]\gt ; ributions to the ?
of KARL PEARSON , Univcrsit'f Coltcge , London .
Communicated Professor H ober 1 [ PLATES 1\mdash ; 5 .
] CONTENTS .
I.\mdash ; On the Dissection of Asymmctrical Freqnency-Cves .
Theory , SS 1-8 .
71-8 .
) Examplc : Professor snrements of the " " Citd \ldquo ; of SS9-10 . . . . . . . .
lI.\mdash ; On the Dissection of Frequency-Curves .
General Theory , SS 11-12 Application .
" " No. 4 SS . . . . .
90-100 III.\mdash ; Invcstigation of an -Curve represcnting Mr. H. measurements of the lapace of .
SS 16-18 . . . . .
Table I. First Six Powers of First Thirty Natnlffi Numbcrs . . . . .
Table II .
dinntes of Normal . . . . . . . . .
107 Note added February 10 , 1894 . . . . .
107-110 I. the of -Cnrves .
( 1 .
) IF measurements be made of the same part or .
in hundred 01 thousand specimens of the same type or family , and a curve be constructed of which the abscissa represents the size of the and the ordinate the number of specimens falling within a definite small range Sx of organ , this curve be ermed a frequency-curve .
The centre or origin for surement of the orgall , if please , be taken at the of the specimens measured .
this c. as the frequency-curve may be looked upon as one in which the tbous , and or per ten thousand , as the case may be\mdash ; of a given small of mean , is plotted up to the mean of that .
Such part in the mathematical theory of evoIution , and have been ] dealt by Mr. F. GALTON , Professor WELDON , and others .
In most ses , in the of errors of observation , they have a fairly definibe shape one that * Symmetrical shapes may of coursc occur which are not of the normnl or -cnrve form .
Scc Parb II .
, S 11 of this paper .
9 .
ON MATICAL T EVOLUTION .
chcs close of tion to the well-known error or probability .
A , which , practical purposes , can be represented by for of paper be curve .
When a eJll c rise to a ) normal curve , we may probably assume serles lethi n. a condition ; here is ) ) and des truction the mean .
In the case of certain biological : , and relnents there is , however , a ) well-marked ation from this it becomes impol.tant to determine the direction and amount of ( lcviation .
The , asymmetry cffise f ct that the units grouped ' in the measured material are not really homogeneous .
It may happen that mixture of 2 , , . .
homogeneous groups , each of which deviates about its mean symmetrically and in manner represented with sufficient accuracy by the normnl curve .
abnormal frequency-curve may be really built up of normal curves havi parallel but not necessarily coincident axes and different parameters .
Even where the material is really homogeneous , but rives an abnormal frequency-curve the amount direction of the abnormality will be indicated if this frequency-curve can bc up normal .
The object of the present paper is to discuss the dissection of frequency-curves into normal curves .
The equations for the dissection of a -curve into normal curves can be written down in the same manner for the special case treated in this paper ; they require us only to calculate moments .
But the analytical difficulties , even for the case of so that it be questioned whether the general theory could ever be applied in to numerical case .
Therc are reasons , indeed , why the esolution into two is of special importance .
probably up first into two species , rather than three or more , owing to the pressure at 1 } !
iven time of some particular form of atural selection ; in attemptto an lbsolutely h material , we are less likely to have got a mlxtuye of ] or more groups of two only .
Lastly , even the eneity may be threefold or more , the dissection into two is likely us , nt t approximat , ion to the two chief groups .
In case of neous , with abnormal frequency-curve , dissection into two normal us the mount and direction of the chief ality .
So bc of the yalue of the special case dealt with here .
A the cases which may thoeoretically occur .
sible to find two components , and these two Unen b were uot , a } of the following kind PROF .
K. PE ARSON ON THE MATHEMATICAL THEORY OF EVOLUTION .
73 Between the six constan ts on either side of this equation an infinite variety of relations can be reached by giving an infinite variety of values , and it seems impossible to satisfy this series by the same set of ] of the constants .
For example , let be very great , and suppose to be the largest of all the quantities , and .
Dividing by and putting great we have whence , proceeding to the limit , unless or The first is impossible by hypothesis , therefore the latter must be true , say .
This gives us at once Returning to the original equation , and making large in it , we see that the first two terms become equal on either side .
Hence , the second two terms must become equal as approaches infinity , or Dividing again by , this leads in the same manner as before to , and , ultimately , to Our original equation may now be written ( ) .
Put , then the left-hand side vanishes and , accordingly , the right must vanish , but this involves either or Similarly , putting , we find that either or . . . . . . . . .
Thus , either the two sets of components are identical , or is true , PROI K. RSON ON TH } EVOLUTION .
, in successinn , and integrate the results respect i betwePll We find ( ) , aid of and to . . . . . . .
( ) , and reducing by aid of , ( ) , and to the two forms , . . . . .
S , . . . . .
Equations , ( ) , S , and four independent equations , which suffice to determine , as definite functions of , and .
But are in general ] ) endent of , and ; hence it follows that cannot in general be true , or we must have is , a curve which breaks up into can break np in one , and one way only .
Now it is clear that in actual statistic practice our abnormal frequency-curve will never be the absolutely true sum of two orlnal-curves ; indeed , if it be not a mixture , but an frequency-curve , it is not necessarily a very close approach to the SU1I1 requency -curves of ormal type , \mdash ; it be the limit to an binomial .
We must not , thel.efol.e , be surprised if more than one solution be iven by any method of dissection .
A mathematical criterion for discrimi ting t " " true , " " easily be given .
For example , in the method of the present , ) define the ' trne or at any rate the " " best solution which for a sixth nearest in value to that of the -curve .
a theoretical criterion , howeyer , may not have much of of nre given in S 5 of this , and permit or fhcbG infcgrations ] ) erformed fornl of the ] to ) mmetric is and LRVC olllv .
lways positive .
( htly given in ) of this \ldquo ; Roy .
Soc. Proc vol. 54 , ] PROF .
K. PL ARSON ON TH MATH 1VOLUTION .
75 practical value .
For after we have made the areas and first lnoments of curves identical , their sixth moments will in general ) like much closer together } either are to that of the curve of observations .
Added to the great labour in volved in the calculation of the sixth moment is sufficierlb to detcr the practical , if any other convenient , results of measurement on other organs\mdash ; suffices in the particular case to discriminate between f , be found .
Thus , while the mathematical solution should be unique , yeb from utilitarian standpoint we have to be content with a compoulld which fits the obseryations closely , and more than one such compound curve may arise .
All we call do is to adopt a nethod which minimizes the gences of the actual statistics a mathematically true npound .
The util i ) roblem is to find likely components of a curve which is not the true curve , and would only be the true cllrve had an infinite number of absolutely accurate uremen t. As there are different methods of fitting a normal curye to a series of obser y , depending on whether we start from the mean or the median , and proceed by " " rbiles , \ldquo ; meanl error or error of mean square , and as these methods lead in some cases to slightly different normal-curves , so various methods for breaking up an abnormal -cnrve lead to different results .
As from the utilital.ian standpoint good results for a simple normal curve are obtained by finding the mean from the first moment , and the of mean from the second moment , so it seems likely the investigation , based on the first five or six moments of the frequency-curve , lead to good results .
While a metbod of equating chosen ordinates of the given Ctll 've and those of the components leaves each equation based only on the measul'enlents of organs of one size .
the method of moments uses the given data in the case of equation for the unknowns , and errors in measurement will , thus , individually less influence .
At tho same time it would be of great interest to discover whetber other nJethods of dissection lead to results identical or nearly identical with the method of moments adopted by the present writer .
Any other lnethod analytically ossible h ot yet , however , to him ; criterion stin solutions so close as tbose of and 2 , ] dopted 1 Professor when he appeals to the measurelnents of a organ .
( 2 .
) In the case of a -curye whose curves , the complete solution ) ends in the method adopted in ding the roots of a of the ?
order .
It is possible that a simpler solution be found , but the method adopted has been chosen after trials each component normal curve has three riables : the positionu of its its ' ' standard-deviation\ldquo ; " " Mean Errol " " of bIeilIl illld ( iii .
) its .
Six relations between the iven f ' and curves would thel.efore suffice to determinc the six ns .
relations of ihis kind can be written down , but , unfcrtunately , the of thetl lead 76 PROF .
K. ARSON ON THE MATHEMATTCAL THEORy OF EVOLUTION .
exponential , the solution of seems more beyond the wit of man than of a numerical equation even of the ninth order .
( 3 .
) In nple the conditions will be sufficient to reduce the suitable roots of this equation very largely , possibly to two or even one .
These limiting conditions } be considered later .
A suitable root of this equation leads to a quadratic for the of the two component normal curves .
This quadratic is fundamental , and rs to highly sestive for the problem of evolution We have two cases : its roots are positive .
In this case the given frequency-curve is the of two normal curves .
The units of the frequency-curve may be considered as composed of definite proportions of two species , each of which stable about its mean .
The process of differentiation here appeal .
S complete .
( ii .
) root ?
positive negative .
The given frequency-curve is now the of two probability-curves .
The probability-curve , with positive area , may possibly now be looked upon as the birth- population ( unselectively dimin ished by death ) .
The negative probability-curve is a selective diminution of units about a certain mean ; that mean may , perhaps , be the average of the less " " fit It possible that in some umerical cases solutions of both the types and ( ii .
) will be found to but I imagine that in most cases of a well-marked and characteristic -curve , either only one type of solution will exist , or , if two types do exist , then one will give a much better agreement with the actual of the curve than the other .
That the two types of solutions should exist side by side is , perhaps , to be expected .
In such cases we have examples of groups , process of diHel.entiation into separate species by the selected mean .
( iii .
) From the nature of the problem , the case of both roots negative does not occur .
We to the bolution of the -curve to it , if possible , into two com-curves , .
into normal ( 4 ) finitions and Problems .
iven any curve ABC , and the line , if we take the sum of the products of every element of by ower of the dance of the element from the line , we form the moment of ) area the line , it ' / ' the strip parallel to its distance from , then the moment , the ation extendi en all over ABC , or from A to in our the is alwa .
; bounded straight line , , perpendicular to PROF .
K. ARSON ON THE IIATHEMATICAL OF EVOLUTION .
77 If be any standard , say 10 or 100 units , then the lnoment is of the order , if be the area of ABC .
It therefore equals , where is a purely numerical factor .
We shall invariably represent it as the product of these three factors .
( ii .
) Given the first moments about the coefficients to find the nth moment about or the coefficient Let the distance between / and be , then or \amp ; c. In particular , since ( 1 ) .
When the line passes through the centroid of the curve , and the curve is about are all .
Hence if in this case we take to the right of , or negative , ( 2 ) .
of the centroid of ABC is the ratio of its first moment to its , and To find the successive moments given curve about a given linle .
For the purposes of the present problem we require only the first five moments of ctlrve like ABC about line passing through its centroid .
The solution may be obtained either analytically or graphically according to the accuracy or rapidity with which we wish to work .
.-.Suppose the frequency-curve to be ined by plotting up the results of 1000 measurements , each unit of length along AC corresponding to an equal change in the deviation .
Starting from the point , beyond which no individual occurs , we may have in practice , perhaps , 20 to 30 equal ranges of deviations we reach the point , which terminates the deviations on the left .
The equal range being taken as the unit of length , let the numbers in the groups at 1 , 2 , 3 , 4 , 5 .
units of distance from be , / ?
Then moment clearly equals very appJoximately .
, or since , and may be couyeniently taken ( 3 ) .
Sufficien values can then be provided we 4 , know the ) the 1 ' 1 le ura nulnbers up to about 20 to 30 .
The values of these powers ) given later in this papel Knowlng , first five lonlCntS about ) evel.tical through find the centroid by aid of ( iii .
) , and then the momeuts about the vertical through the centroid by aid of tions ( .
following to detel.mnle the nlonlents :\mdash ; PROF .
K. PEARSON ON THE MATHF , MATICAL THEORY OF EVOLUTION .
79 ( 4 ) .
The centroid having been found , it may be asked : Why we should not calculate directly ?
The answer lies in the fact that the centroid }will not enerally coincide with a unit division on the deviation axis , and the powers to be calculated , instead of being those of two place ures , become in general powers of nambers containing three or four figures .
Thus the labour of bhe arithmetic is much increased .
.\mdash ; If the figure be drawn on a scale , the moments may be found with a fair degree of accuracy by aid of the process , which has long been of use in graphical statics for ding the first , second , and third moments of plane areas .
It is required to find the monlents about of curve ABC , bounded by the straight line .
Take parallel to and distance .
Take any line , first to from to A ; let the perpeudicular from on meet it , and let meet in ; let the perpendicular from on meet in and let in ; let the perpendicular from on meet it in and let meet in .
In this manner a series of points , are determined .
Let these points be determined .
a sel.ies of positions of taken at short intervals from to , then all the cesponding Q being joined , we obtain curves termed respectively the first , second , third , fourth , and fifth moment* The third moment of a plane area is used in .
graphicalIy of of a spindle about its axis .
The method described is sometimes ibuted to , but secms to have been long in use to find ' ' equivalent \ldquo ; in the case of beam scctions .
curves .
Let the areas \amp ; c. , be read off with a planimeter , and be Then ( 5 ) .
will construct these curves with great readiness , and if on a sufficiently large scale , the results may be read to ilhin the one per cent. error .
' ' Equations ( 4 ) then enable us to complete the problem of finding the moments about a line the centroid .
Or , the first moment being found about , and so the centroid ; we may shift till it passes through the centroid , and then proceed to find , . .
directly in the aboye manner .
In this case care will have to be taken in reading the areas of the -curves , which have now pieces of their areas , to carry the pla nimeter point , in the proper sense , round their contours .
( 5 .
) Properfies of the -curre .
Let the equation to the probabil i-curve be\mdash ; Then will be termed its ( error of mean square ) .
is the total nmnber of units measured , or the area of the probability curve .
To find the second and fourth moments of the prob -curye about bhe axis of Let them be and Then Clearly and are zero .
ator , G U calculated the first .
moments of a number statisti ical frequency-curves , with theob.ect of fi footnote , p. 74 ) .
Tho ethod is cticc , ftnd I hope soon to have an instrument them to generalized probability-curve ( see to construct these cves nlcchan]clly , designed by uary 9 , 1894 .
] PROF .
K. PEARSON ON THhl MATHEMATICAL THEORY OF EVOLUTION .
( ii .
) Now let be a standard area and a standard length .
Let us use Equations ( 2 ) of Art .
4 ( ii .
) , taking as the of sy mmetry of the probabilitycurve , and at a distance to the left , then\mdash ; Now let , and ) Then , and are purely numerical quantities , and we have for the first five moments round . . . . .
( 7 ) .
( 6 .
) We are now i11 a position to write down the equations ich give the general solution of our problem .
Let the deyiation-axis of the asymmetrical frequency-curve be taken as axis of , and let the axis of be a this axis ] the centroid of the frequency-curve .
Let this centroid the first five momentcoefficients about the axis of of ) efrequency-cul.ve , , be found either analytically or graphically by the methods rested in .
Then , if the position and magnitude of the coml ) onelltn ) curves be by the quantities , and , or thel ) , and '- , ?
we have , since moments round the vertical are1 , PROI K. PFAR STICAL THEORY OF ' The first here esents the uality of the areas of the resultant and its components .
to the simplest terms , we have the equations to find the si unknowlls , ( 8 ) .
( 9 ) .
( 10 ) .
( 11 ) .
( 12 ) .
( 13 ) .
give the complete solution of the problem .
After several trials , I find that the from these equations , and the determlnat on of equat giving ) tn lead to a resultiug equation of the lowest ( 7 .
) , have ( 15 ) .
* All m ) .
nttempts to ) in a simplcr ve failed .
ting of selected ordinates , or of selected pol.tion , of { , or of moments l.ound the axis of all { ppear to to exponential equations defying .
It is ossible , , that some other six of a less complex kind may tely PROF .
K. PEARSON ON THE OF EVOLUTION .
quations ( and ( 15 ) clearly the numbers in the component so soon as and are found .
Substituting these values of in ( 10 ) and ( t1 ) , we have two equations to determine and in of them we find .
( 17 ) .
These equations clearly give and efore , the standard-deviations of the component groups when are known .
brevity , put \mdash ; . . . . .
( I 8 ) , \mdash ; \mdash ; . . . . .
( 1 !
while and ( 13 ) we have We must now substitute ( 18 ) and ( 19 ) in ( 20 ) and ( 21 ) .
We find \ldquo ; whence , 84 K. ON ON MATlCAL TUTION .
Write ( 22 ) , and ( 23 ) , then , up , bove equations ( 24 ) , ( 25 ) .
these let us first find in of ] the first by and the secolld .
( 26 ) .
by and add to find .
' soon is ) then will be the oots uf th quadratic ( 28 ) .
, to ( 27 ) , substitute value of , and we have an equation the ) solution of the problem now turns .
This equation is ( 29 ) .
( 8 .
) may be on uation .
Since this equation is of an odd order , one root } ) found .
Further , remembering that and we in the case of a normj curve , for which , while .
and the coefficients of the above equation of the nintb order villish except the first .
Thus 1 should thy e ) will bc zero .
Accordingly , since , with hynnnetl } , ) coefficients become small , it will be needful to work their values out to a greater degree of exactness the ' the degree of asymmetry .
PROF .
K. PEARSON ON THE MATICAL OF .
85 Given that a frequency-curve is compounded of two normal curves , equations ( 29 ) , ( 28 ) , ( 27 ) , ( 14 ) , ( 15 ) , ( 16 ) , and ( 17 ) form the complete solution of the We may throw the whole solution into the following form : Stage I.\mdash ; Find the centroid of the frequency-curve and calculate and Stage II.\mdash ; Solve ( 29 ) for find the corresponding values of from ( 27 ) .
\mdash ; Find the positions of the axes of the component normal curves from ( 28 ) .
Stage IV.\mdash ; The fractions and that the areas of the curves are of the area of the frequency-curve are the roots of the quadratic : \mdash ; \mdash ; . .
( 30 ) .
Stage V.\mdash ; Since ) and the standard-deviations are given at once on substituting in ( 18 ) and ( 19 ) .
( 9 .
) The whole metbod may be illustrated by the following numerical example:\mdash ; Breadth of " " \ldquo ; of .\mdash ; Professor W. F. R. WELDON kind ] given me the following statistics from among his measul.ements on .
They are 1000 individuals from ) .
The abscissie of the are the of bead\ldquo ; to body-lengtl ] , and one unit of abscissa of body-length .
No. 1 of the .bscissae corresponds to of body-length .
The ordinates epresent the umbel of individual crabs corresponding to each set of ratios of to bodylength .
Thus there was one fell into the range , three fell into the range , five , into the range , and so on .
The average length of an imals measured 35 millims. , and measurements were recol.ded to 1 millim. Abscissae .
Ordinates .
Abscissae .
Ordinates .
Abscissae .
Ordinates .
Abscissae .
Ordinates .
1:$ This curve is plotted out as the dark continuous line in Plate 1 , fig. 1 , and is asymmetrical .
1 proceeded to calculate its first five moments in the analytical nlethod suggested on p. 78 , each calculation being made twice independently .
I took , and .
The moments were taken about the vertical through the point , and were calculated by the aid of Table I. of the powers of the first 30 natural at the end of this memoir .
The following results were obtained:\mdash ; 304.923 , since , is clearly the distance of the centroid vertical of the frequencycurve the .
The moments about this centroid vertical now calculated by aid of ( 1 ) , p. 77 .
There resulted:\mdash ; 53.874,770 1576.533,413 85.205,407 where ur given in ternls of the 's by ( 22 ) of p. 84 .
now to the fundamental nonic ( 29 ) , let it be divided by 24 , and written in the form Then the coefficients were calculated ca cu ate , and the following values found:\mdash ; 4,353.742 PROF .
K. PEARSON ON THE MATHEMATICAL THEORY OF EyOLUTION .
Put and divide by we then have for the fimdamental nonic the followinos equation , where on lie three decimal places are retained:\mdash ; After a somewhat laborious calculation , the values of functions , , , , .
, , .
, .
( were ascertained ) the oWing results:\mdash ; 3 changes .
6 changes .
Thus there are real roots .
These three real roots were then localized as follows:\mdash ; Two roots between and and One root between and 1 , As successive approximations , I found:\mdash ; To : : : 5 , 40 , 422 , With sufficient accuracy we may then take for the values of 1st solution , 2nd , , , , 4.170 .
of solutio ) was calculated from ( 27 ) on thenu found .
resulted : The quadratic for are here identical ) ( the distances of the centroid of the celJtroid v of the , is : \mdash ; ence The ' of thus the ] ) of indiyiduals in either group are respectively The values of the standard-deyiations , and , were now determined from ( 18 ) and ( 19 ) , where , since , and 2 At the same time the ximum ordinates of the component probability-curves , and , were found from There resulted 4 .
3.1154 .
the 1 st solubion may be summed up lows : 1st Coml)oncnl .
3 .
4 .
'2nd Component .
2.490 .
3.1154 .
74.976 .
These two curves by aid of the II .
, which was calculated for ) ul.pose from the exponential These curVes are plotted out in fig. 1 , and their dded together give the resultant curve .
It be seen that curve is in ) rkably close eement with the original asymmetrical frequencycurve , an eenlelJt ( close as we could sonably expect from the commit convcuicut to ) the standard-deviation th with the ) robable or modnl , in of which the unction is usually { ttbulated .
PROF .
K. ARSON ON THE MATHEMATICAL THEORY OF UTION .
parative smallness of the number of indiyiduals dealt with , and the fact that the observation-curve can at best only be an approximation to the true resultant .
2nd Solution.\mdash ; Precisely similar calculations were undertaken for the value , and it will , be sufficient to cite the final conclusions here .
Quadratic for 1st Componen 2.769 , 2.878 , 64.764 , 2nd Component .
4.7702 .
44.559 .
These component-curves are drawn in 2 , and their ordinates added togethel We see that we have again broken up our asymmetrical frequency-curve into two probability-curves , whose sum is a very close approximation to the original curve .
3 Solution : While the first two solutions have been additive , this solution makes and of the same sign , or the centroids of the component curves fall both on the same side of the centroid vertical of the -curye .
Accordingly the area of one of them must be negative , and the solution promised to be a subtractive one , i.e. , to represent the frequency-curve as the difference of two normal curves .
Determining and then from ( 27 ) , we find ; hence The roots of this equation are , howeyer , imaginary .
In the case of ' foreheads , therefore , we cannot represent the frequency-curve for their forehead lengtlls as the difference of two normal curyes .
( 10 .
) So far as the nonic is concerned , our work is now accomplished .
the biologist 's measurements and assuming them to be the chance distribution of two unequal groups about two different means , then other of our solutions is the correct answer .
Applying the test of .
the sixth moment , we find for the observations , while for the first solution it is 188,099 and for second solution 192,446 .
According to this test , the first solution is the lired one , but , as we have noticed , the two are themselves mnch closer ) to * The theory of tjon will here , perllaps , confirm It .
tclls the first and not the second solution is in good dance 1 his other nlctsureme]ts .
PR ATlSON ON OF VOLUTION , observations ( see 75 ) .
fact , contours of the compound-curve for both solutions are very close together , and neither differs more from the observations than most ormai curves symmetrical frequency-curves statistical measurenents of this kind .
, ftre so close that , notwithstanding we have demonstrated a theoretical elless f solution of the problem , we see from the of practical statistics , it is possible for the given material to be broken up into more than o1le pair of normal curves .
Thus the problem indeed becomes somewhat arbitrary\mdash ; at any rate till the metry of the frequency-curve becomes much more han is the case with that of the foreheads of Naples crabs .
Indeed , although method adopted leads to only two solutions , it is quite possible that pairs of component normal curves might be tentatively found lying in the neighbourhood of those determined by the above solutions , which would give resultant-curves fairly close to he -curve .
Professor WELDON had , indeed , found by repeated trials one such solution , but this solution differs widely in the third and higher lnoments from the ) tions ; it cannot , thel.efore , be con sidered to have the same those given by the present theory .
Granted that the original observations represent a mixture of two species varying about their mean according to exact normal curves , our method gives solutions , and two only .
Without correlated measure1nents , it be difficult to discrimina between these solutions\mdash ; at any rate from the standpoint of practical statistics .
The perhaps over-fine theoretical test of the sixth moment decides for the first solution .
II.\mdash ; he of Symmefrical Fr-Curves .
( 11 .
) Another tant case of the dissection of a frequency-curve can arise , when the frequency-curve , without being asymmetrical , still consists of the sum or difference of two , i.e. , when the means obout which the component groups are distributed are identical .
Thi case is all the more and important , as it is not unlikely to occur in statistical investigations , and the symmetry of the frequency-curve is then in itscl likely to the to believe that he is dealing with an lple of the normal fi.equeltcy-curve .
It seems to me that without very strong grounds for belief in the reneity of any statistical material , we not to be tisfied 1 its representation by the ordinary normal curve , simply because our results are and fit the normal curve fairly well .
We ought first to ascertain whether or not they would fit still better the sum or difference of two This , at any rate , is a first to demonstrating the homogeneity of onr possibly our test for two may , not because our material is , but because its is multiple ather t double .
* Rrise in the CiSG of compound fi.equency-curves , without identity of the cause of his case , for components we should have for different means , PROF .
K. PEARSON ( -lEORy ) LVOLUTION .
We will now modify the ults of our previous investigation to suit th case of :asymmetrica ] frequency-cnrve which has arisen fronn the superposition of two normalcurves having the same axis .
In } case if we nnite , , in Equations ( 8 ) to ( 13 ) we have ( 11 ) and ( 13 ) identically satisfied , and ( 8 ) , ( 10 ) , ) become . . . . .
, eqnality of component group-totals and of thei ] ' tions .
cquality lcbf ; likoIy than equality of means and divergence of totals and stand deviatioltS .
old it exist , ever , we fall back on a sub-case of the general case we already dealt with .
We need only , in Equ tioltS ( 8 ) , and we have whence or , sibiliby of the solution cleal .
lie depcnds on being } The following is an oxample of this case .
Mr. gives soulc ults f practice , on ) of his Text Book on Lcast Squares .
He does not seem to hnvo ticed t } the resulting-curve } far from a .
I find for these obsel.vations 6 .
rfhc of general symmG lssuming then that hots fired two ) ) with eqnal precision , I find and b almost exactly .
We have accordingly the 1000 hots as a wholo ] Allowing a uniform of deScctivo to 482 , we lind compoulld-culc fitting cIosely Mr. MhRItllIAN'S figure , and indicating that the gull was aimed 'he frcs of and 7 , and not at th of 6 .
Six was possibly white , esulls of would is a change of bout m .
K. ON ON OltY O . . . .
( 32 ) , . . . .
( 33 ) .
Clearly we , more equation .
At first sight it might seem a fourth would come eadil .
fro ] ] ] the fact that the mid-ordinate of the frequencycurve is the sum of the mid-ordinates of the component probability-curves .
leads to or ( 34 ) , if the disadvantage of .
our solution back 011 the correctness with which ] have observed surements of one size only , namely , the mean , the result of eliminating veen ( leads to equation of the eighth order .
To avoid this , it seems easier , as well as more accurate , to take as the fourth equation that , obtained from th sixth moment .
Lct the moment of the given frequeIlcy-curve about its axis of symmetry , then or , . . . . .
( 35 ) .
solution , ) , ) , and is easy .
Elimig we have , whence all the ations .
, if , be the 2 , momellb of a bout ibs axis PROF .
K. PEARSON ON THL TICAL THEORY OF UTION .
Tilus The two roots of this atic are clearly and , so that the complete solution is where and ) are roots of , . . .
( 12 .
) Now we may note seyeral general points about these equations .
Let be the greater root , then if lie between ) and and are both positive , or the frequency-curye is the sum of two normal curves .
( ii .
) is positive and negative , or the greater component ' is positive , we have then a real fference solution .
( iii .
) ) is negative and is positive , again the greater component group is positive , or we have a real diJerence solution .
Obviously if and ) , the coeflicients of the quadratic ( 36 ) all become zero , but these are just the conditions which would be satisfied if the frequency-curve were a true normal curve .
This gives for all practical purposes a very sufficient test of whether a given ical frequency-curve is a true normal curve If be not equal to , and be not equal to ) , then we to a symmetrical -curre refers to ogeneous m. We must then investigate whether a better result cannot be obtain ed by treating it as two superposed normal curves having the same axis .
The quantities and I propose to call the excess and defect of the frequency-curve , The excess sures the excess of one-third of the moment ovel the square of the second moment ; the defect measures the defect of the fourth moment from one-fifth the ratio of the sixth moment to the second moment .
Here " " excess\ldquo ; ' ' defect\ldquo ; are used in the algebraic sense , and may take eithel sign .
They } ) to be introduction of tho into botb cxcos .
is to tYuished f absolute measul t of ivergcnce .
for ) oses of the divergence of given symmetrical frequencylormal type .
We may the quadratic ( 36 ) in terlns of and analyze the .
to aracter of the excess and defect .
The This gives ( 37 ) .
We have the following cases : both .
Then the values of n ) are both real , but they must be both , otherwise and would not be real .
It is necessary , therethat .
( ii .
) both llegative .
Then ?
/ ) will be if , when does not lie between If then must , in orciel that 10 have both values positive , we must or lattel ' condition is clearly satisfied if the other PROF .
K. PEARSON ON THHi MATH MATICAL THEORy OF VOLUT I. 95 it is easy to see is less than Hence , our final then ; but if then either or it must lie between and positive and negative ; if the values of are real , one must be negative , and therefore the solution impossible .
( iv .
) negative and positive ; if the values of are real , one must be negative , and therefore the solution impossible .
Thus we conclude : If the excess and defect are not zero , the frequency-curye , although symmetrical , is not normal .
If the excess and defect are of opposite signs , then the frequency-curve cannot be broken up into the sum or difference of two normal curves with connnon axis .
The frequency-curve , if compounded of llormal-curves at all , is of a higher and more complex character .
If the excess and defect are of the same sign , then , provided certain relations hold between the values of the excess and defect given in and ( ii .
) above , there is a real solution of the equation which resolves the frequency-curve into two components .
( 13 .
) I propose to illustrate this discussion by the consideration of a nu1nerical example .
Professor WELDON has kindly complied with my request for the numerical details of the most symmetrical curve deduced from his measurements of ples crabs by placing the following statistics for a shell measurement\mdash ; No. 4 of his sel.ies --at my disposal .
The resultant-curve and the corres ondin c. normal pictured in fig. 3 ( Place 3 ) .
Clearly , from ordioary statistician 's standpoint , could not expect a more symmetrical , or a closer graphical agreement , the normal curve .
But is this a real or merely an apparent agreement ?
The .
is , as we shall see , vital for the to be put on Professor WELDON'S results .
PTIOI K. ]IiSON ON TH MATH OF EVOLUTION .
No. 4 .
of Crat ) Ordi 1 unit .
r ) Ordi 1 unit .
r ) Ordi 1 unit .
r ) Ordi 1 unit .
r ) Ordi 1 unit .
r ) Ordi 1 unit .
r ) Ordi 1 unit .
r ) Ordi 1 unit .
r ) Ordi 1 unit .
r ) Ordi 1 unit .
r ) Ordi 1 unit .
r ) Ordinates Absciss ; c. ( 1 unit crab ) .
126 1 : ; 72 41 28 8 I7 7 2 The first six moments were calculated exactly as in the previons case of S 9 , by aid of I. , xcept that now equals ) !
we go a stage further to and equals unity before .
We have 0 .
7.5092 3.4751 These results give for the ) osition of the centroid , and the devlation gives the ) odulus , , and the central nate of the normal curve .
The mo , as calculated from the mean error , , so that the reement is very close .
The normal curve in fig. 3 is con .
the values by aid of Table II .
additional quantities were now calculated:\mdash ; lf we had a ) curve , ) , and to be .
This , .
conrse , we should not expect in any actual set of observations but the smallness of metry of in these esults .
' , shows a very fair approximation to PRObl .
K. PEARSON ON MATICAL T OF UTTON .
97 Since ) , we see that the roots ( 37 ) of our p. 94 cTre both positive , accordingly it is possible to break up the observation-curve into two normal curves wit , coincident axes .
Calculating the two values of we have whence from 1 ) .
93 : , or For all practical purposes the second group gives the normal curve of the set of observations ; that a half per cent. of Crabs been removed by selection about the same is not large enough to be significant in measurements of the kind we are here dealing with .
So far , , we that No. 4 of Professor WELDON'S measurements cannot be treated as the sum or erence of two normal curves having their axes coincident with ally substantial improvement on the normal curve peculiar to the original group .
( 14 .
) Hitherto we have used " " Crab Measurements No. 4 " " to illustrate he dissection of symmetrical frequency-curves , but a little consideration shows at once that this judging of symmetry by the eye is very likely to be fallacious , and No. 4 may , after all , break up into two normal curves with non-coincident axes .
Should these two curves correspond to practically the same groups as in the case of the " " Fore- heads then we shall have demonstrated that the asyInmetry of that frequency-curve is in all probability due to a mixture of two families in the Naples Crabs and a result of differentiation going on one homogeneous species .
The parent symmetry of No. 4 weighs nothing in the as may be readily tested by ogeth e two normal curves with not widely divergent axes or What we have been investigating , , in is in which the method of our first investigation old .
to ) the axes of the component llormal coincidence is I , therefore , proceeded to form the nonic for result values of , and already given .
The nonic being * The nearest whole number is herc taken The arithmetic throughout was of colllsc of lost 1ious c ] K. ON THEM ATICA I OF EVOLUTIO ents 977 .
I 2.281,243 , .
, we for the nonic and its first derived the following cxpressions\mdash ; .
nd , S05 .
The functions were now , and with the following results\mdash ; . . .
llonic h one root of between and and no roots between nd .
Tn it inary roots and only 1 real one .
by the PROF .
K. RSON ON ATHEMATICAL THLORY EVOLUTION .
99 This root was now localized .
Putting { he nonic , I easily found to lie between and 1 , then betweenl 15 and 16 , and by succession of approximations to be , and finally15326 .
Thus was then ascertained from equation ( 27 ) of 1 ) .
84 , and finally found to be .
The quadrati for then : which has both its roots imaginary .
Thus , considerably to my surprise , but greatly tu my ion , it that there is solution whatever of the of up the of No. 4 measurements into two normal components .
All nine roots of the fundamental to of ) robtem .
The best and most accurate representation of No. 4 the curve of The result of this investigation seems to most important .
material is homogeneous , and the asymmetry of the " " forehead\ldquo ; clllve points to differentiation in organ , and not to a mixture of two been dredged up .
On the other hand , I think that for the problem of cvolution the dissection of the most symmetrical curve given by the measurements is unnecessary .
will always be the problem : Is the material homogeneous and a true eyolution going on , or is the material a mixture ?
To throw the solution on the judgment of the eye in examining the graphical results is , I fee ] certain , futile .
Whenever in measuring a series of organs the an curve , we must accordingly proceed as follows :\mdash ; .
\mdash ; Break up this asymmetrical curve into components ; if there ] solutions , the theory of correlation or the test of the sixth moment will , enable us to say which is the most ctory .
( ii).\mdash ; Endeavour to break up the most curve ; if it cannot ) broken up , either into normal components with non-coincident or normal ponents coincident axes , the material is homoo e ' and curve points to a true differentiation in he organ to which it refers .
If , on tho other the most frequency-curye does break up , then if the in its component groups be the same or the same as in those ) .
to the curve : we are really with a mixture of lmaterial , and we shadl have ascertained the proportions of the mixtnre .
If the should not be the same , then we cannot assert that we , but we found a case of differentiation in both allS at the same time .
found a double-humped fi.equency- ( for the ) of ) of to the treatment of this portion of the of tion . .
to No. 4 no real negative the ) ' , that near its ompared walues obetween 0 it ched : aroot , and at .
llee then it were , a tendency towar to llle }yhether this tendency " " in any way corresponded to were diHerentiated .
I therefore nvesti gated root of the first derived function of the nonl lying about , and found it to be .
Thi led to equation ( 27 ) being , whence Whence ] or the nuln bers in the two groups are 971 and Clearly cveu eltcy to a roo no way to a root\ldquo ; in no wa fits either solution of the ' ' forehead\ldquo ; se , No. 4 measnrements ) either break up , nor have they even tendencv to break up , in the same manner as the eheads.\ldquo ; Since the nonic have " " tendency\ldquo ; to two real roots at time , we may note that the other root to which it may be said to tend , for which is a minimun ] , lies veen - is just insignificant as iuyestigated above .
We llay that only is the of No. 4 homogeneous , but it has not even a tendency\ldquo ; vards heterogeneity .
111 .
The object of the sent 1 solely to a general method to normal types , not a icalo investigation , it Hic !
to , yhen the rules for the reduction of symmetrical beel and illustrated .
But it must be remembered th ' the solution of a nonic , that the variety presented curvo for the claspers of rwjgs .
tiou ements of uothcr o , it SCCIUH impossible to say whether the { hc of tCCS , Mr. earwigs were our thods of ould the belonging of to each of tho ) PROF .
K. PARSON ON MATHLMATICAL THEORY ]VOLUTION .
101 by the roots of this equation suggests very considerable divergences and peculiarities as likely to arise , when a number of frequency-curves are dealt with .
The discussion of the case of Crabs must nob be taken as indicating that the incidents of case will be generally true for other groups of biological measurements , until a reat variety of such groups of measurements have been mathematically analyzed .
In order to throw more light on the general question , I have added the following analysis for the case of Prawns , the measurements for which were kindly placed my disposal by Mr. H. THOMPSON , who has been making elaborate lleasnrements of 1,000 specimens in the Laboratory of iversity College , London .
.\mdash ; Measurements in 998 ?
specimens dult ) from penultimate to hindmost tooth on the carapace .
Number of length .
4 : ; 117 The novel and somewhat remarkable feature in these results are the " " and 69 .
To neglect these giants , as in some anolnalous , , no doubt be convenient , so far as the analysis is concerned , would lead simpler cdtlctiu of the .
They have , however , been retain ed as among the data given to and their presence affords an interesting illustration of the various ' may arise in the solution of the fundamental nonic .
( 17 .
) The curve ( see fig. 4 ) by the observed numbers will be once en PROF .
K. RSON ON TH1 THEORY OF VOLUTION be distinctly asymmetrical .
Adopting the ) length 31 as the origin of coordinates , using the same notation as before , we have the fo results ' 94 , The standard-deviatio of the group as a is given The obtained from . . .
directly . . . .
, , ( In the case of the " " foreheads\ldquo ; of Crabs , the mean error from was , and directly .
This divergence between the error , as found practically from second first moments , is a very good test of the asymmetry of the frequencycurve .
In the very measurements of ' ' Crabs No. 4 the modulus , as calculated from the standard-deviation and from tlJe m error , had the near values and ) The curve obtained from the observations as a single group ( i.e. , and given in fig. 4 ( Plate 4 ) .
Taking we for the nonic nd its first differential \mdash ; 3773.575 , -- , I C3 , 34 .
34 .
*Theso to of accuracy than in the case of the Crabs , a csult , of this to change in the value of , of , not iu , but in thu senso of ally meau error , PROF .
K. PEARSON ON THE ] ICAL THEORY EVOLUTION .
103 Clearly there is only one positive root .
This was found to ) This gave whence I found 9.669,970 .
Consequently the roots of vere imaginary and no solution involving the ference of two normal components was possible .
The next stage was to find the negative roots .
These were easily demonstrated to lie between and 1 , and then it was shown that the value of only changed sign twice between these values .
Thus the nonic was proved , without functions , to have only three real roots .
The two negative roots and These roots lead to the follo solutions:\mdash ; ( A. ) Firs itive .
for of , 997,856 , 1st Component .
995,860 , 057,6086 , 3.5595 , 002,144 .
2nd 2.140 , ( B. ) Second additive Solufion for of ) 5.163 , 5907 , 147,3614 , 973,0024 , .
104 PROF .
K. RSON ON bl THEORY OF EVOLUTION Component .
1st omponent 5.310,9521 , , 8 .
1.203 , 280 .
To these tions we ( C. ) observat ons .
16.19 3.7572 , ( D. ) ofCurrc dednced by two " " iants\ldquo ; from , 3.6051 , The corresponding to ( A ) , ( B ) , ( C ) , and ( D ) as well as the observationctlrve are give in figs. 4 and 5 , and I shall now proceed to discuss several important points with to them .
( 18 .
) The first point to be noted is the existence of bhe dwarf , , and the giants , carapaces 65 and 69 .
The normal curve has a , and the mean carapace being bout 4 , we have no less than three measurements by more than four times the standard-deviation from the mean ; two of them , indeed , differ by nearly six standard-deviation from mean .
We might expect three such four times t st -deviation to occur in the measurement of .50,000 , they are mely i ) robable in the measurement of 1000 prawns .
occur in measurement of 1000 Prawns , with a deviation six times the , is so ) that it ought to lead us to reject the curve representation of the surements .
We are either dealing mixed of , or possibly there are few deformed dividuals normal tion There is anothe ] point , ] ) vever , in which the normal , based on the total *I exeIn dc { ] ) of { thy ious c of measurcment , having Gason to believe in the great IIich the were made . .
K. ON MATl COF .
1 observations , diverges considerably from the observational result , namely ( see , in the defect of carapaces about 45 .
This defect largely to the appearance of the curye .
I felt very confident that by ectin g the eccentric group of " " \ldquo ; I could find two components , whose resultant vould fit the curve of ) servation as closely the resultant-curves for the case of the } of Crabs .
I was peculiarly interested , however , in whether the method of olution by aid of the nonic would pay more attention to the giants or to the less improbable of individuals about 45 .
I even ined that out of the nine possible ions some might be solutions for the giants some for the 45 defect .
As a matbel of , the two solutions hich have any meaning are entirely taken up with the very improbable outlying eccelJtricities of the observations .
These eccentricities must first be removed from ) observations before the method will be of service in resolving the asymlnetry of the bulk of the observation-curve .
The method in which the nonic deals with the abnormalities is very ] islic , and I venture to think highly suggestive .
In fig. 4 the tlJe two giants is .
It fits the observation-curve , its far go , slightly better than the true normal curve .
But the solution of the nonic tells us not to absolutely reject the .
It gives us two components , the first of which fits the obseryations slightly better than the normal ( iants excluded ) .
It has practically the same area as compared with 996 ) , a slightly less standard-deviation as with , consequently all increased maximum ordinate .
This , with if slirrbtly shifted axis , ves a somewhat better fit .
In addition to this first componenlt we have a component with an area of , and a mean of 70 for the component corresponds closely to the two giants with a meall of 67 .
It ] , however , all standard-deviation .
Clearly the addition of to the f component , if distributed really , could make no sensible change in its ) and we may then sum up the first solution of the nonic in the words:\mdash ; It does not absolutely reject the two giants , but places all imaginary distribution of in their neighbourhood , and thus obtains for the other component itnd the resultant-curve ( which must be practically identical with it ) a better approach to the obseryation-curve than if the giants had been rejected .
It would appear , therefore , that our method of dissection , by of small components with imaginary distributions , a means of obtaining better than by simply rejecting ( or , perhaps , even weighting ) anomalous observations .
The second method by which nonic to account the icities o these carapace Ineasurements , is by mixing a small } ) opulation obont ) ercent .
giants with the normal population .
These giants have mean ) of , while the rest of the population has a meall of .
This population ) , has a very large standard-deviation : with the } ) ' .
K. ARSON ON THE MATIC A THEORY EVOLUTION .
107 TABLE II.\mdash ; Ordinates of Normal Curye .
eviation .
Deviation .
Frequency .
maximum Frequency .
1 9950 9560 0 .
9231 0 .
0 .
7827 0 .
7262 6670 5467 4286 3246 2780 2357 1645 0889 0340 0198 0015 [ NOTE , added February 10 , 1894 .
The importance of breaking up asymmetrical frequency-curves into normal components has been nized for a long time by anthropologists and ists .
Attempts solution have been made by R. LIVI , statura degli ' Firenze , 1883 ( see also Archirio per lologia , ' vol. 13 , irenze , 1883 , and ' Statistica , ' vol. 8 , , pp. 119-56 .
Also by O. AMMON in his recent work 'Die atiirlt t A nslcse Menschen , ' Jena , 1893 .
These attempts can hardly be looked upon as serious .
Professor LEXIS and Dr. have pointed out that the curve of deaths for year for 1000 persons born in the year\mdash ; the true mortalitycurve\mdash ; is also in all probability a compound curve .
Since writing the above memoir I have succeeded in resolving this mortality-curve into components which are not , however , all of normal , but ) econ e we ile mortality , of the skew ( see p. 74 above ) .
O. , in the volume cited above , to demonstrate tion in the length-breadth index of the of South-Germans since .
1Ie does this by comparison of the index as obtained from measurelnents on kulIs from the Row-Graves and on modern skulls .
He has not , however , noticed the frequency-curve for Row-Grave skulls is } ) ? .
I breaking it up into two components , one of ] coincides in and standard-deviation with the frequency-curve for the skulls of modern } ) HEORY OVOLUTIO N , the contain a mixed population , one element of corresponds closely to the mo ern ou dern ] ) ) ulation .
statelnent , ' evolution has taken place in this particular skull index appears to to .
The whole of the compound natnre of sknll frequencycuryes , both in Germany , is a very and fficult one , and I do wish present to anticipate esnlts , which I bope when my stigations to as a whole .
The above may suffice to indicate range of to whjch a resolution of -curves into normal components be applied .
( 2 .
) With regard to the method dopted in the .
itself , I am very conscious of defects under which it suffers\mdash ; the laborious of the arithmetic olved , and the question of what may be the probable error of the solution obtained the llethod of higher .
But I had to deal with the that the problem is hich t tly needed a solution in the case of both economic and biological statistics .
Better solutions than mine may be ultimately found , but although more hematically trained statistician for some tecognized the imporof problen ] , no solution , so I hitherto been forthcoming .
With to the amount of error introduced by the use of higher moluents , word be have not been able to work out the general problem suggested to me by I'rofessor DARWIN : " " Given probable error of every ordinate of -curve , what are the probable errors of the elements of the two normal into hich it may be -sected ?
\ldquo ; I vever , indicate the sort of differences which are likely to occur in on or on low moments .
uppose the distribution of an organ in a group ctually does follow a norni frequency-curve .
Then it is obvious that in of these animals at random and their organs , an error of the in the fi.equency of an organ of a given size is more likely to occul in it ' the mean than in a size the mean .
Now a low moment pays attention than high moment to an error in the frequency near the mean and less ttelltion t a moment to one off In other a frequencymolncnts fits best near the centre ; one calculated from fits neal the tails of the observation-curve .
problem is the owing : error in frequency near the tail is not as probable as an in equel the ; but if it does occur a high moment pay much to it low lnoment ; the othel hand , the low moment attention than the high moment to nor probable rors in frequency .
tendency on the whole will prevail ?
to the result in the foot-note , p. 92 , we have .
the 2 . . .
5 . . .
K. ON [ CAL yOLU ON .
109 Now , ] error Soccur in the quency y ) cling ' to Jet be the of , when calculated from ; then by the S/ S 5.3.1 , Comparing this with the error arising in calculating second moment in the usual manner , we have When is small wil ] be with S , and the has a great advantage .
This is ined until for the fourth moment , for the sixth moment .
But the probability of an organ is less in ] of about 2 in the 100 .
Hence we may take it that errors for ich the 4th noments giye a worse result than the 2nd momen for are , while el.rors for which they give a much better result than the 2 , moment are very ] ) .
Take , however , practically the worst case , an error occurring in the frequency ] organ sponding to , an error on likely to occur about three times in thousand erl.ors supposing errors distributed as normal fi.equencies .
We find The errors from the fourth and sixth moments are thus only and ) errors from the second moment , but errors from ] the second Olel er than 6 and 45 are giyen whenever is less than , or in more ) .
cent. of cases .
It would thus appear an error which will ) ) ] as compared with a low moment is emely rare ; ilc , on the contrary , errors which put a low moment at great disadvantage high moment are frequent .
As a type of the sort of erences we from king ith ] nents cspcct ivcly , loticc t values for the standard-deviation of No. the second , , and sixth moments\mdash ; Practically , it would ) ifficult to say which of these results gives the best fitting theoretical curve .
For atistics of kind they sensibly the same .
Thus , till another lnethod of the roblcm of ) resolution of frequencycurves is proponnded , I ink there is not sufficient evidence against use of higher moments to lead us to discard a method npon them as essentially likely to lead to , \mdash ; K. P.

