When x is a large number, the term divided by 12# be

comes very small, and the series within the brackets may be
considered as equal to unity. In this case, then, the for
mula becomes
1.2. 3 x=*<r-**j2^ ,

which gives a approximation in most cases. If,

sufficient

for example, #=1000, the result will be within a 12000th

part of the truth.

Now, let E and F be two events of such a nature that the
one or the other must happen in every trial ; let p and q be
their respective probabilities, and P the probability that in

1
Stirling s investigation of the theorem, or rather of its equivalent,
to find the sum of the logarithms of a series of numbers in arithme
tical progression, is given in his Methodus Differentialis, p. 135.
 INVOLVING LARGE NUMBERS. 133

w + ft=A trials, E will happen m times and F ra times ; then

by (12) we have

When
~
1 .2.3 ......... 6 _ m n
-2.3 ...... n P *
and A are large numbers, the value of this
w, n,

coefficient may be computed from the above formula, which

gives
1 .2.3 ...... A=
1.2. 3 ...... m=
1.2. 3 ...... 7z=rc
whence

This expression represents any term of the series (p-f- 9)*.

The greatest term, which corresponds to the most probable
result, is (88) that in which m and n are to each other in

the ratio of p to q, or when m=:hp, and n=hq. Let the

greatest term therefore be denoted by P , that is to say, let
P be the chance of the most probable result of h trials, and
we shall have

P =V(h+2nmn),
or P =<v/(l-j-27r^).
This formula shews that the absolute probability of that
last

combination which has the greatest number of chances in

its favour becomes less and less as the number of trials is

increased ; for the fraction 1 -f-/*, to the square root of

which the probability is
proportional, diminishes as h is in

creased.

91. As an example, suppose a shilling to be tossed 100

times in succession. In this case p=q=, hp=50, hq50,
and the most probable result of the trials is 50 times head
and 50 times tail. We have then ^=100, #z=50, and

4/(h-t-2irmri) = 1 -j- \f(5Q-rr) for the measure of the probabi-

134 SOLUTION OF QUESTIONS

lity that the event will happen in this way exactly. On cal

culation, this is found =.07979 ; whence it appears, that

although 50 heads and 50 tails is a more probable result of

100 trials than any other combination which can be named,
its absolute probability is measured by a very small fraction.

The probability of the contrary event, or that there will not

be thrown 5 heads and 5 tails exactly, is 1 07979 = .92021,
so that the odds against the event are about 92 to 8, or 23
to 2. Had the number of trials been 1000, the probability
of 500 times head and 500 times tail exactly, though more
likely to occur than any other combination, would have been
found 1 -r- V(500?r) ; that is to say, V 10 times, or rather more
than 3 times less than in the former case. In general, when
the chances in favour of the simple events are equal, the pro

bability of the combination which is more likely to happen

than any other, is
inversely proportional to the square root of
the number of trials.

92. The formulae in (90) enable us also to determine the

term of the developement of (p-\-q)*
ratio of the greatest

to any other term of the series, and consequently the rela

tion of the probabilities of the different compound events.

Let m : n : :
p :
q, whence m=hp and nhq, and let P* de
note the probability that in h trials the event E will occur

(m x) times, and the event F (n x) times, the probabi +

of the simple events E and F being respectively p and

__
lities

q. By (13) we have
1 2 3 ......... h
P ~
1.2.3 ...... (m a?) x 1.2. 3 ...... (n
which by (90) becomes

_
~~

.^
w ie nce,
}
substituting m-^-h for p, and n-^-
 INVOLVING LARGE NUMBERS. 135

for ^, and leaving out the factors common to the numera

tor and denominator, we find,

P, = ^/Y
Now log (m m +* and
#) *;=:( m-\-x J)log(/w a?) ;

log (m a?)t= log m ---

y
m 2m
-
sc
5
2
&c.

therefore log (m x)~

m + x~* =
x^ \

(x + 2^2 + &C J 5

whence, neglecting terms divided by m m &c., m being

2 3
, ,

supposed to be a large number in comparison with x,

2

-*=( m+x
therefore, on passing to numbers,
.fl
2m * 2m ,

or snce =

2m
In like manner, by changing m into /z, and a? into

we get

Multiplying the first of these two expressions by m n

and the second by w n +-r we have ,

136 SOLUTION OF QUESTIONS

whence, substituting these values in that of P*, and ne

glecting the quantity divided by mn,
it h
Px = v - J
\ _^L._^_

\lTrmnJ
e 1m
(
2n = - -
)
*r
It
/ ( - -h
\27rmnJ
\
)
e
_ ***
2mn.

The term of the series h which corresponds

(p+q) to this

value of P* is that which is x places to the right of the

greatest term ; and it has been shewn, (90), that the great
est term has
for its expression V(h-r-2Trmn) ; therefore the

greatest term being denoted, as before, by P and the term ,

which comes after it x places by P^, we have

that is to say, the probability the event E will happen m

times and fail n times in m -f n trials, is to the probability
of its happening (m x) times and failing (n + x) times in
e~ *
the ratio of 1 to .

Since the numbers m and n enter symmetrically into the

expression ehx^rZmn^ jt js ev icl en t that the result would

have been the same if, instead of seeking the ratio of the

greatest term to that which succeeds it

by x places, we had
sought the ratio of the greatest term to that which precedes
it by x places. Hence if the most probable result of m -f- n
trials be that E will happen m times and fail n times, the

probability that it will happen m x times and fail n + x

times is the same as the probability that it will happen m -\-
x
times and fail n x times.

The following example will suffice to shew the applica

tion of the formula : A
thrown 6000 times, required
die is

the probability that the number of aces turned up will be

exactly 960 ?
Here p, the chance of throwing ace, is J, q=$ 9 and
7i=6000; whence m=hp= 1000, and n=hg=50QQ. We
have first to find P , the chance of the most probable result,
 INVOLVING LARGE NUMBERS. 1 37

or of 1000 aces. By (90), P =\/(h--27rmn) whence, ;

substituting the above values, P =V3 V(5000x 3.14159).

On performing the operation indicated by the logarithmic
tables, we get log P =8.14050, whence P =.0138.
The calculation of
-
e 7u2 f2mn
"

is as follows: Assume
t
2
=hx 2 +2mn. We have x=l 000 960=40.
log 40= 1.60206
2

3.20412
log h log 6000=3.77815
2
log 7** =6.98227
log 2mn= log 10,000,000=7
2
log =9.98227
*

log e=.43429, log .43429=9-63778

2
log (* x-43429)^9.62005
t* x .43429=* 2 log e~ .41692

addlogP =8.14050

log P x =:7.72358
therefore P =.0053,
;E
which is the chance of 960 aces ex

actly. The odds against this event are therefore 9947 to

53, or nearly 188 to 1.

93. When h, m, and n are large numbers, and x is small,

-h*^ 2mn differs little from
the exponential e unity, and it de
creases slowly as x increases, so long as x is small in compa
rison of m and n. Suppose m~n and x^z^/m^ it becomes

e- l = = r
- ; so that if we assume m=100, the

10th term before or after the greatest would still exceed the
3d part of the greatest. But when x becomes greater than
138 SOLUTION OF QUESTIONS

A/n the exponential, and consequently also the terms

-v/m or
which are multiplied by it, begin to diminish with great

rapidity, and the diminution is more rapid as x increases.

Ifw=/z=100, and #=50, then the exponent &r 2 ~2ww=25,

so that fr***-&mn 9 = l_i_g 25
) a quantity which is
altogether
insensible. We may therefore conclude generally that when
h isa large number, the principal terms of the developement
of Q+gO* are those which are near the greatest term, and
thath may be taken so large that the terms towards the
beginning or end of the series may at length become smaller
than any assignable quantity.
94. From the proposition which has now been demon
strated it follows, that although the probability of that par
ticular compound event which has the greatest number of
chances in its favour is very small when the number of trials

is great, yet on account of the rapid diminution of the terms

towards the beginning andendof the series, the sum of a com
paratively small number of terms taken on both sides of the
greatest, may be very much greater than all the remaining
terms of the series ; and, consequently, there will be a very

great probability that the compound event will be repre

sented by one or other of those terms. This consideration
leads us to one of the most important questions in the theory,

namely, to determine the probability that in a large number

of trials, h, an event E, which must either happen or fail in
each trial, and of which the chance of happening in any
trial is p, will happen not less than hp / times, and not

oftener than hp+l times ; or, making hp=zm, hq=n, to de

termine the probability that the number of occurrences of

E will be included between the limits nf=zl.
Let x be any number between and /. Then (92) the

probability that E will occur (m x) times and fail

(n + x)
 i*K_ A^/7tf U/fc-/
yw

INVOLVING LARGE NUMBERS. 139

times isP x =.P e- " ^ 27W "(where P =V(h-s-27rmn). Now if

7 p2

in this expression we make a; successively equal to each of the

numbersO, 1, 2,.../, we shall have the respective probabilities

of E
happening TW, m 1, m 2, m I times in h trials ;

and the sum of these probabilities will be the probability that

E happens not oftener than m times, and not seldomer than
m I times. The same suppositions with respect to x will

give the probabilities of E happening m, m -J- 1 ,

m -f- 2, . . . m -f-
/

times, the sum of which will be the probability that E happens

not seldomer than m times, and not oftener than m-f-/ times.
Adding, therefore, those two sums, and deducting P the pro-

bility which corresponds to x=0, on account of its being in

cluded in each sum, and therefore having been counted

twice, the result will be the sum of the terms of the binomial
h
(p+q) comprised between, and including, the two terms
of which the first has for a factor jo* , and the last p
m ~l
,

and will therefore express the probability that the number of

occurrences of E will fall within the limits m=zl. Let
this probability be denoted by R, and let SP* represent the
sum of all the values of Px obtained by substituting suc

cessively 0, 1, 2, 3,..,/ for x, we then have R=2SP^ P ,

whence, writing for P^ and P , their values,

h \ _*f! / h
27rmn

95, In order to find an approximate value of this expres

sion we must have recourse to a formula first given by Euler
for converting sums of the kind denoted by S into definite
integrals (for which see Lacroix, Traite du Calcul Differen-
tiel et
Integral, torn. iii.
p. 136, or Herschel s Treatise on

Differences, p. 513). Assuming u to denote a function of

x, the formula is as follows :

140 SOLUTION OF QUESTIONS

/* 1 1 du
bw = /udx +
,
u +- ^5-3 k- &c. 4- constant.
t/ 2 2 b ax
1 . .

On making u=Px P er-***42mn we g ntj .._. 2

ax mn
e -hx2~2mn. therefore, if we suppose a? to be not greater
than */m or </n, this differential coefficient is of the order

l-^-h, (as may be easily shewn by substituting hp for m, and

hq for ri),
and may be rejected, since h is supposed to be a
very large number. The above equation therefore becomes

SPX = P e-hx ^ 2mn dx + P e-h * ^ 2mn + constant ;

and on supposing #=0 this gives 0= ^P + constant,

therefore the constant is
equal to |P , and we have

Assume t=x\/(h-t-2mn), whence dt=dx*/(h--%mri) ;

substitute these in the above equation, and it becomes by

reason of P =</(h--27rmn),

whence, from the equation R=2SPa. P we , obtain

The integral in this expression must be taken between the

limits #=0 and x=l. When #=0 we have also =0, and
when x=l, then t=li/(h-+- 2mn) ; therefore assuming r for

the limiting value of t, that is, making T=/t/(A-*-2m), and

results
therefore l=T+/(2mn--h), there

for the probability that the number of occurrences of E will

fall between m=^r^(2mn-^-h) ; or, replacing m and n by

 INVOLVING LABGE NUMBEKS. 141

hp and hq, the probability that the number of occurrences

of E will fall within the limits hpzZT^(V,hpq), or be equal
to one of those limits.

This expression for R has been found by neglecting

quantities of the order of smallness 1 -t-h ; consequently

the greater the value of h the more nearly it
approaches to
accuracy, but it is
only rigorously true when h is infinite.

2
96. The integral^" dt is
computed as follows. Develop
tz
ing the exponential e~ in a series of the ascending powers of
2
t , and integrating the successive terms between the limits

=0 and t=r, we find

a series which converges rapidly when r is less than unity.

In the contrary case, however, or when T is
greater than

unity, the series is divergent, and it is necessary to proceed

by a different method. Let the factor e~^ be multiplied
and divided by t; we have then

r tr**dt=r^<r
and on integrating by parts

Repeating the same process on the last integral, and so

on with the last after each succeeding integration, the fol

lowing series is obtained,

When t= infinity the right

hand side of this equation be
comes ; whence between the limits =r and t
infinity,

we have
142 SOLUTION OF QUESTIONS

a series which converges very rapidly when r is greater

than unity. Now the value of a definite integral between
and infinity is obviously equal to its two parts, of which
the first is taken between and T, and the second between
r and infinity ; that is to say,

r
C
oo
e^dt is well known 1
to have for its expression
/oo
therefore

r
so that the integral may be computed from either of the

above series, according as r is less or greater than 1.

The integral fe^dt is of great

importance in the higher
mathematics. It occurs in the investigation of the path of
a ray of light through the atmosphere, and of the law of
the diffusion of heat in the interior of solid bodies, as well
as in the determination of the degree of reliance that may
be placed on the results of astronomical observations, and
generally in most of the more difficult and important ap
plications of the theory of probabilities. A table of its

values from =0 to =3, for intervals each = 01, was given

by Kramp, at the end of his Analyse des Refractions Astro-

e tz
dt=^-r is ascribed by Gauss (Theoria
/OD >

Motus Corporum Ccekstium, p. 212,) to Laplace. On making

J

et*=z, the integral is transformed into rfzflog - \ between

{J
the limits 2=0 and z=l, the value of which =^\/*> nad b een
given by Euler, long before, in the Petersburg
Memoirs. See Le-
301.
gendre, Exercises du Calcul Integral, torn. i. p.
 INVOLVING LARGE NUMBERS. 143

nomiques, Strasburg, 1799- In the Berliner Astronomisches

Jahrbuchfor 1834, there is also a table of its values from
tQ t2 (for the same intervals) multiplied by
to 2-^-/y/7r,

with their first and second differences, for the purpose of fa

cilitating interpolation. This last table, which appears to

have been derived from that of Kramp, and which is imme

diately applicable in the calculation of the probability R,
we have extended to t=3, and given at the end of the pre
sent article. As the function which is thus tabulated will
occur frequently in what follows, we shall in future, for con
venience in printing, denote it by 0, that is to say, we shall

assume

the two forms being equivalent in consequence of the above

equation.
97. Some very important conclusions follow immediately
from the formula in (95). The quantity R denotes the pro

bability that in a very great number of trials 7^, the event

E, of which the a priori probability in any trial is p, will

occur not seldomer than Tip / times, and not oftener than
hpJf-l times, or that the number of its occurrences will be
included between the limits hpizl, or at least be equal to
one of those limits. Hence R also denotes the probability

that the ratio of the occurrences of E to the whole number

of be included within the limits pi=.l--h. We
trials, will

have assumed T=l\/(h+2mn) ; but m=kp and n=hq ;

whence l=r\/(2hpq), and con
therefore T=l-r-*/(2hpq),

sequently l--hTV(%pq--h). Now, if we suppose r to be

constant, so that the probability expressed by R may remain
the same, then p and q being given, / is
proportional to
144 SOLUTION OF QUESTIONS

the square root of A, and consequently the greater the

number of trials the smaller will / be in proportion to that

number. Thus, number of trials be 1000, and we

if the
have a given probability R that the number of occurrences
of E will not differ more than 10 from the number which is

the most probable of all (that from lOOOjs), then if we

is,

take 100,000 trials, we shall have the same probability R

that the number of occurrences of E will not differ more than
10 X V 100= 100 from the most probable number. But
a difference of 1 in 1 000 is 1 - 1 00th of the whole, whilst a dif

ference of 100 in 100,000 is 1-1 000th of the whole, and thus

the ratio of I to h becomes smaller and smaller, or the ratio
of the occurrences of E to the whole number of trials ap
proaches nearer and nearer to p, as the number of trials is

increased and the experiments may be repeated until the

;

difference between p and p-^=.l-~h, in respect of a given

probability R which may be as great as we please, shall

be less than any assignable quantity.

If, on the other hand, we suppose l-s-h to be constant,

then r is
proportional to the square root of the number of
trials. But as r increases, 0, and consequently R, approaches
nearer and nearer to unity, (and it may be seen, by referring

to the table, that it is

only necessary to have rr=3 in order
to have 0= 9999779) ; whence the number of trials A may

always be increased until we obtain a probability approach

ing as nearly to certainty as we please, that the number of
occurrences of E will be comprised within the given limits

(hp-==.l) ; or, which is the same thing, that the ratio of the

number of occurrences of E to the whole number of trials,

from p, the probability of E in a single trial?

shall not differ

more than a given quantity l-r-h which may be less than

any assigned fraction. This is the celebrated theorem
 INVOLVING LARGE NUMBERS. 145

which was demonstrated by James Bernoulli in the Ars

Conjcctandi.
98. The application of the preceding results to numeri
cal examples, is rendered extremely easy by means of the
table of the values of 9. From the formula in (95) we
have the probability

that the occurrences of E in h trials will fall within the

limits 7</?rz/, the relation between / and r being given by
the equation lT*/(2hpq). If, therefore, we suppose / to

be given, r becomes known, and the corresponding value

of is found from the table ; and,
conversely, if be as
sumed, r is
given by the table, whence the corresponding
limits / are deduced. With respect to the quantity P e~ rZ ,

we may observe that it denotes the probability that the

number of occurrences of the event E will be hp -{-/or hp I
precisely (92), and is therefore
always a very small fraction
when is a large number (90). It may be regarded as a
h
correction of 0, which in most cases might be omitted with
out sensibly affecting the result ; but when h is not very

large, or / is a small number, becomes necessary to take

it

it into account. In such cases its value may be computed

directly as in the example in (92) ; but this labour may be

avoided by increasing r, so as to include it within the

limits of the integral 0. Thus, let R be the probability

that the number of arrivals of E will be included within
the limits lip^=.x, and R the probability of the limits be

ing hp^=i(x-\-l), and let and be respectively the

corresponding values of the integral. We have then,

giving P.,. the same signification as in (92), the two equa
tions
146 SOLUTION OF QUESTIONS

and the difference R R of these two probabilities is ob

viously the double of the probability that the result of the

trials will be either (hp -|- x -f- 1 ) times E, or (hp x 1)
times E, exactly. But the chance of either of these events
being P^i, we have therefore R -^-R=2P.r+ i. Now, when
h is
large, P* and P^+i are very small, and very nearly equal
to each other, (their difference is in fact of the order of quan
tities omitted) ; hence R R= 0, and also 2P^ +1 2P X ,

and consequently ;
0^2?^,, or P x =(> e). Substi
tuting this value of P^ in the equation R e-f-P^., we get R=J
(0 _j-0) ; so that if we take from the table the values of
and corresponding to /and /+ 1, half their sum will give R.
But as the interval between and in the table is always
small, half their sum will not differ sensibly from the value
of corresponding to /+^> whence this value of is

equal to R, and we have the following rule for determining

the limits corresponding to a given probability, or vice
versa :
When the limits are assumed, find r from the equation

l-\-jjr\/(2hpq) ; then the value of in the table, corre

sponding to r is the probability that in h trials the number

of occurrences of the event E, the chance of which in a sin

gle trial is p, will lie within the limits hp-==.l both inclusive.

Conversely, when isassumed, find the corresponding va

lue of t in the table, by means of which the limit / will be
given by the equation/+ ^=rV(2/ipq). It is obvious, that
if and the probability
the limit / be both assumed, then
h may be determined from the same equation,
r 99. We will now give some examples of the application

of the preceding formulae.

Suppose pq=^ and A=200, and let it be proposed to

assign the limits within which there is a probability =J

 INVOLVING LARGE NUMBERS. 147

that the number of occurrences of E will fall. In this case

the equation 1+^ =T</(2hpq) becomes /-f
J r^/ 1 00= 1 Or*
Now, it is
easily found from the table that for e=l we have
r 4769, whence /+ =4-769, and J=4-269. On tossing
a shilling 200 times, it is more than an even wager
therefore
that head will turn up net seldomer than 95 times, and not
oftener than 105 times.

Suppose p=q=i, /*=3600,and let it be proposed to assign

the probability that the number of occurrences of E will not
exceed the limits 1800r:30. In this case the equation

/4-i = T-v/(2/^) becomes 30-5 = rv/(2x 900) = 30rv/2,

whence = 30-5-f-30-v/2 = -7l89; and the table gives
T

=-6907= ff nearly. Hence in tossing a shilling 3600

times, the odds are 28 to 13 that head will not turn up
oftener than 1800 + 30=1830 times, nor seldomer than 1800
301770 times. Neglecting the second term of R (95)
and taking simply /rrlOr, the table gives e= 6827, which
is the solution given by Demoivre, p. 245.

Suppose />=|> y = $> an d let it be proposed to determine

how many trials must be made in order that it
may be one
to one that the number of occurrences of E will not differ

more than 10 from the most probable number.

For J-
we have r= 4769 therefore the equation
;

l+l=rV(Zhpq) becomes 10-5=-4769vXlO&-f-36), whence

2
/*=3-6(10-5-:--4769) On computing this formula h is found
.

t=t:l745-2. Say 1746, .J of which is 291 ; and it follows that

if a die be thrown 1746 times it is an even wager that the

number of aces will fall between 291=^10, that is, be

tween 281 and 301, or be equal to one of those numbers.
In (92) we found the probability to be -0053, that in
6000 throws of a die the number of aces will be exactly
148 SOLUTION OF QUESTIONS

960. Let itnow be proposed to assign the probability 0,

that in 6000 throws the number of aces will lie between
960 and 1040, that is, between 1000qp40. Here ^6000,
/>=, <? >
and /=40 ; the equation of the limits therefore

becomes 40-5=7V(10000-5-6), whence T=-405 v 6=-992,

corresponding to which the table gives 0=-8394.
The following question is discussed by Nicolas Ber
noulli in the Montmort s Analyse des Jeux de
Appendix to

Hazard, and is noticed by Demoivre and Laplace. From

the observations of the births of both sexes in London dur

ing 82 years (from 1629 to 171 1) it was found that the aver
age number of children annually born in London, was about
14,000, and the ratio of the number of males to that of fe

males, was nearly as 18 to 17, the average number of male

births being 7200, and of female births 6800. In the year
in which the greatest difference from this ratio took place,
the actual numbers were 7037 males and 6963 females, so
that the difference from the average amounted to 163.

Assuming, then, the comparative facility of male and female

births to be as 18 to 17, required the probability that out of

14000 children born, the number of males shall not be greater

than 73M3, nor less than 7037.
This question is evidently equivalent to the following :

Let 14000 dice, each having 35 faces, 18 white and 17

black, be thrown ; what is the probability that the number
of white faces turned up, will be comprised between the
limits 7200dtl63. We have therefore =1400, j=J,
<7=i|. /=163, and the formula l+%=T</( 2hpq) becomes
l63,5=V(2x 14000X 18 X !7)-=-35, whence 1-955. T=
The corresponding value of is found from the table= 9943,
which is the probability that the number of white faces shall
 INVOLVING LARGE NUMBERS. 149

not be greater than 7363, nor less than 7037. The odds
in favour of the event are therefore 9943 to 57, or about
175 to 1.

100. We now proceed to consider the case in which the

probabilities of the simple events are not known, a priori,

but inferred from the results of experience. It was shewn
in (52) that the probability n of an event happening in
f
times, and failing n times in h trials, (h -f n ), when =m
it has been observed to happen m times, and fail n times in

^previous trials, is expressed by this equation

Now, when m, n, m , ri, are large numbers, an approxi

mate value of n, more accurate in proportion as those

numbers become larger, is obtained from Stirling s theorem
x
(90), which for any number x gives \_x~\=x?e~ \/(2iTx}.

Applying the theorem therefore to the expressions within

the brackets in the above equation, and assuming

-
/
l*
we obtain, in consequence ofm-\-n=k,

n
j
(+
Let m =0m, n ~Gn, and consequently h =.6h ; then taking

, for 7 /" , ^h
h .
,., (which mav be done with-

out sensible error, since h is

by supposition a large num
ber,) the equation becomes
m(\ + 6) m + mfn

^-w =A
/
or, since m-}-n=h, / ,
150 SOLUTION OF QUESTIONS

r^TiTT

Making the same substitutions in the expression denoted

by K, we get, after reduction, K= 1-^(1 + 0); whence,

m n \n
_ (m\ (
Vr+0U7 \f)
The value of n now found, is the probability that in a
future series of trials the ratio of the occurrences of E to
those of F will be the same as in the
preceding trials, which
are supposed to have been very numerous. If the chances

of E and F had been given a priori equal to m-r-h and

n-i-h respectively, the probability of m f
times E, and n
times F in m -\-n future trials would have been P==
m\m / n \ n/

T) (T)
n ~-h between the probability P of
z=n-i-h), the relation
that combination of simple events which has the greatest
number of chances in its favour, when the chances of the

simple events are known a priori, and the probability of

the same combination when the chances of the simple events

are only presumed from previous trials, is

expressed by
this equation,

101. When h is very small in comparison of h, 6 be

comes a very small and may be neglected, and we

fraction,

have then n P . But when h f is a number comparable

with h, n is less than P ; and it diminishes rapidly when 6
exceeds 1. The reason of this is obvious. If the con

tents of the urn are not known a.priori, however numerous

the trials may have been there is only a presumption that the
chance of drawing a white ball in a single trial is measured
by WH-A; whereas, in the case of the ratio of the balls being
 INVOLVING LARGE NUMBERS. 151
*

previously known, the measure of the probability is cer

tain. As an instance of the manner in which the proba

bility of an assigned series of future events diminishes,

when the probabilities of the simple events are inferred
from experience, let us suppose h =h, whence 0=1, and
consequently n=P -r- -v/2=-7071 XP - Now it was shewn
in (91) that if a ball be drawn at random 100 times from
an urn which contains an equal number of black and white
balls, the probability P that the result will be 50 white
,

balls, and 50 black, precisely, is -07979- It follows there

fore, that if the contents of the urn be unknown, and

we can only judge of the relative numbers of the two sorts

of balls it contains from having observed that in 100 trials

there have been drawn 50 white balls and 50 black, the

probability n of that combination in 100 future trials, be

comes -07979 X -707 1 05642.
102. The result obtained in (100) enables us to de
termine the probability that the number of occurrences of

E in h* future trials, will not differ in excess or defect from

the most probable number, by more than a certain given

number /. It has been shewn (95) that in the case of the

probabilities and q of the simple events being given a

p
priori, if we determine r from the equation l=
the formula

gives the probability R that m will be comprised within the

limits hp-=*=.rV(2hpq) ; or, dividing by h, the probability

that the ratio of m to A be comprised within the limits

will

h). Conversely ^whenp and q are not known,

1
This inference, though admitted by both Laplace and Poisson,
is not strictly correct. In a paper published in the Transactions of
the Cambridge Philosophical Society, (vol. vi. part iii.) Mr. De Mor-
152 SOLUTION OF QUESTIONS

but the event E has been observed to happen m times in h

triAls, then
R -f V(h-lTTmn)e-
r-

gives the probability R that p is

comprised within the limits

m_^_T /2mn
~Ti~h*l ~h
These limits approach more nearly to each other as h in
creases ; and when A is a large number, the ratios m--h,
n-i-h maybe assumed, without sensible error, as the chances
of E and F in computing the probable result of a future

series of A trials, provided, however, that h (though abso

lutely a large number) be small relatively to h. When this
condition is not fulfilled, the assumption of m--h and n-~h
as the E and F, might lead to consi
a priori chances of
derable error but an approximation to the limits corre
;

sponding to a given value of R may be obtained from the

following considerations :

Suppose a large number h of events to have been observ

ed, and that the result of the observation gave times E m
and n times F. Let a new series of hf trials be made, and
that in this new series p is the real chance of E

gan has shewn by a direct analysis that in the case of p and q not
being known a priori, but made
equal to the observed ratios m-S-h,
w-f-A, the presumption of the true value of p lying within the limits
I ^
stated in the text is not, as there inferred, -f-

but ---4n----- e The last correction to is smaller

6hpq \/ TT
than the former, and being divided by A, is of the order of quantities
that have been rejected in the approximations. It is right to state

that the method of simplifying the calculation of R in the direct case, .

by taking the integral between limits corresponding to l+\ instead

of /, is noticed, for the first time so far as we are aware, by Mr. De
Morgan in the same paper.
 INVOLVING LARGE NUMBERS. 153

and q of F ; we have then a given probability R that

the number of occurrences of E will fall within the limits

kfpz=rV(%Jt pq), Now, for p and q substitute the ratios

observed in the first set of experiments, namely, m-^-h and

n~h, and the limits corresponding to R become

which, therefore, are the true limits on the hypothesis that

the chance of E in a single trial is m~h. But as this
chance not certain, but only presumed, the limits require
is

to be extended in order that may preserve the same va R

lue. Confining our attention to 0, the first term of the

expression for R (the second may be disregarded in the pre

sent approximation), let h =m -}-n and m! n =.m : : n, then
hf
is the sum of the terms of the binomial (p-}-q) from that
in which the exponent of p is m 4- / to that in which the

exponent is m /. Now, when p and q are given apriori,

the chance of m times E and n times F in h trials is P ;
f

and when p and q are only presumed from the results of

previous trials, the chance of the same combination is n ;
and (100) n is less than P in the ratio of 1 to ^/(l -{-6). In
likemanner, the chance of each of the other combinations
of E and F included in the integral will be less in the

case of p and q presumed, than in the case of p and q given,

in the same ratio of 1 to */(\ +6). But it has been seen

r

(93) that when h is a large number, the terms of the de-

h
velopement of (p -f- q) which are nearest the greatest term,
diminish at first very slowly ; and, further, that only a small
number of terms on each side of the greatest are required

to be taken, since I is less than +/m or */ri (95) ; we may

therefore, without sensible error, assume to be proportion
al to the number of terms included in the summation, or
154 SOLUTION OF QUESTIONS

that the value of not be changed if we include in the

will

summation a number of terms greater in proportion as the

value of each individual term is less. Hence it follows that
the limits must be increased in the ratio of y^ !-}-#) to 1,

and the value of corresponding to r will give the proba

bility that the number of events E, in /* future trials, will

be included between

103. The following question may be proposed as an ex

ample of the application of the last formula. Out of a given

number h of individuals taken at the age A, it has been ob
served that m are alive at the age A + a; required the pro
bability that out of h other individuals taken at the same age
A the number who survive at the age A -f-
a will be includ
ed between m :=/, the ratio of in to h 1 being the same as
that of M to h.
To solve this question, we have to find r from the

/= and the corresponding

equation ^ \/(2h mn(l-\-6)) ;

value of in the table, will give the required probability.

From the table given in the article MORTALITY, vol. xv.

p. 555, it
appears that out of 5642 individuals taken at the
age 30, the number surviving at the age 50, according to
the Carlisle Table, is 4397. Taking those numbers as an
example, we have /*=5642, ^=4397, ^=1245; and as

suming also 7^ = 5642, whence 0=1 and \/(\ + 6)=^/2,

the equation of the limits becomes I=T X 62-30, Let it be
proposed to determine I from the condition 0= J. In this
case the table gives T=*4769, and we have consequently
fc=29 7. Hence it
appears, that if it has been observed
that of 5642 individuals taken at the age of 30, 1245 die
 INVOLVING LARGE NUMBERS. 155

before reaching the age of 50, it is an even wager that out of

5642 other individuals also taken at the age of 30, and sub
jected to the same chances of mortality, the number who die

before reaching the age of 50 will lie between 1245r:30,

that is, between 1215 and 1275.
104. The following experiment recorded by Buffon, in
his Arithmetique Morale, affords an example of the ap
plication of the preceding formulae to the determination
of the probable existence of a physical cause from the
results of a large number of observations. A piece of
money was tossed 4040 times successively, and the result
was head 2048 times, and tail 1992 times. Supposing the

piece to have been perfectly symmetrical, the most pro

bable result would have been the same number of heads
and tails. Let it now be proposed to assign the probability

afforded by the experiment that the piece was not symme

trical, and that its form or physical structure was such as

to render head an event, a priori, more probable than tail.

In this case =4040, ?w=2048, rc=1992; and by (102)

we have the probability R (or 0, neglecting the correc

tion) that p, the unknown chance of head, is comprised

m m
between the limits
-^^. r /2mn Now - = 2048

=.50693, and^Y?^
~ =r X fi
.011124 ; therefore if we

assume r x011124=.00693, we shall have the probability

that
p comprised between the limits .50693rr.00693,
is

that is, between two limits of which the least is .5, or one-

half. This assumption gives T=.00693~.011124=.623 ;

and the corresponding value of is found from the table

=,62170. Now if p lie between the above limits, its value

is evidently greater than ^ ; but the probability of its

lOO SOLUTION OF QUESTIONS

between those limits is not the whole probability that p is

greater than -J ; for there is a chance of its exceeding the

greatest limit, in which case also its value will be greater
than l. The probability that p is not comprised between
the assumed limits is 1 .62170 =.37830 ; and if it is not

comprised between these limits, there is an equal chance of

its being greater]than the greatest limit, or less than the least ;

the probability of its exceeding the greatest limit is conse

quently J X .37830=. 189 15. Hence the whole probability

that p is
greater than .5, or that the chance of head is

greater than that of tail is .62 170 -f .189 15 =.81 085 ; and
the odds are therefore 81 to 19, or rather more than 4 to 1
that the piece was not perfectly symmetrical.
105. The formulas which have been demonstrated in the

present section are immediately applicable to the determina

tion of the probable limits of the gain or loss which may
arise from undertaking a great number of risks with a given

expectation in respect of each. The following question has

important practical applications. interested in a great A is
number of similar enterprises, in each of which E or F must
necessarily happen. When E happens he receives the sum
a, and when F
happens he pays the sum ; required the

probability that his gain or loss shall be comprised within

given limits ?

Let p be the chance of the event E, q that of F, and k

the number of enterprises. Suppose E happens m times,
and F n times ; the sum to be received will be ma, and the
sum to be paid will be w/3, and therefore his gain will be
ma n&. Let m=/ip, n=hq, then m times E and n times
F is the most probable result, and in this case the gain
ma w.,3 becomes Ji(pa q&). Find T from/-j-
J TV(2//p,y),
then (98) is the probability that the number of occur-
 INVOLVING LARGE NUMBERS. 157

rences of E will lie between the limits f}p+.l. But if E

happens hp I times, and consequently F hq + l times, the

corresponding benefit is
(hp l)a (hq-\.l)p=h(pa ^,3)

/(a-f-/3) ; and if E happens hp -f / times, and F hq / times,

the benefit is
(hp^-l}a(hql)^=/t(p:iq3)-{-l(a + ^);
whence is the probability that his gain, that is, the diffe

rence between what he receives and what he pays, will be in

cluded within the limits h(pa <jr,3)rp:/(a-f-/3) both inclusive.
106. The following conclusions follow immediately from
this solution.

(1). If pa be greater than qfr so that A has a mathema

tical advantage (however small) in each risk, the risk may
be repeated a sufficient number of times, or h may be taken
a sufficiently high number, to give a probability as nearly

equal to certainty as we please, that A s gain shall exceed

any given sum, however great.
(2). Let there be two players A and B, whose chances
of gaining a game p and q, and let /3 be
are respectively

the sum staked upon each game by A, and a the sum staked

by B, then pa is the mathematical expectation of A in re

spect of a single game, and qft that of

B ; and if pa be
greater than q$ (however small the difference) the game
may be repeated so often as to give rise to a probability
approaching as nearly to certainty as we please, that A s
gain shall become equal to the whole of B s capital, and,
consequently, that B will be ruined.

(3). If the mathematical expectations

of the two players
be equal, then pa ;3zrO, and the most probable individual

result of a large number of games, is that the gains and losses

on either side shall be the same. But if /be supposed constant,

then T is
inversely proportional to Vh- and consequently the
game may be repeated until 0, the probability that the
158 SOLUTION OF QUESTIONS.

gain or loss /(a+) shall be comprised within given limits,

shall become as small as we please. Hence 1 0, the pro
bability that the gain or loss shall not be comprised within
given limits, may be rendered as great as we please ; and it
follows that although the play may be on terms of perfect
equality, it
may be continued until a probability shall be
obtained, approaching as nearly to certainty as we please,
that one of the two players shall be ruined.

(4). The number of games which must be played, to

afford a given amount of probability that one of the parties
shall lose the whole of his fortune, depends on the magni
tude of the stakes (a-{-/3); but whether the stakes be large
or small, the final result is the same. When the stakes
are small, a greater number of games must be played.
107. As an example of this class of problems, we may
take the following question : A and B engage in play
with equal chances of winning, and stake five sovereigns on
each game how many games must they undertake to play
;

in order that it may be two to one that one of them shall

lose at least 100 sovereigns?

Here>=^, q=^, a=5, /3=5, and Z(a-j-/3)=100, whence

/=10. The equation l+=r\/(2hpq) therefore becomes
10-5=iV(-i-2), whence h2 x (10-5) 2
-f-T
2
.
Now, the
odds being 2 to 1 against the limits of the gain or loss
not exceeding 100, the probability 9 of the limits not ex

ceeding 100 is
^=-33333, corresponding to which the
table gives by interpolation r=r -30458 ; substituting which
in the above equation we find A=r2376 8; so that if 2377
games
are played, the odds are 2 to 1 that one of the players shall

have gained 10 games more than half that number,

at least

and, consequently, that the other shall have gained at least 10

less than half, or that one of them shall have gained at least
 INVOLVING LARGE NUMBERS.

20 games more than the other, and consequently have gain

ed at least 100 sovereigns.
It is to be carefully observed that this question supposes
the account between A and B not to be balanced until
2377 games have been played. If the condition of the play
had been that it should cease as soon as A or B should have
lost 100 sovereigns, the question would have been of an

entirely different kind, and a much smaller number of games

would have given the same probability of an equal loss.

108. The question just alluded to belongs to a class of

problems connected with the Duration of Play of extreme >

difficulty, and which have given rise to some of the most

abstruse and refined researches in the modern analysis. In

order to give an idea of the subject, we may take the fol

lowing question, which has been frequently consideced.

A and B, whose chances of winning a game are respec
tively p and q, play on these terms A has m counters, and :

B has n counters ; when A loses a game he gives a counter

to B, and when B loses a game he gives a counter to A,

and the play is to cease when one of them has lost all his
counters. What is the probability that the play, which

may go on for ever, shall be finished before more than k

games shall have been played.

To take a simple case, suppose each to have three coun

ters, and let the probability be required that the play shall
be concluded with or before the ninth game. As the play
cannot end with less than three games, let the binomial
5
(p-\-q) be developed, and the terms
P s + 3/> 2 7 + 3/^ 2 + q *
give the respective probabilities of all the cases which can
arise in three games. The first term is the probability of
A gaining all the three games, the last term is the proba-
160 SOLUTION OF QUESTIONS

bility of B gaining them, and the sum of the remaining two

terms is the probability that neither will win all the
games,
or the chance that a fourth will be played. Now, if the
fourth game be played, p is As chance of winning it, and

q Bs chance ; but these chances

will only exist in respect

of the fourth game, provided the play be not concluded with

the previous one, the probability of which is
3p-q + 3pq*.
Multiplying, therefore, 3p-q + 3pq 2
by p + q, the product

of the different ways in

gives the respective probabilities
which the four games may be gained by A and B, except
in which the play would have terminated
ing the two ways
with the third game. But the play cannot end in any of
these ways ; for, taking the first term example, if B
for

gains a
counter before A gains three, the play cannot ter
minate until A
gain back that counter, and three others
besides, so that five games must be played. In fact, it is

obvious that there is no way of gaining an odd number of

counters in an even number of games, or vice versa. The
the chance of the 5th game
last product therefore expresses
2
being played ; and by +
reason ofp gf=l it is equal to 3p q
2
the chance of the 4th being played, as it obviously
4- 3pq >

ou^ht to be, since the play cannot terminate with the 4th.
be played, p is A s chance of gain
Again, if the 5th game
ing it, and q B s chance of gaining it; multiplying there

fore the last product by p + q, the different terms of the

result, namely,

of all the cases which can

give the respective probabilities
arise by the 5th game. The first term is the probability of
A gaining 4 games and B gaining 1, and the last term is the
 INVOLVING LARGE NUMBERS. 161

probability of B gaining 4 and

gaining 1. A
These
terms therefore are the probabilities of the play ending in
favour of A and B respectively with the 5th game, and the
sum of the other two terms is the probability that the play
will not terminate with the 5th
game, or the chance of
another game being played.
By pursuing the same reasoning it will be evident that
on rejecting the two extreme terms of the above product,
and multiplying the remainder byjo-fg, there will result
the probabilities of the different ways in which six games

may be played without the one player gaining all the

counters of the other. But as the play cannot termimate
with the 6th game, multiply again by p-\-q> and the result

will indicate the probability of the different cases that can

arise out of the 7th game. Rejecting the two extreme
terms, which give the respective probabilities of the play

being concluded in favour of A or B, and multiplying the

remaining two first by p -\-q to obtain the different proba
bilities in respect of the 8th game, and again by p-\-q, as the

play cannot terminate with the 8th, we have the product

27 jy + 81 jy + 81 />y + 27/>y,
of which the first and last terms give the respective chances
of A and B winning at the 9th game, and the sum of the
other two terms the probability that the play will not be
concluded by the 9th.
If we now collect the terms which have been set aside

in the successive products, and denote by a and b the

respective probabilities of A and B gaining at the 9th

game, or sooner, we shall have

162 SOLUTION OF QUESTIONS
3
a=p + 3p*q + 9/>Y + 27 jo

where the law of the series is evident.

It is easy to see that this process may be applied
whatever be the number of counters which A and B have
at the commencement, and whatever be the number, h,

of games to which the play is limited. The general rule

is as follows : of the two numbers m and n, let m be that

which is not less than the other. Raise p -\-qto the power

n, and reject the first term (which gives the chance of

A winning n games in succession), and also the last if
mn. Multiply the remainder (1i n) times in succession
by (p-\-q\ rejecting at each multiplication the first or

last term of the product when it gives a combination which

would terminate the play in favour of A or B ; the sum
of the terms rejected from the left-hand side of the dif
ferent products gives the probability in favour of A, and
the sum of the terms rejected from the right-hand side the

probability in favour of B. As the coefficients of the suc

cessive products are obviously formed by adding the coeffici

ent of the corresponding term in the preceding product to

that of the term immediately before it, the products may
be written down at once without the trouble of multiplica
tion; but it is evident that when m, n, and h are large num

bers, it would be quite impracticable to sum the series

formed of the rejected terms by the ordinary methods.
From the manner in which the series are derived, they are
called recurring series ; a general theory of which was first

given by Demoivre in his Doctrine of Chances, and forms

the most remarkable portion of that work.
109- The general problem is reduced to an equation of
finite differences as follows i Let yXit represent As expec-
 INVOLVING- LARGE NUMBERS. 163

tation when x games have been played, and he has still t

counters to win, or B has t counters in his hand. If A gain

the next game the value of his expectation will become
y,+i,f_i, and the chance of his gaining it
isjo; therefore
his expectation in respect of that event is
py x +\, ti- On
the other hand, if A loses thenext game his expectation
will become yx + \,t+i-> and the chance of losing it is q ;
therefore his expectation in respect of that event is
qyx +i, +i
t

Hence, according to the principles laid down in (32),

a linear equation of finite differences, with three independ

ent variables. It is therefore on the integration of an equa

tion of this kind that the problem of the duration of play

ultimately depends, but the subject is of much too compli

cated a nature to admit of its.

being satisfactorily explained
in this place. We must therefore content ourselves, with

referring the reader to the treatise on generating func-*

tions, which forms the first

part of the Theorie Analytiquz
of Laplace. 1

1
Lagrange, in vol. i. of the Memoirs of the Society of Turin, was
the first who shewed that the investigation of the general term of
a recurring series depends on the integration of a linear equation of
finite differences. In vols. vi. and vii. of the Memoires presentes a
f Academic des Sciences of Paris, Laplace proposed a general method
for the summation of recurring series by the integration of such equa

tions,and in the latter volume gives a number of examples of their

use in the more complicated questions in the theory of chances,
amongst which is the problem enunciated in ( 108). The subject
was afterwards resumed by Lagrange in the volume of the Berlin
Memoirs for 1775, where he has given a more direct method than that
of Laplace, for the integration of the class of equations in question,
and also applied it to the solution of the principal problems proposed,
in the works of Montmort and Demoivre. A
general solution of
the problem in the text is given by Ampere in a Tract entitled Con
siderations sur la Theorie Mathematique du Jeu> (Lyons, 1802.)
164 RESULTS OF DISCORDANT OBSERVATIONS,

SECTION IX.

OF THE MOST PROBABLE MEAN RESULTS OF NUMEROUS

DISCORDANT OBSERVATIONS, AND THE LIMITS OF PRO
BABLE ERROR.

110. In the preceding section we have considered a

class of questions which apply to events depending on con
stant causes, and supposed to be of such a nature that they

necessarily happen or fail in each experiment, and have

given formulae by which approximate results can be ob

tained when the numbers involved are so large that they
cannot be conveniently treated, or cannot be treated at all,

by the ordinary methods of calculation. We come now to

a more difficult problem, namely, to investigate the pro

bable result of a large number of observations which have
reference not to the simple occurrence or failure of a cer
tain event, but to the magnitude of a thing, susceptible,
within certain limits, of a very great or an infinite num
ber of different values, equally or unequally probable, the
chance of any particular value being also supposed to vary
in each experiment. On account of its immediate applica
tion to the determination of the most probable values of
astronomical and physical elements from the results of ob

servation, this is, perhaps, in reference to practical utility,

the most important question in the theory.
111. Let A represent a thing of any sort (as a line, or
 AND LIMITS OF PROBABLE ERROR. lo
an angle, or a function of any quantity) which may have

every possible value within given limits, or which may

be constant in itself, but of such a nature that its real mag
nitude can only be observed within certain limits of accu

racy,and suppose a great number of observations to be

made. The object is, in the first place, to assign the pro

bability that the sum of the observed values shall fall within

given limits, supposing the chances of the different values

ofA to be known a priori ; and, in the second place, when
the law of the chances is unknown, to determine from the
observations themselves the most probable mean value of

A, and also the limits within which there is a given amount

of probability that the difference between such mean value,
and the true, but unknown value of A, shall be contained.
112. Let a and b be the limits of the possible values of

A, x a value of A between a and b, and P the probability

that the sum of the values of A given by h observations
will be s exactly, s being a given quantity between ha
and lib. Assume the values of A to be equidifferent, and
multiples of a certain constant e, and make
ae=, Pf=l), o-e=S, ie^X,
where a, /3, and o- are whole numbers (which may be posi
tive or negative), whole number proportional
and i is also a

to x, and varying between the limits e=a, and ^=/3, and

which, therefore, may be positive or negative, or zero. If

the different values of A are supposed to be equally proba
ble, the chance of obtaining any given one of them, as x,
in a single trial is unit divided by the number of possible
values, or equal to l-i-(/3 a-f-1) and if we assume an
;

indeterminate quantity w, then (20) the number of combi

nations which give the sum of h values of A equal to o-e is

the coefficient of that term of the multinomial

lt>6 RESULTS OF DISCORDANT OBSERVATIONS,

(or of the developement of (Sw )* from z rra to z:=/3) in

which the exponent of w and consequently the pro
is a- ;

bability P that the sum of the values of A will be s ex

actly is that coefficient divided by (/3 a -|-l)\

113. If the chances of the different values of A are un

equal, and also vary in each trial, let p\ be the probability
of the observed value of A being x in the first trial, p% the
probability of its being x in the second,p 3 that of its
being
x in the third, and so on. Now when h= 1, or when there
is only a single trial, then s # e e, and we have P=p l
.

If 7i=2, then, assuming z e to be the value of A in the first

trial, and if its value in the second, (i and i

being any two
numbers between a and /3), the two observations may give
the sum of the two values equal to o-e in as many different

ways as it is
possible to satisfy the equation z-f-z =o-; and

consequently, according to the theory of combinations, P is

the coefficient o . that term of the product (arranged ac

cording to the powers of w) of the two series represented

1 ts
bySpjW* and 2/? 2 ?# ,
in which the exponent of iv is
equal
to ere. In like manner, if A=3, then the sum of the ob

served values of A may be equal to ere in as many different

ways as the equation i-\-i -\-i"=.<r admits of different so

lutions, and consequently P is the coefficient of the term of

2p 3 w
11 li
the developement of the product Sp^w .
Sp^v" . ,

in which the exponent of w is

equal to ere.
Generally, when
the number of observations is h, the probability P of the

sum of the observed values of A being s, or ere, exactly, is

the coefficient of w ff *
in the developement of the product

the sums "2,

including all values of i from i= a to e =
 AND LIMITS or PROBABLE ERROR. lt)7

6 "" l
Assume (e being the base of the Napierean
iu-=.e

logarithms), and let the above product be denoted by X.

We shall then have

Now since P is the coefficient of the term of the develope-

*~~1
ment of this product which contains the factor e 9
if

we conceive the developement effected we shall have

P + &c.

a series in which all the terms are of the same form. Mul
ff
*~ we1
tiplying both sides of the equation by e , get

Now by a well known theorem in trigonometry, ( AL.GEBRA,

r ~ ) ^-
art. 269), e<
|
= cos (<r <r)^+V lsin((/-o-)fj

substituting therefore this value, and multiplying by dd,

the equation becomes

Xe-*t V
I=
i"^=P^+P [cos (r
/
_-0*+\/^T sin (r *) } <#+
&c -

The factor which multiplies P in this equation will evi

dently become zero when integrated from 6= n to 6= -j-rr,

(TT being the semicircumference to radius 1), the positive
and negative elements of the integral being equal, and con

sequently destroying each other. The same thing also

takes place with respect to the following terms, which are
all of the same form. Integrating therefore between those
limits, and observing that /e?#=27r, we t
find

114. This value of P denotes the infinitely small chance

that the sum of the values of A in A trials will be s exactly.

Let p.
and v be two integer numbers between ha and h$,
and let Gt denote the probability that s will be comprised
168 RESULTS OF DISCORDANT OBSERVATIONS,

between the two limits ju and j/e, (these limits being in

cluded between ha and kb) ; then Q, will be found by sub

stituting successively y, /*-f 1, ^-j-2,...!/ for a- in the above

value of P, and taking the sum of all the resulting terms.
This substitution gives the following series multiplied by X
under the sign of integration :

On multiplying the series now found by

(=2/^7 1 sin 10), all the terms of the product, excepting
the first and the
destroy each other, and the
last, sum of
the terms becomes simply

therefore on making the substitution, and performing the

multiplication now indicated, and dividing by 2*J 1
sinj#,
we obtain for the value of Q the equation

115. In order to simplify the expression for Q, let the

number of possible values of A within the given limits be
conceived to be infinite, in which case the constant be
comes infinitely small, and therefore, since the limits are

finite, p,
and v infinitely great, Let the following substi

tutions also be made :

p=^ 8, vc=. ^-J-5, 6=fZ,

6 being positive in order that v may be greater than /t,

agreeably to what has already been assumed. On substitut

ing these expressions in the above equation, the limits of
the new variable z will be cr infinity; for e having been
supposed infinitely small, z must become infinitely great
when 6=7r. Now since p and v are infinitely great, /* J
and v-j- become sensibl /*
and i/,
whence we have
 AND LIMITS OF PROBABLE ERROR. 163

IsinSz. Again, by reason of 6=ez, we have

d6=edz ; and e being infinitely small, 6 must be a very
small arc, therefore ^6 may be taken for sin ^0, whence
d6+ sin ^6=2dz--z. By means of these transformations
the expression for Q, becomes

i/ AC
\t> YZ v L
am
o; n *~
o^ ,

and denotes the probability that the sum of the h values of

A will lie between i^^iS.

116. It is now necessary to assign a value to the product

denoted by X. Since thenumber of possible values of A
between a and b has been supposed infinite, the chance of
obtaining any given one of them, as x, in a single trial, is
infinitely small. Assuming this chance to be a function
of X) and to vary in the different trials, let it be represented
by <
na? in respect of the nth trial. In order to preserve

continuity in the values of A, this must be understood as

signifying that $ nxdx is the infinitely small chance that the
value of A given by the nth observation will lie between
x and x + dx. The function <J> n x, therefore, represents the
law of the facility of the different values of A. It is posi
tive for all values of x between a and b, and vanishes for

all values of x less than a or greater than b ; and it is im

portant to remark, that whatever number n may be, the

integral fcf) n
xdx taken from a? a to xb is
always equal to
unity ; for since every observation gives a value of A be
tween a and b, the sum of all the probabilities in respect
of each observation must be unity or certainty. From this

assumption, then, we have (f> l xdx=p li (f) 2 xdx=p 2

l
* l
<p h xdx=pM whence the sums Zp n e (11 3) are changed
17(7 RESULTS OF DISCORDANT OBSERVATIONS,

into definite integrals ; and therefore, since 6=ez, x-=if?

and consequently i0=xz, we obtain for the value of X,

the limits of the integrals being x=a and xb.

zx ~~ 1
By reason of e * =r cos zx-}- \/ 1 sin zx, each of

these integrals may be expressed in terms of the cosine and

sine of zx. The nth, for instance, becomes f(j> Hx cos zx.dx
-f d lf(j) nx sin zx.dx. Now since /</><&?= 1 , (from x=a
to x=z b), and $ x can have only positive values, each of
n

the integrals is less than 1 whence we may assume

;

zx.dx = B x = Rn sin rn
fan x cos cos rn ;
f<p n
sin zx.dx ;

En being a positive quantity, and r n an angle having al

ways a real value. This gives

pz _ "
_
xdx=iK n (co$rn---^~\ sinrn)=R ne^ V- 1

J e
*J-\$ n 5

whence substituting successively for TZ the numbers

3...^, and for the sake of brevity making
1, 2,

r^r, ............ +r w
we get XsYe*^"" 1
; and the expression for Q becomes
0= - >+C
w y/ co

117. The integral in this last expression is equivalent to

two others, namely

/Ycos(y-^) sin 8 z .
+ */ 1 /Y sin (y-tyz) sin 8 2; . .

Now, on attending to the nature of the quantities repre

sented by Y and y, it will be manifest that according as z is

r
positive or negative, n ,
and consequently y is
positive or

negative, while Y is positive in all cases, since Rn is always

 AND LIMITS OF PROBABLE ERROR. 17 1

positive. Hence cos (y ^z) is always positive, and the ele

ments of the first of the above integrals having thus the same
value and the same sign for the same value of 2, whether z

be positive or negative, the value of the integral from oo

to -J- is the double of its value from to oo On the .

other hand, since y and z have both the same sign,

sin (y tyz) is positive or negative according as z is positive
or negative, and the elements of the integral into which it

enters being equal for z and -J-z, but having contrary

signs, destroy each other, and the integral from #= oo to

to #= -j-
oo vanishes. The expression for Q, is therefore
transformed into
00
fa
Y cos (y TJSZ) sin 8z. .
z
118. The formula newfound cannot in general be integrat
ed by any of the known methods, but in the present case the

quantities denoted by Y and y are such that an approxi

mate value of Q may be obtained, which will always be
more nearly equal to the true value as h, the number of
observations, is increased. On adding the squares of the
two quantities represented by Rn cos rn and Rn sin rn we
get
R B 2 =(/0 nar cos zx.dx)
2
+ (fcf> n
x sin zx.dxf.
If 2=0, this becomes R n =/</> n#er, whence by (1 16), R n = 1 .

When z has a real value, then it

may be shewn that RM
is less than 1 x f be any value of A different from
; for let

x, then as x can only vary from as a to b, we have obviously,

f f
f$ nx and/ftX B>mzx .dx =sjfy nx sin zx.dx,
cos zx .dx =fip n x cos zx.dx,

and the above equation may be put under this form,

R & n =f<p nx cos zx.dx-f$ nx cos zx dx +f$ nx siii zx.dx-fQ nx sin zx .dx,
whence
f f
cos z(xx )dxdx .
 172 RESULTS OF DISCORDANT OBSERVATIONS,

Now, excepting the case in which 2=0, this double

integral is
always than ff$ nx$ nx dxdx or less than
less >

2
(f(f) n xdx) , and consequently R,, is less than f<J> n xdx, that
is, less than unity. Since, then, it has been shewn that
R n is equal to unity when z=Q, and less than unity for all

other values of z, and since Y is a quantity of the order R nA ,

it follows that Y must diminish with great rapidity when z,

or its
equal #-f-e differs sensibly from 0, and even for very small
values of z becomes insensible when h is a large number.

We may therefore assume Y=e~~* , an expression which

is equal to unity when 6=0, and diminishes rapidly as 6 is in

creased, and becomes zero when 6 is infinite.

119- For the sake of abridging let us assume
k n =fx<f> n xdx, k 2
n =fx <}> nxdx, k"=fx*4>jcdx, &c.
(the integrals in respect of x being always from x-zza to

#= b). From known formulae we have

cos zx=
t t a TC

substituting these series for cos zx and sin zx in the inte

grals f<f) nx cos zx.dx audf(p n x sin zx.dx, and also k n k , n> k" n ,

&c., for the values they have now been assumed to repre

sent, then, from (116) we have

R n cos rn = 1 k n + k "n &c.

and be seen presently that all the terms involving

it will

higher powers of z than the cube may be neglected. Add

ing together the squares of these two equations, we get
2
JR n s=l z\k n k\) +z 4/ &c.; whence

f being independent of z. On dividing the second by the

 AND LIMITS OF PROBABLE ERROR. 173

there results tan rn =zk n z f

first, lz k" n +z*kji n &c. ;

whence by reason of r n tan r n = tan 3 r n -f- &c.,

r n =zk n 1*3 (" 3k n k H + 2k n *)\+ &c.

If, therefore, we make

the values of Rn and rn become respectively

R m =l &c.; r n =zk n
2
* cB +*4/ z* ffn + &c.

Now, by hypothesis (116) YsR^R^xRa ...... R ;

therefore log Y=2 log R =2log (1

n z 2 cB + s 4/ &c.)=
2
S {z c tt z*(f ;K) &c.} (by reason of the formula
log R B =R B
2
1
J(K n 1
) + &c.) But we have also assumed
(117) Y=e ; hence log Y= 2
, and consequently
6
2
=2 \z c 2
n z^ (/ icB ) &c.} In like manner, since

y=r^ +r 2 +r 3 ... +rA =Srn , therefore ^=2z^ B--S 5 ^ + fl

&c. Now, the sums S include all values of cn ^ n gnt from , ,

^=1 to w=/i ; let the mean values of those quantities,

therefore, be denoted by c, k, g, that is to say, let

-c n =hc, 2/ n =M, 2^ 7l =%,

hf"=2(f ^C B ), and we have 6 =z /ic
2 2
and make also

z^hf Jf &c. By reverting the series the value of z is found

in terms of 6;

But the second term of

namely *=
this series,
-^
+ ^^
being divided by h*/ft 9
+ &c.

and h being by supposition a large number, is very small in

comparison of the first, and may be neglected as insen
sible. All the succeeding terms of the series are divided

by higher powers of A, and may therefore be rejected a for

tiori. Confining the approximation, therefore, to terms of
the order \--/h, and rejecting all those into which h or

its powers enters as a divisor, we have z=6--i/(hc), and

likewise dz-~-z=dd-^-0.

From (116) we have also y=2rn =zZk n z*2g n + &c.,

 174 RESULTS OF DISCORDANT OBSERVATIONS,

therefore in consequence of the above transformations, y=

zliR z*hg-\- &c. ; and on substituting for z its value just

found in terms of 6, y=k&^(h--c) gb* -i-c\/(hc), and

consequently?/ tyz = (hk ^tf-j-y^Ac)
z
g6 c-r-\/(hc).
In order to deduce from this an expression for cos (ij \/AZ),

let u and v denote any two arcs, then by trigonometry,

cos (u t?)=cos u cos v-{- sin u sin v. Suppose v to be small,
and let its cosine and sine be developed in series and sub
stituted in this equation ; it will become

cos (u y)=cos u --- 1 1 -v

cos u+ &c. -{-v sin u -- fc
-
J
sin u+ &c.

/ 3
whence, making ^)0-i-v (^ c ) v=yd -+-c</(hc),
u=(hk
and_rejecting as before terms of the order 1 -r-h, we have

If we now substitute the values of Y, z, dz, cos (jj ^r)

found in the last three paragraphs in the value of Q (1 17)
we obtain the following expression in which the largest
terms omitted are of the order 1 -r h, and which therefore
is more accurate as h is a higher number, viz.

2 r _* c ,, . 6

sn
A/ (he)
1 20. As no restriction has yet been made with respect to
the value of ^, excepting that it is a mean between /*eand j/e,

and therefore included between ha and hb (115), let us

now assume ^l/=hk. This gives cos (hk ^)=1, and sin
; and the equation becomes
2
= - //"<
o
.

sin
77^7 V(hc)
 AND LIMITS OF PROBABLE ERROR. !

175

which is the probability that the sum of the observed va

lues of A will fall between hk=+=..

121.The last step in this investigation is to reduce the
to a known form, which may^be accom
integral now found

plished as follows Let u be a new variable, then by means

:

l u
of the trigonometrical formula cos u\e ^~^-\-\e~ ^~\

But ^-fwfl^/ZT
= "
( % ; assunie

therefore, v=d ^Hi (whence dv = d6), then

When 0=0, then v =: i^^/ 1, and when 6 is infinite,

v is infinite ; therefore, if the integral in respect of 6 be

taken from 6=0 to 6 = oo, the integral in respect of v must

be taken from v == \u ^HTl to v =: oo. ;

If we now suppose u to be negative, we shall have in like

manner
^ J^de = dv, the
f^ %?-**fe~** limits in

this case being from v = +~2 M y/ 1

to v == Hence

But the sum of the two on the right-hand side of

integrals
this equation, the first being taken from v M
a */- 1 =
to infinity, and the second from v -j-
AM / \
=
to infinity,

isobviously the double of j*e~* dv from v=0 to v== x, or

(96) equal to \^TT ;

and we have therefore

e cos

Let both sides of the equation be multiplied by du, and

integrated from w=0 to =-=- \/(hc)=u ; then observ

ing that/^cos (u6)dd sin (z^)-j-0, we shall have

176 RESULTS OF DISCORDANT OBSERVATIONS,

Comparing this equation with that in (110), we find

u
Q=(l -i- v/7ry*o V-* w2 G?w. Now, let u=2t, and let r be what*
becomes when u=u =8~\/(hc) ; then iw 2 = 2
, du=2dt,
8-r-\/(hc)=2T, or d=2r^/(hc), and we have, finally,

2 /v _>2 2
Q= e d^ or, Q=l
for the probability that s, the sum of the observed values of

A, will be comprised between the limits ^ 8 and ^-}-&,

that is, between Mrp2r-v/(^c); or, that the arithmetical mean
of all the observations, namely s-f-A, will lie between

122. The expression now found for Q is that which in

(96) was denoted by 0, and of which the table gives the

values corresponding to the different values of r. The ge
neral result of the investigation is, therefore, that whatever
be the nature of the function $ nx which represents the law
of the facility of the different values of A, if a large num
ber of observations be made, the sum of the values of A,
divided by the number of observations, approaches continu
ally to a certain special quantity k (which is the true mean
value of A) as the number of observations is increased, and
that by multiplying the number of observations, a probabi
lity may always be obtained, approaching as nearly to cer
tainty as we please, that the difference between the arith
metical mean or average of the observations and the true
mean value of A, will be comprised within limits which may
be made as small as we please.
The analysis employed in the preceding articles, (113
to 121), for the purpose of establishing this very impor
tant result, belongs to Poisson, and is
given in nearly the
same form in the Recherches sur la Probabilite des Juge-
 AXD LIMITS OF PROBABLE ERROR. 177

ments, chap, iv., and in the Additions to the Connaissance

des Terns for 1832. We have preferred it to the method
followed by Laplace in the Theorie Analytique, as being
somewhat simpler and also more general.
1 23. In order that the limits 2r\/(hc) may be real, it is

necessary that the special quantity c be positive, a condition

which has hitherto been assumed. Now, since c=Zcn /<,

it is obvious that c will be positive if c n n k 2 n ) be po =(k

sitive. On writing for k n and k n their values
(119) we have

the limits of the integrals being always from x=a to x=b.

But it is evident that no change will be made in the values

of these definite integrals (the limits continuing the same),

by substituting in them any other of the possible values of

A, as x
f
We have therefore jx ^ x dx ==fx(f) xdx, and since
.
H
f
n
t

in all cases J dx =l, the above equation may be other

(f>*x

wise written

whence 2cn =flfy nx<p nx (x 2

or 2c n =ff<p n xcf) nx (x 2 x x)dxdx f .

Adding together the two last equations, there results

a quantity which is
necessarily positive, and can never be
zero so long as x can have different values.

124. The special quantity k to which the average of the

values of A continually approaches, is connected with the

centre of gravity of the area of a curve by the following rela

tion.Let x and y be the co-ordinates of a curve, of which the

equation is y=<p nx then the element of the area is n xrJx.

; <$>

But (116) <p n xdx is the infinitely small probability that the
value of A in the nth observation will lie between x and
jr
-f-
dx ; therefore the element of the area of the curve
 178 RESULTS OF DISCORDANT OBSERVATIONS,

represents this probability, and the curve itself represents

the law of the probability of the different values of A in re

spect of the nth trial. In like manner, the curve whose co

ordinates are x and ( 1 -r-^)2ip n .r, represents the law of the mean
probability of A in respect of the whole series of observa

tions. be the absciss of the centre of gravity of

Now, if xi

any curve whose co-ordinates are x and y, the well known

formula of mechanics gives x^fyxdx-i-fydx; therefore,

applying this formula to the curve of the mean probability,

and making the whole area (fydx from x=a to #=&)== 1,
the absciss of the centre of gravity Xi=(\-^-h)^fx^ n xdx.
is

But denoted by k (1 19} hence the spe

this is-the quantity ;

cial quantity to which the average of a large number of ob

servations indefinitely approaches is the absciss of the cen
tre of gravity of the area of the curve which represents the
law of the mean chances of A.
125. It has been assumed in the foregoing analysis that
A is
susceptible of an infinite number of values, increasing
continuously from a to b. The results, however, are easily
adapted to those cases in which the number of possible va
lues ofA is finite. Suppose A to be a thing susceptible of
only X different values, represented by a v 2,
as # /?
and let the chances of these values, which may be different

in the different trials, be respectively y lf y 2 y. yx ,

in respect of the wth trial. Now, suppose q> nx to be a dis

continuous function, which vanishes for all values of x, of
which the difference from one or other of the above values
of A exceeds an infinitely small quantity c ; then the whole

integral J$ xdx from

n x=a to x=b, be made up of a
will

series of X partial integrals Jty n xdx taken between the limits

<z<=b:c, the sum of which will be unity, since one or other of

the values of A must necessarily be given by the trial. But
 AND LIMITS OF PROBABLE ERROR. 1 , 9

the integral f<p n xdx between the limits r/,:re is the expres
sion of the chance that the value of A given in the nth trial

will lie between a.rtze ; whence for those \im\ts f<p n i xdx=y i
.

Now the difference x a must be

t infinitely small, since it
cannot exceed e ; we may therefore substitute a for x, f and
a? for x 2 under the sign of integration, when the limits are
r/ t =Z,so that for those limits we ha.vej $q>, xdx=ia t i J<p H Jcdx==

yat t
. On writing for i all the different numbers 1, 2, 3 ...... X,
and observing that the X partial integrals thus formed make
up the whole integral Jx<p n xdx from to x=b, and that xa
sum is k H we have, in respect of the nth trial,
therefore their ,

2
In like manner, for k n -=zfx n xdx (from a to >),
we have
2 2 a 2
A =7ii +yBfl +y 3 j
...... -f /A. ;

so that the two special quantities k and k become

f

the sums 2 extending to all the h values of n, or to all the

trials, the chances denoted by y 15 y ,

e. being supposed
to vary in the different trials.

126. When the chances of the different values of A are

equal and constant, then y<= 1 ^-X, and the above values of h
and k become

so that k is the arithmetical mean of the possible values cf

A, and k the mean of the squares of those values. On this
f

f
hypothesis, therefore, k and k may be computed a priori,
and consequently the limits determined within which there
is a given probability that the average of h observations
2
will fall, thelimits
being /fcnp2r v /(c-r- A,) where c=J( A ).
180 RESULTS OF DISCORDANT OBSERVATIONS,

When the chances of the different values of A are un

equal, but constant in the different trials, then A=# n , and
k -=.k n and we have
i

In this case the special quantity k to which the average of the

observed values continually approaches, is the sum of the

possible values, each multiplied into its

respective proba
bility ; and k is the sum of the products of the squares of
those values into their respective probabilities.

127. Resuming the consideration of the general formula

in (121), we shall now

give an example of its application
when the function which represents the law of facility
of the different values of A is supposed to be known a

priori.
Of all the hypotheses which may be made respecting the
law of facility, the simplest is that which supposes the

chances of all the possible values of the thing observed to

be equal, and to remain constant during the series of trials.
This supposes <>#= <>#=: a constant; whence ftpxdx, be
tween the limits x=a and #=&, becomes (b a) Qx. But
between those limits we have also f$xdx-=. 1 therefore ;

<px= l-t-(b a). From this value of <px it is easy to deduce

the special quantities k and k . On the present hypothesis
kzs.k n and k k n ; therefore, the limits of the integral being

x=a and x=zb, we have A= IxQxdxiz If*xdx =

J b
- a
b2
-
2(b
a2
-

a)
2
a), whence k^=^(b -j- a) . Inlike manner k =fx*qixdx

-
/xdxa=.i(6 -|-a4-a
j
b
2 2
) ; whence c=\(k f
k~)

8- Hence b 121 we have

 AND LIMITS OF PROBABLE ERROR. 181

the probability e that the average value of A given by h

observations, or the sum of the values of A divided by their

number, will lie between

128. This formula may be applied to the following ques

tion. Of the comets which have been observed since the

year 240 of our era, the parabolic elements of 1 38 have

been computed, and the mean inclination of their orbits to the
ecliptic is found to be 48 55 Now, supposing every possi
.

ble inclination of an orbit to be equally probable, let the

probability be demanded that the mean inclination of 138

orbits will not differ from 45 (the mean of the possible in

clinations) more than 5 in excess or defect.

In this case the limits of the possible values of the phe

nomenon are and 90. We have therefore =0, =90,
/j=138, and the above limits of the error of the average,
become 45 rpr x 90-7--v/(6 X 138). In order that
the limits may not exceed 5, we have to determine T from
the equation T x 90 -r-
V(6 X 138) =5, which gives

T=^V23 ;whence T=l-6 very nearly. The tabular value

of corresponding to r=l-6 is 97635, or nearly fj and ;

the odds are therefore 41 to 1 that on the supposition of all

inclinations being equally probable, the mean inclination of

138 comets would fall between 45r^r:o , that is, between

40 and 50. The mean of the inclinations actually com
puted falls within those limits (being 48 45 ) ; there is

therefore a very great probability that whatever may be

the nature of the unknown causes which determine the
positions of the cometary orbits, it is not such as to render
different inclinations unequally probable.

If the question had been to assign the limits within which

it is as probable that the mean of the inclinations will fall
182 RESULTS OF DISCORDANT OBSERVATIONS,

as not, we should have had 0^, and consequently (from

the table) T -476936, and the limits would have been
45:zp90 X-476936-rV(6x 138), which is found on cal
culation to be 45;qpl -5. On the supposition, therefore,
that all inclinations are equally probable, it is one to one
that the mean of 138 inclinations will fall between 43^ and
and 46J, or at least not exceed those limits.

129. On the same hypothesis of an equal probability of

all
possible values, if we suppose the mean value of A to be
0, we have then = b, and $x becomes l-r-2, whence

the limits corresponding to a given value of (127) become

0=+22 T b :
V(6h). Let 0=J, whence T=-476936, and

suppose 7i=600. With these values the limits become

Or+z 0166 nearly ; that is to say, it is an even wager that
the average of 600 observations will not differ from the
true mean value of A more than the sixteen-thousandth part
of a or b, what is the greatest possible difference.
130. As a second hypothesis, suppose the chance of a
given value of A to decrease uniformly as the magnitude in

creases from to zt a ; then <px will be found as follows :

Let <p#:=/3 when =0 we have then by the hypothesis

;

<px:$(n x) :
a, whence <px=i(a #)/3-r-, and conse-

cpientlyftpxdz^px
2
/3# -7-2#, which, from #=0 to x -f-#>

becomes J/3. But f$xdx from x= a to x=-{-a is 1

(errors beyond those

limits being supposed impossible),

therefore from x=Q to x=. + a,f<xdx=^, and consequently

2
l, or/3 1 -J-tf. Hence $x(a #)-7- , from which the
-J/Sa

value of c is easily deduced, that of k being 0, as in the

former case.
131. Although the function Qx which represents the law
of facility of the different values of A is in general unknown,
its form may be assigned if we assume that it is
subject
 AND LIMITS OF PROBABLE ERROR. 183

to certain conditions, which, from the nature of the

thing, must be very nearly, if not absolutely true, in most

practical cases: 1st, That the chance of an error dimi

nishes as the magnitude of the error increases, and for er

rors beyond a certain limit vanishes altogether ; and, 2d,

that positive and negative errors, of equal magnitude, are

equally probable. The last condition is

equivalent to the
assumption that the average of the observed values is the
true mean value. For simplification, we suppose the chance
of an error of a given magnitude to remain constant in all the
trials.

132. Let x, x , of values of A, the sum

x", &c. be a series

of which is s, and the numberand make m=s~-h, then h,

m is the arithmetical mean or average, which by hypothesis

is the true value of the phenomenon A. Let x m = A,
x m=& ,
x" w=A", &c., so that A, A , A", &c. are the
errors of x x
} , the most probable single er
x", &c. Now,
ror is ; and the probability of obtaining an error of a given

magnitude A in any observation is

obviously the same as
that of obtaining a given value of x ; therefore <px=

<p(x m)= <pA ; so that <pA is the probability of a single error

being exactly A. In like manner, the probability of an error

ror rzA is <pA7 ; and if we take P to denote the probability

of a given system of errors, A, A , A", &c., then the errors

being supposed independent of each other, we have (7)

<pA P=pA
<pA", &c. . .

Let this system be assumed to be the most probable result

of the observations, then P is a maximum, and its differen

tial co-efficient zero. Taking the logarithms of both sides

of the equation, differentiating, and making d log. (pArr^
and c?P-f-c?A=0, we obtain
1 84 RESULTS OF DISCORDANT OBSERVATIONS,

an equation which may be otherwise written

This is the conditional equation of the most probable sys

tem of errors. But the hypothesis of the average being
the true value, furnishes this other equation,

0=(ar w) + (a? i) + (ar" &) + ,

&c.

or, which is the same, 0=A-{-A -{-A"-f-, &c. ; and on com

paring with the above conditional equation, it is evi

this

dent that they can only be both true simultaneously on the

A A A
supposition of
<Z>

A
=
A7
= A =, &c. Hence it follows

that p A--A is independent of any particular value of A, or

is equal to a constant, which we shall call K. We have then
A _d.log. 0A
A AdA
=K
2
The integral of this expression is
log. <pA s|KA -{- const.,

which, making the last constant = log. H, and passing to num

A2
bers, gives <pA=He* . It now only remains to deter
mine the two constants H and K. With respect to K, as
we suppose the most probable value of A to be 0, and that
<pA diminishes as A increases, it is obvious that K must be

negative. Assume K = y, and the formula becomes

^A=He yA*. p or t ]ie determination of H we have the

equation y*<pAe?A= 1, the limits of the integral being a f and

-fa , where a =^(b a), a and b being the limiting values

of x. But it is to be observed, that as all values of A ex
ceeding the limitszi a are supposed to be impossible, or
be so improbable that it is unnecessary to take
at least to

account of them, the value of the integral /<pAc/A from

A= a to A=-f-fl will not be altered by extending the
limits from infinity to -f- infinity. We have therefore
 AND LIMITS OF PROBABLE ERROR. 185

l. Let A=-rVy, then d=dt+*/7 , and

pA=He? ? A2 =He-* 2 , on substituting which in the last

equation, and observing that from t = oo to = +00

we have/e- ^rr-v/Tr (96), we find (H-sVy) -/"=!, and
Al1
H=V(y-s-ir). Whence, finally, pA=V(y-Mr>T-* .

1 33. The general properties of the function now found

may be illustrated by means of a curve line. Let aNb be

a curve of which <pA is the ordinate corresponding to the

absciss A. Let AB be its axis, and MN its greatest ordi
nate. Suppose the origin to be placed at M, draw PQ, an
ordinate at any point P, andpq indefinitely near to PQ, and
make MB^=a , MA=_a , MP=A, and PQ=?A ; then
as was shewn element of the area, repre
in (124), PQ<7/>, the

sents the chance of an error lying between A and A-j-e?A, that

is,ofan error greater than MP

but less than M;;. Now, if <p A
y ^ tne function
2

VCv-r" 71
")
6 will not be changed by chang

ing A into A therefore <pA =#(; A), and the curve is

symmetrical on both sides of MN, as it

obviously ought to be
according to the hypothesis making MP =MP, then
; for on

positive and negative errors of equal magnitude being equally

2
probable, we must have P Cy^PQ. Again, since e *^
diminishes rapidly as A increases, the curve at a short dis
tance from MN must approach
very near to its axis AB ;
but as the function only vanishes when A is infinite, the curve
will not meet the axis at any finite distance from MN. This
curve, therefore, can only represent approximately the law of
186 RESULTS OF DISCORDANT OBSERVATIONS,

facility, inasmuch as it is
supposed that errors beyond a cer
tain limit are impossible ;but on account of the rapid dimi
nution of the ordinate at a short distance from MN, the
chance of an error exceeding a small value of A, as MB,
becomes insensible. Hence the limits of the integrals in

respect of A may be extended without sensibly altering their

values from Azra to A=r=op .

134. It is now necessary to find the special quantities k,

k andc.
, Substituting A
and observing that as the for x,

law ofthe chances is here supposed to remain constant, we have

k=k n k =k n, , the formulae in (119) become k=f<p Ac/A,
k 2
Hence on making $=-*/(y-+--n)e * A2 , we
=fA <pd.
have

When A becomes infinite, this becomes 0, therefore from

A oo to A= -J- GO, k=0. This is an obvious conse

quence of the symmetry of the curve, for the centre of gra

vity is necessarily in the straight line MN.

With respect to k f we may proceed thus. We have
A2 fZA.
7/rr/A (/)A^A=:-v/(y-r-7r)/ A
2 2
e--y But from the

principles of the differential calculus,

d Ae~ y^=e i
Y An~dA 2
e * A VA,
.
1y
therefore, integrating and transposing,

C
J
Now, from A = ootoA= + 00, the term of this equa
tion which is not under the sign of integration vanishes, and
1
ey^-d\=\/(TT-^-y) (from (96), on substituting t for
2 * A V.A =
7A
2
), therefore/A *?
(1 -f-
^y)^/(^ -f- 7); and
consequently k =l 1y.
 AND LIMITS OF PROBABLE ERROR. 187

2
In (119) we assumed c=^(k k ) ; therefore in the pre
sent case c=kwhence c=l-f-4y, or y=l-r-4c.
9

135. The expressions which have now been found for

the function which represents the probability of an error,

and the limits corresponding to an assigned degree of pro

bability, are given in terms of the indeterminate constant y

(or c), which depends on the nature of the observation, and

therefore, where instruments are requisite, on the goodness

of the instrument and the skill of the observer. This con
stant is called by Laplace the modulus of the law of facility.
cannot, in general, be assigned a priori ; but if we assume
It

that positive and negative departures from the mean are

alike probable, which is the most plausible hypothesis the
nature of the thing admits of, an approximation to its value,
in respect of observations of a given kind, maybe deduced
with great probability from the results of a large series of
observations of the same kind already made. We now pro
ceed to give the analysis by which this is
accomplished, fol

lowing the method ofPoisson. The approximation is carried

only to quantities of the order l-^-y/A; terms having A for

a divisor are neglected on account of their smallness, h

being supposed a large number.

136. In the expression for Q in (119)> suppose ^-=5,
and consequently ^ 5=0,5=25, and write
\p--f-
also z for
; the equation then becomes

- r
2 f 50
Q /
e
e~ cos (hkz bz) sin bz .

sin (hkzSz) sin

and Q is the probability that s, the sum of the values of A
given by all the observations, will lie between and 23.

If therefore, we suppose 5 to be variable, the differential

of
188 RESULTS OF DISCORDANT OBSERVATIONS,

this expression taken with respect to S, will express the

infinitely small chance of the sum of the values being 25

exactly. Differentiating, and observing that if u and v
denote any two arcs, the trigonometrical formulae give

sin (u v) sin v-\- cos (u v) cos v=* cos (2v u),

cos (u v) sin v -f-
sin (u v) cos v= sin (2v u),
we shall find
2dd r* _ 2 zdB

Let t be a variable quantity, and assume 28=hk+2t^(hc),

whence <2&=<#V(Ae), and let the corresponding value of

d& be denoted by qdt, we shall have, on substituting

ao

these values, and replacing z by B-^-V(Jic), qdt =

27fe
x
/ e

The two
_/>*
cos (2t6)d6
TrC*/(tlC)J
/*
/o

integrals in this equation are found from the

e
__ S 2
sin

formula in (121). Writing 2t for w, that formula gives

? cos

and if this last equation be differentiated in respect of /,

three times in succession, the result will be

x

whence, if we make V= TT-\ (%t

^
2t z \ we shall have
2c^/(hc)
qdt= (1 -=-v/7r) (lV)<r*dt,
where V
a quantity containing only uneven powers of (,
is

and of the order l-^-yV*, so that when multiplied by another

of the same order, the product will be of the order 1-f-A,
 AND LIMITS OF PROBABLE ERROR. 189

and will therefore be rejected in the present approxima

tion. This value of qdt is the probability that s will be pre

cisely 28 or hk +
2t^/(hc) or it is the infinitely small pro
t

bability of the equation

137. In order to apply this result to the determination

of the probable limits in terms of observations actually

made, it is necessary to remark that the analysis by means

of which it has been obtainedgrounded on the very
is

general supposition that the thing to be measured may be

any function whatever of the quantity observed for the in ;

finitely small chance of a particular value of the function

is evidently the same as that of the corresponding value of
the quantity, and is
consequently <$> n
xdx. Let X therefore
be a function of x, and let K, C, T, be what k, c, t become
when X is substituted for x, the above equation then be
comes

the symbol 2 including all the h values of X ; and the pro

bability of this equation is an expression of the same form
as that which is
represented by qdt.
made with respect
138. Hitherto no restriction has been
to <ja ; we now
introduce the hypothesis that positive and

negative departures from the mean of equal magnitude are

equally probable, and consequently that the curVe repre
senting the law of facility is symmetrical, but shall
sup
pose the chances of a particular value, or a particular error,
to vary in the different trials. Let the origin be transferred
to the centre of gravity, the absciss of which =.k, and let x k
A, x k= A c. We have then by (132) <#=$ A,
fx(f>xdx=f&. A d A fx $xdx=f&
2
</)
2
A d A the integra-
, </> ,

tion in respect of A being from oo to -f- oo. The special

100 RESULTS OF DISCORDANT OBSERVATIONS,

quantities k and k then become 7e=(l-j-/i)syA0 n Ac? A

= 0, A =(l-i-/Os/A 2 <
n A</A, whence

The object is now to eliminate <

n A, and determine c in

terms of the observations.

139. Let A M =# be the observed value of A in the wth ob
servation, then \ n =A is the true error of the observa
tion. Let the function denoted by X in (137) be (X n k)
2

=A 2
,
and the corresponding value of K (since in this case,
2
K=(l-5-A) sfXfaxdx) becomes K=(l -f-A) 2/A <
n A6?A.

Comparing this with the value of c found above, we have

K=2c; therefore on substituting these values of X and
K in the equation (137), and assuming t and cf to be the
2
values ofT and C when X=(A n k) , we get 2 (A n &)
2

=2/ic-f 2*V(^0 whence

c=(l-h2A)2(X B )
2
* U (1)
(U being a quantity of the order 1 -f-\/^) 5 an d the probabi

lity of this equation is

where V r
is a function containing only uneven powers of
t 9
and of the order 1-t-Vh.
In the equation (137) suppose X=#=A n , and let /"and
c" be the corresponding values of T and C, then since on
this supposition K=k, the equation becomes 2X n =
/"U", (2)

being of the order l-t-^7*); and the probability

of this
(U"
is
equation

where V r/
, like V 7
and V, contains only uneven powers of
i" and is of the order l-s-\/h.
140. The two equations (1) and (2) may be regarded
 AND LIMITS OF PROBABLE ERROR. 191

as two distinct events, having the respective probabilities

now assigned to them, and therefore the probability of their

being true simultaneously is the product of their respective

probabilities, and is accordingly (neglecting the product

V V" which is a quantity divided by A),

q q"dt dt"=(\+Tr)(\ V">-<*<r-<" V 2

^W.
Let the value of k given by equation (2) be substituted
in (1), and the expression now given will accordingly be

the probability of the resulting equation, namely,

c= U ) 2 -<"U".
i 2(X "~T .+<
Let wz=(l-f-A)2X n then , m is the average or arithmetical
mean of the observed values, and X n m the reputed error
of the observation. The last equation will then become c=:

(l^-27i) 2 (X fJ m+t U y *"U"; or, rejecting <VU )

a

which is of the order 1 -4-^,

2
c=(l -s-2A):B {(A n m) + 2(A n mXU } t"\J".

For the sake of abridging let us also assume

/ *=(1^)Z(X B m)\ =(! -4-A)i(A n _mX,

so that /i
is the mean of the squares of the errors, or mean
square of the errors, and the equation becomes
c-\^v\j t"U", (3)
f f
the probability of which is
q q dt dt".
141. Now, by (121), we have the probability that s-^-h }

or SA u -i-A z=.m the arithmetical mean of the observed va

lues of A, will fall within the limits k-f=2rV(c^-h). Sub
stituting in those limits the above value of c, and observing

(^+*U --*"U )^= VCW + N^U *"U")+ &c.,

/ // 7
that
and that U and U being of the order \~ ^/h when di
7/
t

vided again by Jh are to be rejected, the limits become

^2rV(i/^), or fc+rrV^-f-A),
and the probability of these being the true limits is mul-
192 RESULTS OF DISCORDANT OBSERVATIONS,

tiplied into the probability of the equation el/A-fj/U

*"U" ; and is therefore (140)

(1 +7r)(l V \")<r**<r**dedt .

142. The expression now obtained is the infinitely small

probability of the limits Azfzr x/(2/i-4-^) of the average m,

in respect of the particular value of s, for which we have de
duced the equation (3). But for every value of s between
the limits and 2S, there will be an equation correspond

ing to (3) ; have the whole proba

therefore, in order to

bility of those limits, the integral of the expression must

be found for all values of t and t". From the nature of
the expressions e~~*"* and e~ trri as well as the consideration

that errors beyond a certain magnitude, though possible,

are wholly improbable, it is evident that the integration may
be extended without sensible error from oo to -J- oo ; and
since the functions V 7
and V" contain only uneven powers
of t and t", the terms into which they enter, disappear in
the integrations between those limits. (SeeLacroix, Calcul
f
Diff. et Integral, torn. iii.
p. 506). Now, from l =. oo

to t = + oo we have (96)/e~ W= V * and/e~

;
r W= V * ;

therefore

The result of the preceding analysis is therefore that on

the hypothesis of positive and negative errors of equal mag
nitude being equally probable, and on rejecting terms di
vided by h (the number of the observations may be always
so great as to render such terms insensible), we may sub
stitute \\i for c in the limits of the error to be apprehended,
without sensibly altering the probability, and consequently

there is the probability 0= Je ^dt that the true

 AND LIMITS OF PROBABLE ERROR. 193

mean value k of the phenomenon A will lie between the limits

.^- or

which contain only quantities given by observation.

On this hypothesis we have also (138) c=(l_ 2h)
2
2/A n Ae?A, or, supposing the law of facility to remain con
<

2 2
stant during the trials, c=4/A
$Ae?A, therefore /*==/A
$Ac?A; that is to say, the mean of the squares of the actual
errors may be taken for the sum of the products of the

squares of the possible errors multiplied by their respective

probabilities.
It is important to remark that as the obser
vations become more numerous, the quantity p, the mean
of the squares of the errors, converges more and more to a
constant quantity, and finally becomes independent of the
number of observations.
1 43. The limits now found may be otherwise expressed.

By hypothesis, m=i (l-r-A)SA =thera

arithmetical mean of
the observed values, and p.=( 1
-4-^)2(X n m)
2
= the mean
of the squares of the reputed errors. m) Now (X n
2

2 2 2
XB 2X,,m-j-w and (l-^h)^2\ nm=2m(l-^-k)^\ n =:2m ;
,

therefore /m=(l-7-A)SX n 2
w that is to say, the mean of
2
,

the squares of the observations minus the square of the

mean. Hence the limits, corresponding to a given proba

bility 0, of the difference between the average of all the

observations and the true value, are expressed by either of

these formulae

rzr^/ {(2 X mean square of errors-4-Aj,

2
dtrViSx mean square of obs. (mean of obs.) }-f-V^
h being the number of observations, and the relation be
tween and T being given by the table. Generally speak
ing, the first of these formulas is the most convenient for
calculation.
K
194 RESULTS OF DISCORDANT OBSERVATIONS,

144. Let / be the limit of the error to be feared in tak

ing the average of the observations as the true result, then

l=Tj(2n -t-h), and r=/ x/(A-j-2 /i.) Now when T is con
stant, that is, for a given probability 0, the determination
willbe more exact in proportion as / is a smaller number,
and the precision will therefore be proportional to *J (h-^-2p.).
Hence ^/(h-i-2p) is called by Gauss the measure of the
precision of the determination. Suppose two series of ob
servations to have been made for the determination of an ele

ment, the comparative accuracy of the results will depend on

two things, the number of observations in each series, and the
amount of the squares of the errors in each. If the num
ber of observationsis the same in both series, the
precision
of each result will be inversely as the square root of the
sum of the squares of the errors, and the presumption of
accuracy is in favour of that result with respect to which
the sum of the squares of the errors is less than in the other.
On the other hand, if the mean square of the errors is the
same in both then the observations are alike good
series,

in both, and their relative values of the two results are di

rectly as the square roots of the number of observations in

each series. Hence, in order that one determination may be

twice as good as another, it must be founded on four times
the number of equally good observations. These considera
tions are very important, in comparing tables of mean va
lues of whatever kind, for example, of the probabilities of

life at the different ages, and in estimating risks which de

pend upon them.

145. Astronomers employ the terms, weight, probable

error, and mean error, of a result, to denote certain func

tions of f*,
the mean square of the errors. The square of
the quantity which measures the precision of the result, is
 AND LIMITS OF PROBABLE ERROR. 195

called thetveight of the determination. Denoting the weight

by w, we have therefore

or the weight is
equal to the square of the number of ob
servations divided by twice the sum of the squares of the
errors. Substituting this in the expression of the limits,
we have l^r^-Jw, and T=l*/w ; that is to say, for a given

probability 0, the limits of the error to be apprehended in

taking the average as the true result are reciprocally pro

portional to the square root of the weight. When obser
vations of different kinds, or results deduced from observa

tion, are compared with each other, their relative weights

(supposing the number of observations the same) are inversely

as p, and are expressed numerically by taking the weight
of a certain series of observations as the unit of weight.
146. %
The probable error of the determination is that

which corresponds to the probability 5.

For 0= J we
have r=-476936; whence rV2=-674489, and the formula
mdb3r x/(2/i-f-/) becomes mzi 674489*J(n-i-h) ; whence

probable error =-674489 V (/*-*- h }-

147. The mean error of the result of a large number
of observationsmay be deduced from the general formula
in (136) as follows. That formula gives qdt(\-^- JTT)

(1 V^e-^dt for the probability that the sum of the ob

served values will be 2S =M+
2t*/(hc) exactly. Di

viding the sum by h, qdt is also the probability that the

average value given by all the observations will be exactly

k+2t J(c-t-h ).
/> Now, on the hypothesis that positive and

negative departures from the mean are equally probable,

and supposing the origin of the co-ordinates to be trans

ferred to the centre of gravity of the curve of mean proba-
196 RESULTS OF DISCORDANT OBSERVATIONS,

2
bility, we have =0, and qdt=(l-s-Jvr)(l V)*r-< <& is the

infinitely small chance of the average error being 2 A/(c -4- ft)

exactly. Multiplying therefore this error into the chance

of its taking place, and integrating the product from t=0
to t=. oo, we shall have the mean error, or mean risk

of all the possible average errors affected with the positive

sign. Now, observing that V represents a quantity divid

ed by fjh, and therefore when multiplied by 2 A/(c-j-A)
becomes of the order 1 -f-^, and may consequently be re
jected, the product of the average error 2.t>J(c-+-h) into its

probability is 2*J(c-i-7rh) Xte t

*dt; and since fte-^dtzz
^fd.e-
t2
^e-**, which from tfrrO to t= oo becomes
simply^,
the integral of the above product from t=0 to =oo is

^(c-s-nh). Substituting for cits value

(142) =J/i, this re-
sultbecomes ^/(p^-2?rh) whence on computing ^/(l -r-27r)
;

we obtain

mean error of series =.398942^(^1-:- fi).

This is the mean error or mean risk in respect of posi

tive errors alone, or on the supposition that negative errors

are not taken into account. But as positive and negative
errors are equally likely, the mean error in respect of nega
tive errors is the same quantity, whence the mean error in

respect of errors of both kinds is .797884 J^^fi). This

is usually called the average error. The mean error differs
from the probable error in this respect, that it
depends on the
magnitude of individual errors, as well as on the proportion
in which errors of different magnitudes occur. The proba
ble error is
independent of the magnitude.
148. When the quantity p (the mean square of the er

rors) has been found from a series of observations, the

precision, weight, probable error, and mean error, of a com-
 AND LIMITS OF PROBABLE ERROR. 197

ing observation of the same kind are found by supposing

k~\ in the above expressions, and are respectively

. .
precision :

weight . .
nl-r-2/z
probable error . =.674489 V/* 1 C
mean error . =.39 8942 Jp.

149- The preceding formula? give the limits of the error

to be feared in determining the value of a quantity from a
series of observations, when the thing to be determined is

that onwhich the observations are immediately made. We

have now to apply the formulae to the cases in which the

quantity sought is not observed itself, but is a function of

several others, which are separately determined by obser
vation. The following problem is important :

Let u be a given function of a number of unknown quan-

titities, x, x 9 x", &c. ; it is required to assign the limits of
the probable error in the determination of U, and the

weight of the result, when values of or, # , x", found from

observations independent of each other, and respectively af
fected with the probable errors Jp., gvV> */ /-""? & c< ($ ==
.674489) are adopted instead of the true but unknown
values of those quantities.
Let u=f(x, x x", &c.) be the given function, X, X X",
, ,

&c. observed values of #, x x", &c. and make X x=e, ,

X x =e ,
X" x"=e", &c. so that e, e , e", &c. are the

errors of observation, supposed to be so small that their

du du du
squares maybe rejected. Make =a, -7-,
= ,
7-^7
= ,

&c., then , , a", are given quantities ; and on substitut

ing x-\*e, x -\-e

l
, x" J^e" for a?, a? , x", respectively, in the

equation u=f(x, x , x"> &c.), and supposing u to become

198 RESULTS OF DISCORDANT OBSERVATIONS,

when the substitutions are made, so that E is the

corresponding error of u, we have, on expanding u by

Taylor s theorem,
E=ae+aV-f a e"-f &c.

in respect of a single observation of each of the quantities.

Taking the square of both sides of the equation, we have

Now since positive and negative errors are supposed equal

ly probable, the sums of the products ee , ee", e , e", &c. or

their mean values, become each = ; therefore

SE 2 =a 2 Ee 2 + a /2 2e/2 H-a :Se

/2
+ &c.

Taking the mean value of each of these sums, and observing

thatp.
the mean value of Se 2 is independent of the num
ber of observations (142), and assuming M to be the mean
value of sE we
2
, get

M=aV+a Y-i-"V + &c.

This equation contains the solution of the problem, for

all the functions of the error are given in terms of M. The
probable error is -674489 V M -

150. Let W be the weight of the determination, and

w, itf, w", &c. the weights corresponding to p., p. , // , &c.
then by the definition of weight, w is reciprocally propor
tioned to p., and to ; W M
and we have by substitution,

If the weights are supposed all equal, this becomes

W-~ a 2 f2
+a + a" 2
+8tc.
Suppose the errors e, ef e", &c. to be respectively multi

plied by numbers proportional to the square roots of the

weights, (which is
equivalent to supposing all the observa-
 AND LIMITS OF PROBABLE ERROR. 199

tions to have the same degree of precision measured by

*J(nio)),
then the value of M becomes

But w being reciprocally as /a,

we have pw=p u/=p w", &c.

=1, therefore
*
w= .. *2 I s* " \ rt
 200 OF THE METHOD

SECTION X.

OF THE METHOD OF LEAST SQUARES.

151. In the determination of astronomical and physical

elements from the data of observation, the thing which is

actually observed is for the most x part not the element

which is sought to be determined, but a known function of
that element. Thus, if V
be a given function of X deter
mined by the equation V=F(X), the quantity observed
may be a value of V, whilst the element sought to be de
termined is X. If the observation could give the value of
V with absolute accuracy, then X would also be absolutely
known ; but as all observations are affected with certain
errors of greater or less amount, owing to the imperfections

of instruments or of sense, or the ever varying circum

stances under which they are made, an exact value of X
cannot be found from any single observation ; and in order
to obtain the utmost precision, it is necessary to employ a

great number of observations, repeated under every variety

of circumstance by which the result can be supposed to be
affected.

152. The observed quantity V, instead of being a func

tion of a single element X, may be a function of several
elements X, Y, Z, &c. for
example,
; V
may be the posi
tion of a planet, in which case it is a function of the six

elements of the orbit, for the determination of which the

 OF LEAST SQUARES. 201

observation is made. Each observation gives rise to an equa

tion of this form, V=F(X, Y, Z, &c.) ; therefore when the
number of equations is just equal to the number of un
known quantities, the problem is determinate and suppos ;

ing F be an algebraic function, the values of X, Y, Z,

to

&c. may be found by the ordinary methods of elimination.

If the number of equations is less than the number of un

known quantities, the problem is indeterminate ; but if

greater, it
may be said to be more than determinate, inas

much as the equations may be combined in an infinite

number of ways, each combination giving a diffe

distinct

rent value of the elements. becomes a ques

It therefore

tion of the utmost importance to the perfection of the

sciences of observation, to assign the particular combination

which gives the most advantageous results, or values of X,

Y, Z, &c. affected with the smallest probable errors.

153. As approximate values of the elements are in all

cases either already known, or can be easily found, the cb

ject of accumulating observations is the correction of the

approximate values. Let V be the true value of the thing

observed, V an approximate value, however found, X the

true value of the element sought, X an approximate value,
corresponding to V , so that we have the two equations

V=F(X), V =F(X ) ; also, let the observed value of V

in any observation be L, and make
v=VL, l=V L,
then v is the true but unknown error of the observation,

and reputed error, that is to say, the difference between

/ its

the computed value of the function and the result of the

observation. Now if we assume x to represent the true

correction of the approximate element, so that X X^^

then, on substituting X +# for X in the function F, we
202 OF THE METHOD

get V=F(X -{-#) ; whence, expanding the function by

Taylor theorem, and rejecting terms multiplied byx 2 and
s

higher powers of x, because x is a very small quantity

T.TH.g.
Let us now denote the differential coefficient, which is a
known quantity, by a ; then, observing that V V =v I, the

equation becomes #=/-{-#; that is to say, the true error

of the observation is a linear function of the correction of

the element.
154. In like manner, when there are several elements,
dV dV dV ~
-^ =
,

X, Y, Z, &c.; on making c
?
^y ~d%
>

&c. a single observation furnishes the equation

v=l^.ax-\-bi/--cz+ &c.,
and a series of observations, whose errors are respectively
v, v &c. gives a system of linear equations equal in
j v",

number to the number of observations ; namely,

v=l -f-## by c2-f- &c. + +
&c. (1)
+ &c.
&c.
and the object is to give such values to #, y> z, &c. that the
f
errors v, v , v", &c. in respect of the whole of the observa

tions, be the least possible. The equations being

shall

supposed independent of each other, if their number is just

equal to that of the unknown quantities, the errors v, v v", ,

&c. can be made all zero ; but if, as is usually the case,
there are more equations than unknown quantities, it is
impossible by any means whatever to annihilate the whole
of them, and therefore all that can be accomplished is to

find the system of values of x, y-> z, &c. which most nearly,

 OF LEAST SQUARES. 203

and with the greatest probability, satisfies the whole of the

equations. If the observations are not all equally good, the

equations are supposed to be each multiplied by a number
proportional to the square root of the presumed weight of
the observation on which it
depends, in order that they may
all have the same degree of precision.
155. As the question is to find the most probable values
of x, y, z, thing necessary is to express each
&c. the first

of these elements in terms of the observations. Suppose

k y K, k", &c. to be a system of indeterminate quantities,

independent of x y, z9 &c. and let the first of the above

conditional equations be multiplied by , the second by k ,

the third by k", and so on ; then adding the products, if

k, k , k", &c. be determined so as to make the coefficient

of a? equal to unit, and those ofy, z, &c. each equal to j that

is to say, so as to satisfy the equations

ka + k a + k"a"+ &c. =1
kb+ k b +k"b"
f
-f-
&c. =0 (2)
kc +k + c k"c" + &c =0
&c.
we shall then have x=K + kv+Ki/+k"i/ + &c. where
K is a quantity independent of v, v , v", &c. Hence x is

=Aw-J-^V+A V 4&c. and

/
found =K, with an error ; the

weight of the determination, by the formula in (150), is

The wei &ht of the Determination is

&c
2 2 f2
consequently greater in proportion as k -\-k -\-k -{- &c.

is smaller ; and hence of all the possible systems of inde

terminate coefficients, k , k", &c. which satisfy the equa

,

tions (1), the system which gives the most probable value
of x> or the most advantageous result, is that for which
. is an absolute minimum.
204 OF THE METHOD

156- We have now to find, in terms of known quanti

ties, values of the indeterminate coefficients A, k f , k", &c.
which satisfy the condition of the minimum. For the sake
of abridging, let us denote the aggregate of the products
aa +a +a a a"+ &c, by S(aa), that of ab+a b +a"b u -f
&c. by S(ab), and so on, and also assume

4. &c. V
v"+ &c. (3)
=ct;-fcV -fcV -f&c.
On substituting in these equations the values of v, vf t v",

&c. given by the equations (1), there results

+ ;2:S(ac) + &c.
<2&)

xS(ab) +y$(bb) + z$(bc) + &c. (4)

=S(c/) + x$(ac) +y$(bc)+zS(ccj+ &c.
a system of equations equal in number to the number of
elements x, y, z, &c. and from which, consequently, those
elements would be determined absolutely if the observa
tions were perfectly exact, that is, if the errrors v, i/, v",
&c. were individually zero, and consequently , 77, , &c.,
were each zero. On eliminating y, z, &c. from the last sys

tem, the value of x is given in terms of , rj, , and known

quantities by a linear equation of the following form :

ar=A+ /J+^ + Af+i

&c. (5)
where / g,&c. are co-efficients independent of x, y>
h, z,

&c. and also of ^, 77, &c.

If we now substitute in equation (5) the values , 77, f,

Sec. given by equations (3), and also assume

a=:fa+gb+hc+ &c.
==/ H-^ +^-t- &c. (6)
a
/
=/ a -f-^ //
/ /

+ V + &C.,
/i
/ /

we shall have by addition

=A-f-av4-aV+aV
/
+ &c. ;

whence it
appears that a, a , a", &c. are a system of multi-
 OF LEAST SQUARES. 205

pliers by which y> z, &c. are eliminated from equations

(1) ; they must therefore satisfy the equations (2), whence
atf-faV-f-a V-f&C. =1
&C. =0 (7)
&C. =0.
Subtracting these from the equations (2) we obtain

&c .

&C.

&C.,
on multiplying which respectively by f, g, k, and adding
the products, we get by reason of the equations (6),
0=( k a)a + (& a X + (k"a")a + &C. f

This equation may be put under the form k

2
-f-
k 2
-f-
k" 2 -f~ & c *

from which it is evident that k 2 + k 2 -\-k" 2

-\- &c. will be a
minimum when A=r a, k =a ,
f
k =a", &c. Hence it- fol

lows that the most probable value of x which can be de

duced from the equations (1), is x=A ; and by (150) the
weight of the determination is 1 -4- (aa -f aV + aV -f &c.)=

This quantity S (aa) is equal tof the co-efficient of | in

the equation (5); for on multiplying the first of equations

(7), by^ the second by g, and the third by h, and adding

the products, we obtain by reason of equations (6),

aa-f aV + a"a"+ &C. f.

157. The method explained in the
last paragraphs two
of determining the most advantageous combination of a

system of linear equations, of the form of those in (154), is

given by Gauss in his Theoria Combinationis Observationum
erroribus minimis obnoxice, (Gottingen, 1823). The prac
tical rule to which it leads is as follows :
Having given a
near value V of a function of several elements, X, Y, Z,
206 OF THE METHOD

&c. and also a series L, I/, L", &c. of observed values of

V, make (V L)Vw*=, (VL )Jw =v f

(V L")^*
,

v", &c. and form the equations in (1). From these equa
tions (4) are easily deduced ; and from these, again, by
elimination, are found the values of a?, y, z, &c. the correc
tions of the approximate elements X, Y, Z, &c., in equa
tions of the form (5), which, for the sake of symmetry, may
be thus written :

C.

&C.
&C.
then the most probable values of #,y, 2, &c. are respectively

A, B, C, &c.; the weights of the determinations respectively

- &c. and the probable errors of the se-

r-, 7^-rr, 7 r-, ;

(aa) (ffi) (yy)

veral determinations are p^(oa), p^(/3j3), p*/(yy), &c.,
where p=-476936.
158. The values of a?, y, 2, &c. now deduced are obtained

immediately, by supposing the sum of the squares of the

errors of observation to be a minimum. Thus, forming
the squares of the equations (1), and making Qsrt^-f-t/ -fi

t/ s-j- &c., the differentiation ofii in respect of each of the

variables #, y, z, &c. produces the quantities denoted in

(156), by , T), & &c. that is to say, it

gives

-= 2 - =2,,^ =2
do.
|)
do, dQ,
f, & C.
therefore if Q be a minimum, , ?, ^, become severally
zero, and the equations (4) give by elimination, x=A,

3/=B, z=C, where A, B, and C denote the same quantities

as above. Now from equation (5) the general value of x is

and the most probable value being #=A, it follows that

 OF :LEAST SQUARES. 207

the most probable values of the corrections x, y, z, are found

by making the differential coefficients of Q equal to zero,

2 2 /2
that is, by making v -\-v +t/ -f- &c. an absolute mini
mum. Hence this method of combining equations of con
dition is called the method of least squares ; and it follows
from the preceding analysis, that it gives the most probable
values of the corrections, or the most advantageous results.

159. As an example, let us suppose there is only one un

known element X, of which X is known to be an approxi
mate value, and L, L/, L", &c. are observed values, the
weights of which are respectively proportional to w, #/, w",
&c. and that it is
required to determine the most probable
value of X from the observations, and also the weight of the
determination. Make(X L)-v/w=r#, (X L)^/M?=:/, and
let x be the correction of X, so that X=X x. On sub

stituting this in (X L) </w=.v, we have (X x L)-v/^=v,

or vljw XA/W. Each observation gives a similar equa
tion, and the equations (1) in (154) consequently become

, &c.

therefore, multiplying each by the coefficient of its own x, we

have=S(&?) xS (w), whence consequently x=S(lw)-^-
S(0) |-r-S(w),thatistosay,the most probable value of a? is

&c.
&c.
and the weight of the determination is
proportional to the
/
reciprocal of.w-f-/+/ -f. &c.
Since X L=# we have
I, also

/
&c.
hence this proposition : If a series of values of an element
208 OP THE METHOD

are found from observations which have not all the same

degree of precision, the most probable value of the element

isfound by multiplying each observation by a number pro

portional to its weight, and dividing the sum of the products

by the sum of the weights ; and the comparative weight of

the result is unit divided by the sum of all the weights.
If the weights be all
equal, and the number of the obser
vations be h, then X=(L-f I/-J-L"-}- &c.)-r-A ; that is to

say, the average of a series of equally good observations

gives the most probable value. The average may, there

fore be considered as a particular case of the method of

least squares.

160. To illustrate the method of proceeding when there

are several elements to be corrected from the observations,
we shall take the following numerical example from Gauss

(Theoria Motus). Suppose there are three elements, and

that three observations, of equal weight, have given the

equations x y-}-2z=3, 3x +
2y 5z=5, 4 x -\-y-\-4z~21 ;

and that a fourth observation, of which the relative weight

is one-fourth, or its precision one-half, of that of the others
has given 2#-{- 6^+6^=28. The first step is to reduce
this last equation to the same standard of weight with the
others, for which purpose it must be multiplied by ^ it ;

then becomes x -{- 3y -f- 3z= 1 4. Now, as x, y> and z can

not be determined so as to satisfy four independent equa

tions, we suppose each observation, or equation, to be affect

ed with an error v, and accordingly obtain the following

system of equations, corresponding to equations (1), viz. :

vf = 5 -f 3x -f 2y5
v"=
t/"= .
 OP LEAST SQUARES. 209

from which the most probable values of x, y, and z are to

be deduced. Let each equation be multiplied by the co
efficient of its own ,v, taken with its
proper sign, namely,
the first by 1, the second by 3, the third by 4, and the

fourth by 1 ; the results added together give the value

of , namely, ==
88 -j- 27 x -\- 6y. In like manner, let the
first be multiplied by 1, the second by 2, the third by 1,

and the fourth by 3, the sum of the products will give 77.

Lastly, let the equations be multiplied respectively by the

coefficients of z, and the sum of the products made equal to

f ; we have then the following equations corresponding to

the equations (4)
=88 H- 27x fy ++
17= 70 + 6* + 15y+z

From these we get by elimination

737y =26 1 71 2 + 5477 f

397982=76242 + 12-4i7 + 1473&
whence (157) A, B, C, the most probable values of x, y, z,

are respectively

A = 49154=2.^TAT*
and the
470 B=
19899
relative weights w,
W
2617

w
=3-ii
,
551 r
C=
ieP,
^
76242
39798
are respectively
==1 916

W
=-m9--
19899
=M 6 W,737
=~54
,~
= > 13 6 >-
39798
"147 3=
27 >

whence the probable errors (-476936-=- Jiv) are respectively

096, -129, -092.

The method of least squares, to which modern astrono

my is indebted for much of its precision, was first proposed

by Legendre, in his Nouvelles Methodes pour la Determi
nation les Orbites des Cometes, (Paris, 1806,) merely as
210 OF THE METHOD

means of avoiding inconvenience and uncertainty arising

from the want of a uniform and determinate method of

combining numerous equations of condition, and without

The same method,
reference to the theory of probability.

however, had previously been discovered by Gauss, and a

demonstration of it, deduced from the general theory of
chances, was given by him in his Theoria Motus, (1809 )

It may be shewn in various ways, that this method of com

bination gives values of the unknown quantities affected
with the smallest probable errors ; but it is to be observed,
that all the demonstrations are subordinate to the hypo
and negative errors of equal magni
thesis, that positive

tude are equally probable, or that the average of a large

number of results gives the most probable value, and con
sequently that the function which represents the probabi
lity of an error has the form assigned to it in (132).

The limits of this article will not permit us to enter into

further details respecting the applications of the method of

least squares. On the general theory of the probable
errors of results deduced from observation, and the most
advantageous methods of combining equations of condition,
the reader may consult the Tfieorie Analytique des Proba-
bilites of Laplace ; the Theoria Motus of Gauss ; the The
oria Combinationis Observationum, and the Supplementum
Theories Combinations, &c. (Gottingen, 1828) of the same
author ; the Recherches sur la Probabilite des Jugements,

with the two Memoirs of Poisson in the Connaissance

des Terns for 1827 and 1832 ; and three masterly papers,
by Mr. Ivory, in the Philosophical Magazine for 1825. In
the volumes of the Berliner Astronomisches Jahrbuch for

1833, 1834, and 1835, M. Encke has treated the subject

at great length, and given a number of formulae calculated
 OF LEAST SQUARES. 211

to facilitate the labours of the computer. We may also

refer, in conclusion, to a very recent and remarkable dis

quisition on the theory of probable errors, by the celebrated

astronomer Bessel, forming Nos. 358 and 359 of Schuma
cher s Astronomische Abhandlungen, Altona, October 1838.
 TABLE
OF THE

VALUES OF THE INTEGRAL

2

for intervals each = 01, from r=0 to r=3, with their first
and second differences.
TABLE, CONTINUED. 213
214 TABLE, CONTINUED.
TABLE, CONTINUED. 215
 ERRATA.

Essai sur les Probabilites, read Theorie

Page 40, note, for

Analytique des Probabilites.

43, for adopt, read adapt.

with. In the note, read
49, line 13, read beginning

Traite Elememuire.

89, line 9, for mm, read m+m .

 7 DAY USE
RETURN TO pM WHICH BORROWED
f OY.
,
# 7* M *i ,^cc
: ,
*

This publication is due on the LAST DATE

and HOUR stamped below.

^
.SI

RB 17-50m-7 65 General Library

(F5759slO)4188 University of California
Berkeley
 4--

609735

UNIVERSITY OF CALIFORNIA LIBRARY