% LaTeX source for Todhunter on Bayes

\documentclass{article}

\usepackage{amstex}

\begin{document}

\begin{center}
{\Large CHAPTER XIV.}

\bigskip

{\large\textbf{BAYES.}}
\end{center}

\bigskip

539.  \textsc{The} name of Bayes is associated with one of the most
important parts of our subject, namely, the method of estimating the
probabilities of causes by which an unknown event may have been
produced.  As we shall see, Bayes commenced the investigation, and
Laplace developed it and enunciated the general principle in the form
which it has since retained.

\medskip

540.  We have to notice two memoirs which bear the following titles:

\smallskip

\textit{An Essay towards solving a Problem in the Doctrine of Chances.
By the late Rev.\ Mr.\ Bayes, F.R.S.\ communicated by Mr Price in a
Letter to John Canton, A.M.\,F.R.S.  A Demonstration of the Second Rule
in the Essay towards the Solution of a Problem in the Doctrine of
Chances, published in the Philosophical Transactions, Vol.\
\textsc{liii}.  Communicated by the Rev.\ Mr.\ Richard Price, in a
Letter to Mr.\ John Canton, M.A.\,F.R.S.}

\smallskip

The first of these memoirs occupies pages 370--418 of Vol.\
\textsc{liii}.\ of the \textit{Philosophical Transactions}; it is the
volume for 1763, and the date of publication is 1764.

The second memoir occupies pages 296--325 of Vol.\ \textsc{liv}.\ of the
\textit{Philosophical Transactions;} it is the volume for 1764, and the
date of publication is 1765.

\medskip

541.  Bayes proposes to establish the following theorem: If an event has
happened $p$ times and failed $q$ times, the probability that its
chances at a single trial lie between $a$ and $b$ is
$\frac{\int_a^b x^p(1-x)^q\,dx}{\int_0^1 x^p(1-x)^q\,dx}.$

Bayes does not use this notation; areas of curves, according to the
fashion of his time, occur instead of integrals.  Moreover we shall see
that there is an important condition implied which we have omitted in
the above enunciation, for the sake of brevity: we shall return to this
point in Art.\ 552.

Bayes also gives rules for obtaining approximate values of the areas
which correspond to our integrals.

\medskip

542.  It will be seen from the title of the first memoir that it was
published after the death of Bayes.  The Rev.\ Mr Richard Price is the
well known writer, whose name is famous in connexion with politics,
science and theology.  He begins his letter to Canton thus:

{\small Dear Sir, I now send you an essay which I have found among the
papers of our deceased friend Mr Bayes, and which, in my opinion, has
great merit and deserves to be preserved.}

\medskip

543.  The first memoir contains an introductory letter from Price to
Canton; the essay by Bayes follows, in which he begins with a brief
demonstration of the general laws of the Theory of Probability, and then
establishes his theorem.  The enunciations are given of two rules which
Bayes proposed for finding approximate values of the areas which to him
represented out integrals.  Price himself added \textit{An Appendix
containing an Application of the foregoing Rules to some particular
Cases}.

The second memoir contains Bayes's demonstration of his principal rule
for approximation; and some investigations by Price which also related
to the subject of approximation.

\medskip

544.  Bayes begins, as we have said, with a brief demonstration of the
general laws of the Theory of Probability; this part of his essay is
excessively obscure, and contrasts most unfavourably with the treatment
of the same subject by De Moivre.

Bayes gives the formula by which we must calculate the probability of a
compound event.

\smallskip

Suppose we denote the probability of the compound event by
$\frac{P}{N}$, the probability of the first event by $z$, and the
probability of the second on the supposition of the happening of the
first by $\frac{b}{N}$.  Then our principle gives us
$\frac{P}{N}=z\times\frac{b}{N}$, and therefore $z=\frac{P}{b}$.  This
result Bayes seems to present as something new and remarkable; he
arrives at it by a strange process, and enunciates it as his Proposition
5 in these obscure terms:

\smallskip

If there be two subsequent events, the probability of the 2nd
$\frac{b}{N}$ and the probability of both together $\frac{P}{N}$, and it
being 1st discovered that the 2nd event has happened, from hence I guess
that the 1st event has also happened, the probability I am right is
$\frac{P}{b}$.

\smallskip

Price himself gives a note which shows a clearer appreciation of the

\medskip

545.  We now pass on to the remarkable part of the essay.  Imagine a
rectangular billiard table $ABCD$.  Let a ball be rolled on it at
random, and when the ball comes to rest, let its perpendicular distance
from $AB$ be measured; denote this by $x$.  Let $a$ denote the distance
between $AB$ and $CD$.  Then the probability that the value of $x$ lies
between any two assigned values $b$ and $c$ is $\frac{c-b}{a}$.  This we
should assume as obvious; Bayes, however, demonstrates it very
elaborately.

\medskip

546.  Suppose that a ball is rolled in the manner just explained;
through the point at which it comes to rest let a line $EF$ be drawn
parallel to $AB$, so that the billiard table is divided into the two
portions $AEFB$ and $EDCF$.  A second ball is to be rolled on the table,
required the probability that it will rest within the space $AEFB$.  If
$x$ denote the distance between $AB$ and $EF$ the required probability
is $\frac{x}{a}$: this follows from the preceding Article.

\medskip

547.  Bayes now considers the following compound event:  The first ball
is to be rolled once, and so $EF$ determined; then $p+q$ trials are to
be made in succession with the second ball: required the probability,
before the first ball is rolled, that the distance of $EF$ from $AB$
will lie between $b$ and $c$, and that the second ball will rest $p$
times within the space $AEFB$, and $q$ times without that space.

We should proceed thus in the solution: The chance that $EF$ falls at a
distance $x$ from $AB$ is $\frac{dx}{a}$; the chance that the second
event then happens $p$ times and fails $q$ times is
$\frac{|\underline{p+q}}{|\underline{p}\,|\underline{q}} \left(\frac{x}{a}\right)^p\left(1-\frac{x}{a}\right)^q;$
hence the chance of the occurrence of the two contingencies is
$\frac{dx}{a}\frac{|\underline{p+q}}{|\underline{p}\,|\underline{q}} \left(\frac{x}{a}\right)^p\left(1-\frac{x}{a}\right)^q.$

Therefore the whole required probability is
$\frac{a|\underline{p+q}}{|\underline{p}\,|\underline{q}} \int_b^c\left(\frac{x}{a}\right)^p\left(1-\frac{x}{a}\right)^q dx.$

Bayes's method of solution is of course very different from the above.
With him an area takes the place of the integral, and he establishes the
result by a rigorous demonstrations of the \textit{ex absurdo} kind.

\medskip

548.  As a corollary Bayes gives the following: The probability, before
the first ball is rolled, that $EF$ will lie between $AB$ and $CD$, and
that the second event will happen $p$ times and fail $q$ times, is found
by putting the limits 0 and $a$ instead of $b$ and $c$.  But it is
\textit{certain} that $EF$ will lie between $AB$ and $CD$.  Hence we
have for the probability, before the first ball is thrown, that the
second event will happen $p$ times and fail $q$ times
$\frac{a|\underline{p+q}}{|\underline{p}\,|\underline{q}} \int_0^a\left(\frac{x}{a}\right)^p\left(1-\frac{x}{a}\right)^q dx.$

\medskip

549.  We now arrive at the most important point of the essay.  Suppose
we only know that the second event has happened $p$ times and failed $q$
times, and that we want to infer from this fact the probable position of
the line $EF$ which is to us unknown.  The probability that the distance
of $EF$ from $AB$ lies between $b$ and $c$ is
$\frac{\int_b^c x^p(1-x)^q\,dx}{\int_0^1 x^p(1-x)^q\,dx}.$

This depends on Bayes's Proposition 5, which we have given in our Art.\
544.  For let $a$ denote the required probability; then
$z\times\text{probability of second event}= \text{probability of compound event}.$

The probability of the compound event is given in Art.\ 547, and the
probability of the second event is given in Art.\ 548; hence the value
of $z$ follows.

\medskip

530.  Bayes then proceeds to find the area of a certain curve, or as we
should say to integrate a certain expression.  We have
$\int x^p(1-x)^q dx=\frac{x^{p+1}}{p+1}-\frac{q}{1}\frac{x^{p+2}}{p+2} -\frac{q(q-1)}{1.2}\frac{x^{p+3}}{p+3}-\dots$

This series may be put in another form; let $u$ stand for $1-x$, then
the series is equivalent to
$\frac{x^{p+1}u^q}{p+1}+\frac{q}{(p+1)(p+2)}\frac{x^{p+2}u^{q-1}}{p+3} -\frac{q(q-1)(q-2)}{(p+1)(p+2)(p+3)}\frac{x^{p+3}u^{q-3}}{p+4}+\dots$

This may be verified by putting for $u$ its value and rearranging
according to powers of $x$.  Or if we differentiate the series with
respect to $x$, we shall find that the terms cancel so as to leave us
only $x^pu^q$,

\medskip

551.  The general theory of the estimation of the probabilities of
causes from observed events was first given by Laplace in the
\textit{M\'emoires\dots par divers Savans}, Vol.\ \textsc{vi}.\ 1774.
One of Laplace's results is that if an event has happened $p$ times and
failed $q$ times, the probability that it will happen at the next trial
is
$\frac{\int_0^1 x^{p+1}(1-x)^q\,dx}{\int_0^1 x^p(1-x)^q\,dx}.$

Lubbock and Drinkwater think that Bayes, or perhaps rather Price,
confounded the probability given by Bayes's theorem with the probability
given by the result just taken from Laplace; see \textit{Lubbock and
Drinkwater}, page 48.  But it appears to me that Price understood
correctly what Bayes's theorem really expressed.  Price's first example
is that in which $p=1$, and $q=0$.  Price says that there would be
odds of three to one for somewhat more than an even chance that it
would happen on a second trial.''  His demonstration is then given; it
amounts to this:
$\frac{\int_{\frac{1}{2}}^1 x^p(1-x)^q\,dx}{\int_0^1 x^p(1-x)^q\,dx} = \frac{3}{4},$
where $p=1$ and $q=0$.  Thus there is a probability $\frac{3}{4}$ that
the chance of the event lies between $\frac{1}{2}$ and 1, that is a
probability $\frac{3}{4}$ that the event is more likely to happen than
not.

\medskip

552.  It must be observed with respect to the result in Art.\ 549, that
in Bayes's own problem we \textit{know} that \textit{a priori} any
position of $EF$ between $AB$ and $CD$ is equally likely; or at least we
know what amount of assumption is involved in this supposition.  In the
applications which have been made of Bayes's theorem, and of such
results as that which we have taken from Laplace to Art.\ 551, there has
however often been no adequate ground for such knowledge or assumption.

\medskip

553.  We have already stated that Bayes gave two rules for approximating
to the value of the area which corresponds to the integral.  In the
first memoir, Price suppressed the demonstration to save room; in the
second memoir, Bayes's demonstration of the principal rule is given:
Price himself also continues the subject.  These investigations are very
laborious, especially Price's.

The following are among the most definite results which price gives.
Let $n=p+q$, and suppose that neither $p$ nor $q$ is small; let
$h=\frac{\sqrt{(pq)}}{n\sqrt{(n-1)}}$.  Then if an event has happened
$p$ times and failed $q$ times, the odds are about 1 to 1 that its
chances at a single trial lie between $\frac{p}{n}+\frac{h}{\sqrt{2}}$
and $\frac{p}{n}-\frac{h}{\sqrt{2}}$; the odds are about 2 to 1 that its
chance at a single trial lies between $\frac{p}{n}+h$ and $\frac{p}{n}-h$;
the odds are about 5 to 1 that its chance at a single trial lies between
$\frac{p}{n}+h\sqrt{2}$ and $\frac{p}{n}-h\sqrt{2}$.  These results may
be verified by Laplace's method of approximating to the value of the
definite integrals on which they depend.

\medskip

554.  We may observe that the curve $y=x^p(1-x)^q$ has two points of
inflexion, the ordinates of which are equidistant from the maximum
ordinate; the distance is equal to the quantity $h$ of the preceding
Article.  These points of inflexion are of importance in the methods of
Bayes and Price.

\bigskip

\noindent
[Chapter XIV, Articles 539--554 of I Todhunter, \textit{A History of the
Mathematical Theory of Probability from the time of Pascal to that of
Laplace}, Cambridge: University Press 1865; reprinted New York, NY:
Chelsea 1949.]

\end{document}

%