LII .
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. LIII .
Communicated by the Rev. Mr. Richard Price , in a Letter to Mr. John Canton , M. A. F. R. S. Dear Sir , Nov. 26 , 1764 .
I Send you the following Supplement to the Essay on a Problem in the Doctrine of Chances , hoping that you may not think it improper to be communicated to the Royal Society .
I should not have troubled you again in this way had I not found that some additions to my former papers were necessary in order to explain some passages in them , and particularly what is hinted in the note at the end of the Appendix .
I have first given the deduction of Mr. Bayes 's second rule chiefly in his own words ; and then added , as briefly as possible , the demonstrations of several propositions , which seem to improve considerably the solution of the problem , and to throw light on the nature of the curve by the quadrature of which this solution is obtained .
' ' Perhaps , there is no reason for being very anxious about proceeding to further improvements .
It would , however , be very agreeable to me to see a yet easier and nearer approximation to the value of the two series 's in the first rule : but this I must leave abler persons to seek , chusing now entirely to drop this subject .
Read Dec. 6 , 1764 .
The solution of the problem enquired after in the papers I have sent you has , I think , been hitherto a desideratum in philosophy of some consequence .
To this we are now in a great measure helped by the abilities and skill of our late worthy friend ; and thus are furnished with a necessary guide in determining the nature and proportions of unknown causes from their effects , and an effectual guard against one great danger to which philosophers are subject ; I mean , the danger of founding conclusions on an insufficient induction , and of receiving just conclusions with more assurance than the number of experiments will warrant .
I am , under a sense of the value of your friendship , heartily yours , Richard Price .
O FDC s HlthfA WW QR ART .
I. If the curve ADH be divided into two parts by the ordinate Dh making Ah to Hh as p is to q ; then taking a= p/ n and b= q/ n the ratio of the Area ADh to HO will be ap+1~b q/ p +1~ 1+ q/ p +2~ p/ q +q~q1~ p2/ p +2~p+3~ q2 +q~ q-1 ~ q-2 ~p3/ p +2~p+3 ~p +4~ q3 + &amp;c. For the series xp +1 rq/ p +1+ q/ p +1~ xp +2~ rq 1/ p +2 + &amp;c. in Prop : 10 .
Art .
2 .
of the Essay , which expresses the ratio of ACf to H O , becomes this series when x=a= p/ n , b=r= q/ n ; that is when Cf has moved till it coincides with Dh and ACf becomes AD h. In like manner , from Art .
3 .
in the Essay , it appears that the ratio of HDh to HO is pq+1a b/ q +1X1+ p/ q +2X q/ p + p/ q +2X p-1/ q +3X q2/ p2 + &amp;c. From hence it follows that the ratio of the difference pq between ADh and HDh to HO is a b/ n multiplied by the difference between the series p/ p +1+ q/ p +1 X p2/ p q+2q+qXq1X p3/ p +1Xp+2Xp q2 +3 q2 + &amp;c. and the series q/ q +1+pX q2/ q +1Xpq+2p+ pXp1X q3/ q +1Xq+2X p2 q+3 p2 + &amp;c. the former series being to be substracted from the latter , if HDh is greater than AD h , and vice versa .
2 .
The ratio of any term in the former of the two foregoing series to that which next but one follows the correspondent term in the latter is pq+ p/ p XqX pq+2 p/ p XqXpX q/ q p+qXpq+3 p/ p qqXp q/ p q+2qXpq+4 p/ p q2q Xpq p/ p q+3qXpq+5 p/ p q3qXpq2 p/ p q+4qXpq+6 p/ p q4q &amp;c. taking twice as many terms and four over as there are units in the number which expresses the distance of the term in the former series from its first term ; which ratio if q be greater than p , it is evident must be greater than the ratio of equality .
Wherefore , if from the second series you substract the two first terms which together are less than two , the remainder is less than the former series ; and of consequence , the former series substracted from the latter cannot leave a remainder so great as two .
And therefore in this case , that is , when q is greater than p , by the preceding article , the ratio of HDhADh to HO pq cannot be so great as 2a b/ n. 3 .
The curve ADH being as before divided into two parts ADh and HD h , let the ordinates Cf and Ft be placed on each side of Dh and at the same distance from it , and let z be the ratio of hf or ht to A H. Then if y , x and r be respectively the ratios of C f , Af and Hf to A H , by the nature pq of the curve y=x r. But because the ratio of Ah to AH is a , and that of hf to AH is z , the ratio of A h-h f ( =A f ) to AH is a-z .
Wherefore a-z=x .
And in like manner b+ z=r .
But pq y=x r , and y is the ratio of Cf to A H. Wherefore the ratio of Cf to AH is a z~p Xb+ z~q .
And in like manner the ratio of Ft to AH is a+z~p X b-z~q .
And consequently Cf is to Ft as a-z~p X b+z~q is to a+ z~p Xb z~q .
4 .
If q is greater than p , a+z~p X b-z~q is greater than a-z~p Xb+ z~q , and the ratio between them increases as z increases .
For the hyperbolic logarithm of that ratio taken as usual , and then instead of p and q putting na and nb because a= p/ n and b= q/ n ( Vid. Art .
1 .
) you will find to be 2n multiplied by the series b2 a2/ 3 b2 a2 X z3 + b4 a4/ 5 b4 a4 X z5 + b6 a6/ 7 b6 a6 X z7 &amp;c. which logarithm when q is greater than p , and therefore b greater than a has all its terms positive , and so much the greater as z is greater ; and therefore it is the logarithm of a ratio greater than that of equality , and which increases as z increases .
5 .
By Art .
3 .
Ft is to Cf as a+z~p X b-z~q is to a-z~p X b+z~q .
And by Art .
4 .
a+z~p X b-z~q is greater than a-z~p Xb+ z~q , and the ratio between them increases as z increases , if q is greater than p. Wherefore , upon this supposition , also Ft is greater than C f , and the ratio between them increases as z or ht and hf increases , and consequently this will be true also concerning the areas described by them as their equal abscisses ht and hf increase .
Wherefore , when q is greater than p , DhtF is greater than Dhf C , and the ratio between them increases as hf=ht increases .
6 .
Because Ah is to Hh as p is to q , when q is greater than p , Hh is greater than A h. In Hh therefore taking hl equal to A h , by the preceding Art .
the part of the figure HDh which insists upon hl will be greater than AD h , and the ratio of that part of HDh to ADh will be greater than the ratio of DhtF to Dhf C. Consequently , much more ( q being greater than p ) the whole figure HDh is greater than AD h , and the ratio of HDh to ADh is greater than that of DhtF to Dhf C. 7 .
When q is greater than p , 1 n2 z2/ p q~n/ 2 is greater than 1n z/ p~p X1+n z/ q~q and less than 1n z/ q~q X1+n z/ p~p .
For the fluxion of 1 n2 z2/ p q~n/ 2 is n3 z z~/ p qX1 n2 z2/ p q~n/ 2 1 and the fluxion of 1n z/ p~p X1+n z/ q~q ( because p+q= n ) is n3 z z~/ p qX1n z/ p~p 1X1+n z/ q~q 1 .
Wherefore 1 n2 z2/ p q~n 2 is to 1n z/ p~p X1+n z/ q~q as the fluxion of the former multiplied by 1 n2 z2/ p q~ to the fluxion of the latter multiplied by ( 1 n z/ p X 1+n z/ q or ) 1n z/ p +n z/ q n2 z2/ p q. From which analogy , because q is greater than p , it is plain that 1n z/ p~p X1+n z/ q~q varies at a greater rate in respect of its own magnitude than 1 n2 z2/ p q~n/ 2 does .
And , because their fluxions as found out before have a negative sign before them , they both decrease as z increases ; creases ; consequently , if they are ever equal , as z increases the latter must be the largest .
But when z =0 they are each equal to 1 .
In like manner the other part of this article appears .
And hence , since a= p/ n and b= q/ n , it is manifest that ap bq X1 n2 z2/ p q~n/ 2 is greater than a-z~p Xb+ z~q and less than a+ z~p Xb z~q , when q is greater than p. 8 .
Suppose now further that the curve RQW be described meeting the lines D h , F t , ht produced in R , Q , W , in such manner that F t , which is to Cf as a+ z~p Xb z~q to a z~p Xb+ z~q ( Vid. Art .
3 .
) shall be to Qt as a+ z~p Xb z~q to ap~ bq X 1 n2 z2/ p q~n/ 2 wherever the points t and f fall at equal distances from h. And it is manifest by the foregoing Art .
that Qt must always be greater than C f , and less than F t. And of consequence the same must be true concerning the areas described by their motion while their equal abscisses increase .
Wherefore RhtQ is greater than Dhf C , and less than D ht F. 9 .
Since Ft is to Qt as a+ z~p Xb z~q to ap bq X1 n2 z2/ p q~n/ 2 ; and a+ z~p Xb z~q ( by Art .
3 .
) expresses the ratio of Ft to A H ; the ratio of Qt to AH must be ap bq X1 n2 z2/ p q~n/ 2 , and as has been all along supposed z is the ratio of ht to A H. Wherefore , by squaring the curve Rht Q , it will appear that the ratio of RhtQ to HO is ap bq Xz n3 z3/ 2.3 pq+ n-2/ 4 X n5 25/ 2.5 p2 q2 n 2/ 4 X n 4/ 6 X n7 z7/ 2.7 p3 q3 + &amp;c. which ( if m2 =n 3/ 2 p q ) is ap bq X~2p q/ n ~nXmz m3 z3/ 3 +n 2/ 2 nX m5 z5/ 5 n 2/ 2 n Xn 4/ 3 nX m7 z7/ 7 +n 2/ 2 nXn 4/ 3 nXn 6/ 4 nX m9 z9/ 9 &amp;c. Which last series when n2 z2/ p q= 1 , and consequently the ordinate Qt vanishes , becomes B B3/ 3 + B2 1/ 2B2 X B5/ 5 B2 1/ 2B2 X B2 2/ 3B2 X B7/ 7 + &amp;c. taking B2 = n/ 2 .
10 .
If B2 = n/ 2 the ratio of the whole figure RQWh to HO is ~2p q/ n ~n X ap bq XB B3/ 3 + B2 1/ 2B2 X B5/ 5 &amp;c. Now , ( by Prop. 10 .
Art .
4 .
of the Essay ) the ratio of ACFH to HO is 1/ n +1 x 1/ E , E being the coefficient of that term of the binomial a+b~n expanded in which occurs ap bq : Wherefore , the ratio of RQWh to ACFH is n+ 1/ n x~2p q/ ~ nxE ap bq xB B3/ 3 + B21/ 2B2 xB 5/ 5 &amp;c. Put G now for the coefficient of the middle term of the same binomial , and if p=q= n/ 2 , E= G , a = 1/ 2 =b the area RQWh is equal to half ACF H ; for then Q t , F t , Cf are all equal , and consequently the areas RQW h , HDh and AD h. Wherefore , the series B-B3/ 3 + &amp;c. is equal to ~2 n/ n +1x 2n-1/ G. But the series B B3/ 3 + &amp;c. ( because B2 = n/ 2 ) does not alter whatever p and q are , whilst their sum n remains the same .
Wherefore , in all cases , the ratio of RQWh to ACFH is ~p q/ n xE ap bq/ G x 2n .
11 .
By Prop. 10 .
Art .
4 .
of the Essay , the ratio of ACFH to H O* is 1/ n +1x 1/ E ; and by Art .
9 .
the ratio of R ht Q to HO is ap bq x~2p q/ n~n xmz m3 z3/ 3 +n 2/ 2n x m5 z5/ 5 &amp;c. Wherefore , the ratio of * It is hoped that the imperfection of the figure all along referred to will.be excused .
The lines Rh and Dh should appear equal ; and it will be found presently , that the curve line ACDFH should have been drawn from F and C convex towards A H. R ht Q to ACFH is n+ 1/ n x~2p q/ ~ nxE ap bq x mz-m3 z3/ 3 +n 2/ 2 nxm5z 5/ 5 n 2/ 2 nxn 4/ 3 nx m7 z7/ 7 + &amp;c. Likewise , by Art .
10 .
the ratio of RQWh to ACFH is ~p q/ n xE ap bq/ G x 2n .
Wherefore the ratio of RhtQ to RQWh is n+ 1/ ~ nx~ 2/ 2 nxG xmz m3 z3/ 3 + &amp;c. 12 .
By Art .
2 .
6 .
When q is greater than p , the ratio of HDhADh to HO is less than 2 ap bq/ n. And by Prop. 10 .
Art .
4 .
of the Essay , the ratio of ACFH or HDh+ADh to HO is 1/ n +1x 1/ E. Wherefore , the sum of these two ratios , or the ratio of 2HDh to H O , is less than 1/ n +1 x 1/ E +2 ap bq/ n ; and the difference between them , or the ratio of 2ADh to HO is greater than 1/ n +1x 1/ E 2 ap bq/ n. Wherefore , the ratio of 2HDh to 2AD h , or that of HDh to AD h , is less than that of 1/ n +1x 1/ E +2 ap bq/ n to 1/ n +1x 1/ E 2 ap bq/ n , which is equal to the ratio of 1x2E ap bq +2E ap bq/ n to 1 2E ap bq 2E ap bq/ n. But the ratio of HDh to AD h , by Art .
6 .
is greater than the ratio of DhtF to Dhf C , when q is greater than p. Wherefore , much more when q is greater than p , the ratio of DhtF to DhfC will be less than that of 1+ 2E ap bq +2E ap bq/ n to 12E ap bq 2E ap bq/ n. And because , by Art .
8 .
RhtQ is a mean between DhtF and Dhf C , the ratio of DhtF to RhtQ will be less than that of 1+2E ap bq +2E ap bq/ n to 12E ap bq 2E ap bq/ n. And the ratio of DhfC to RhtQ will be greater than that of 12E ap bq 2E ap bq/ n to 1+2E ap bq +2E ap bq/ n. RULE II .
If nothing is known of an event but that it has happened p times and failed q in p+q or n trials , and q be greater than p ; and from hence I judge that the probability of its happening in a single trial lies between p/ n and p/ n + z , ( if m2 = n3/ 2 p q , a= p/ n , b= q/ n , E the coefficient of the term in which occurs ap bq when a+ b~n is expanded , and ~= n+ 1/ n x~2p q/ ~ nxE ap bq xmz m3 z3/ 3 +n 2/ 2 nx m5 z5/ 5 n 2/ 2 n~n 4/ 3 n~ m7 z7/ 7 + ~c .
) my chance to be in the right is greater than ~ , and less than ~~ 1+2E ap bq +2E ap bq/ n/ 1 2E ap bq 2E ap bq/ n. For by Art .
11 .
compared with the value of ~ here set down , the ratio of RhtQ to ACFH is ~ .
But by Art .
8 .
DhtF is greater than Rht Q , and by Art .
12 .
the ratio of DhtF to RhtQ is less than that of 1+2E ap bq +2E ap bq/ n to 12E ap bq 2E ap bq/ n. From whence it is plain that the ratio of DhtF to ACFH is greater than ~ , and less than ~~1+2E ap bq +2E ap bq/ n/ 1 2E ap bq 2E ap bq/ n But , as appears from the 10th Proposition in the Essay , the chance that the probability of the event lies between p/ n and p/ n +z ( that is , between the ratio of Ah to A H , and that of At to A H ) is the ratio of DhtF to ACF H. Wherefore , the chance I am right in my guess is greater than ~ and less than ~~ 1+2E ap bq +2E ap bq/ n/ 1 2E ap bq 2E ap bq/ n. In like manner , 2dly , the same things supposed , if I judge that the probability of the event lies between p/ n and p/ n z , my chance to be right is less than ~ , and greater than ~~12E ap bq 2E ap bq/ n/ 1 +2E ap bq +2E ap bq/ n. This is manifest as the other case , because DhfC is lessthan Rht Q , but the ratio between them is greater than that of 12E ap bq 2E ap bq/ n to 1+2E ap bq + 2E ap bq/ n. 3dly , If , other things supposed as before , p is greater than q , and I judge the probability of the event lies between p/ n and p/ n + z , my chance to be right is less than ~ , and greater than ~~ 12E ap bq 2E ap bq/ n/ 1 +2E ap bq +2E ap bq/ n. But if I judge it lies between p/ nand p/ n z , my chance is greater than ~ , and less than ~~1+2E ap bq +2E ap bq/ n/ 1 2E ap bq 2E ap bq/ n. And if p= q , which ever of these ways I guess , my chance is ~ exactly .
This may be proved in the same manner with the foregoing cases , where q is greater than p , or may be proved from them by considering the happening and failing of an event , as the same with the failing and happening of its contrary .
4thly , Other things supposed the same , whether q be greater or less than p , and I judge that the probability of the event lies between p/ n +z and p/ n z , my chance is greater than 2 ~/ 1 +2E ap bq +2E ap bq/ n , and less than 2 ~/ 1 2E ap bq 2E ap bq/ n. This is an evident corollary from the cases already determined .
And here , if p= q , my chance is 2~ exactly .
Thus far I have transcribed Mr. Bayes .
It appears , from the Appendix to the Essay , that the rule here demonstrated , though of great use , does not give the required chance within limits sufficiently narrow .
It is therefore necessary to look out for a contraction of these limits ; and this , I think , we shall discover by the help of the following deductions ; which , for the sake of greater dinstinctness , I shall give as a continuation of the foregoing Articles .
13 .
The ratio of the fluxion of ~1 n2 z2/ p q n/ 2 to the fluxion of ~1 +n z/ p px ~1 n z/ q q is ~1 n2 z2/ p q n/ 2 1/ ~ 1+n z/ p p-1 x ~1 n z/ q q-1 ; and the ratio of the fluxion of ~1 n z/ p px~1n z/ q q to the fluxion of ~1 n2 z2/ p q n/ 2 is ~1 n z/ p p1x ~1 +n z/ q q 1/ ~1 n2 z2/ p q n/ 2 1 .
This will immediately appear from Art .
7 .
14 .
While z is increasing from nothing till n2 z2/ p q becomes equal to unity , these two ratios are at first greater than the ratio of equality , and increase as z increases , till they come to a maximum .
Afterwards they decrease until they become first equal to the ratio of equality , and then less .
This is proved by finding the hyperbolic logarithms of these ratios .
The hyperbolic logarithm of the first is the series q 1/ q p 1/ p xnz+ q 1/ q2 +p 1/ p2 n 2/ p qx n2 z2/ 2 +q 1/ q3 p 1/ p3 x n3 z3/ 3 +q 1/ q4 +p 1/ p4 n 2/ p2 q2 x n4 z4/ 4 +q 1/ q5 p 1/ p5 x n5 z5/ 5 +q 1/ q6 +p 1/ p6 n 2/ p3 q3 x n6 z6/ 6 + &amp;c. The hyperbolic logarithm of the second ratio is the series q 1/ q p 1/ p xnzq 1/ q2 + p 1/ p2 n 2/ p qx n2 z2/ 2 +q 1/ q3 p 1/ p3 x n3 z3/ 3 q 1/ q4 +p 1/ p4 n 2/ p2 q2 x n4 z4/ 4 + &amp;c. It will appear from examining these two serieses ( q all along supposed greater than p ) that while z is small the value of each of them is positive , and increases as z increases till it becomes a maximum , after which it decreases till it becomes nothing , and after that negative ; which demonstrates this article .
15 .
The former of the two ratios in Art .
13 .
( q being greater than p ) is at first , while z is increasing from nothing , less than the second ratio ; and does not become equal to it , till some time after both ratios have been the greatest possible .
Upon considering the two serieses in the last Art .
it will appear that the first term of the first series is always positive , the second negative , the third also negative , after which the terms become alternately positive and negative .
On the other hand , it will appear that in the second series the two first terms are always positive , and all that follow negative .
But as the serieses converge very fast when z is small , the second term being negative in the first series and positive in the second , has a greater effect in making the first series less than the second , than can be compensated for by the terms being afterwards alternately negative and positive in the one , and all negative in the other .
It will further appear from considering these serieses , that the first must continue less than the second 'till z becomes so large as to make the fourth term equal to the second , in which circumstances the two serieses are nearly equal .
Afterwards , as z goes on to increase , the value of both lessens continually ; but the second now decreasing fastest , as before it increased fastest , becomes first nothing .
After which , the other series becomes nothing ; and after that both remain negative .
From hence it is easy to infer this Article .
16 .
What has been now shown of the ratio of the fluxion of ~1 n2 z2/ p q n/ 2 to the fluxion of ~1+n z/ p px ~1 n z/ q q compared with the ratio of the fluxion of ~1 n z/ p px ~1 +n z/ q q to the fluxion of ~1 n2 z2/ p q n/ 2 is also true of the ratio of the fluxion of ap bq x ~1 n2 z2/ p q n/ 2 ( or Qt in the figure ) to the fluxion of ~a +zpxbzq ( or F t ) compared with the ratio of the fluxion of ~a zpx ~b +zq ( or C f ) to the fluxion of ap bq x ~1 n2 z2/ p q n/ 2 or Q t ; the latter quantities being only the former multiplied by the common and permanent quantity ap bq .
It appears , therefore , that if we conceive F t , Q t , Cf ( Vid. Fig. ) to move with equal and uniform velocities , from Dh and Rh along A H , in order to generate the areas HD h , RW h , AD h ; Cf will at first not only decrease faster than Q t , and Qt than F t ; but the ratio of the rate at which Cf decreases to the rate at which Qt decreases , will be greater than the ratio of the rate at which Qt decreases to the rate at which Ft decreases .
It appears also that after some time , first Cf and Q t , and then Qt and Ft will come to decrease equally ; after which , Qt will decrease faster than C f , and Ft faster than Q t. 17 .
The curves DF H , RQ W , DC A , have each of them a point of contrary flexure ; and the value of z , or of the equal abscisses at that point , is in all three ~p q/ ~ n3 n2 .
This may be found in the common manner , by putting the second fluxions of the ordinates equal to nothing .
In the single case , when either p or q is equal to unity , one of these points vanishes , or coincides with A or H. 18 .
At the points of contrary flexure ( q being greater than p ) the ratio of the fluxion of Qt to the fluxion of Ft is a maximum ; and the same is true of the ratio of the fluxion of Cf to the fluxion of Q t. This is found by making the fluxions of the logarithms of these ratios , or of ~1 n2 z2/ p q n/ 2 1/ ~1 +n z/ p p1~ ~1 n z/ q q 1 , and ~1 n z/ p p1~ ~1 +n z/ q q 1/ ~1 n2 n2/ p q n/ 2 1 1 equal to nothing : which will give the value of z equal to ~p q/ ~ n3 n2 , or the same with the value of z at the points of contrary flexure .
19 .
At the points of contrary flexure , the ratio of the fluxion of Cf to the fluxion of Q t , is greatest in comparison of the ratio of the fluxion of Qt to the fluxion of F t. This is proved by finding the value of z when the fluxion of the former ratio divided by the latter , or of ~1 n2 z2/ p2 p1~ ~1 n2 z2/ q2 q 1/ ~1 n2 z2/ p qn2 is nothing , which will still give z=~p q/ ~ n3 n2 .
The reason , therefore , in the nature of the curve , which , as the ordinates flow , keeps at first the excess of Ft above Qt less than the excess of Qt above C f , operates with its greatest force at the points of contrary flexure .
20 .
The greatest part of the area RQWh lies between R h , and the ordinate at the point of contrary flexure .
By Art .
11 the ratio of RhtQ to RQWh is n+ 1/ ~ n~~ 2/ 2n ~G~ ~m z m3 z3/ 3 +n 2/ 2 n~ m5 z5/ 5 &amp;c. Substitute here ~p q/ n3 n2 for z , and 2n/ ~ nk~H * for G ( K being the ratio of the quadrantal arc to * This is always the true value of G ; but it would be too tedious to give the demonstration of this here .
radius , and H the ratio whose hyperbolic logarithm is 3/ 12 n 15/ 360 n3 + 63/ 1260 n5 * &amp;c. ) and the ratio of RhtQ to RQWh at the point of contrary flexure , will be n+ 1/ ~ n~~ n-1 ~.797884/ H ~ ~1 n/ 2 .
3 .
~n 1+ ~n ~ ~n 2/ 2.5.4 .
~n-12 n. ~n 2 .
~n 4/ 2.3.7.8 .
~n 13+ n. ~n 2 .
~n 4 .
~n 6/ 2.3.4.9.16 .
~n 14 &amp;c. Now when n is little , the value of this expression will be considerably greater than.6822 .
It approaches to this continually as n increases ; and when n is large , it may be taken for this exactly .
Thus when n= 6 , this expression is equal to.804 .
When n= 110 , it is equal to.6903 .
If we would know the ratio of RhtQ to RQW h , when Cf comes to decrease no faster in respect of Q t , than Qt decreases in respect of F t ; that is , when the excess of Qt above C f , is greatest in comparison of the excess of Ft above Q t , it may be found ( by putting the fourth term of the series in the 14th Art .
equal to the second term , and then finding the value of z ) to be about.8426 , when n , p , and q are considerable ; and in other cases greater .
Coroll. It is easy to gather from hence that in like manner the greatest part of the area ADH lies between the two ordinates at the points of contrary flexure ~ .
* Vid. the Second Rule in the Essay , Phil. Trans. Vol. LIII .
~ From this Article may be inferred a method of finding at once , without any labour , whereabouts it is reasonable to judge the probability of an unknown event lies , about which a given number of experiments have been made .
For when 21 .
RhtQ is greater than the arithmetical mean between DhtF and Dhf C. This appears from the latter part of Art .
19 .
for what is there proved of the ordinates must hold true of the contemporary areas generated by them .
And though beyond the points at which the ratio of the decrease of Qt to the decrease of Ft comes to an equality with the ratio of the decrease of Qt to the decrease of C f , the excess of Ft above Qt begins to grow larger than before in respect of the excess of Qt above C f ; yet as it appears from the last article , that above five sixths of the areas RQWh and ACFH are generated before the ordinates come to these points , and as also beyond these points the said ratios , 'till they become neither p nor q are very small , or even not less than 10 , it will be nearly an equal chance , that the probability of the event lies between p/ n + ~p q/ ~2 n3 2 n2 and p/ n ~p q/ ~2 n3 2 n2 .
It will be the odds of two to one that it lies between p/ n + ~p q/ ~n 3 n2 and p/ n ~p q/ ~n 3 n2 ; and the odds of five to one that it lies between p/ n + ~2 p q/ ~n3 n2 and p/ n ~2 p q/ ~n3 n2 .
For instance ; when p= 1000 , q= 100 , there will be nearly an equal chance , that the probability of the event lies between 10/ 11 + 1/ 163 and 10/ 11 1/ 163 ; two to one that it lies between 10/ 11 + 1/ 115 and 10/ 11 1/ 115 ; and five to one that it lies between 10/ 11 + 1/ 81 and 10/ 11 1/ 81 .
negative and for some time afterwards , are but small ; the effect produced before towards rendering the excess of DhtF above RbtQ always less than the excess of RhtQ above Dhf C , will be such as cannot be compensated for afterwards .
A further proof of this will appear from considering that even when RhtQ is increased to RQW h , it is but little short of the arithmetical mean between ADh and HD h. For from Art .
11 .
and 20 .
it may be inferred that the ratio of the whole area RQWh to this mean , or to ACF H/ 2 , is h~ H , which is never far from the ratio of equality , but when both p and q are of any considerable magnitude , it is very nearly the ratio of equality .
For example ; when n= 110 , q= 100 , p= 10 , it is .9938 .
22 .
The ratio of DhtF to RhtQ is less than that of 1+2E ap bq +2E ap bq/ n to one .
For by Art .
12 .
the ratio of DhtF to DhfC is less than that of 1+2E ap bq +2E ap bq/ n to 12E ap bq 2E ap bq/ n. But by the last Art .
RhtQ is greater than the arithmetical mean between DhtF and Dhf C , and 1 is exactly the arithmetical mean between 1+2E ap bq +2E ap bq/ n and 12E ap bq 2E ap bq/ n. From whence this Article is plain .
23 .
The ratio of DhtF to ACFH is greater than ~ , and less than ~~1+2E ap bq +2E ap bq/ n. For DhtF being greater than Rht Q , the ratio of it to ACFH must be greater than the ratio of RhtQ to ACF H , or greater than ~ .
Also ; since the ratio of RhtQ to ACFH is equal to ~ ; and the ratio of DhtF to RhtQ is less than the ratio of 1+2E ap bq +2E ap bq/ n to 1 ; it follows that the ratio compounded of the ratio of RhtQ to ACF H , and of DhtF to Rht Q , that is , the ratio of DhtF to ACFH must be less than ~~ ~1 +2E ap bq + ~2 E ap bq/ n. 24 .
The ratio of DhtF+DhfC to ACFH ( that is , the chance for being right in judging that the probability of an event perfectly unknown , which has happened p and failed q times in p+q or n trials , lies somewhere between p/ n +z and p/ n z ) is greater than 2 ~/ 1 +2E ap bq +2E ap bq/ n , and less than 2 ~ .
The former part of this Art .
has been already proved , Art .
12 .
The latter part is evident from Art .
21 .
For RhtQ being greater than the arithmetical mean between DhtF and Dhf C , 2RhtQ must be greater than DhtF+Dhf C ; and consequently the ratio of 2RhtQ to ACF H , greater than the ratio of DhtF+DhfC to ACFH * .
It will be easily seen that this Article improves considerably the rule given in Art .
12 .
But we may determine within still narrower limits whereabouts the required chance must lie , as will appear from the following Articles .
25 .
If c and d stand for any two fractions , whenever the fluxion of c~Ft is greater than the fluxion of d~Cf ( Vid. fig. ) c~Ft+d~Cf will be greater than Q t. For in the same manner with Art .
6 .
it will appear that c~Ft+d~Cf is to Q t , as the fluxion of c~Ft~1+ ~n z/ q ~ ~1 n z/ q together with the fluxion of d~Cf~ ~1 n z/ p ~ ~1 +n z/ q to the fluxion of Qt~ ~1 n2 z2/ p q. Now since 1 n2 z2/ p q is the arithmetical mean between ~1 +n z/ p ~ ~1 n z/ q and ~1 n z/ p ~ ~1 +n z/ q , it is plain , that were the fluxion of c~Ft equal to the fluxion of d~C f , c~Ft+d~Cf would decrease in respect of its own magnitude at the same rate with Q t ; and , therefore , since at first equal , they would always continue equal .
But the fluxion of c~Ft being greater than the fluxion of d~Cf by supposition , and ( since q greater than ) ~p 1+n z/ p ~ ~1 n z/ q , * This Art .
is true , whether p be greater or less than q. also greater than ~1 n z/ p ~ ~1 +n z/ q , it follows that the fluxion of c~Ft~ ~1 +n z/ p ~ ~1 n z/ q added to the fluxion of d~Cf~ ~1 n z/ p ~ ~1 +n z/ q is greater than these two fluxions multiplied by ~1 n2 z2/ p q ; and , therefore , greater , than the fluxion of Qt~ ~1 n2 z2/ p q ; and , therefore , c~Ft+d~Cf greater than Q t. 26 .
If we suppose three continued arithmetical means between Cf and Ft ( 3 Cf+F t/ 4 , Cf+F t/ 2 , 3Ft+C f/ 4 ) Qt will be greater than the second , and less than the third , if p is greater than 1 .
That Qt will be greater than the second has been already proved ; and that it will be less than the third , will be an immediate consequence from the last Article , if it can be shown that the fluxion of 3F t/ 4 is greater than the fluxion of C f/ 4 .
This will appear in the following manner .
The ratio of the fluxion of Cf to the fluxion of Ft is by Art .
7 .
and 14 .
~1 n z/ p p1~ ~1 +n z/ q q 1/ ~1 +n z/ p p1~ ~1 n z/ q q 1 .
The hyperbolic logarithm of this ratio is ~1/ p 1/ q ~2nz ~1/ p2 1/ p3 1/ q2 + ~1/ q3 ~2 n3 z3/ 3 ~1/ p4 1/ p5 1/ q4 + 1/ q5 ~2 n5 z5/ 5 , &amp;c. This ratio by Art .
18 .
is greatest at the point of contrary flexure , or when z= ~p q/ ~n3 n2 .
Substitute this for z in the series , and it will become ~1/ p 1/ q ~2 ~p q/ ~n 1 ~1/ p2 1/ p3 1/ q2 + 1/ q3 ~2 p3/ 2 ~ q3/ 2/ 3 ~ ~n 1 3/ 2 - , &amp;c. which , therefore , expresses the logarithm of the ratio when greatest , and will easily discover it in every case .
It is apparent that the value of this series is greatest when p is least in respect of q. Suppose then p= 2 , and q infinite .
In this case , the value of the series will be 1.072 , and the number answering to this logarithm is not greater than 2.92 .
The fluxion , therefore , of C f , when greatest , cannot be three times the contemporary fluxion of F t ; from whence it follows that the fluxion of 3F t/ 4 must be greater than the fluxion of C f/ 4 .
It is easy to see how these demonstrations are to be varied when q is less than p , and how in this case similar conclusions may be drawn .
Or , the same conclusions will in this case immediately appear , by changing p into q and q into p , which will not make any difference in the demonstrations .
In the manner specified in this Article we may always find within certain limits how near the value of Qt comes to the arithmetical mean between Ft and C f , which limits grow narrower and narrower , as p and q are taken larger , or their ratio comes nearer to that of equality , 'till at last , when p and q are either very great or equal , Qt coincides with this mean .
Thus , if either p or q is not less than 10 ; that is , in all cases , where it is not practicable without great difficulty to find the required chance exactly by the first rule , Qt will be greater than the fourth , and less than the fifth of seven arithmetical means between Cf and F t. 27 .
The arithmetical means mentioned in the last Article may be conceived as ordinates describing areas at the same time with Q t ; and what has been proved concerning them is true also of the areas described by them compared with Rht Q. 28 .
If either p or q is greater than 1 , the true chance that the probability of an unknown event which has happened p times and failed q in ~p +q or n trials , should lie somewhere between p/ n +z and p/ n z is less than 2 ~ , and greaterer than ~+ ~~ ~12E ap bq 2E ap bq/ n/ 1 +E ap bq +E ap bq/ n. If either p or q is greater than 10 , this chance is less than 2 ~ , and greater than ~+ ~~ ~12E ap bq 2E ap bq/ n/ 1 + 1/ 2 E ap bq +E ap bq/ 2 n. This is easily proved in the same manner with Art .
12 , 23 , 24 .
That it may appear how far what has been now demonstrated improves the solution of the present problem , let us take the fifth case mentioned in the Appendix to the Essay , and enquire what reason there is for judging that the probability of an event concerning which nothing is known , but that it has happened 100 times and failed 1000 times in 1100 trials , lies between 10/ 11 + 1/ 110 and 10/ 11 1/ 110 .
The second rule as given in Art .
12 .
informs us , that the chance * for this must lie between.6512 , ( or the odds of 186 to 100 ) and.7700 , ( or the odds of 334 to 100 ) .
But from the last Art .
it will appear that the required chance in this case must lie between 2 ~ , and ~+~~1E ap bq 2E ap bq/ n/ 1 + 1/ 10 E ap bq +E ap bq/ 10 n ; or , between .6748 and.7057 ; that is , between the odds of 239 to 100 , and 207 to 100 .
In all cases when z is small , and also whenever the disparity between p and q is not great 2~ is almost exactly the true chance required .
And I have reason to think , that even in all other cases , 2~ gives the * In the Appendix , this chance , as discovered by Mr. Bayes 's second rule , is given wrong , in consequence of making m2 equal to n3/ p q , whereas it should have been taken equal to n3/ 2 pq as appears from Article 8 .
true chance nearer than within the limits now determined .
But not to pursue this subject any further ; I shall only add that the value of 2~ may be always calculated very nearly , and without great difficulty ; for the approximations to the value of E ap bq , and of the series mz m3 z3/ 3 +n 2/ 2 n~ m5 z5/ 5 , &amp;c. * given in the Essay , are sufficiently accurate in all cases where it is necessary to use them .
* In the expression for this last approximation there is an error of the press which should be corrected ; for the sign before the fourth term should be and not + .

