]\gt ; PHILOSOPHICAL TRANSACTIONS .
I. Mathematical Contributions to the Theory of Evolution.\mdash ; XI .
On the Influence of Natural Selection on the Variability and Correlation of Organs .
KARL PEARSON , F.R.S. , University College , London .
Received December 20 , 1901 , \mdash ; Read January 23 , 1902 .
CONTENTS .
Page ( 1 ) Introductory .
On the influence of seleotion ) correlation . . . . .
2 ( 2 ) On the fundamental theorem in multiple correlation . . . . .
3 ( a ) Mean value of one variable and its variability when other variables are absolutely fixed . . . . .
8 ( b ) Correlation of two variables when / Qther variables are absolutely fixed . . . . .
9 ( 3 ) General theorem in selection . . . . . .
10 With normal correlation selection of size is uninfluenced by selection of variability . . . .
13 ( 4 ) Distinction between directly and indirectly selecbed characters . . . . .
19 ( 6 ) Influence of selection of organ A on the mean and variability of , and on the correlation of A and . . . . .
22 ( 6 ) Influence of selection of organ A on the correlation of organs and C. 25 strations I. to . . . . .
27 ( 7 ) Influence of selection of organs A and on organs and and on their relation to A and 29 ( 8 ) Illustration .
Influence of selection of long bones of the leg on the long bones of the arm 32 Illustration II .
Influenco of selection of and humerus on the stature . . . . .
35 Illustration III Influence of selection of stature in modifying femllr and humerus . . . .
37 ( 9 ) Application to problems in heredity . . . . . . . . .
38 stration 1 .
Influence on direct heredity of the selection of parents . . . . . . . .
39 .
Influence on collateral heredity of a selection of parents . . . . . . .
42 Illustrafion III .
Influence of selecting two organs in parent on the correlation of the same organs in the offspring . . . . .
43 ( 10 ) Selective correlation coefficient and how it directly arises in the case of selection . . . . .
45 ( 11 ) On the probability of survival and the surface of survival-rates . . . . . . . . . . .
46 ( 12 ) Survival-rate curve for selection of one organ . . . . . .
51 .
Death-rate which would reduce the cephalic index of the French to that of the Naqada race . . . . . . . . . .
51 ( 13 ) Surface of survival-rates for selection of two organs . . . . . . . . . . . .
52 ( 14 ) Illustrahon .
Selection of the Aino , as far as femur and humerus are concerlled , from the French and rice vers4 ; constants and diagram . . . . . . . . . . . . .
54 Positive and negative selection . . . . . . . .
59 ( 15 ) General consideration of the surface of survival-rates .
Types of rate . . . . . .
61 VOL. CC.\mdash ; A 321 .
21.11.02 ( 1 .
) Introductory .
On the Influence of Selection upon AT an earlier stage in the development of the statistical theory of evolution it was suggested that the coefficient of correlation ( GALTON'S tion ) might be found constant for all races of the same species\mdash ; in fact , it was considered possible that this coefficient might be the long-sought-for criterion of identity in species .
Professor WELDON , following up this suggestion of Mr. GALTON'S , then made the elaborate series of measurements on crabs with which his name will always be closely associated .
To a first approximation these researches seemed to confirm the possibility of GALTON'S function being a true criterion of species .
When , however , a finer mathematical test was applied to Professor WELDON'S observations as well as to other statistical series for organs in man , ; it became clear that the coefficient of correlation varied from local race to local race , and could not be used as a criterion of species .
A slight investigation undertaken in the summer of 1896 convinced me that the coefficient of correlation between any two organs , is just as much peculiar and characteristic of a local race as the means and variations of those organs .
In fact , if local races be the outcome of natural selection , then their coefficients of correlation must in general differ .
The object of the present paper is to show , not only that natural selection must determine the amount of correlation , but that it is probably the chief factor in the production of correlation .
If selection , natural or artificial , be capable of producing correlation , then it seems impossible to regard all correlation as evidence of a causal nexus , although the coriverse proposition that all causal nexus denotes correlation , is undoubtedly the most philosophical method of regarding causality .
In dealing with the influence of selection on correlation , I shall suppose the distribution of complex groups of organs to follow the normal correlation surface\mdash ; the generalised Gaussian law of frequency .
I shall further assume the selection surfaces to be normal in character .
Neither of these assumptions is absolutely true , but the Gaussian ] in a good many cases describes the frequency sufficiently closely to enable us to obtain fair numerical results by its application .
Probably in all cases it will enable us to reach qualitative if not accurate quantitative theoretical deductions .
I have the less hesitation in asserting this , as .
G. U. YULE has recently succeeded in deducing the chief formulae for correlation and regression as given by the Gaussian law from general principles , which make no appeal to a special law of frequency .
" " Mathematicul Contributions to the Theory of Evolution.\mdash ; III .
, \ldquo ; ' Phil. Trans , vol. 189 , pp. 266 and 280 .
See also E. WARREN , " " Variation in Portunus depurator 'Roy .
Soc. Proc vol. 60 , pp. 233-4 .
See a series of letters in 'Nature , ' vol. 54 , 1896 , arising from a discussion upon a paper by A. R. WALLACE .
'Roy .
Soc. Proc vol. 60 , p. 477 .
SELECTION ON THE yARIABILITy AND CORRELATION OF ORGANS .
In particular , if a selected group be not given with very great accuracy by a normal frequency surface , still we may , I think , consider ourselves justified in supposing that the effects of the actual selection , and those of a normal selection with the same means , the same amounts of variation , and with correlations of the same intensity will be at least qualitatively alike in character , if they be not indeed exactly the same quantitatiyely within the limits fixed by the probable errors of the constants .
My plan in this memoir will be as follows : shall first state the fundamental theorem in multiple correlation with a new proof , so that the formulae required may be once for all collected for reference .
* I shall then give the algebraic investigation of the new formulae for selection .
I shall afterwards consider special simple cases , and illustrate them by examples .
Finally I shall draw attention to the nature of the selectiye death-rate as indicated in cases of this kind , and consider at length its algebraic theory .
Throughout I shall endeavour to illustrate the somewhat complex algebra by arithmetical examples .
( 2 .
) On the Fundamental Theorem in Multiple I have shown in my memoir on " " Regression , Heredity , and Panmixia\ldquo ; ( ' Phil. Trans , vol. 187 , p. 261 ) that if the variables of a complex be functions of independent variables with frequency distributions following the normal law , and such that the principle of superposition holds for the deyiations from the means supposed small ; then the frequency of the complex with deviations from the means of the variables lying between and and and will be where : expt .
\mdash ; . . . . .
Here . .
ars constants , and denotes a summation for every value of and for every pair of values of and in the series from 1 to In the same memoir ( p. 302 ) I have determined the values of in terms of the correlations and the standard deviations and of the variables .
This point had already been considered by Professor EDGEWORTH ( ' Phil. Mag vol. p. 201 , 1892 ) , and some further results by Mr. A. BLACK , reached before his death in 1893 , were published in the 'Camb .
Phil. Trans. ' ( yol .
16 , p. 219 , 1897 ) .
The present investigation is , I think , novel , and adds to results already reached others required in the present memoir , so that it thus places together with a fairly simple proof all the fundamental results to which I shall have occasion to appeal later .
* We have used theseformulae for several years , but they do not appear to have been hitherto published in a collected form .
Let us consider the quadric of the order constant , and fix our attention on two of the variables , say the first two , and .
If these be considered constants , the quadric of the remaining variables will not now be referred to its " " centre But its centre , , will be giyen by the equations : . . . . . . . . . . . . . . . .
The following expressions will not be zero , but will be written and : . . .
Now , if be the discriminant , . . . .
( iii .
) .
' and the minor corresponding to , we have by solving the above linear equations , whence : Generally also : .
, say , .
SELECTION ON THE yARIABILITy AND CORRELATION OF ORGANS .
This determines the central co-ordinate for any variable for a given value of and Now let us transfer the quadric to as origin .
It may be written Making use of linear equations ( ii .
) and ( iii .
) : . . . . .
For arranging vertical columns in rows , the remaining terms are , each line of which vanishes by the equations ( ii .
) for the centre .
Accordingly : where is a quadratic function of , . .
, not involving and at all .
Hence : Now integrate with respect to all the variables , . . .
from - to keeping and constant .
Then , although the origin is a function of and cannot involve and but only , \amp ; c. ; let the result be .
Then : must be the correlation surface for and , for all values of , . . . .
We may write it in the form expt .
\mdash ; ( viii .
) .
Comparing this with the known form of the correlation surface for two variables , * expt .
\mdash ; ( ix .
) , we have at once . . . . . . . . . . .
Whence . . . . . . .
, .
Or , generally : ( xiii .
) , .
Thus correlations and variations are fully determined in terms of the discriminant and its minor for the constants , . . .
We have next the inverse proposition to find the 's in terms of the 's and We have , by well-known propositions in the theory of determinants : Or , Hence solving ( xv .
) , * PEARSON , ' Phil. Trarls vol. , p. 264 .
SELECTION ON THE VARIABILITY AND CORRELATION OF ORGANS .
where and is the minor corresponding to the constituent But . . . . . . . . .
where is the determinant , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , and is the minor corresponding to the constituent .
Thus we have ( xvii .
) , or , generally , . . .
( xviii .
) .
Thus may be expt .
\mdash ; ( xix .
) .
It remains to determine from the fact that the volume of the surface Or , expt . . . . . . .
expt .
\mdash ; ) .
( xx .
) , by writing , \amp ; c. Now , integrate first with regard to writing first * Phil. Trans , vol. 187 , p. 302 .
, is the sum for all values of , and for all pairs of values of and , from 2 to inclusive , where .
For if , or the determinant of the correlation coefficients , omitting all involving the first variable , i.e. , the first row and column of , the determinant corresponding to the minor of the constituent in or to the second minor of is given by . . . . . .
( xxi .
) , and for . . . .
( xxii .
) .
Hence , integrating between the limits and , we have expt .
\mdash ; This is of precisely the same form as before , except that we have the factor , and the multiple integral is reduced by one integration and by the disappearance of all correlations involving the first variable .
Now , integrate with regard to .
The sole effect will be to multiply by a factor , where is the nlnor of not involving correlations of the second variable .
Thus , by repeating the process , we have ultimately or , ( xxiii .
) , which gives the constant of the surface .
The preceding investigation enables us al es us aso to deal with two further points .
Given variables , what is the mean value of the first variable and its variability for definite values , of the other iables ?
SELECTION ON THE yARIABILITY AND CORRELATION OF ORGANS .
9 .
Clearly , when . .
are constants , the distribution of is of the form constant expt .
\mdash ; Or , re-introducing the , . .
we have a distribution about the poin given by xxiv . .
with standard deviation ( b ) Given variables , what are the mean values , the standard deviations , and the correlation , of two of them , when we give definite values / to the remaining variables ?
In this case we have from expt .
\mdash ; involving and } . . . .
Writing for the coefficient of , and for that of , we have for the centre , , by transferring to the minors of and the .
Or , , xxvi . .
Here is the determinant formed by striking out the first two rows and columns of is the minor obtained by striking out the second row and column from , and then the first row and column ; the minor obtained by striking out the first row and column , and then the second row and column .
But a comparison with ( xxiv .
) shows us that these values for and are precisely what we should have obtained for the regression equations of the 1st and 2nd variables respectively alone on the other variables .
Thus the existence and the correlations of have no effect on the value of , nor those of on the value of VOL. CC .
\mdash ; A. PRoETSSOR K. PLARSON ON THE INFLUENCE OF ATURAL Returning to , we remark that the terms of the second order in ) , on which the correlation and variations depend , are not altered by a transfer to the centre and of the array Hence by and ( xvii .
) we have .
( xxvii .
) .
is the elation of the 1st 2nd organs for the remaining with constant values .
Again , Whence we easily find from ( xviii .
) and ( xxvii .
) or , ( xxviii .
) , where is , as before , the determinant without its first two rows and columns .
These by ( xxiv .
) , what we should have reached by ignoring in finding , and in finding ( 3 .
) Theorcm in To the selected , the and selected correlcctions , when organs selected , naturally or rtifcially , out of complex of organs .
Let the selected group of organs have their means raised , .
( some of these quantities may be negative ) ; their standard deviations changed from , . .
to , . .
, and their mutual correlations from , . .
, . .
to , . .
, . .
The whole system of organs before selection will be defined by the means as origin of measurement for each organ , by the standard deviations , . .
and by the coefficients of correlation , . .
) , . . . . .
Let be the determinant 1 , , . . .
, 1 , , . . . . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 , . . .
1 and the minor corresponding to the constituent .
Then the unselected population is given by the frequency surface of equation , SELECTION ON THE VARIABILITY AND OF ORGANS .
constant expt .
\mdash ; ( xxix .
) .
For brevity , we can also write this in the form constant expt .
\mdash ; Now consider for the time only organs\mdash ; namely , the first organs and the , and let us write , if be , for the determinant .
: 1 , , 1 , , 1 , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , 1 , , ' 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , 1 , ' , ( xxx .
) .
Then if be the minor corresponding to the constituent , and if , the distribution of the organs will be given by the frequency distribution constant expt .
\mdash ; .
( xxxi .
) .
being a sum for every value of throughout the organs , and for every pair of values .
Now let the first organs be given values , . .
, then the mean value of will be given by , ( xxxii .
) .
Now these coefficients ca , be found once if be known .
For example : and - and can be written down by symmetry ( xxxiii .
) .
Now suppose in the expression xxix .
we were to ut , . .
equal to , . . .
respectively , then the system of equations to find the means of , . .
, . .
for this array would be the equations ( xxxiv .
) , where are the ' co-ordinates of the centre\ldquo ; of the array of the organs .
If we were to solve these equations we ought to get precisely the solution for that we have found for in ( xxxii .
) above , where none of the coefficients involve correlation-coefficients other than those of the first organs among themselves and with the organ .
This result flows from the pretty obvious law that the mean of the organ for an array determined by values of the organs cannot be in any way dependent on our considering the relation of this selection of organs to any additional organs beside the : see p. 9 .
Thus the solution of ( xxxiy .
) is simply obtained by putting , . .
successively in ( xxxiii .
) .
Let us now select the first organs not with absolnte values : but varying about means , . .
, with standard deviations , and with mutual corre1tions , . .
, . .
, . . .
Ws have then to multiply in xxix .
) an exponential quadratic function of the .
, . . .
, i. e. , the selective tion surface , and divide it by another exponential quadratic surface , i. e. , the correlation surface of the organs This follows , since the frequency of each complex of organs must be reduced in the ratio of the selected to the primary frequency of the complex of selected organs .
But it will be clear that such a reduction must give us a result of the following form for the final frequency surface of the organs : constant expt .
\mdash ; { and and and inear terms in and ( xxxv .
) , where the 's denote the changed SELECTION ON THE AND CORRELATION OF ORGANS .
1,3 Now if we differentiate the quadric to find its " " centre\ldquo ; we have equations in , but the utions of these , if were eliminated , are known to be the " " centre\ldquo ; .
Hence we require only equations involving and we can put the 's for the remaining values .
Let us take the differentials of the quadric with regard to , then the resultin equations involve none of the , but only the .
They reproduce in fact ( xxxiv .
) .
But the values of . . .
found from xxxiv .
) are , we have seen , identical with the values of ( xxxii .
) .
Thus we have ' ( xxxvi .
) .
These give the most general form of -a theorem proved for a particular case in ' Phil. Trans , vol. 187 , p. 300 , ( ii .
) .
If systcrns of organs be selected with clny arbitrary variations and correlations out of complexes of , then the mean stzes of the remaining organs precisel.u ralues as if the of all the of organs been to one size not varied mean value. .
The variations of the sele , cted ' about these mean values , well the no influence the mean changes of the organs .
Returning to equation ( xxxv .
) we know that if the determinant ' , .
' , .
' , .
' , .
( xxxvii .
) be formed , its constituents and not the linear terms in the exponential 01 ( xxxv .
) determine all the standard deviations and correlations .
Let be the variation after selection of the organ ; then if be one of the selected organs , if be for one of the unselected organs has still to be found .
Let be the correlation coefficient of the and organs after selection , then if and 1 , be selected organs , ; if they be not both selected organs , then it to be found .
Let be the minor corresponding to the constituent in the above determinant , then by and ( xiv .
) ( xxxviii .
) , and ( xxxix .
) .
Now let us write for brevity and .
Clearly as long as and are less than and will both be known , i.e. , are equal to and Now by a well-known property of determinants . . . . . . . . . . . .
( x1 .
) .
Comparing these equations for the unknowns with ( xxxiv .
) for finding , we see that they are absolutely identical if we change in the latter into Accordingly the solution is given by ( xxxvi .
) , or we have Provided be we might equally well have used we conclude that if be and be , '1 , '2 , Hence ( xli .
) .
Or , substituting the known values of . .
, we have for and CTION O THE yARIABILITy AND CORRELATION OF ORGANS .
The next stage in work is to find when both and are , and also when and both are .
This is done at once by substituting the minors , . .
in the equations formed from the last lines of the determinant ( xxxvii .
) .
We obtain the following system : . . . . . . . . . . . .
, . . . . . . . .
These are identical with equations ( x1 .
) , except that the equation with the coefficients , . .
has unity instead of zero on the right-hand side .
Hence we see that will be the same function of , . .
that is of , . .
in equation ( xxxvi .
) , but it will add to this a function of the last system of , the 's , . . .
, . . .
, . . .
Whatever this function may be we will represent it for the time by ; we notice that it is independent of the selected variations , . .
, the selected means , . .
, and the selected correlation coefficients , . .
It depends only on the characters before selection .
We thus have Now the system . .
can be found from ( xli .
) , since whatevel and be .
Hence we have Here denotes a sum from to , and a sum for every pair of values of 'and out of 1 , 2 , , . .
When we have simply It only remains to determine and This we can do by putting all the 's , or selecting our organs of one size only .
We see at once that and are the values of and that is , of , and , when we select organs of definite values and seek the correlation and the variabilities of two others , the and the .
These values have already been found on p. 10 .
Or : , , ( xlii .
) .
, The notation of that page has been changed so that now stands for the determinant 21 , 1 , , 1 , 21 , 1 , , 1 , 21 , 1 , , 1 , 21 , 1 , , 1 , 21 , 1 , , 1 , 21 , 1 , , 1 , 21 , 1 , , 1 , 21 , 1 , , 1 , 21 , 1 , , 1 , 21 , 1 , , 1 , 21 , 1 , , 1 , 21 , 1 , , 1 , 21 , 1 , , 1 , ( xliii .
) .
1 , , 1 , is the minor corresponding to the constituent the minor corresponding .
the constituent at the meet of the column and row ; and the determinant with the last two rows and two columns struck out .
For example : SELECTION ON VARJABILITY AND LATION OF ORGAN , , , , 1 .
, , 1 , , 1 , , 1 , , 1 , 1 , \ldquo ; We can now collect our complete results The variability of a llon-selected orgar given by . . . .
, ( xliv .
) .
is after the selection of organs The correlation of two non-selected organs and both is after the selection of organs giyen by .
correlation between a non-selected organ and a selected is given by or , ( xlvii .
) .
* The expanded valucs of these determine , ' Phil. ans , vol. 187 , p. 294 , , 1 ) Here in takes every value from 1 to , and in and every possible air of values from 1 to .
Equations xlvii .
fully determme all the requlred quantities and form the full solution of the problem of selection .
Before we see the remal.kably simple forms they take for simpler cases , we may draw some general conclusions of a most important character .
In the first place we must distinguish between directly selected and what we have termed non-selected organs .
It would be better to term the latter indirectly selected organs .
Suppose the recruiting sergeant were to pay attention only to stature and seek to form a regiment of men of about 5 feet 10 inches .
He might have a real range of stature about 6 or 7 inches , but he would strive to get men of this height from the population .
We will suppose that he did not consider chest-breadth , head-length , , Iungs or any other character .
The distribution of these " " non-selected\ldquo ; characters in the regiment would not be the same as in the general population .
Their means would have changed by ( xxxvi .
) and their variabilities and correlations be given by xlvii . .
In other words , an indirect selection would have taken place .
A selection by stature would change foot-length and head-length and indeed every other correlated organ .
Much the same result must occur in natural selection .
If it be eous for a species to have a certain group of its organs of definite size , falling within a definite range , and related to each other in a definite manner , then these changes cannot take place without modifying not only the size , but the variability and correlation of all the other organs correlated with these , although these organs themselves be not directly selected .
Practically this means all the other organs , for so far one can hardly say with cerbainty that we have come across any two characters in an organism which are uncorrelated .
Many of those investigated are highly correlated , all appear to have some correlation , even if it be very small or negative .
We may therefore conclude as follows:-The selection of complex of characters or organs in an organism changes all the other characters and organs not directly ] ected .
( b ) If the change in the complex be continuous and progressive , the other characters will continue to be modified until the change in them is so considerable that selection begins to act directly upon them also .
( c ) The noted here are not confined to the average value of a nondirectly selected character and to its val.iability ; the correlations between non-directly selected characters and the ] ations between directly and non-directly selected characters are also both changed .
If local races have been produced by selection from a common stock , it will be impossible to look upon correlation as a criterion for species .
Every selection will modify such correlation , and it has no greater fixity than either type value ( mean ) or yariability ( standard deviation ) .
The whole of these statements will become more manifest as we apply our ] ELECTION ON VARIABILITY AND LATION OF ORGAN theorem to special cases , but we must note that if two organs were uncorrelated with each other , it still might be possible by selecting a third , a third and fourth , to produce correlation between them .
Further , by selection of one or more organs , two non-directly selected organs can have their existing correlation increased , lessened or even changed in sign .
( 4 .
) A primary difficulty will of course arise in the case of natural selection .
How are we to determine which are the directly and which are the indirectly selected organs ?
With artificial selection by man , we know which organs have been selected fairly well ; attention has been paid to colour , size , proportion of parts , \amp ; c. Even in the case of the medical examination of the recruit , it is chest , lungs , heart , stature , \amp ; c. , which form the basis of the acceptance or rejection .
If the head or foot be not absolutely deformed , little if any attention is paid to them , so with hair-colour , probably eye-colour , and a mass of other details .
No doubt the direct medical selection indirectly selects these , but could roughly class the selected and nonselected organs or characters and investigate the changes in the correlations of the latter owing to the indirect selection .
But how are we to form these classes in the case of natural selection ?
The investigations may look difficult , and even the standpoint of arithmetic appallin , but it seems to me that the differentiation of organs directly and indirectly selected classes is the keynote to the problem of evolution by natural selection .
Let us look at a simple case and see whether it will throw any light on the problem of distinguishing between directly and indirectly selected organs .
Suppose we have two organs only , with means , standard deyiations correlation and let the first be selected so as to have a mean value , and standard deviation .
Let be the standard deviation of the second organ and the correlation of the two organs after selection , and the mean of the nonselected organ .
Then by ( xxxii .
) : and it will be show1l later see that and Hence we have : ROFEbSOR K. ARSON oN NATURAL In other words the regression coefficie1lt non the non-selected organ on the selected remains unchanged , while that of the selected organ on the non-selected will , as a , be widely modified .
Further , let be the mean value of the second organ before se lection to a value of the first ; let and be the means of the organs aer se fter selection , and be the mean value of the second organ to a va ue lue of th first .
Then the equation to the regression line before selection after selection it is , But this is identically the same line as the regression line before selection .
Hence llot only the slope ( regression coefficient ) of the line , but its position is identical , and we have the following result : \mdash ; If two local ) from stock by the selection of one organ only , then the regression for the two of non-directly .velected on the selected will be the same direction position ; the of the selected organ -selected organ will di.ffer for the two Of course the means , standard deviations and fltions , not only of the selected organ but of all the non-selected organs also , will probably have changed .
It is only certain of the regression lines which remain unchanged and serve as a criterion to enable us to distinguish between directly and non-directly selected organs .
Of course the problem in Nature will not be as simple as this , for differe1ltiation of the two local races may have arisen from the selection of more than one organ , 01 may have al.isen from the selection of two different organs , but the illustration will , I think , indicate the nature of the investigation we are proposing .
We can easily generalise our theorem by considering the form of the selection surface given on p. 12 .
Any result obtained from ( xxxv .
) which does not involve any of the 's will be a result unaffected by the selection that has gone on .
Now to obtaiu a egression equation we put any .
of the 's equal to constallts , to 's The in this simplc that the regression line is umchanged is quite obvion , and , indeed , may serve as a proof .
SELECTION ON THE VARIABILITY AND CORRELATION OF ) say , and find the " " centre\ldquo ; of the quadric of the remaining , the co-ordinates of this centre , expressed in terms of the , are the regression equations .
Now it will be clear , that if we put all the selected to , the differentials of the quadric with regard to the remaining.or non-selected 's can contain no or the coefficients of the regression equations thus found will not be modified by selection .
Further , we might have given not only the selected organs , but any number of the non-selected organs constant values , and the resulting regression equations would involve only the 's and not the Hence we have the following general theorems : \mdash ; If organ has been modified only by indirect selection , then iis cocfficie is on any complex of other organs , however large or small , provided it includes the directly lected organs , voill unchanged by the selection .
The same organ in two rent local which have been derie from a stock by the selection of ) complexes of organs , some of which my or may be common , will , if it not been selected in either case , give the same regression coefficients for any grou of organs which includes the of both complexes and non-directly selected organs besides .
If the partial regression equations have changed coefficients , then we cannot at once determine whether\mdash ; We are dealing with a non-directly selected organ , and have not included the directly selected organs in the group upon which we are calculating the regres- sion ; or ( b ) We are really dealing with a directly selected organ .
In this case , we have also certainly not included at least one directly selected organ in the regression group .
Theoretically , however , and ( ii .
) suffice to find out which , if any , the nondirectly selected organs in the differentiation of local aces .
Practically , however , the number of organs and characters may be so great , and our ignorance of those probably selected so complete , that the arithmetic of determining so extensive a series of partial regression coefficients may be quite beyond our powers .
Still , where the divergence between local races is not too great , and the source of the differentiation not too obscure , it is probable that the above theorems will lead to results of great interest .
* Without laying too great weight on these theorems , I would still venture to suggest that if the criterion of a species bs the discovery of any numerical constant * Mr. L. BRAMLEY-MOORE has been working with this end in view ut the longbones in .
But even here the direct selection of parts of the vertebral column\mdash ; for which , present , we ve no correla- tion values either among themsclves with the ) \mdash ; and of the and foot , which Dr. W. R. MACDONELT .
has just shown , are very highly with the ) ones , may render nugatory all attempt to ascertairt which , if any , long-bone has beeu only indirectly , any rate , least directly selected .
or of constants , which is the same for all local races , then these constants must not be sought in the values of mean characters , degrees of variability , or of correlation , but in a system of ] ) al.tial regression coefficients , and the of these is therefore of first class biological importance ; it is the classification of the characters into directly and non-directly selected groups , i. e. , it is the discovery of the modus of the factors by which the tiation has taken place .
We are a long way from solution yet , but we may venture , perhaps , to admit a faint glimmer of light in the direction of what might seem the culminating problem of the mathematical method as applied to evolution\mdash ; the piecing together by quantitative analysis of the stages of descent .
( 5 .
) I will take now the application of the above results to simple cases ; but for the benefit of those cannot easily follow the main principles of our investigation through the stages of determinant analysis , I will prove directly the proposition that : the selectaon of organ A alters the mean and variability of correlated organ the correlation A B. Let the correlation surface for the two organs be , where is the number of individuals in the general population before selection , and the subscripts 1 and 2 refer organs A and respectively .
Let the distribution of the population after selection of the A gaIl be where is total destruction , the mean and the variability of the population with regard to A aftel ' selection .
selection this distl.ibution was Hence , the selection being random with regard to the of espon ding to any , we have for the surface after selection for each array must be altered in the atio of ths esponding z to This gJves for the surface in full 1 .
expt .
\mdash ; \mdash ; \mdash ; ( xlviii .
) .
ECTION ON THE VARIABTLITY AND LATION OF ORGANS .
Integrate this for every possible value of and we shall have the distribution .
and of or after the selection of A. After some reductions we find for the frequency where . . . . . .
( xlix .
) , which gives the standard deviation of the indirectly selected organ , and the " " centre " " of this organ is given by . . . . .
( 1 .
) .
These are special cases of our results ( xly .
) and ( xxxvi .
) aboye respectively .
Further , returning to the correlation surface ( xlyiii .
) , the coefficient of correlation is the coefficient of divided by the product of the coefficients of and Hence we find .
Let , then we have ( lii .
) .
This shows us that decreases with , or that decreases with , that is to say , the more intense the selection the less is the correlation .
This in broad terms demonstrates the general principle that intensity of selection connotes a lessening of correlation .
It is this principle which very possibly accounts for the fact that the more civilized races of man appear to be not only more variable but more highly correlated than the less civilized , among whom the struggle for existence is more intense .
It may , perhaps , also account for the skeletons of women of the civilized races having their parts more highly correlated than the parts of those of men .
* Lastly , it may well throw some light on the markedly plastic character of races which have been stringently selected with regard to one or a few organs only .
As an illustration , suppose that the correlation between femur and tibia were and let us investigate what would be the effect of reducing the variability of the tibia by direct selection 50 per cent. We find at once on substituting and in above that , or a reduction of about 37 per cent. This wilt , perhaps , be sufficient to indicate what immense changes must be made in the correlation of highly correlated )rgans whenever selection , artificial or natural , is stringent .
It is important to notice that the change in the size of the organ in no Se9 and PEARSON , ' Roy .
Soc. Proc vol. 61 , p. 354 ; and LEE , ' Phil : Trans , vol. 196 , p. 231 , SELECTION ON THE VARIABILITy AND CORRELATION OF ORGANS .
Further , taking AP as our unit , , and .
Hence , from ( lxix .
) : Therefore if in the diagram be taken equal to and be drawn parallel to , we shall have , or we can scale off the reduced variability .
Thus the diagram enables us to see at a glance the reduction in correlation and variability .
( 6 .
) Let us now write down the results when an organ A is selected out of a group of three organs , , whose constants are marked by the subscripts 1 , 2 , 3 , respectively .
Let , and be represented , when required , by .
Then we find from ( xlv .
) ( xlvii .
) : .
( liii .
) , ( liv .
) , , .
where , as before , we write .
Let us also write , and Then we can replace the above results by cus .
VOL. These equations admit of easy interpretation by spherical Let be the pole of the great circle DEFG .
Take DG } , Join to and .
the small circle of radius ound P meet , PF , PG in respectively .
Draw the arcs De and Let the small circles with centres and and radii De and respectively meet in Q. Join DQ and Then the quantities required are : DQG .
For DE ; hence : De Ee , or ; similarly Next , De , or ; similarly Lastly , from the triangle DQG : , and DG , but DG DQ DQ QG DQG ; or Thus all the relations can be expressed in terms of the sides and angles of a simple system of spherical triangles .
For the degree of accuracy generally possible in biological and sociological investigatious these triangles can be solved by a spherical trigonometer , such as that sold by , of Prague .
* The changes , however , which undergoes for various values of , indeed , far more difficult to appreciate as a whole than those of or .
In order that they may be followed easily , and in order to solve directly to a degree of approximation sufficient for many practical pmposes problems in the influence of selection on correlation , my assistant , .
L. N. G. , has kindly drawn up the tables which accompany this memoir .
It will suffice fairly well for all but a few special values of SELECTION ON THE VARIABILITY AND CORRELATION OF ORGANS .
Of course to bring them within any reasonable compass we have had to limit the values taken .
In the first place we have considered only eleven grades of selective stringency given by , 1/ 10 , 2/ 10 , 3/ 10 , 4/ 10 , 5/ 10 , 6/ 10 , 7/ 10 , 8/ 10 , 9/ 10 , 1 , the corresponding values of in the tables are entered as The tables are calculated for both positive and negative , bnt and are always supposed positive .
If and be both .qative , then will be the same as if they were both positive .
If and be of opposite signs , then all we have to do is to look out in the table in whioh has a sign the reyerse of its actual value , and having found the corresponding value of , then change its sign to obtain the actual coefficient of correlation after selection .
This follows , if be the negative coefficient , by writing : Lastly , it would clearly be very laborious to tabulate for a very great serles of values of .
Accordingly a selection had to be made of these coefficients of correlation .
They were given the values , 25 , 5 , 75 , and 1 .
These may be spoken of as zero , small , medium , large , and perfect correlations , and the ranges to 25 , 25 to 5 , 5 to 75 , and 75 to 1 , as the ranges of little , moderate , considerable , and high correlation respectively .
There would thus appear to be 15 combinations of values for ; hese are given in the key to the tables as , ( b ) , , , , ( p ) , see p. 63 .
If these 15 values had to be combined with the 10 values ( 5 positive and 5 negative ) of and the 11 values of , we should have 1650 entries in our tables .
But this number is much reduced by the consideration that the expression has for the real correlation of three characters to be always positive .
can also never be greater than unity .
Accordingly all values of , which do not satisfy these conditions , have been excluded from the tables ; they cannot arise in nature .
A few impossible values of have been included in the tables , but these are placed there solely for the purpose of finding by interpolation values of , which are less than unity .
The following purely hypothetical illustrations of formula ( lvi .
) and the tables will serve to indicate their use .
I.\mdash ; Suppose the correlation of tibia and femur with each other to be 8 , and of both with the stature to be 6 .
How would their correlation be altered if the variation in stature were reduced by selection to half its present value ?
Let as before , and suppose ; then let be the values of the constants next below the required values occurring in the tables , and giving ; let be the true value of , corresponding to the values .
Thus we have , as far as first differences : In our case .
Further , we look up Table IV .
, and the nearest case is ( j ) under which gives .
We then see that difference between ( j ) and ( m ) cases difference between result in IV .
and ; and lastly difference between and columns of ( j ) row of Table Iy . .
Thus we find : The valu6 by straightforward calculation of formula ( lvi .
) is 6981 , the two results giving substantially the same value 7 .
Thus we see that such a selection would reduce the correlation of tibia and femur by per cent. Illustration ( II.).\mdash ; Suppose the correlation of humerus and femur to be 5 , and of those with stature to 7 and 8 respectively .
How would the correlation of humerus femur be modified by a selection of stature given by In this case , .
We turn to Table III .
and take out ( k ) under , which gives us .
We have and , whence we find , but the.differences of the table are too great at this point for the result to be very trustworthy .
* Suppose we take and as before , but , and therefore , to be found from ; then and as before , .
Hence we deduce .
The mean of these two values of is 1377 , the true value calculated from ( lvi .
) is .
Taking 14 for the practical value , we see that the correlation of humerus and femur has been reduced by this comparatively moderate selection of stature upwards of 70 per cent. !
lstration ( III .
) .
\mdash ; Suppose a case in which humerus and femur were not correlated , that both were correlated 7 with stature .
What would be the effect of selecting stature with the same intensity , i.e. , .
Our best results from the tables will be to take from Table I. , which glves .
We have then , hence is to be found from ( m ) and ( h ) and , and The actual value by formula *Second differences ought to be used , and the process indicated is practically equivalent to using them .
SELECTION ON THE VARIABILITy AND CORRELATION OF ORGANS .
Now this again is a remarkable result ; by selecting an organ correlated with two others , neither of which are correlated with each other , we have produced a considerable correlation , and what is more , one of a negative sign .
In other words , if humerus and femur were unrelated to each other , but were related to stature , then a selection of stature would result in men of long femur having a short humerus , and vice versed .
I V.\mdash ; Suppose the correlation between greatest length and breadth of the skull to be 25 , and between both and the auricular height to be 5 .
Now let a stringent selection , , of height take place .
What modification will there be of the length and breadth correlation Table II .
, , case ( j ) gives us at once the result\mdash ; In other words , the correlation between length and breadth would be sensibly destroyed by such a selection .
Thus correlation can be created or destroyed or reversed by selection .
The above illustrations , hypothetical though they may be , will suffice to indicate how entirely dependent correlation is upon selection .
We must look upon coefficients of correlation , in fact , as just as much the outcome of selection as coefficients of variation , standard-deviations , or even the mean size of organs .
No selection can take place , in the sense in which it has usually been understood to take place\mdash ; i.e. , by a change of mean and of variability , without at the same time the means , variabilities and the of all correlated , but not directly selected , organs being varied .
This is true whether the non-selected organs be initially correlated or not among themselves .
We must always bear in mind this all-important funda- mental conception , that natural or artificial selection , or even random sampling , are in themselves active factors in the modification ( i.e. , creation , destruction , or reversal ) of correlation .
Thus not only is the impossibility of the constancy of correlation for local races obvious , but the primary importance of insuring that our samples are repre .
, and not accidentally selected samples , in all observations or experiments on heredity , homotyposis , or organic correlation becomes more and more manifest .
We must not lay too much stress on two heredity constants\mdash ; differing , for example , by more than the probable error of their unless we are convinced , which practically it will be difficult to be , that all modification of correlation by unintentional and unmarked selection has really been avoided .
( 7 .
) Let us now take the next most simple case .
lf be four mutually correlated ( in , either the .
same or different individuals ) , and a selection of A and , to the in the racters of the non.selected organs .
.30 PROFESSOR K. PEARSON ON THE INFLUENCE OF NATURAL Let subscripts 1 , 2 , , 4 mark the organs respectively ; let be the constants which determine the selection of A and ; and let us apply to ( xlvii .
) .
Here the coefficients will be obtained from the set for in ( xxxiii .
) and ( xliv .
) .
Hence we find : ( lviii .
) .
( lix .
) . .
( lxi .
) . .
( lxii .
) . .
( lxiii .
) . .
( lxiv .
) .
Finally , for the cfiange of means of the non-directly selected organs , we have : ( lxv .
) .
( lxvi .
) .
If we write ( lxv .
) and in the form\mdash ; SELECTION ON THE VARIABILITY AND CORRELATION OF ORGANS .
lxvii .
the 's are the partial regression coefficients , and the whole solution can be expressed in terms of them .
Thus : .
( lxviii .
) .
( lxix .
) .
( lxx .
) .
Thus the whole series of results can be easily calculated , if the regression coefficients are first calculated .
I may make some remarks upon these results .
A formula equivalent to ( lxviu .
) was first given by me in my memoir on " " Heredity , Panmixia , and Regression\ldquo ; Phil. Trans. , and used for certain problems of inheritance , and conclusions drawn from ( lxix .
) or ( lxx .
) have been cited or indicated in other memoirs .
Some interesting results follow at once .
If the selection be very stringent , and sensibly , then all correlation between a selected and non-selected organ is destroyed .
But .
( lxxi .
) .
This is what I have termed a correlation coefficient\mdash ; i.e. , the correlation between and when fixed values are given to A and B. So far as I am aware , such coefficients were first directly used by Mr. G. U. ] in certain economic problems .
* They are of very considerable interest , but for natural or artificial selection are not quite so important as the generalised form ( lxix .
) , for we generally select about a mean value , and not absolutely at it .
It will be noticed that the coefficient of correlation of two non-selected organs differs from the corresponding partial correlation coefficient by terms of the square order in , bnt the coefficient of correlation of a selected and non-selected organ .
Soc. Proc.\ldquo ; ] .
, pp. 485 , 488 ; conomic J ' ecember , , and )e ; cennber , 1896 .
differs from zero by terms of the first order in .
Hence , when selection is intense ( small ) , we may neglect the er as compared with the latter , an we ave we thus the basis of a method of approximation very useful in some cases .
I have not yet succeeded in giving a geometrical interpretation to the above formulae , but have little doubt that it would be quite easy if the " " spherical geometry of four dimensioned space were more familiar to .
It will suffice to observe that it is easy to find cases in which the correlation of a non-directly selected organ with a directly selected organ , or with another of own class , can be reduced , destroyed , increased , or reversed .
In fact , all our previous warnlngs as to the caution necessary in avoiding unintentional selection collecting material for testing correlation remain in force , and , indeed , are emphasised .
The following .
illustrations will indicate the kind of problems which may be attacked by such formulae as the above:\mdash ; ( 8 ) .
ILLUSTRATION I.\mdash ; Influence of a Selection of the Long Bones of the Leg on the Long Bones of the Arm , and on the of the Leg to the Arm .
A numerical example will throw light on the application of the above formulae , and effectively illustrate the manner in which a selection influences size , variation , and correlation .
Consider the long bones femur , tibia , humerus , and radius , indicated by the subscripts 2 , 3 , 4 respec ively , and let be the mean values in centimetres .
Then the following numerical values are given in a memoir by Miss ALICE LEE and myself : * .
Soc. Proc vol. 61 , p. 343 seq. The correlations are not work out for exactly the same lengths in the case of the two races , but the numbers will serve quite well for the purposes of illustration .
SELECTION ON THE VARIABILITy AND CORRELATION OF ORGANS .
Now let us select from the French population a group having the same characteristics of the long bones of the leg as the Aino population , and then compare the characteristics of the arm bones of this selected group with those of the Aino population .
Our selection is given by : , 1.67 .
The following constants must now be determined arithmetically : ' If and be the mean humerus and radius 01 a femur-tibia selection from the French population , we have from These would give the effect of selecting any femur and tibia defined by and from the mean values of the humerus and radius .
For the particular selection indicated above : both of which are about a centimetre in excess of the Aino population .
By selecting , efore , from the French , a population with a mean leg like the Aino , we should still find the average arm of this population some two centimetres greater in length V0L .
cc .
than the Aino .
The variabilities 23 and of humerus and radius for a population selected from the French by femur and tibia are obtained from ( lxiii .
) and ( lix .
) .
We have : These give for the particular case : Turning to the correlation of humerus and radius , we hay by : giving for the particular case : It will thus be seen that if we selected the a group with the same variabilities and correlation of femur and tibia as the Aino , the variabilities and correlation of the humerus and radius of this group would not be very different from those of the .
On the other hand , the correlations between upper and lower members would be very significantly different .
Generally we have by ( lxi .
) for selection from the French : These yield for our particular case : These are all smaller than the corresponding French values , the selection has reduced the correlation , but the population has in all the cases but a greater SELECTION ON THE VARIABILITY AND CORRELATION OF ORGANS .
correlation than the French .
We must accordingly conclude that by a leg selection from the French aimed at reproducing the proportions of the Aino leg , we should not obtain an arm equivalent to the Aino arm .
The diyergences are indicated in the accompanying tab le- : \mdash ; \ldquo ; , umerus aumerusMean hradius , ffemur aection.from trench.tibia aadius 7tibia aumerus 8 There is , of course , no special reason for supposing that the French and Aino differ merely by an evolution which has acted by selection of femur and tibia .
We might have obtained a race out of the French more nearly akin to the Aino by a selection of femur and humerus , the process would numerically be exactly similar .
The illustration here chosen is taken merely as an instance , to indicate how the methods developed in this memoir enable us to ascertain with quantitative certainty how far racial differences may be due to the more or less stringent selection of a limited number of organs in the one race .
If we consider that local races have been differentiated from a parent stock by the selection of the chief or more markedly divergent organs , then we have in processes such as that just illustrated a method of ascertaining , at least tentatively , whether two races are to be considered as merely local varieties , and further the particular organs through selection of which the rentiation has taken place .
II.\mdash ; Influence of Selection of Femur Hurnerns in Modifying Stature .
The foHowin data have been calculated for me by Miss ALIC } LEE from ROLLET'S measurements on the French : * They have been undertaking , with the view of more scientifically than appears to mc yet to have been done , the mean stature of a race from a measurement of the long bones found in , \amp ; c. ROLLET'S measurements are given in 'De la Mensuration des Os longs des Membres , ' Lyons , 1889 .
I hope shortly to publish a memoir on the subject .
[ The memoir in question was pubIished in ' Phil. Trans , vol. 192 , pp. , 1898 .
] Now let us select from among French males a group having the same variability , correlation , and mean size of humerus and femur as French females , and let us ask how this would alter the variability ( 2 ) , mean size of stature in French males , and also the correlation between stature and humerus and stature and femur .
We have at once from the second column\mdash ; whence we find , From we deduce , This rmula gives the stature of any group of males selected from the French and having their femur and humerus respectively and centims. longer than the average .
For the special selection referred to above , and , hence This example shows us that if we selected French men with the same femur and humerus as French women , it would be the selection of the humerus which would an contribute mostly to the reduction of stature\mdash ; a somewhat singular result .
Furth such a selected group of French men would be still some centims. taller than er , average of French women ( instead of about Probabl had an the selected the tibia as well , the greater tion of remainin advanta ahei we would have disappeared .
avan age eight SELECTION THE VARIABILITy AND CORRELATION OF ANS .
To find the variability in stature of the selected group we must use ( lviii .
) .
We deduce : In our particular case this gives : The actual variability in stature of French women is measured by a standard deviation of .
Hence our selected group of men would be sensibly equally variable with French women , as far as absolute variation is concerned .
Lastly , from ( lxi .
) : which give in our particular case : Such a selection , therefore , would accordingly only increase insensibly the correlation between stature and humerus , while leaving that between stature and femur the same .
The sensible reduction of correlation between stature and humerus to 771 ) , which is found as we pass from male to female , does not arise when we select a group of males with their femur and humerus of the same length , variation , and correlation as those of the females .
ILLUSTRATION III .
jluence of Selection of Stature in Modifying } and Humerus .
Let us select a group of French men having the same height and variability in height as French women , and calculate the changes which will arise in their femur and humerus .
Here the selection is given by : We now need only the earlier formulae of this memoir .
From ( 1 .
) we find These give for our : Thus of males , selected to have the same stature as the females , would have a htly s femur and a htly 1 humerus .
A slightly longer femur in woman and a slightly longer humerus in man would thus appear to be sexual Turning , these are given by ( xlix .
) .
We find : This shows us that while the selection would give the same yariability of humerus to the men that women have , it would fail to produce the reduction of variability in the femur , which is characteristic of the women .
we deduce while from we have Thus we see that very small changes would be made in the COl.relations , stature and femur , stature and humerus , femur and humerus , if we selected French men to have the same size and yariability of stature as French women .
The explanation of this lies in the nearly equal absolute variability of the two sexes with regard to stature , for , as we have seen , it is the selection of variability which modifies correlation .
Looking at the table of values p. 36 , we see that the largest difference of ility in the two sexes lies the femur , and accordingly it is from a selection of femur that we should expect the greatest differences in the variability and correlation of the two sexes to have arisen , but even this rence alone would not account for the observed sexual ences in the correlation .
Indeed , it would be surprising if it did , for the selection of other , notably the pelvis , must have played a conin the differentiation of sex .
* ( 9 .
) I shall now proceed to a series of problems , which will show the application of esults , such as those obtained in this memoir , to questions which arise in dealing with inheritance and selection .
If we suppose a general population to have statistical \ldquo ; elnain constant at any rate for a moderate interval , we still want to know not only the error may arise from a random sampling , but also the sort of effect which results from our sample being too much drawn from one kind of environment , from a rather limited class , or from any other practically necessary or unconsclously lIltroduced limitation of the random character of our sample .
imitivo l more nearly equal in size , variability , and correlation than highly civilized man and " " Proc .
SELECTION ON THE VARIABILITY AND CORRELATION OF ORGANS .
ILLUSTRATION I.\mdash ; To find the Influence on tloe Intensity of Heredity of the Selection of Let the subscripts 1 , 2 , 3 refer espectively to father , mother , child .
Let us first select one parent\mdash ; say , the father\mdash ; very stringently , i. e. , is small .
Then we need only equations ( xlix .
) and .
These give us : and .
The first may be written ' or , we see that will decrease , as decreases .
Thus if we have for It is clear , therefore , that the correlation of parent and child will be much reduced by such a selection .
On the other hand , the regression coefficient will not be altered , i. e. , , as we have seen .
Hence in problems of heredity , where we suspect a parent to have been highly selected , we should seek for the regression of son on parent rather than for the correlation .
Thus in the case of Basset Hounds , ae some if not all the reduction in correlation between sire and offspring may be due to selection of the sire .
A test of whether the reduction in correlation is due to selection of a parent ought to be given by a comparison of and .
We cannot , I think , suppose , unless natural selection be very stringent , that differs much from .
Hence it follows that ought generally to be large , if there be selection of a parent .
We can hardly test this point effectively in the case of the Basset Houmds , owing to the nature of the classification .
In racehorses , although the sire appears to be far more selected than the dam , there is not a great reduction of the coefficient of correlation between sire and offspring , appears to be less than , but not so greatly and certainly less , that we can be surprised that the correlation of sire and offspring is not much less than we have found it for material in which selection of the father is certainly far less marked .
We must accept the warning as to the reduction 'Roy .
Soc. Proc vol. 66 , p. 157 .
Se , e ' Phil. Trans , vol. 195 , p. 93 .
in correlation produced by the stringent selection of one parent , but we must remember the complexity of the factors\mdash ; the variety of other influences at work in selecting and modifying selection\mdash ; before we lay much stress on this source of alteration in parental correlation .
Now let us deal with the case of both parents selected , and suppose their selection by , and the change of their coefficient of assortative mating from to .
We have from formulae ( lxviii .
) and ( lxx .
) by a little rearranging . . . . .
( lxxii .
) , , .
Now let us take special cases to bring out points .
Let us suppose , or no assortative mating to exist , and let us enquire what change would be made in parental correlation if we selected parents who had assortatively mated , without altering their variability , i.e. , let us take , we have at once . . . . .
( lxxiii .
) , . . . . .
( lxxiv .
) , whence we find for and Thus the general effect of assortative pairing of parents is to increase the correlation between parent and offspring sensibly , but not to very rapidly increase the variability of the offsprmg .
Thus marriages within a class would , if heredity statistics were collected for a class , tend to show increased parental correlation .
Very high assortative mating no doubt occurs with some forms of breeding , and we may well find in such cases higher values of the parental heredity than we should obtain for a populaion of .
same species with random mating .
I think this may be an effective factor the ralsmg of the parental correlation in the case of coat-colour in thoroughbred horse , SELECTION ON THE VARIABILITy AND COREELATION ORGANS .
Now let us see what happens if we select both parents moderately .
As test cases , let us take and 5 , and for the extreme We have at once ( lxxv .
) , . . . . .
( lxxvi .
) .
Hence we deduce This table is very instructive .
It ws us that selection and assortative mating are factors of opposite influence ; that if selection be only moderate , then with considerable assortative mating the coefficient of parental correlation may be increased , but if selection be stringent , then assortative mating cannot counteract , even if as high as 5 , its destructive influence on parental correlation .
For example , if we take parents remarkable for some intellectual or physical character , say with a variability only a very small fraction of that of the general population , then , however proportionately we might pair them , we should find their relationship to their children , as measured by the coefficient of correlation , very sensibly reduced below that of the general population .
I think we have here the reason why Mr. GALTON'S Family Data , which were drawn from a rather narrow class , and had only a small coefficient of assortative mating , give so much smaller parental correlation than my own Family Data , which seem to me drawn from a wider class , and have a considerably higher assortative mating It will be clear that with factors like ortative mating , natural selection , artificial selection of breeders , unconscious selection of material from one class or one environment , modifying our coefficients of heredity in one or another direction , we can hardly hope for more in practical statistics than an approximation to the strength of the pure inheritance factor by dealing yith the average of as many races and characters as possible .
* The work for Mr. GALTON'S Family Data is given , 'Phil .
Trans , vol. 187 , p. 270 .
My own results are as yet unpublished .
The is bont .
as compttYc with VOL. CC.\mdash ; A. ILLUSTRATION II.\mdash ; To Find Influence of Parental Selection on Modifying ternal Correlation .
Let the subscripts 1 , 2 represent the parents , and 3 and 4 two of their offspring .
Let us first select one parent only , the selection being given as before by Our formulae will now be ( liy .
) and .
So far as the change in variability is concerned , we have already discussed it under our first illustration , so we need only consider : ( lxxvii .
) .
Now , if the offspring are of one sex ; hence : ( lxxviii .
) .
If we take and as reasonable values , we have ( lxxix .
) .
Thus will be greatest when is greatest , i.e. , when there is no selection , and will decrease with until it reaches4048 , when , or there is selection of fathers of one value of the character only .
* The selection of one parent only does not , therefore , immensely modify the correlation of brothers .
Still , if we work sensibly with one class of the community\mdash ; say , men of should expect to find their sons rather less like each other than if we worked with the general population of brothers .
Now let us select both parents .
Here again the variability of the offspring has already been dealt with .
We are concerned with equation ( lxix .
) , and we shall put , or make parental influence equipotent for the two sexes .
Hence ' where is the parental correlation , and .
the coefficient of assortative mating .
Hence we find . . .
( lxxx .
) .
To reduce to numbers , suppose , and for the general population .
We have ( lxxxi .
) * In general the value of ranges from down to SELECTION ON THE VARIABILITY AND CORRELATION OF ORGANS .
Hence if we put , and ( lxxxii .
) .
The following table will suffice to indicate the changes which take place , when we give a series of values to and .
Thus the first rowgives the influence of selecting parents without any assortative mating .
We see that with increasing stringency of selection the reduction of correlation is very considerable , and that with such selection the influence of assortative mating becomes less and less .
Nevertheless , assortative mating can produce quite sensible results , if there be little or no selection .
I am , indeed , inclined to think that a good deal of the high values found for the fraternal colour correlation in the thoroughbred foals*is due to much assortative colour mating in sire and dam .
Of course it cannot be all due to this source .
VALUES of Fraternal Correlation with Parental Selection.3162 27832647 On the whole , I think , we may conclude , so .
as the elative influences of sexual selection in the form of assortative mating and natural selection go , that : Both sexual and selection can sensibly modify the intensity of as measured by the of correlation , the former tends to , the latter to lower , its intensity .
But the effect of the latter , if stringent , is to mask the effect of the former .
In f'act , we may write .
Hence the smaller , the smaller will be fraternal correlation .
This varies as the square of and only as the linear power of .
Thus we see at once why stringency of selection is far more potent than assortative mating .
ILLUSTRATION III.\mdash ; To find the influence of selectin.two organs and parent , on the correlation of the like organs in the offspring .
Let the organs in the parent be denoted by 1 and 2 , and in the offspring by 3 and 4 .
Suppose the organic correlation of the two organs in the general population to be * Phil. Trans , p. 93 .
PROFESSOR K. PEAlffiON ON THE INFLUENCE OF NATURAL that before any selection takes place .
Let , the correlation of the pring be supposed to be the same for both organs ; then organs in the paren an offs .
Finally we have the coefficients of cross-heredity , an These must vanish if there be no heredity and no organic correlation , and should be perfect if both perfect .
Hence we will take ] as a hypothesi With these values of the correlation ients we easily find Hence from ( lxviii .
) and ( lxix .
) we : , ( lxxxiii .
) , .
( lxxxiv .
) .
For simplicity , suppose the stringency of the selection to be the same for both organs , then : If , and ( lxxxv .
) .
The following table indicates the value of :\mdash ; It will be clear from this table that if the selection be at all stringent , no reduction of organic correlation in the parents will affect substantially the organic correlation in the offspring .
On the other hand , if bo , we can have considerable modifications in the value of the correlation , even if the selection be stringent , * See 'Roy , , Proc vol. 62 , p. 411 .
I a good deal of data on the of these heredity correlations now reduced and soon to be published ) SELECTION ON THE VARIABILITY AND CORRELATION OF ORGANS .
Thus we have the following values of Lastly , : or , is given by : Thus , even if there were no correlation between the organs A and in the general population , still a selection of parents in which such organs were correlated would lead to offspring with correlated organs and .
The amount of such correlation would only be 1600 , if the variability of the parent were not selected , and would diminish rapidly with stringent selection of variability .
Still 1600 is quite sensible , and would , if the selection continued for a few generations , continue to increase .
Thus we see how selection of a pair of organs in a parent may increase or even create correlation between the like organs in the offspring .
The reader will find other interesting illustrations in tracing the influence of an absolute selection of one parent only on the correlation of the offspring , , relation between pairs of foals which all have a common sire , the influence of selecting an organ A in the sire and an organ in the dam on the correlation of the organs A and in the offspring , the influence on assortative mating of selecting parents of men of genius , in many other problems .
( 10 .
) It is not without value to oonsider how arises in the case of natural or artificial selection .
Suppose we have two organs , A and , then we shall endeavour to give these definite values , say and , but we shall not be able to get all our individuals with such absolute values , we shall select with certain deviations from and , given by and , say .
Further , we shall endeavour to * This is a peculiarly interesting case , for if ws select men of remarkable intellectual ubility , we should expect to find both parsnts above the average of the general population , but with a negative correlation between them amounting at a to 1905 .
make some function of , say or by TAYLOR'S theorem .
But and small , so that our attempted relation will be of the form : the slope of a line to we endeavour to confine the selected organs .
But we shall not be quite able to exactly hit this off ; when , we shall find , where is a small unavoidable error in selection of uncorrelated with .
Thus , if and be the selected variabilities , we shall have : .
Therefore : ) .
Furthel : and therefore : ) .
Or , is at once obtained from the slope of the line , by which we endeavour to fix the relationship of the organs A and B. Or , again , we may look upon as a quantity measuring the mean divergence of the organ from that absolute fulfilment of the relationship between A and which we are striving to attain .
Thus is a quantity which naturally arises in every attempt , whether artificial or natural , to select organs having a definite relationship to each other .
Much the same considerations arise when we select three or more organs .
In each case the selected coefficients of correlation are constants which enable us to express to a first approximation the form of relationship we are aiming at , and ( ii .
) the average degree of divergence from absolute fulfilment of this relationship .
Thus , without regard to any particular distribution of frequency , the 's and the 's are the appropriate constants to express approximately the nature of any form of natural or artificial selection .
( 11 .
) ?
the Probability of and the Surface of Survival In the course of the present paper I. have assumed that when measurements are made on any population for a complex of organs , the frequency surface may be taken as approximately normal .
If this holds for the population before and after selection , and measurements made on many groups at different periods of life seem to indicate that it is proximately true , it follows that can determine the form of the probability of survival as a function of the means , variations , and correlations of the selected and unselected populations .
SELECTION ON THE yARIABILITy AND CORRELATION OF ORGANS .
Let the unselected population be given by expt .
\mdash ; ( lxxxvi .
) .
Let the probability of survival be given by lxxxvii .
, where is at present an unknown function , which is to be a maximum for , . . .
and , if the selection be at all stringent , to take rapidly decreasing values as , .
take increasing large negative or positive values .
It will be clear then that the individuals who are " " fittest to survive i.e. , have the smallest death-rate , are those whose organs are defined by : , . .
and fitness generally will be measured by the closeness of the individual to these ' ' fittest\ldquo ; individuals .
In order to find the surface of survivors , immediately after the selection if growth be taking place , at any later stage if growth have ceased , we have only to multiply by , or : . . . . .
( lxxxviii .
) , is what in the earlier part } this memoir I have termed the selection surface .
Now if this selection surface be itself normal , it will be of the form : expt .
- . .
( lxxxix .
) .
Here , as in the value of , all the constants , . .
ars known in terms of the variations and correlations .
If there be selection of organs only out of the , then , will all be zero .
Since by Equation ( lxxxviii .
) , it follows that the function which defines the probability of survival must be of he normal exponential type , or propose to deal in another memoir with the important problems of slow selection during rapid growth , and of secular selection during generations .
expt . . .
( xc .
) .
Thus , to determine the probability of survival , we require to know the values of the 's and in terms of the , and .
The shortest method of finding is to put , and then note that : .
( xci .
) .
Since ) we see that the relations for , given and , and for , given and , are cyclicly interchangeable if at the same time we change to to ) to to .
If , . .
be the standard deviations of the unselected population , this amounts to changing , .
to .
respectively .
Thus the results which give tIje probability of survival iu terms of unsolected and selected populations can always by an easy interchange used to obtain the selected population from a knowledge of the unselected population and of the probability of survival .
Let the unselected population be defined by ) , . . . . . .
and . .
Let the seleoted population be defined by , . . .
, and , . .
Let the constants of the probability of survival function , or , . . .
, be expressed by , .
, . .
as if it were a normal correlation surface .
* Then the problem will be solved , if we know the , and 's in terms of the , and r 's .
quating the squares , products , and linear terms in the 's in the equation , we have at once the system : ( xcii .
) , all values of and from 1 . . .
( xciii .
) , for all values from 1 to * These mnt not be confused with the , . . .
, . . .
of the previous discnsion .
Tho new in part constants of the sclected SELECTION ON THE VARIABILITY AND CORRELATION OF If we now substitute from the first two equations for the 's we find : . . . .
( xciv .
) .
Now let be the determinant ' , .
and the minor of its constituent .
Then we have at once : .
Or : ( xcv .
) .
Thus is determined in terms of , . . .
, which define the maximum frequency of survival .
In a similar manner by making the proper changes indicated above we have : ' .
( xcvi .
) , where the determinant / PROFhSSOR K. ON OF TURAL , , . .
, , . . .
the minor of its constituent Now it will be clear from these results that as a.general rule it is impossible for to be equal to H In other words : The frequentty met with in any given setected co }nunity , i.e. , tl , mediocre ) the individual fittest to survive .
It is only in the limiting case of natural selection being so stringent that one type individual alone is able to survive , that the fittest class has a numerical majority over any other class of the community .
This seems to me an important , algebraically almost self-obvious truth , and yet one which is very much obscured by the use of such a phrase as the " " survival of the fittest Of course , if there be continuous selection , or an environment so stable that the probability of survival remains constant for a long period , there will be a gradual approach , never theoretically an actual iden.tification of the mediocre and the fittest .
But in actual nature the environment , at any rate so far as it depends on climatological conditions , must have a long period as compared with the vital and reproductive periods of innumerable forms of life .
A hard winter , a drought , a flood , a famine , a plague or epidemic of any kind , even if fairly stringent , will rarely , if ever , render the most frequently surviving individual identical with the individual who is fittest to survive .
* Still less will this identity take place in the many processes of artificial selection , which are becomin and will more and more become valuable laboratory aids in our appreciation of the aetion of natural selection .
The divergence between the most frequently surviving and the fittest individual is measured by the above formulae for the 's in terms of the To complete the solution , the be from the equations of type and then from the 's the and 's follow by the well-known determinants for multiple correlation : see our Equations ( xi .
) ( xii .
) .
Throughout the earlier part of this memoir I have used only the surface of selection , but the ) investigation will enable us whenever desired to replace it by the probability of survival .
I will illustrate this by obtaining the formulae suitable to the simpler cases .
* We badly want a name the selection which acts for a short time and rapidly modifies the adult population .
It is practically the type of selection considered in this paper .
It is epidemic or catastrophic in cter .
The point is of considerable importance , moro than one influential writer has spoken of the resull of natural selection as the preservation of the type the mortality of which is least under the giyen conditions .
SELECTION ON THE VARIABILITY AND CORRELATION OF ORGANS . . .
\mdash ; Selection of Single Organ only .
The original population is given by and the curve of survivors by The probability of survival is where we easily find , if ( xcvii .
) , . . . . .
( xcviii .
) , and .
( xcx .
) .
As all ation consider a selection from modern French peasants , which should reduce the mean and variability of their cephalic index to those of the Libyan race .
French peasants :\mdash ; Libyans:\mdash ; Hence : These give : Thus for such a change as 7 points in the cephalic index to take place by selection* we should have to make the " " fittest to survive\ldquo ; of such a ridiculously low cephalic index as , and such a high variation as We find and accordingly the probability of survival given by ' where are the number of Frenchmen converted into Libyans so far as cephalic index is concerned .
I have purposely taken a somewhat extreme case of selection in order to illustrate how widely the most frequently surviving individual can diverge from the fittest .
In this case , if the chances of survival of the fittest , ( ii .
) of the individuals most frequent after selection , and ( iii .
) of the individuals most frequent before selection , be , and respectively , we have : * This is , of course , supposing the change to occur by catastrophic selection and not by a continuous secular selection , see footnote preceding page .
: : : : 1 or , the chances of survival of individual of the fittest type would be about eight times as great as those of an individual of the most frequent type after selection and about 700 times as great as those of an individual of the most frequent type before selection .
If ) , be the numbers after selection in the three classes*of the fittest to survive , the most frequent after selection and the most frequent before selection , we find : In other words , the most numerous type before selection is still after selection about times as numerous as the type with the least mortality , and this latter type is only about as numerous as the type to be most frequently met with after selection has taken place .
Thus , although there would have been a very great evolution in cephalic index , due to a stringent selection , the fittest to survive would always have formed but a small fraction of the dominant type .
Even if we were to replace the selection here considered by a gradual evolution spread over several generations , we should still reach in the main the same conclusion , i.e. , that natural selection never proceeds the of the fittest , or he those with the least .
These will always remain a small fraction of the community\mdash ; they are the goal , but often the very distant goal , to which selection tends to shift the population .
( 13 .
) CASE ( ii.)\mdash ; Selection of Two Organs .
In this case let the surface of survivors bs : expt .
, original populati : expt .
\mdash ; \mdash ; ( ci .
) , and the curve of probability of survival : expt .
\mdash ; \mdash ; . .
Since : * By individuals of a type or class is meant here , as elsewhere in this section , all the group falling within some small definite range of variation lying round a particular value of the organ ( e.g. , or which drfines the typc or SELECTION ON ?
VARIABILITy CORRELATION OF ORGANS .
we find at once ( ciii .
) . . .
( cvi .
) . .
( cvii .
) .
Let measure ths stringency of the selection , and measure the change in correlation .
* Then solving the above equations find : ( cviii .
) , , , . . .
( cxii .
) , where Similarly , if the original population and the curve of probability of surviving or of survival rates } ) given , we have to find the selected population : ( cxiii .
) , * If .
The quantity has been conveniently termed the " " divergence \ldquo ; by Mr. SHEPPARD .
Hence is the ratio of the sines of the selected and unselected divergences .
The above formula for can be at once changed into one suitable for trigonometrical logarithmic calculation .
Let , and ; then , if be the side of the spherical triangle , of which are the other sides and the i1lcluded angle : where .
( cxiv .
) , , .
( cxvi .
) , ( cxvii .
) , where ) ( l4.)Illustration.\mdash ; The following results are taken from the paper by Miss ALICE and myself already cited:\mdash ; Aino 8 Femul centimR .
40.770 centims. Humerus , , 29.502 , , 1.343 Correlation Femur and } .
Humerus indicated by the symbols above , let us select from the ench a ulatio having the same femur and humerus relations as the Aino .
popu aon We have at once : Whence we find : 9027 , 3.4870 , where are the number of Ainos which can be obtain ta ned from N Frenchmen .
We have accordin 1 gy the .
following for the surface of fitness to survive : SELECTION ON THE VARIABILITY AND CORRELATION expt .
\mdash ; Now it is clear that if we wanted by a " " catastrophic\ldquo ; selection to convert the French into something resembling the Aino , we should have to give the least deathrate to those French with femur corresponding to and hulnerus to or to the dwarfs with femur centims. and humerus centims. !
By no other means could we shift the modal value of the French population down as low as the Aino modal value .
The physical meaning of this is that we have been compelled to put on an excessive death-rate for the bigger Frenchmen .
An interesting point of our work is whence we see that while a selective reduction of humerus is far more effective in reducing both femur and humerus centres of survival than a reduction of femur , selective reduction of femur occurring contemporaneously with that of the humerus actually tends to raise the centre of ' the humerus , i.e. , the coefficient of is Now let us consider the frequency of survivors per unit length , say centimetre of femur and humerus , at different points .
The surface of survivors , , the Aino population , is If we put we have the frequency after selection of the original population type ; if we put we have the frequency after selection of the new population type ; and if we put , we shall have the iency after selection of those best fitted to .
If these frequencies be respectively , we find on substituting the numerical values that Thus the most frequent type of the new population is now about thirty times as frequent as the old most frequent type , while the type most fitted to survive has practically no existence at all .
It probably lies outside the actual boundary of the French population .
Here really arises the question as to how we are , in any actual problem , to fix the ratio of to ; or , what amounts to the same thing , to fix a practical boundary to a given population .
Such a boundary must be conventional , but I think that for practical purposes we are quite safe if we assume that an individual who occurs only .
thousand produce no effect on the physical evolution of the population as a whole .
Now the form of a correlation-surface for two organs , and , is us write ; then constant gives a series of ellipses the contour lines of the surface , or lines of equal frequency , i.e. , giving individuals with equal probability of occurrence .
Let the equation to these contour lines referred to their principal axes be Tben we have at once : or , AB Further , if ) the angle the A principal axis makes with the axis of , we .
These fully determine the principal axes of the frequency surface .
Now consider the fi.equency batween the elliptic cylinders corresponding to and ; we have it Hence , if be the frequency outside any contour ( cxviii .
) .
For to of we have simply , whence * For easy calculation put .
Then we have at once if : where : sln SELECTION ON THE VARIABILITY AND CORRELATION OF This will enable us to determine our conventional boundary to effective population .
Now let us refer our non-selected and selected populations to their centres and principal axes .
We find for the contour curves:\mdash ; Referred to its principal axes , the rate of survival is now Suppose we require to get at least 100'0 Aino out of the French population , , then .
Now suppose the limiting ellipse drawn , then the French population must be sufficiently large to give the individuals inside this ellipse .
Now gets smaller as we go further from the centre of the survival surface .
Hence the contour line of the survival surface corresponding to must be touched externally by the limiting contour of the Aino population , in ordel ' that we may get at least 1000 Ain out of N Frenchmen .
Now , by a graphical construction , I find the major axis of the elliptic contour line of the survival surface which touches the Aino limiting ellipse , is about .
This gives for the parameter of this ellipse , , or .
Whence : gives the greatest possible value of and the least possible value of N. Numerically this gives us about , or we should want more than 7,000,000 of Frenchmen to obtain our 1000 Aino by a catastrophic selection .
The actual bounding contour line of this least possible number of Frenchmen*has for its major axis centims. , and it touches the Aino limiting ellipse at the point where it is touched by the survival contour Now let us turn the problem round and ask what is the least population of Aino * The least ible to reproduce the Aino , as far as femur and humerus are concerned , to 1 in 1000 of the poputation .
VOL. CC.\mdash ; A. I SSOR ON TH INFLUENCL NATURAL from which we could produce 1000 Frenchmen by a catastrophic selection .
In this case the surface of survivals is simply obtained by inverting , and , if be the number of 1000required number of FrenclJlnen , we have : HerG j ' gcts larger as we go away from the centre of the surface of survivals , we must make the Fren ch limiting ellipse just touched internally by the contour line of the sul.face of survivors for which .
The major axis of this contour line for was found by a graphical process to be about .
This gives for , or : leads to about , or we should want upwards of 32,460,000,000 of Aino to ) roduce the 1000 Frenchmen .
bounding contour line of this number of Aino has a major axis of centims. bout , and touches the French limiting contour in the point in which it is touched by the contour of the surface of Now the difference between these two unselected populations is very great .
We see that to get the Aino a very great number of Frenchmen would have to be externlinated , bout 7000 for each Aino selected ; but to get the Frenchmen from the appalling number of Aino would have to be destroyed , upwards of 32,000,000 for each nchman selected .
Even if the selection were not catastrophic spread over centuries and centuries , we must nise what a large consumption of life there must be\mdash ; individuals destroyed without progeny*\mdash ; if we are to suppose any civilized race like the French produced by selection from an apparently imitive type like the Aino .
Indeed , the return journey in this case seems much easier the upward ascent .
Beyond all this we have only made French and Aino alike two organs , and only for one character of each of !
Allowing for our conventional limit to the population , allowing for the fact that our Aino data are drawn from a very limited population of remarkably small variability , it seems very improbable that the French have ever been produced by selection from a primitive race at all resembling the Aino .
The fact that the Aino could be so much more easily obtained by selection from the French seems to indicate that they are rather *Of with secular selection spread over ma eneratiotlS , it is largely the potentiality and not the actnality of Iife which is destroyed .
Still , while the gross unber killed among a small primitive community not be , the death-rate must still be immense .
I hope bo return to these points wheIl d wibh solection o.s distinguished from catastro } ) selection .
SELECTION ON THE VARIABILITY AND OF some egenerate offshoot of a race superior to themselves than ) of the primitive people from which the Circassian be ) posed to have ) The whole of this discussion is , of course , very tever is to be laid upon it except of method , ] ) of the amount of elimination which must be evolve one second in the case of organs we know measurement to continuity and only saltatory changes in no influence on the mass-evolution which ) duced the local But given mples of material ' method will enable us whether a race A\mdash ; for ( course a limited lllmbel of ) with less ) deduced from a from A. will ) follow path of least selection is ich necessarily used .
l'ossibly A and have been ched by far less expenditul'e of from .
Still it is something definite in the midst of our ) ) aftel ' truth in ) of descent have even a rough appreciation of the lnount of elective duction w would arise from alternative suggestions .
is why this specinl numerical of the surface of survival has been given .
The reader will possibly find matter somewhat by the The femur is measured along the ] vertical .
A is the type or mean femur-humerus of ) .
Within in continuous ellipse round A the whole Aino ) ulationnI ) 1000 wonId is the type of the French population the ) th area within which up to 1 in 1000 of the ) fall .
Since tlJe drawn to centimetres of the bones , it will be seen ) variation within both populations .
is the of the .
the selection of Aino from French it makes the " " fittest ) survive In ) selection of the French from the Aino , is no the centre of fitness , ) the " " centre of unfitness\ldquo ; ; the Aino killed intensity which incr the closer we to P. Now it seems to me that two cases , distinct in theory , ought to manifest themselves in Nature require names .
A race may be modified because a complex ) organs with a system of values is good for it , because it is it .
race may bc modified a certain element of it is fittest or } ) ecause it is unfittest ) survive .
th formel ' case we select for survival round centre , in the latter case we select for uction .
I propose to call these cases positive and negatioe selection respectively .
It may be that if there be positive selection in ) the population will ) negative in another .
But the kernel of the matter is in either case the existence of a centre , a definite set of most fit or of most unfit organs , while in positive selection ] fit organs , and in negative selection the more fit organs are distributed 'wid ) of the field , and do ot reach a maximum of unfitness or a maximum of respectively for any definite individual .
SFLECTION ON VARIABILITY AND LATION OF ORGANS .
the nature of the contour lines of the surface of survivals shows that the contour lines above referred to , and marked\ldquo ; boundary\ldquo ; in the diagram , must touch the Aino 1 in 1000 limit and the French 1 in 1000 limit respectively at the points in which they are touched by the contour lines and of the corresponding surfaces of survivals .
I have ah.eady indicated that the major axes of these boundaries are for the French and for the Aino .
The corresponding values of the parameter are respectively given by Hence by ( cxviii .
) we can easily find the fi.equency of population outside the and ; if these be and we have : Thus the French population would have to be extended to a boundary in which only about 1 in six millions was excluded , and the Aino population to a boundary excluding only 15 in the billion !
The boundaries of what we may thus term the selection populations are far larger than our conventional boundaries of 1 in 1000 for representative populations .
In fact , it would be impossible to select a representative population from a conventional representative French population and vice \mdash ; in either case the very exceptional members of French or Aino populations are required to complete the conyentional representative populations of Aino or French by selection .
( 15 .
) I have devoted most of my consideration of the surface of survivals to a particular case in which two organs have been selected , and we consider the nature of which determines the fraction of each group of individuals which survives .
I have done this partly because normal surfaces are at best only approximate representation of our selectable and selected distributions , and partly because I have thought a concrete case would best bring out the general points of tions of this kind .
But some little indication of the properties of the surface of survival-rates ought to be indicated here , or it may appear that they have been overlooked .
While the contour lines of the correlation frequency surfaces for two orga1ls must be ellipses , this does not follow in the case of the surface of survival-rates .
In our illustration they were ellipses , but they may be also parabolas , hyperbolas , or even straight lines .
We must not therefore expect to find always a " " centre\ldquo ; of positive or negative selection .
We may come across a " " saddle-back system " " of contours with the rate of survival constant along two intersecting lines , but rising in one pair of opposite angles and falling in the other pair .
In this case we have fields of negative and positive selection separated by two dependent relations between the two organs , which are linear and for which the survival-rate is the same , they may be termed the " " critical lines .
For one pair of angles the centre is now a " " centre of fitness for the other pair of angles a " " centre of unfitness It seems to me that these critical organic relations may possess considerable biological importance .
If the contour lines of the surface of survival-rates are parabolas , we have really only a ) limiting case of the centre at a very great distance .
It is one in which the fittest ( or most unfit ) has no practical existence , but there is a direction towards which the rate of survival will be found to be always increasing or decreasing .
If the contour lines of the surface of survival-rates are parallel straight lines , then so as the deviation in one organ has a certain definite relation to that in the other , the survival-rate will remain constant .
In this case the survival-rate will fall uniformly in one direction and remain constant in the direction at right angles to it .
All the cases I have given here can occur just as easily as the elliptic contour system of our illustration and diagram .
Each is marked by quite definite biological characteristics , and we may , perhaps , class them as elliptic , hyperbolic , parabolic , and linear selection .
Even if the surface of survival-rates be not of the exponential quadric type discussed in this paper , yet to the neighbourhood of each part of it this classification of selection types will apply .
If we pass to more than two organs , then similar considerations will apply ; we shall only be reproducing the geometry of quadric surfaces in space of three and higher dimensions .
But before we allow ourselves excursions into the higher geometry of the surface of survival-rates , it seems desirable that we should obtain quantitative determinations of this surface by experiments in artificial selection .
We shall then be better able to see what part of our geometry will really be of service for the problems of natural selection .
The field is too large to be cultivated for merely theoretical interests .
We must first determine what parts of it are likely to have practical application to life as we find it , but of death-rates in the .
case of any living but man , we are at present sadly ignorant .

