XIX .
A Letter to the Right Honourable George Earl of Macclesfield , President of the Royal Society , on the Advantage of taking the Mean of a Number of Observations , in practical Astronomy : By T. Simpson , F. R. S. My Lord , IT is well known to your Lordship , that the method practised by astronomers , in order to diminish the errors arising from the imperfections of instruments , and of the organs of sense , by taking the Mean of several observations , has not been so generally received , but that some persons , of considerable note , have been of opinion , and even publicly maintained , that one single observation , taken Read April 10 , taken with due care , was as much to be relied on as the ~ of a great number .
As this appeared to me to be a matter of much importance , I had a strong inclination to try whether , by the application of mathematical principles , it might not receive some new light ; from whence the utility and advantage of the method in practice might appear with a greater degree of evidence .
In the prosecution of this design ( the result of which I have now the honour to transmit to your Lordship ) I have , indeed , been obliged to make use of an hypothesis , or to assume a series of numbers , to express the respective chances for the different errors to which any single observation is subject ; which series , to me , seems not ill-adapted : but this I shall submit entirely to the judgment of your Lordship , who have made so great a number of observations , at your seat at Shirburn ; where , to the best collection of mathematical books , your Lordship has added a more complete set of astronomical instruments than ( perhaps ) are to be found in the possession of any nobleman in Europe .
Should not the assumption , which I have made use of , appear to your Lordship so well chosen as some others might be , it will , however , be sufficient to answer the intended purpose : and your Lordship will find , on calculation , that , whatever series is assumed for the chances of the happening of the different errors , the result will turn out greatly in favour of the method now practised , by taking a mean value .
But I shall no longer detain your Lordship with general observations , but proceed to the matter proposed ; which I shall consider in the following propositions .
PROPOSITION I. Supposing that the several chances for the different errors that any single observation can admit of , are expressed by the terms of the progression r-v r-3 , r-2 , r-1 , r~ , r1 , r2 , r3 rv ( where the exponents denote the quantities and qualities of the particular errors , and the terms themselves the respective chances for their happening ) : it is proposed to determine the probability , or odds , that the error , by taking the Mean of a given number ( n ) of observations , exceeds not a given quantity ( m/ n ) .
It is evident , from the laws of chance , that , if the given series , r-v + r-3 + r-2 + r-1 + r~ + r1 + r2 + r3 + rv , expressing all the chances in one observation , be raised to the nth power , the terms of the series thence arising will truly exhibit all the different chances in all the proposed ( n ) observations .
In order to raise this power , with the greatest facility , our given expression may be reduced to r-v x 1-r2v+1/ 1-r : whereof the nth power ( making w= 2v + 1 ) , will be r-nv x 1-rw~n x 1-r~-n ; which , expanded , becomes r-nv nrw-nv + n/ 1.n-1/ 2r2w-nv -n/ 1.n-1/ 2.n-2/ 3r3w-nv + , &amp;c. multiplied into 1+ nr + n/ 1.n+1/ 2r2 + n/ 1.n+1/ 2.n+2/ 3r3 + , &amp;c. Now , to find from hence the sum of all the chances whereby the excess of the positive errors above the negative ones can amount , precisely , to a given number m , it will be sufficient ( instead of multiplying the former series by the whole of the latter ) to multiply by such terms of the latter , only , as are necessary to the production of the given exponent m , in question .
Thus , the first term ( r-nv ) of the former series is to be multiplied by that term of the second , whose exponent is nv + m , in order that the power of r , in the product , may be rm .
But it is plain , from the law of the series , that the coefficient of this term ( putting nv+m= q ) will be n/ 1.n+1/ 2.n+2/ 3 ( q ) , q being the number of factors ; and consequently , that the product under consideration will be n~1.n+1/ 2.n+2/ 3 ( q ) x rm .
Again , the second term of the former series being nrw-nv , the exponent of the corresponding term of the latter will be w+nv+m ( = q w ) , and therefore the term itself equal to n/ 1.n+1/ 2.n+2/ 3 ( q-w ) x rq-w : which , drawn into -nrw-nv , gives n/ 1.n+1/ 2.n+2/ 3 ( q-w ) x-nrm , for the second term required .
In like manner , the third term , of the product , whose exponent is m , will be found n/ 1.n+1/ 2.n+2/ 3 ( q-2w ) xn/ 1.n-1/ 2rm : and the sum of all the terms having the same given exponent ( m ) will consequently be +n/ 1.n+1/ 2.n+2/ 3 ( q ) x rm -n/ 1.n+1/ 2.n+2/ 3 ( q-w ) x nrm +n/ 1.n+1/ 2.n+2/ 3 ( q-2w ) x n/ 1.n-1/ 2rm -n/ 1.n+1/ 2.n+2/ 3 ( q-3w ) x n/ 1.n-1/ 2.n-2/ 3rm &amp;c. &amp;c. From which general expression , by expounding m by 0 , + 1 , 1 , + 2 , 2 , &amp;c. successively , the sum of all the chances , whereby the difference of the positive and negative errors can fall within the proposed limits , will be found ; which , divided by r-vn x 1-rw~n x 1-r~-n , will give the true measure of the probability required : from whence the advantage of taking the Mean of several observations might be shown : but this I shall exemplify in the next proposition ; which is better adapted to the purpose , and to which this is premised , as a Lemma .
Remark .
If r be taken = 1 , or the chances for the errors in excess and defect be supposed exactly the same ; then our expression , by expunging the powers of r , will become the very same with that showing the chances for throwing n+q points with n dice ; each die having as many faces ( w ) , as the result of any one single observation , can come out different ways .
Which may be otherwise made to appear , independent of any kind of calculation , from the bare consider sideration , that the chances for throwing precisely the number m , with n dice , whereof the faces , of each , are numbered v 3 , 2 , 1 , 0 , +1 , +2 , +3 , + v , must be the very same as the chances whereby the positive errors can exceed the negative ones by that precise number ; which last are , evidently , the same as the chances for throwing precisely the number v+ 1 .
n+m ( or n+ q ) with the same n dice , when they are numbered in the common way , with the terms of the natural progression 1 , 2 , 3 , 4 , 5 , &amp;c. : because the number upon each face being , here , increased by v+ 1 , the whole increase upon all the n faces will be expressed by v+ 1 .
n ; so that there will be now the very same chance for the number v+ 1 .
n+ m , as there was before for the number m ; since the chances for throwing any faces assigned will continue the same .
however those faces are numbered .
PROPOSITION II .
Supposing the respective chances , for the different errors which any single observation can admit of , to be expressed by the terms of the series r-v + 2r1-v + 3r2-v +v+ 1.r0 + 3rv-2 + 2rv-1 + rv ( whereof the coefficients , from the middle one ( v + 1 ) , decrease , both ways , according to the terms of an arithmetical progression ) : it is proposed to determine the probability , or odds , that the error , by taking the Mean of a given number ( t ) of observations , exceeds not a given quantity ( m/ t ) .
Pursuing the method laid down in the preceding problem , the sum of the series here given will appear to be r-v ~ 1-r v+1 2/ 1-r 2 ( being the same with the square of the geometrical progression r -1/ 2v ~1+r+r2+r3+r v ) , and the t th power thereof , by making n=2 t , and w=v+ 1 .
will therefore be given =r -tv ~ 1-r wn~ 1-r -n = r -tv nr w-tv +n/ 1 .
n-1 / 2r 2w-tv +n/ 1 .
n-1 / 2 .
n-2 / 3r 3w-tv + , &amp;c. multiplied into 1+ nr +n/ 1 .
n+1 / 2r2+n/ 1 .
n+1 / 2 .
n+2 / 3r3 + , &amp;c. Which series being the same with those in the preceding problem ( excepting only that the exponents of the former of them are expressed in terms of t , instead of n ) , it is plain , therefore , that if q be made = tv +m ( instead of nv + m ) , the conclusion , there brought out , will answer equally here : so that the sum of all the chances whereby the excess of the positive errors above the negative ones , can amount to a given number m , precisely , will be truly represented by +n/ 1 .
n+1 / 2 .
n+2 / 3 ( q ) ~rm n/ 1 .
n+1 / 2 .
n+2 / 3 ( q-w ) ~nrm +n/ 1 .
n+1 / 2 .
n+2 / 3 ( q-2w ) ~n/ 1 .
n-1 / 2rm n/ 1 .
n+1 / 2 .
n+2 / 3 ( q-3w ) ~n/ 1 .
n-1 / 2 .
n-2 / 3rm &amp;c. &amp;c. But But this general expression , as several of the factors destroy each other , may be transformed to another , which is more commodious .
Thus n/ 1 .
n+1 / 2 .
n+2 / 3 ( q ) , in the first line , will , by breaking the numerator and denominator into two parts , become n. n+1 .
n+2 .
n+3 .
q~ q+1 .
q+2 .
q+3 .
q+n-1 / 1 .
2 .
3 .
4 .
n. n+1 .
n+2 .
n+3 .
q : which , by equal division , is reduced to q. q+1 .
q+2 q+n-1 / 1 .
2 .
3 .
n-1 = q+n-1 .
q+n-2 q/ 1 .
2 n-1 = p-1 / 1 .
p-2 / 2 .
p-1 / 3 ( n-1 ) ; supposing p ( = q+ n ) = tv + m+ n. In the very same manner , making q~ = q-w , and p~ ( =q~+n ) = p-w , it appears , that n/ 1 .
n+1 / 2 .
n+2 / 3 ( q-w ) = p~-1 / 1 .
p~-2 / 2 .
p~-3 / 3 ( n-1 ) , &amp;c. &amp;c. and consequently , that our whole given expression ( supposing p~ = p-2w , p~ = p-3w , &amp;c. ) will be transformed to + p-1 / 1 .
p-2 / 2 .
p-3 / 3 ( n-1 ) ~rm p~-1 / 1 .
p~-2 / 2 .
p~-3 / 3 ( n-1 ) ~nrm + p~-1 / 1 .
p~-2 / 2 .
p~-3 / 3 ( n-1 ) ~n/ 1 .
n-1 / 2rm p~-1 / 1 .
p~-2 / 2 .
p~-3 / 3 ( n-1 ) ~n/ 1 .
n-1 / 2 .
n-2 / 3rm &amp;c. &amp;c. Which expression is to be continued till the factors become nothing , or negative ; and which , when r= 1 , will be the very same with that exhibiting the number of chances for p points , precisely , with n dice , having each w faces : and in this case , where the chances for the errors in excess and defect are the same , the solution is the most simple it can be ; since , from the chances above determined , answering to the number p , precisely , the sum of the chances for all the inferior numbers ( inclusive ) may be readily obtained ; being given ( from the Method of Increments ) equal to +p/ 1 .
p-1 / 2 .
p-2 / 3 .
p-3 / 4 ( n ) p~ / 1 .
p~-1 / 2 .
p~-2 / 3 .
p~-3 / 4 ( n ) ~n + p~ / 1 .
p~-1 / 2 .
p~-2 / 3 .
p~-3 / 4 ( n ) ~n/ 1 .
n-1 / 2 p~ / 1 .
p~-1 / 2 .
p~-2 / 3 .
p~-3 / 4 ( n ) ~n 1 .
n-1 / 2 .
n-2 / 3 &amp;c. &amp;c. The difference between which and half ( w n ) , the sum of all the chances , ( which difference I shall denote by D ) , will consequently be the number of the chances whereby the errors in excess ( or in defect ) can fall within the given limit m : so that D/ 1/ 2 wn will be the true measure of the required probability , that the error , by taking the Mean of t observations , exceeds not the quantity m/ t , proposed .
To illustrate this by an example , from whence the utility of the method in practice , may clearly appear , it will be necessary , in the first place , to assign some number for v , expressing the limits of the the errors to which any observation is subject .
These limits , indeed , depend on the goodness of the instrument , and the skill of the observer ; but I shall suppose here , that every observation may be relied on to 5 seconds ; and that the chances for the several errors , -5~ , -4~ , -3~ , -2~ , -1~ , 0~ , +1~ , +2~ , +3~ , +4~ , +5~ , included within the limits thus assigned , are respectively proportional to the terms of the series 1 , 2 , 3 , 4 , 5 , 6 , 5 , 4 , 3 , 2 , 1 : which series seems much better adapted than if all the terms were to be equal , since it is highly reasonable to suppose , that the chances for the different errors decrease , as the errors themselves increase .
These particulars being premised , let it be now required to find , what the probability , or chance , for an error of 1 , 2 , 3 , 4 , or 5 seconds will be , when ( instead of relying on one ) the Mean of six observations is taken .
Here , then , v being =5 , and t=6 , we have n ( =2t ) =12 , w ( =v+1 ) =6 , and p ( =tv+n+m ) =42+m : but the value of m , if we first seek the chances whereby the error exceeds not I second , will be had from the equation m/ t=~ 1 ; where either sign may be used , but the negative one is the most commodious : from whence we have m ( = -t ) = -6 ; and therefore p=36 , p~=30 , p~=24 , p~=18 , &amp;c. which values being substituted in the general expression above determined , it will become 36/ 1.35/ 2.34/ 3 ( 12 ) 30/ 1.29/ 2.28/ 3 .
( 12 ) ~ 12 + 24/ 1.23/ 2.22/ 3 ( 12 ) ~ 66 18/ 1.17/ 2.16/ 3 ( 12 ) ~ 220 = 299576368 : and this subtracted from 1088391168 ( = ~~ 612 ) , leaves 788814800 , for the value of D corresponding .
Therefore the required probability , that the error , by taking the Mean of six observations , exceeds not a single second , will be truly measured by the fraction 788814800/ 1088391168 ; and consequently the odds will be as 788814800 to 299576368 , or as 2 2/ 3 to 1 , nearly .
But the proportion , or odds , when one single observation is relied on , is only as 16 to 20 , or as 8/ 10 to 1.To find , now , the probability , that the result comes within 2 seconds of the truth , let m/ t be made = 2 ; so shall m ( = 2t ) = 12 ; and therefore p=30 , p~=24 , p~=18 , &amp;c. And our general expression will here come out = 36079407 ; and consequently D= 1052311761 : whence 1052311761/ 1088391168 will be the true measure of the probability here sought ; and the odds , or proportion of the chances , will therefore be as 1052311761 to 36079407 , or as 29 to 1 , nearly .
But the proportion , or odds , when one single observation is relied on , is only as 2 to 1 : so that the chance , for an error exceeding 2 seconds , is not 1/ 10 part so great from the Mean of six , as from one single observation .
And it will be found , in the same manner , that the chance for an error exceeding 3 seconds , will not be 1/ 1000 part so great from the Mean of six , as from one single observation .
Upon the whole of of which it appears , that the taking of the Mean of a number of observations , greatly diminishes the chances for all the smaller errors , and cuts off almost all possibility of any great ones : which last consideration , alone , seems sufficient to recommend use of the method , not only to astronomers , but to all others concerned in making of experiments of any kind ( to which the above reasoning is equally applicable ) .
And the more observations or experiments there are made , the less will the conclusion be liable to err , provided they admit of being repeated under the same circumstances .
Other examples , and particulars might be added , in confirmation of what is here determined ; but as I would not appear tedious to your Lordship , I here conclude , who am , My Lord , Woolwich , March 4 , 1755 .
Your Lordship 's most obedient humble servant , T. Simpson .

