]\gt ; - On th of common .
The oltc is of increftsG , as exemplified ) of populntion , herc to become a one undel similnl circumstances , a ed bllsiness increases in to size .
is that of " " lnil i \ldquo ; olten called , period 01 a field ] er yields rentcr its ) ield thflll Most of the causes of thosc ( liHclcllces are not ) ) , ns discussed , be sified u under as are in olmost the } ) like vital subjc.cf to the colldition of nlcfm .
of of spollds , 110 doubt , of vital and , to bc to but it is .
their , and it to dity plied to deviatio1l It that iollS in exc.css be balallced by of in deticiency fore , if the former be cater thttll ] attcl .
be le .
, that is , bc .
This is impossiin cases , to which } ) statisno , so to tion arro , the is ) ) to ) it sscrts that the hosc h]ltorc do ble rncc , implics sibi 1 of of is Icss It is , otll ) of ] but of use , to thorou , hly of the some of ) octed 1ults may perl ) be ap parent ( first Vitlt t is the .
Donalcl the X[II .
' of \ldquo ; By DONALD ' by ived Octobcl 21 , lltluburcnlell .
may bc oximations to of a unknowll , or may refer to the several mcmber ; of .
D. McAlistcr on / meaSUl.ements will disagree among themselves , but on in order of they show a tendency to cluster round some vaIue .
We naturally inclined to infer that the true value of the ullknown , or typical nlember of the class , is not far from this value .
How to define and determine the appropriate in ious classes of measurement becomes thus a natural object of inquiry .
On we find that there is no strict and final criterion applicable to all cases .
We have to start with an rical ssunll)tion , more or less justifiable on general grounds , but not capable of rigid proof .
In the ordinary Theory of Errors , which deals primarily with discrcpant obscrvafions , the assumption made is reto this:\mdash ; that , on the whole , the best we can take is the mchc lfean of the discordant measures .
This is equivalent to statement that errors ( or tces from the truth ) of equal alnount in excess or defcct are equally likely to occur .
In the class of measures ferred to by .
Galton , which are of the nature of vatiollS , reason is iven for thinking that , on the whole , a bett value is likely to be obtained by the of the discrepant measures .
This is equivalent to the assumption that measures are equally likely which bear to the truth atio of equal lnlOtlnt in excess or defect , so to speak .
example , that measul.es which respectively and of oqually likely to occur .
This paper seeks to develop some consequences of fundamental principle .
outcome is a method for the trealment of a of sures w ltaturally group themselves round their Gcomc.tric Mean .
; method may be briefly presented as the bc ( 1 , bolic)logarithm of each measure , thus , say , / ' ; . . .
; where ] the Arithmotic Mean ( A.lf .
) of { he , so that ( A.M. ) metic lfean of the : this will be the of the comctric lIean ( G.M. ) of the oe 's , so hat If , we have in amcasul't of \ldquo ; of the which we call the " " ?
\ldquo ; be a ' the the eometric 1 tl , ' to it , scribed to Whcn ] is so eated , we can at once from it to ir-sel.ies .
, \amp ; c. , of as for the : and any term interpolated will ) of esponding term in 1 .
it ranted thnt the Mean of the mcasul .
CS is mosf of the quantity measured , it measurcs which differ from the lnenn 1 a bility tvbich less as the becomcs , eithcr in cxcess or defect .
llatulally scek some hich shall the respccti prob , aility to .
to ench llleasure of any given serics .
Thns , if ) one of the mensures , and be the Geometric llca ] of if to , such is tlbilit of , the curve will { he scale , will be the a .
to the first lnethod of Gauss Iotns , , it be that \mdash ; - ftlld constants .
is determined from the that ) sum of the ) ilitics of all the mcasures must bo depends rouping oasurcs , being they closely illld small when they nerftlly d is ] as sure orecision of the tltd is , or S0tncfimcs ' \ldquo ; .
( Sco fig. 1 .
) FIG. formino a ) be ) Mr. D. istel .
on [ Nov. 20 , .
to ally series of measures of which the mean and the weight are , we must suppose value of ( to be given ( either grapbior by table ) for all possible values of .
In other words , we must suppose the number of entries to be indefinitely reat and indefinitely close ether .
It is clear that the actual probability of any individual measure becomes in such a series infinitesimal , though the relative nitudes of any .
of probabilities remain unaltered .
This is indicated in the formula of S2 by the constant becoming infinitesimal .
To avoid- this inconveniencs we are led to modify our problem , and ask , not " " what is the probability of a measure " " but " " what the probability of a measure lying veen t very close limits and " " 4 .
the principle indicated above , , that on the whole there as mally measures less than the mean as there are greater the mean , we -ucc that the required probability is the constant factor results from the condition that the sunl of all the values of this function from to must be unity .
And it will be found that , should be the case .
The cocfficient of , I have ventured to from the prcvious function by .
it the " " of \ldquo ; function may bc defined in words " " the ratio the chance that a measure lies within given s1llall bears to the magnitude of the the cve represents the " " Law of Facility , ' ' find that its form resembles that of the curve of frequency , but the maximum ordinate occurs at in the former , and of coursc at in the latter .
The latter result implies the mcnll most probable value : the former means that there is chance of a measure lying within a given small fraction of .
than within the same small fraction of any obher value that can be named .
C. We graphically represent our results in 1 race t , ( tlIe mean taken as uIlity ) .
Lct the space between the cur and be the boulldal of a lalnina of rying density .
Take the along the of th .
ilIecm .
at to Then ordi is ) to robability of the of tion bonnded the bility of measure .
If dealing very , this mass ill bc to cscnt the mbor of lie bet ]lllss of the nlina i by the at , .
FIG. the thnt of cylindcr .
Lct it cnt by itnother 1in , at it .
The solid included ( the ) of 1 .
read for .
thus 1 the .
D. llcAlister on .
The law of cility alone will be epresented by the curve .
Here the area included vcen any two ordinates , the curve , and the axis of , the probability of a measure lying between tlIe COl.responding abscissae .
The area of the curve is bisected at the maximum ordinate is at ( fig. 4 ) .
FIG. 4 .
; ( Quetelct 's method .
) Suppose the sures we are to be a heries of cstimates of the magnitude of some object ( such as the depth of a givcn tint ) .
forming such an estimate mind is acted on by many small causes tending to mak-e us err either in cxccss or defect .
Suppose each cause acts in such a way as to makc us nnder- or over-estimate in a fixed ratio ; and as a particular let this fixed ratio be the for all the small causes .
If , then , the causes tend to over- and half to under-estimation , we nate rightly .
Our in any given case depends on the par- ticular combination of the causes which has ( so far as our ; conccrned ) fortuitously arisen .
Let , then , the number of uses be fixed ratio .
The total ]lumber of possible combinations will be .
The possible ratios are , . .
, .
chance of a corl.ect stimftte ( ratio 1 ) is by the theory of combinations the chance of estimate bearing a ratio to the truth will be .
Let now ?
become , and near to unity ; while remains finite .
Then , 187 th .
37 a approximation , the uItimately takes the .
If of a ratio \ldquo ; 1VC put , and thus find ; the chance required is then , corrcs o ' to the of ency .
Next .
tbc the shall lie ) narro limits.t ftnd , , find ve to nlultiply the exprcssion by , if takc , the cxexactly ) found for the la of facility .
This is , for it depends 011 suppositions , the secms light on the gcnesis and the of the odulus h S. method of exhibiting the by Mr. Galtoll Mcthod of Intercompaliso]l , is fhc following .
Lct the sures 1( by a series of ange these by sidc small tances n in ordel of itudc .
will then lic on clll of { lexure , Mr Galton calls an rive ; } may ) it as " " of Its equation bc to , constant the ordinates athey stand / .
D. McAlister on [ Nov. 20 , 9 .
We may next show that if the Arithmetic Mean ( A.M. ) of all the be formed then .
follows stl.ictly when the number of measures is indefinitely reat .
It still ves a good approximation to the value of when the number is considerable .
10 .
Among the reater than the mean there is one which may bc called middlcmost : i.e. , such that it is an even wager that a meRsure ( greater than the mean ) lies above it or below it .
A similar lniddlemost measure exists among those that are less than the mean .
As hese two measurcs , with the mean , divide the curve of facility four equal parts , I propose to call them the " " higher quartile\ldquo ; and the " " qucxrtile\ldquo ; respectively .
It will be seen that they respond to the ill-named " " probable errors\ldquo ; of the ordinary theory .
If be the quartiles we can show that that .
Thus .
Similarly between zero and lower quartile we place a mid-measul.e which we call the " " octc le.\ldquo ; The " " her ochle " " will subdivide interval between the higher quartile and infinity .
If , be octiles , we : before and It bc observed that the curve of ibution Q ordiIlates cquidislant the mean and the terminal ordinates .
is from and the asymptote .
The nalogues ested by mean error , and squftre , of the ordinary theory have no very practical for us .
It will be renlombered that they are to obviate the difficuIties arising from el'rors .
In problem tbcse somewhnt artificial have no special placo .
* qual .
is a .
The geomctric nlcan of greater tha)Jnca is a .
the the The of the ivcn 1 ) ] quartilcs and suflicc .
11 .
Let ible measures be givell )hosc respecve wcights arc , and let oduct ) formed .
] ) roduct will b dlible 1 like its factol .
S. It be ) roduct is jcct to of facility of the form lrendy oilled , and furthcr that , as in the theory , the w is such the square of is cqual to sulll of the reciproc of the the Iore generally , , .
bc whose . . .
, bc collltccted by the of ivcn b These ults a strict ; that } is closely and applies to whose is llot Let bo any function of me . . .
, each of may be the meau of some considcrable number of mcasures , so the to bc to the of the function from these fallible of its able to if be the question .
It cmains to S 1 .
ve is the } ) tho of ) ound their hnlctic m It , for examl ) ) , lUlnbcr of ) \ldquo ; be p.ptional to : is can bc cadily e means of ) dinnry t ctiovt .
htill .
\ldquo ; Mr. .
Chambers .
Note .
[ Nov. 7/ -herics corresponds to " " quartile\ldquo ; of the .
For , by ry theory ( probable error ) . .
prob .
error term which differs the Arithmcic Nlean by bable error in the -series is the .
of the quartile in the -series .
AVc infcr no ncw tables are necessary for the of nlethod .
We require only the tables of the error-function and of crbolic logarithms .
Of course common rithms may be , if we remember to introduce the appropl'iate modnlus into conclusion , I de.sirc to indebtedness Mr. llot only for the cstion of problem I have here pted to Hve , but also for many valuable practicnl ints in the XIV .
( Received Septcml ) Colaba , Bomb , I ret in my paper on the " " Diurnal of Wind , " " published , p. of the " " 0of Royal Society , " " I haYe inadvertcntly referred the Bermuda vilttl o ) ations to civil time illsteRd of to astronomicnl cction of this error shows that the } of the ] itics of the components at is durin otlle hours when the sun is hnu below inter omctric variations , on the also be reforred to astronomical of tim easolling of the paper remains the of the ) riations ( , tLosc hich ] poriod ) ffected .
F. of the Royal Society .

