]\gt ; [ 229 ] VIL Mathematical Contributions to the Theory of Erolution . .
On tloe Probable Errors of Frequency and on the Influence of Selection on Variation and Correlation .
KARL PEARSON , F.R.S. , and L. N. G. FILON , B.A. , University College , London .
Received October 18 , \mdash ; Read November 25 , 1897 .
CONTENTS .
PAGE I. Introductory . . . . . . .
229 II .
On the probable errors of a system of frequency constants.\mdash ; General Theorem 231 On the determination of the probable errors and the error correlations of the frequency constants . . . . . . . . . .
236 .
On the probable errors and the coefficients of correlation between errors made in the determination of the constants of a normal frequency tion . . . . .
237 Probable error of a regression coefficient for two organs . . . . .
244 Probable rors of variation and correlation for three organs . . . . . .
245 Case of four or more organs . . . . .
250 IV .
On the probable errors and the coefficients of correlation between errors made in the determination of the constauts in skew variation . . . . .
265 In the case of the curve Numerical illustration : Incidence of enbel'ic fever . . . . .
279 In the case of the curve . . . . .
Numerical illustration : Glands of the forelegs of swine . . . .
289 In the case of the curve . . . . .
297 Numerical illustration : Stature of children . . . . . .
303 Conclusion . . . . . . . . . . .
309 Note on the probable error of the criterion . . . . . . . . . .
310 L \mdash ; GENERAL THEOREM .
1 ) In earlier memoirs by one of the present authors , methods have been discussed for the calculation of the constants ( a ) of variation , normal or skew , of correla* ' Nathematical Contributions to the Theory of Evolntion. .
On Skew V riation ' Phil. Trans , vol. 186 , pp. 343-414 .
tion , when normal .
* The subject of skew correlation would now naturally present itself , but although several important conclusions with regard to skew correlation have been worked out , there are still difficulties which impede the completion of the memoir on that topic .
Meanwhile Mr. G. U. YuLE 1 shown normal relation are , if not completely , eyffl in case of skew correlation .
It seems desirable to take , somewhat out its order , the subject of the present memoir- , partly because the formu'.lae involved have be once or twice cited and several times used in memoirs by one of the present writers , partly because the need of such formulae seems to have been disregarded by various authors in somewhat too readily drawing } sta$istical data Differences in the constants of variation or of correlation have been not infrequently asserted to be significant or non-significant of class or of type , or of race differences , without a due investigation of whether those differences are , from the standpoint of mathematical tistics , .
greater or less than the errors of the differences .
Notwithstanding that every artificial or even random selection of a group out of a changes only the amount of .
variabion , but the amount of correlation of ffie gans of its members as compared with those of the primitive group , it has been supposed that correlation might be a racial constant , and the approximate constancy of coefficients of correlation of the same organs in allied species has been used as a valid argument .
In the like manner differences in variation have been used as an argument for the activity of natural selection without a discussion of the probable errors of those differences .
In dealing wit , variation and correlation we find the distribution described by certain curves or surfaces fully determined when certain constants are kqown .
These are the so-called constants of variation and correlation , the number of which may run up from two to a very considerable figure in the case of a complex of organs .
If we deal with a complex of organs in two groups containing , say , and individuaIs , we can only ascertain whether there is a significant or insignificant difference between those groups by measuring the extent to which the differences of corresponding constants exceed the probable errors of those differences .
The probable error of a difference can at once be found by taking the square root of the sum of the squares of the probable errors of the quantities forming the differ.enoe .
Hence the first step towards determining the significance of a group difference\mdash ; i. , towards ascertaining whether it is really a class , race , or type difference\mdash ; is to calculate the probable errors of the constants of variation and correlation of the individual groups .
This will be the object of our first general theorem .
: " " Mathematical)Contributions to the Theory of Evolution .
mixia ' Phil. Trans , vol. 187 , On the Significance of BRAVAIS ' for Skew Correlation Roy .
Soo .
Proo vol. 60 , pp. This will be snfficiently indicated in the latter part of the present dealt with in a paper on the ' Iuflnence Select.ion on Correlation written , bnt not yet publisbed .
MATHEMATiCib UONTRIBUTIONS TO THII THEORY OF EVOLUTION .
IL ON THE PROBABLE ERRORS 0F A SYSTEM or FREQUENCV CONSTANTS.\mdash ; GENERAL THEOREM .
( 2 ) Let there be a group of individuals , for each of whom a complex of organs is measured , and let . .
be the frequency with which individuals having a plex of organs lying between , occur in the total group of .
Here shall measure the deviation of any organ from the mean ofalllike organs in the roup .
Let , . . .
, be the mean measurements on the organs , so that , . . .
, is the system giving the actual measurements on any individual .
Then , . . .
, are the first set of constants of the fi.equency ; they determine ) " " origin\ldquo ; of the frequency surface .
I.et this frequency surface be given by , where , . . .
frequency constants , which define the form as guished from the position of the frequency surface , and which will be functions of standatd deviations , moments , skewnesses : coeilicients of corl e \amp ; c , , \amp ; c. , of indilual organs , and of pairs of organs in the complex .
The problem before us to find the probable errors of the 's and the , which constants fully determine the position and shape of the frequency surface .
Let , . .
, . . .
, be the standard deviations of the quantities , . .
, and , . . . .
Then a knowledge of these standatd deviations will give us at onee the probable errors of the frequency constants , for we have only to multiply the former by the numerical factor to obtaiu the latter .
Let us now suppose that the value of the frequency constants had been , . . .
, . . .
, instead of the observed values .
Then the frequency of any ohserved individual would have varied as instead of as .
Hence on this hypothesis the probability , , of the set of individuals observed in the group occurring is to the probability , , of the set occurring when the constants are , . .
, . . .
, iu the ratio of the product of all quanfities like for all values of , . .
, to the like product of all quantities Iike , .
; , . . .
, or PROFESSOR K. PEARSON AND .
L. N. G. Taki arithms , the products become sums , or logf Let the first summation now be expanded by TAYLOR'S theol.em , and typical terms up to the second order be written down .
Then we have Iogf ) Iogf ) . .
cubic terms in and where stands for .
Here is to be given every value from 1 to .
and to be given every value from 1 to , but and in the third and sixth sums are only to be given values from .
1 to and 1 to other than and respectively .
In the above formula we may replace the sums by integrals , if we ember that the frequency of the system , . .
, is simply Writing .
, we will investigate the values of the constants separately .
First , . . . . . .
the integrals now not including one with regard to and denoting that to be taken between the extreme limits of for .
Now in most cases of frequency the frequencies for extreme values of any organ are zero .
* Hence equals nothing .
Thus we have In most cases , but not invariably , as , for example , in the oase of some florets and .
In the cases , however , in which does not vanish , the oonclusions finally will be the same , as , only marks a change of origin for the constant frequency distribotion .
MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
233 Secondly , . . . . . .
where is , he total number of individuals measured , which is independent of Therefore Thirdly , .
This will not as a rule vanish .
If the frequency be normal , still a rule vanish .
It will vanish either if the frequency be about , or if there be no correlation between the and organs , , if be of tho form .
Fourthly , .
( ii .
) .
This will not as a rule vanish .
It will vanish , howeyer , if the of , be symmetrical about its mean .
Fifthly , . . . . .
( iii .
) , . . . .
, all of which will generally be finite , bnt , like , and of calculation lvhell the form of frequency is given .
Hence expt .
\mdash ; . .
expt .
( terms in cubic and higher ol.ders of the ) .
This represents the probability of the observed ullit , i.e. , the individnals sets ) , occurring , on the assumption that errols , . .
V0L .
CXCI .
PROFESSOR K. PEARSON AND MR. L. N. G. FILON , . .
, have been made in the determination of the frequency constants .
In other words , we have here the frequency distribution for errors in the values of the frequency constants .
( 3 ) Conclusions to be drawn from the form of ( vi .
) .
The distribution of the errors of frequency constants , if treated exactly , will generally be skew , for the cubic and higher terms in the 's do not vanish .
If , however , the cubic terms are small as compared with the square terms , the frequency distribution of errors will approximate closely to a normal correlation surface .
It would be impossible to evaluate the remainder after the second power terms in TAYLOR'S series for any general expression for frequency .
In special cases we have found that of the third order amount in the most unfavourable circumstances to 4 per cent. of the terms of the second order , generally to a good \amp ; al less .
* Probably series in most cases converges with consid-erable rapidity .
The fact , however , that we are dealing with the first terms of a series should be borne in mind .
It does not seem to have been sufficiently emphasised when the probable error of the standard .
deviation is taken to be per cent. of the standard deviation .
The usual proof of this result , however , involves { he same assumption as to the smallness of the cubic terms .
Supposing the errors so small that we may neglect the cubic terms , we conclude .
the errors made in calculating the constants of any frequency distribution are\mdash ; Themselves distributed according to the normal law of ( ii .
) related among themselves .
Both these conclusions are of the utmost importance .
The first enables us to obtain the probable errors of the frequency constants ; the second depends upon the fact that , and are in general not zero .
The standard deviations of , and the correlations between , the frequency constant errors can now be calculated by the ordinary theory of normal correlation .
Before , howe v , proceeding to these calculations , we may draw one or two other conclusions of considerable generality and wide significance .
Consider a race fulIy defined by the variations .
, . .
, \amp ; c. of the organs of its members and their correlations , . .
Now let a random small selection be made of this race , defined by , .
, . .
, where the magnitudes of , . . .
, are quantities depending * Soc , for the relative ordel of two terms of the second and third ordel .
S , ' Regression , Heredity , Panmixia ' Phil. Trans , vol. 187 , p. 266 .
MATHEMATICAL CONTRIBUTIONS TO THE THEORy OF EVOLUTION .
235 on the nitude of the probable errors of the variation constants , and therefore on the size of the selection .
Then the system , . .
, Sr , .
, is not a system of independent variations , but the changes in variation and correlation are correlated together , since the terms do not generally vanish .
We therefore conclude that if a random selection be made out of a population with regard to one organ , there will be tendencies for the variation of all other organs and the correlation between all organs to change also in certain directions , which can be definitely indicated so soon as the general population has been measured , and the effect of the random selection on one organ has been ascertained .
What is proved here of random selection will be shown to be true still more intensively for artificial selection , , every selection of one organ modifies in a elated manner the variation and correlation of all other organs .
It is impossible to alter one organ without altering all other organs and their each other .
S The remarks in ( ) are not only true for a random selection of variation , but equally well apply to selection of size .
This follows because the terms and do not as a rule vanish .
If a random selection of 100 or 1,000 individuals be made out of a general population , then the mean size of any organ in this sub-group will probably differ from that of the general population .
The result is that the size of all other organs , their amounts of variation and ation , will probably have values differing in directions from those of the general population .
Take two random selections out of a general population ; the probability is that they will have different means for any one organ , and a result will be that they will have correlated systems of changes in the sizes , variations , and correlations of all other organs .
In other words , random selection produces a differentiation of all characters , which differentiation will be the more marked the smaller the random ( ) This principle\mdash ; that a random selection gives a system of ] ated changes in the deviations of all the characters of a species\mdash ; seems , to some extent , capable of explaining the small but systematic diflerences to be found occasionally between closely allied species .
It is not necessary to suppose them due to a long process of natural selection acting on a variety of organs ; a small random selection , or possibly a natural selection of one organ , might suffice to produce the systematic differences of character in all organs .
How far a succession of random selections would give an evolution biassed by the first random selection requires further consideration ; but it seems impossible for the characters of a race to remain fixed under the influence of a heavy but non-selective death-rate .
They will vary from year to year , although this systematic change of * ' Memoir on the Influence of Selection on relation SeIection not only modifies correlation , but the selection of one organ can create corl.elation bet1veen organs previously uncorrelated .
Continuous artificial selection of an organ produces a still more marked differentiation of all other characters , this is treated of in another memoir .
ers be not always in the same direction .
Systematic change of characters produced by random selection may be spoken of as random evolution .
Random evolution is theoretically a possible cause of systematic change ; experiment only can determine how great is its effectiveness in differentiating local races .
In the case of a normal distribution of variation defined by the mean and the standard deviation , it has been usual to suppose that the error made in the mean is independent of the error made in the variation .
In other words , it has been that vanishes , although no proof has been given , or possibly it has not been realised that a proof was .
In this case is of the form expt .
and , whence expt .
O. Thus there is no correlation between error in the mean and error in the standard deviation .
This assumes that we stop at the square terms in ( vi .
) .
If , however , we include the cubic terms , \amp ; c. , product terms in and do arise , and we cannot state straight off that no correlati6n exists , although it may be very small .
In the case of all skew variation , such as is so frequent among and animals , a correlation will ' be found between deviation or error in the mean and the like in the standard deviation .
In other words , to alter the mean by selection ( artificial or random ) is to alter the variation of an organ .
With the exception of the stffiements in this paragraph ( ) , the whole of our general clusions in this section are independent of any particular law of frequency .
( 4 ) On the Determination of the Probable Errors Correlations of the Frequency Constants .
Let , . . .
, be the frequency constants , whether they be the means , standard deviations , or correlations of a complex of organs .
Then if we ] cubic and higher terms in the deviations , . .
, the frequency surface giving the distribution of the variations in the deviations is expt .
\mdash ; whel . . .
( vii .
) .
It is required to find 2 , , the standard deviation of , and , the coefficient of correlation between and MATHEMATICAL CONTRIBUTIONS TO THE THEORY VOLUT 1 .
237 Let be the discriminant : where Further , let be the minor of the row and column ; then and . . . . . .
( viii .
) give the required values of 2 , .
and Further , the standard deviation of } .
for selccted values of all othel 's is This value can often be of service .
Thus suppose a considerable number of skeletons found , but that only in a tively few cases is it possible to pair together the femur and tibia of the indiyiduals .
Then the variations of femur and tibia in the race will be known with great exactness , but the probable error of the correlation between femur and tibia will be given with close approximation by a form l. In fact , whenever we have obtained from a large number of observations the values of the freque1lcy constants race with great exactness , then , usino a values to obtain an additional variation or correlation constant from a few observations , the probable error will be of the form just indicated , and not of the form . .
It is needless , perhaps , to remark that the former is far easier to calculate than the latter .
III .
ON THE PROBABLE ERRORS AND THE COEFFrCIENTS 0F CORRELATTON BETWEEN ERRORS MADE 1N THE DETERMINATION 0F THE 0F A NORMAL DISTRrBUTION .
( 5 ) In order to exhibit more ] the method of tion , it is desirable that a ) case be first taken .
Accordingly we will start with the following problem :\mdash ; To find the Probable Errors Error of the of requency Distribution for Let , be the means , , the standard deviations , the coefficient of correlation , the total number of of obseryations , , any pair of corresponding deviations of the organs .
Then the frequency is given by the surface expt .
\mdash ; ( ix .
) .
* See " " Regression , ) , and Heredity ' Phil. Trans , vol. 187 , p. .
We require to find , the standard deviation of errors in which is he coefficient of regression of , the coefficient of correlation between errors in and : the standard deviation of errors in , th coefficient of correlation between errol .
S in and It follows that all zero , since , by ( ii .
) of p. 233 , will vaoish when the distributions of and are symmetrical .
These correlations would not , however , vanish for the ske frequency distributions , hich are of most frequent occurrence in problems of heredity and fertility in man , \amp ; c. The first stage in the investigation is to write down the second differentials of the logarithm of for all quantities occurring in it .
We find ' * See the memoir on " " edity , Regression , and Panmixia p. 268 .
MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
239 Now these must be multiplied by and integrated for from These integrations follow at once , if we remember that by definition of , and ?
Thus we obtain the following system : ' , where the 's are those of Eqn .
( vii .
) , p. 236 , obtained by taking the above differentials of the logarithms of in order .
We can no write down the correlation surface , giving the frequencies of errors in the constants : expt .
-expt .
\mdash ; .
( xi .
) .
Now several important conclusions follow at once from this result : For the case of normal frequency ( but in general only for this case ) the errors in the means are uncorrelated with the el.rors in the variations and lations .
The error correlation surface breaks up into two parts , of which the first part we have written down involves only the , and would coincide exactly with the correlation surface ( ix .
) , with which we started , if we write in ( ix .
) for , and , for and respectively .
It follows lgly that the standard deviations of the errors in the means are . . . . .
( xii .
) , and the correlation of the errors made in the means is . . . . .
( xiii .
) .
Further , the standard deviations for errors in when the error in is known , or for errors in when the error in is known , are respectively .
The results are , of course , well known ; the results ( xiii .
) and ( xiv .
) are , we believe , novel and important .
An illustration may be of service .
Suppose the stature and arm-length of a population to be under consideration .
Let us suppose the mean stature of the population known from a great number of observations .
Now let the arm-lengths be determined for a random selection of the population ; then , if the stature of these individuals so selected differs in excess from the mean stature of the whole popuIation , the arm-length of the random selection will most probably differ from that of the whole population by quantity fixed by ( xiii .
) , .
or rather by the coefficient of regression , i.e. , Thus with a probable error .
In other words , if the arm-length is to be found from a selection of the general population only , and the stature of this selection differs from that of the general population , it is most reasonable to take the arm-length of the general population to be the mean arm-length of the selected population less the quantity Or , again , if a selection from a general population show a mean organ in excess of that of the general population , the whole system of correlated organs will exhibit changes of which the nlagnitudes are most given by the type bearing of this on what we have termed random evolution will be obvious .
Turning to the second part of the error correlation surface , we note at once that , if two organs be correlated , selection will give system of devicptions in their and their correlation .
Random selection fortiori added , artificial natural selection ) , which alters an organ 's variability , alters the variability and correlation of all other .
In fact , when it is once realised how two random selections from a general population will as a ] organs of different means , of different variabilities , and of different correlations , the means among themselves and the variabilities and correlations among themselves forming groups , it becomes obvious how any assumption of the coefficient of correlation as a constant for local races runs wide of the mark ; and this , whether natural or random evolution , is to be ]ooked upon as the source of the observed differences in character .
What chiefly concerns the biologist in this matter at .
present is this , that even a random selection of one organ MATHEMATICAL CONTRIBUTIONS TO THE THEORY VOLUTION .
241 will produce changes in all other organic characters , which , if small , still sensible and capable of quantitative expression .
* ( 6 ) Returning now to the algebra of our investigation , we have to discuss the second part of ( xi .
) by aid of the formulae given in ( viii .
) , p. 237 .
We require , in the first place , to evaluate the determinant Now ' Divide the first row by , the second row by ) ' the third by , the first column by , and the second by .
Hence Subtract the second column from the first , and then add the first row to the second ; we have The minors are now easily found to be ' ' ' .
fake ( xvii .
) for example ; it for the lirst timo qualltitatively the ological principle that , if a bc selected at andom from the general population , and it has more variability in one aracter , it will be more variable thaJl the general population in other VOL. CXCI . .
2 I From these expressions the valnes of the standard deviations and error are at once found by ( viii .
) .
We ( .
) , , . . . . . . .
( xvii .
) , .
( xviii .
) .
The result ( xv .
) is , of course , old ; the resulcs ( xvi .
) to ( xviii .
) are novel , and lead to interesting conclusions , which are considered in the following paragraphs .
The error of a coefficient of colrelation is 67449 .
If , therefore , the correlation between two organs is less than once to twice they cannot be safely assumed to he correlated at all .
If we know definitely the errors made in and \mdash ; if , for example , we know the variations \mdash ; then the probable error of is that of an array of 's for definite and .
It is given at once by the coefficient of in as 67449 . . . . .
( xix .
) .
This is the value given for the probable error of by oue of the present authors in a former paper .
* At that time he had not fully realised the importance of the principle of the correlation of errors made in determining the magnitude of frequency constallts .
The following table will enable the reader to appreciate the difference in magnitude between the " " absolute\ldquo ; and " " partial\ldquo ; probable errors of : * ' ' Heredity , Regression , and Panmixia ' Phil. Traus , vol. 187 , p. 266 .
MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
It will thus be seen that when is small the absolute and partial probable errol.s are both large and nearly equal ; that when is large the absolute and partial probable errors differ more widely , but , as both are small in this case , their difference is not of much importance .
Bearing these facts in mind , it will be found that reasoning based on the partial error in previous memoirs remains valid , eVen if the partial error be replaced , as it generally should be , by somewhat larger absolute error .
* If we know definitely the variability of any orga and we take a definito group of the general population to find the variability of a second correlated organ , then there will be correlation between the deviation of the variation of this group with regard to t , he first organ from the variation of the first organ in the general population and the deviation of the variation of the group with regard to its second organ from that of the general population 's variation for the second organ .
This correlation is measured by , and thus , if the organs be slightly , it is small ; but if the organs be dosely correlated , it is large .
Suppose , for example , we know the variability of the tibia , and require to find that of the ulna from a comparative]y few specimens .
Let be the variability of the tibia in the specimens for which the ulna can be measured , and the correlation observed between the ulna and tibia in these specimens ; then , the variability of the ulna being observed as , the most probable variability of the ulna in the general population is or or , since the second term is small , we may write for , and the above expression equals .
For the long bones , roughly , and therefore we have the of bility o ' the ulna in the general population to the variability observed in the group It is clear that this expression .
also measures the change in the variability of the ulna due to a random selection of ( 8 ) Although the correlation between deviations in the variability of two organs their mean variabilities only varies as the square of their correlation , the correlation between the deviations in the variability of an organ and in its correlation with a second organ varies as the firsb power of the correlation of the two organs .
In other words , * See , for example , the l'easo1ling as to the non.constancy of the coefficient in ' Phil. Trans , vol. 187 , pp. 267 and 278 .
while a selection of variability may produce only a small or moderate on the variability of correlated organs , a selection of correlation or a ectio of variability is likely to produce considerable changes on variability and correlation respectively .
Let , be the mean values of the standard deviations and coefficient of correlation for any three organs ; let , be the like quantities for a group seIected at random .
the rinciple sion tells us that most probably Substituting the values given ( xv .
) to ( xviii .
) , we find .
Now these equations lead us to some important conclusions .
In the first place , if the correlation be very small or very large , then a random selection of variability ) makes only a small in correlation .
The change in correlation for a selection of variability is greatest when ) then is tely , or over 6 per cent if were as .
On the other hand , the change in variability due to a selection of correlation is small if the correlation be small , but increases rapidly if the correlation become nearly pe.rfect .
Of course , for perfect correlation the probable error of is zero , accordingly it is infinitely improbable that a selection can be made with differing from zero .
But if be not unity , then a selection in which is large , however improbable , will give very large changes in the variability , if be very large .
Our conclusion is accordingly that considerable changes in variability are likely to be produced whenever there is a correlation selection ainong hi ] organs .
( 7 ) To find the Probable Error of the Regression Coefficient for Two Organs .
The regression coefficient is given by Its standard deviation is given by the summation equation CONTRIBUTIONS TO THEORY OF EVOLUTION .
To find its value we adopt a method , which we give on this occasion at length , as it will be frequently used in the sequel .
' logarithmic differentials Square and divide by after summing Now , remembering the definitions of standard deviations and coefficients of correlation , this may be written Now all the quantities on the right have already been found in equations ( xv .
) to ( xviii .
) .
Hence , substituting , we have 1 [ ence ( xxi .
) .
the probable error of a regression coefficient This is of fundamental importance for testing the significance of results obtained by applying the theory of regression to problems in eredity , panmixia , \amp ; c. The probable percentage error in a regression coefficient ) , and hence is small if the correlation be close , and increases rapidly if the correlation be small .
This again illustrates the point to which reference has been made in another memoir , *namely , that when only a few individuals can be measured , the most reliable results -for the purposes of the quantitative theory .
of evolution are to be found from the measurements of the most highly correlated organs .
* PEARSON and LEE : ' Correlation in Civilised and Uncivilised Races ' Roy .
Soc. Proc vol. p. 345 .
Attention should be drawn to the fact that we have replaced errors by differentials .
This is only legitimate so long as product tertns in the errors are negligible pared with linear terms .
This is the assumption almost universaly made by writers on the theory of errors .
It will not lead us astray , so long as we take care in any practical applications to the smallness of , as compared with , and respectively .
8 ) To the Probable and Error Correlations of the Constants of a Normal Frequency Distribution for Three Organs or more .
It will scarcely have escaped the attentive reader that our investigation hitherto , only involving two organs , has left several important problems untouched .
For exampIe , it has dealt only wibh the direct effect of random selection .
But we may ask such a question as this : What is the change in the correlation of two organs when the variability of a third is randomly selected ?
Or again : What is the change in the correlation of two organs when the correlation between one of these and a third , or between a third and a fourth , is randomly selected All these are important problems in the theory of evolution .
The general equation to a normal frequency-surface for organs expt .
\mdash ; .
where is the determinant and , is the minor of the term in the row and column .
We require first to find the quantities like ( vii .
) of our Art .
( 4 ) .
, ( xxii .
) , * GAUSS , admittedly , 'Theoria Combinationis Observationum ' . . .
p. 53 , Problema ; LAPLACE , actuaJly but obsculely ; see 'Theorie analytique des Probabilites , , chap .
IV .
, and ' Recborches sir la Probabilit6 des Jugements , ' chap .
IV .
; more clearly in TODHUNTBR'S account , ' of beoJ y of obability , .
1,002 et seq. FurtlIer , , Article 'Probabiliby , ' S48 , for a like assumption .
MATHEMATICAL CONTRIBUTIONS TO TH EVOLUTION .
247 ( xxiii .
) .
Differentiating the first of these with regard to , and sumtning for all possible values of , we find But .
Hence ( xxiv .
) .
Differentiating ( xxii .
) with regard to and , we have at once . . . . .
Differentiating ( xxii .
) with regard and summing , we have ' or . . . . .
( xxvi .
) .
Differentiating ( xxii .
) with regard t , o and summing , we find .
or . . . . . . . .
( xxvii .
) .
Our next step is to differentiate ( xxiii .
) with regard to and sum .
We have . .
, Now , since , and therefore .
Substituting , we find .
But and hence .
Thus finally ( xxviii .
) , where , 12 is the second minor , found by out from the first and second rows and columns .
In the next place let us differentiate ( xxiii .
) with regard to and sum .
We in precisely similar manner , or , , or ( xxix .
) .
Similarly , we deduce .
( xxx .
) .
This completes all the possible types .
MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF We are now in a position to write down from ( vii .
) of Art .
( 4 ) the complete frequency surface for errors in the constants of a normal frequency surface for organs .
It will suffice to write a type of each term .
We have exponent\mdash ; . . . . .
( xxxi .
) .
Now this result again seems at once to give conclusions of considerable importance .
Thus : Since there are no terms of the type , we infer that the random selection of variation in one organ will most likely only vary the correlation between two other organs by terms of the second order ; and that ( ii .
) the random selection of correlation between two organs will in all probability only the variability of a third organ by terms of the second order .
The selection of correlation between any two organs ill most probable vary the correlation between a second pair , i.e. , terms exist in \amp ; c. ( ) The selection of the variation for any organ varies the correlation between that organ and a third organ , and vice ersd the selection of correlation between two organs changes the variability of both organs .
And lastly ( 8 ) The selection of correlation between two organs varies the correlation between eitl organ and any other organs .
We may exhibit these results more by taking four special organs , say , femur , tibia , humerus , and radius .
Then a group having the variability of its femur from that of the general population , will also have , in all probability , the variability in its tibia , humerus , and radius different ; the ations femur-tibia , femur-humelus , and femur-radius different ; but those of tibia-humerus , humerus- radius , and radius-tibia only slightly different .
Further , a group having the correlation of its femur-tibia different from that of the general population , will have all the other correlations , humerus-radius , femur-humerus , femur-radius , humerus , tibia-radius , different from the values for the general population .
Further , the variability in femur and tibia will be changed ; in all likelihood the variability in humerus and radius only slightly changed .
These general conclusions , which seem to considerable light on the manner in which selection influences the variability and correlation of organs , must now be reduced to quantitative expression .
VOL. OXCI.\mdash ; A. 2 MATHEMATICAL COh'TRIBUTIONS TO THE THEORY OF EVOLUTION .
251 Divide each row by , the first , second , third , and fourth rows by respectively , and the first , second , third , and fourth columns by respectively .
Multiply the fifth column by and subtract it from the first and second ; the sixth column by and subtract it from the first and third ; the seventh column by and subtract from the first and fourth ; the eighth column by ) and subtract it from the second and third ; the ninth column by and subtract it from the second and fourth ; the tenth column by and subtract it from the and fourth .
Remembeling that \amp ; c. , and that , , \amp ; c. , we find , writing .
for , \amp ; c. , for brevity , Here the signs of the terms in the last six rows have been changed from minus : plu Now multiply the first four rows by and add them to the fifth row , the first four rows by and add them to the sixth row , the fi ] st four rows by and add them to the seventh row , and so on .
Then , remembering that \amp ; c. , \amp ; c. , and , further , that does not contain a factor out of each row , The form of is now clear for the case of any number of correlated organs .
( 10 ) Case ( i ) .
Let us evaluate and its minors for the case of three correlated organs .
* The must be introduced if we deal wibh organs .
We may reduce this determinailt as follows to one of the 6th order .
vide the 1st , 2nd , 3rd , 5th , 6th 7th , , 9th , 10th columns by espectively .
After this division , subtract the first column from the , 6th , and .
The determinant reduces to one of the 9th ; subtract the first column of this new determinant from , 7th , and 8th UATHBNATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
258 In this case and we have whence we ' Divide the first three columns by 2 ; multiply the last row by and subtract from the first row ; the fifth row by and subtract from the second ; the fourth row by and subtl'act from the third ; we find column , it again reduoes by one degree .
Repeat the process twice more , and after a slight earrangement , the fact that being remembered , we have , if the general run of terms being ious .
In precisely.simila manner the value of for gans oan be written down , its degree being than the form given in .
We have not succeeded in reducing for the general case [ since writing this , Mr. ARTHUR BERRY , of College , Cambridge , has succeeded in reducing the delerminant for , and also in showing its relation to elliptic ] , bub we feel fairly confident it value will be found to be .
Add the first column multiplied by to the fourth ; add the first multiplied by to the fifth ; and subtract the first multiplied by from the sixth , the determinant will reduce to the minor of the first row and column .
Continuing this process twice more , we ultimately deduce ( xxxiv .
) .
We will now proceed to calculate such of the minors of as will give us results beyond those obtained for two correlated organs .
* We require the correlation of and , and of and Taking and we must strike out in ( xxxii .
) , as we have taken only three organs , the 4th , 7th , 9th , and 10th rows and columns straight off , and for the required minor the 1st row and the 8th column .
We have then for the minor , ' ' .
To reduce this expression take the third row multiplied by from the first , and the fifth multiplied by from the first .
rhen take the fourth row multiplied by from the second , and the fifth row multiplied by from the second , then that a matter of all the minors wel.e WOl.ked out and the resuIts of to ( xviii .
) thus ified .
MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF .
255 and that generally , .
We find : Multiply the first two columns by and subtract their sum from the fourth ; multiply by the second two columns by and subtract their sum from the fifth ; divide out by the factor .
We obtain Now remembering that substitute the various telms , and we find : 256 .
PROFESSOR K. PEARSON AND .
L. N. FILON Subtract times the fifth from the second , and times the fourth from the third column : Add times the second column to the fifth and times the third the fourth , we have ' 1 ( xxxv .
) .
Now we have seen that by ( viii .
) of Art .
( 4 ) , and further and are given by ( xv .
) and ( xvi .
) of Art .
( 6 ) .
Hence ( xxxvi .
) .
In the next place we will determine .
The minor is given by NATHENATICAL CONTRIBUTIONS TO THE OF EVOLOTION .
257 , , ' , , ' ' Add the first three rows together , multiply by and subtract from the fourth , and by and subtract from the fifth , we find rl Multiply the fourth column by , and the fifth by , and subtract the third column .
, 1 Multiply the third row , and the fourth by or , and subtract the latter ; the determinant now reduces to one of the fourth order , and we find : Add the third column , multiplied by to , and subtract the fourth tiplied by from , the second , and then subtract the first Divide out the 2 in the second column , add it to the first , and subtract the third column multiplied by Add times the last row to the first Subtract times the first row from the last , and remember that , \mdash ; 1 , , ES , Multiply the first row by and add it to the second ; the determinant reduces to the third order , NATHEMATICAL CONTRIBUTIONS TO THEORY OF EVOLUTION .
259 Expanding this determinant from its first column , remembering that R Or , since , and Hence , since we have by ( xvi .
) of Art .
( 6 ) and ( xxxiv .
) of Art .
( 9 ) ( xxxvii .
) .
To complete the theory for the errors made in an investigation of the constants for a system of three correlated organs , we require to determine the probable error of a regression coefficient for a partial regression of a first organ on a second , the third organ being constant .
This coefficient is given by Take logarithmic differentials Let this be squared and divided by , and the values found above for the standar deviations of the errors in and , and for the correlations of errors in these quantities be substituted .
After some lengthy algebraic reductions , which it seems unnecessary to reproduce , there results or .
( xxxviii .
) .
The percentage probable error in a partial coefficient of regression is accordingly 67 . .
Before discussing the significance of these quantitative results for three it seems ( iesirable to complete the general case by investigating the con.elation between the errors made in the correlation coefficients of a first pair of organs and a second different pair of organs .
( 11 .
) Case ( ii .
) .
Case of Four or more Organs.\mdash ; In the case of four or more organs the only new error will be that of a partial regression coefficient , but this can theoretically always be found by the method of the preceding paragraph , provided we know all the error correlations .
The only novel correlation among the errors will be that of and this we shall now proceed to investigate .
The qcovery of an error correlation coefficient of this type completes the theory of the errors of normal frequency constants .
Instead of evaluating of ( xxxiii .
) , which in the case of four organs appears to be very laborious , we may proceed as follows:\mdash ; If be written in the form and denote , minor of the corresponding in , we have by a well known property of the determinant , . .
Divide by MATHEMATICAL IONS TO THE OF EVOLUTION .
261 Now , . .
, are the correlations between the errors made in the various quantities , every one of which is known by the previous investigations except that of and or Hence the above equation will suffice to find the latter quantity , since and are known .
We have by ( xxiv .
) and ( xxv .
) , , , by ( xxvii .
) , , by ( xxvi .
) .
Further , , by ( xvi .
) and ( xviii .
) , by ( xxxvi .
) , by ( xxxvii .
) , , by ( xvi .
) .
Now substitute and divide out by common factors , and we find ( xxx ix Now Multiply the third of these by and the second by and add , we have Hence , by the first , , while By means of these relations , let us get rid of the terms in and in ( xxxix .
) above .
-arranging we have , after some reductions , But Hence we can divide out by accordingly , Noting that , we have , after substitution and rearranging , Or , ( x1 .
) .
MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
263 If we put in this result and remember that , we find , after some reductions , which agrees with ( xxxvii .
) , and may be taken as a verification of this result .
* ( 12 .
) We may draw several conclusions from the results ( xxxvi .
) , ( xxxvii .
) , and ( x1 .
) .
While errors in the correlations of a first organ with a second and a third have a correlation themselves of the first order , errors in the variation of a first organ and the correlation of two others , or in the correlation of two organs and in the correlation of a second two , have only correlation of the second order .
Thus a selection of the correlation between two organs modifies the variation of all organs correlated with one or other or both of the first , but only in the second degree .
Again , a selection of the correlation between two organs modifies the correlation of every other pair of organs , one or both of which are correlated with one or both of the first pair ; but this is only in the second degree .
If two organs be entirely uncorrelated a random selection of the variation of a third organ correlated with both of them will tend to generate correlation between the hitherto uncorrelated organs , i.e. , put in ( xxxvi .
) , and we have If a variation be made in the probable value of .
is which may clearly be of sensible magnitude .
Thus correlation may be generated by selection of variation , and vice versed .
If two organs be each entirely uncorrelated with a third , yet a random selection , which produces a correlation between one of these organs and the third , will produce a correlation of the first order between the other of these organs and the third , i.e. , put in ( xxxvu .
) , we have a correlation of the first order between the probable changes .
( 8 .
) Consider four organs of which the first is alone correlated with the third and the second with the fourth , the third and fourth being themselves uncorrelated .
Then any random selection which produces a correlation between the first and * The probable error of a partial regression coefficient for organs has not been worked owing to the labou1 .
involved , but judgiug by the cases on pp. 245 and 260 , it may safely be taken as , where is now the determinant of the degree .
second will tend to produce a correlation between the third and fourth , , if , we still have from ( x1 .
) We may further illustrate these principles by one or two hypothetical examples drawn from actual organs .
Let the actual organs be ( 1 ) ique of father , ( 2 ) artistic sense of mother , ( 3 ) physique of offspring , ( 4 ) artistic sense of offspring .
Suppose in the general population there is no correlation between physique of father and artistic sense of mother , or between physique or artistic sense of parent , and artistic sense and physique respectively of offsprin .
Then , and , presumably , .
Hence is the product of the two coefficients for inheritance of physique from father to child , and for inheritance of artistic sense from mother to child .
Now let a random selection be made out of the general population in which assortative mating between physique in the male and artistic sense in the female presents itself , i.e. , let be sensible ; then we have , most probably , or , a correlation between physique and artistic sense in the offspring will tend to be developed .
Generally , when and do not start from zero , we have , or , any increase of sexual selection in a group tends to empbasise the correlation of the selected qualities in the offspring .
Let the three characteristics be artistic sense ( 1 ) in a man , ) in his mother , ( 3 ) in his wife .
Then if we suppose to be zero .
Hence any selected group with a higher coefficient of maternal inheritance of heredity will have a less coefficient of sexual selection than the general population , and vice versed .
The tendency is , of course , independent of the magnitude of , and really of the particular character .
Supposing likeness of faculty or character to be a rough measure of ' ' sympathy we might conclude any population with inheritance and sexual selection , that on the average a selected sub-group of men having greater sympathy with mothers than the general population will have less sympathy with wives , and vice versed .
MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
265 ( 13 .
) like propositions may be stated with regard to the action of selection on the correlation of characters .
They require but little modification to sbate them for ficial or natural selection , as they are here stated for what have termed random selection .
The above will , however , suffice to indicate how every form of selection of variability or correlation influences in a manner capable of quantitative expression the variability and correlation of all other directly and indirectly correlated organs .
Selection cannot be of service in altering one organ only , it alters at the same time the whole inter-relationship of a complex of organs .
Evolution by natural selection can never be the change of one organ to suit a environ- meant ; it is the balance of advantage and disadvantage produced by the change of all organs involved in the attempt to select one of them .
The moment the intimate correlation of organs in animal or plant life has been fully realised\mdash ; and this realisation owing to recent statistical investigations has become fairly \mdash ; then the conception of natural selection as mouldin any single organ to what may be fittest to its surroundings must be discarded .
The selection of the " " fittest\ldquo ; in one organ would probably mean the selection of the unfit in other organs , and a general balance of fitness in the complex of organ is all that is possible .
* IV .
ON THE PROBABLE ERRORS AND THE OF CORRELATION BETWEEN ERRORS MADE IN THE DETERMINATICN OF THE CONSTANTS IN THE CASE OF SKEW VARIATION .
( 14 .
) The case of Skew Variation has been dealt with at length by one of the present authors in the second paper of this series .
He has shown that in a variety of cases it can be dealt with by a series of curves having three principal algebraical types , each defined by a certain number of constants .
The probable errors of the determination of these constants were not then investigated , but it is dearly of great importance for the practical use of these curves to know how far these constants can , for any given number of observations , be depended upon to give an accurate measure of the skewness and its special features .
At the same time au investigation of the probable errors of these constants leads us to a number of novel properties which are connected with the theory of evolution in the frequent case of skew variation .
* Take , for example , result ( xxxvi .
) ; as faras terms of the second order are concelned Hence , with positive correlation between three organs , the sffect of trying bo get a group stable in oue organ , i.e. , with a negative , is to reduce the correlation between every other pair of organs !
other words , we have to reduce variation at the expense of correlation , ased sbilit of one organ is guined at the expense of decreased stability in the elationship of her organs .
may possibly be illustrated by the long bones of the Freuch , where the lesser of the male relative to the female connotes also a lesser correlation .
See ] and PEARSON : " " On the Relative Variation and Correlation in Oivilised and Uncivilised Races ' R. S. Proc vol. 61 , pp. VOL. CXCI.\mdash ; A. 2 We shall deal first with the skew curve of Type .
( ' Phil. Trams .
' p. 373 ) , because its treatment is less ] and teads at once some principles which must be borne in mind , whenever natural selection acts upon an organ exhibiting skew variation .
( 15 .
) Probable Errors and Error Correlations of the Constanls of the Generalized Curve of Type This is the equation of the curve referred to its mean as origin , where ( xli .
) .
Further , the moments about the centroid vertical are given by , ( xlii .
) .
or , . . . .
( xliii .
) .
The criterion for the application of this curve to any frequency distribution is or , if we write . . . .
( xliv .
) .
Lastly , skewness ( xlv .
) , and the modal frequency , xlvi .
We require to know the probable errors of , and the skewness .
We must discover the best physical constants to describe such skew frequencies and we shall at the same time succeed in deducing certain\mdash ; we believe\mdash ; novel properties of.normal frequency distributions as limiting cases of this skew type of distribution .
* See ' Phil. Trans , vol. 186 , pp. MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
267 ( 16 .
) The first sGage in the investigation is to apply the general proposition of our Art .
2 , to We find : , if , , .
Let , then we easily find I and By aid of these we can at once write down the integrals of the above expressions multiplied by , since .
We find with the notation of p. 243 , , Before we consider the determinant and its minors , we may note that where the 's are the BERNOULLI numbers .
Hence and we have the convenient form , where is the semi-convergent series , \amp ; c. Now we have at once , ' ' ' ' ' .
Divide all three columns by ; the first row and first column by 7 , and then the last row and column by ; we find , .
Divide the second row and column by ; add half the last row to the and we find , 1 , 1 , \mdash ; 2 MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
269 Subtract the first column from the second , and add the double of it to the last , we deduce or , ( xlvii .
) .
The minors can now be written down at once .
whence or ( xlviii .
) .
whence or ( xlix .
) .
, whence , or ( 1 .
) .
whence .
whence .
Lastly , whence ( liii .
) .
This completes the direct series of probable errors and error correlations .
By aid of the above correlations and standard deviations we can now find a further series .
From ( xlii .
) we have for the standard deviation ( about the mean ) , or .
Hence ( liv .
) .
Square both sides of this , divide by and sum , we have at once from the definition of a coefficient of correlation Hence , using ( xlix .
) , ( 1 .
) , and , we find , after reductions , .
Multiply ( liv .
) by , sum and divide by , we have Whence , by ( lii .
) , or , reducing by ( 1 .
) , ( liii .
) , and , .
Next , if be the skewness , we have from ( xlv .
) MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
271 or lvii .
Similarly , since or ( lviii .
) .
We easily obtain , by multiplying ( liv .
) by , the result ( lix .
) .
We can now obtain , for every is negatively proportional to the corresponding .
Hence . . . . .
We next pass to the mean and modal frequencies as given by ( xli .
) and ( xlvi .
) .
We have , by taking logarithmic differentials , where .
* If \mdash ; \amp ; c. , then it is easy to show that ' S. Remembering that , , we easily deduce . .
as the power of the skewness inclusive .
Very generally the probable error of the skewness may be taken as equal to and it is always less than , its value in the case of a normal frequency .
Squaring , introducing the standard deviations , and rearranging , we find We must now evaluate .
This is easily shown from the BERNOULLI number expansion for to be given by where Thus we determine ( lxi .
) .
Expanding the expression in brackets in inverse powers of we find ( lxii , ) .
Result ( lxi .
) , however , with and calculated to , gives a better value than ( lxii .
) .
To find the modal frequency error we must take the logarithmic differential of ( xlvi .
) and proceed in the same way .
We find almost at once .
Whence on squaring and completing the square of the factor of , we find and ( lxffi .
Expanding as far as powers of exclusive we obtain , a very simple expression for the probable error of the modal frequency , MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF VOLUTION .
273 In like manner we shall now determine the probable errors of the moments and their error correlations .
Take the logarithmic differentials of ( lxii .
) Squaring each of these in succession and using the known values of we find ( lxv .
) , , , Now multiply by and we find , after some reductions , .
viii . .
Next multiply by , and we ultimately have .
Lastly , multiplying by , we deduce , after some reductions , . .
VOL. CXCI.\mdash ; A. We may add to these.results the values of and and \amp ; are given by ( xliv .
) ; we find .
( lxxi .
) .
The distances from the mode to the mean , , and from the mean to the end of the range , , are given by and .
Hence . . .
( lxxii .
) , , and further ( ]xxiv .
) .
results ( xlvii .
) to ( lxxiv .
) must be now considered at length .
( 17 .
) The frequency curve of the type considered is fully described by the three con stants , the mean , , and .
But , since three constants would do equally well\mdash ; for example , what may be termed the three physical constants : mean , standard deviation ( variation ) , and skewness\mdash ; it becomes of some importance to inquire which constants have the least percentage of probable error .
Now ( xlviii .
) shows us that the probable error in the mean is precisely the same as in the case of the normal curve and Thus , the percentage error in the mean coefficient of variation , and will certainly be small whenever the coefficient of variation is small. .
Its value is quite independent of the order of MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
275 On the other hand , the probable errors of and are from and respectively .
Here is equal to the series which tends to zero as incoeases . .
The errors in and thus tend to increase indefinitely as increases .
It may then be asked how the form of the curve can be determined with any degree of accuracy .
: The answer is simple : Equation shows us that the correlation between errors and tends , as increases , to become " " perfect i.e. , unity .
But as increases indefinitely , it has been shown that the frequency curve of this type passes over into the normal form.$ It is the high correlation between errors in and which renders the curve , when plotted to observations , such an excellent fih If the errors in and 7 were independent , this would not be so .
At the same time it renders and unsuitable for tabulation as physical or biological constants of the frequency .
Turning to and ( lvii .
) we see that the standard deviation , , and the skewness , Sk .
, are suitable constants for tabulation .
Their probable errors do not tend to increase indefinitely with , and will always be small , if be large .
Hence a frequency distribution of this type is best defined by its mean , its standard-deviation , and its skewness Sk .
These are constants characteristic of the group , for they are given with small probable errors .
If it be desired to draw the form of the frequency-curve , then its algebraic constants , and , may be found from and the possibly considerable errors in and will not vary largely its actual shape .
The nature of the probable errors of the other allied constants may now be considered .
The mean and modal frequencies per unit variation of organ , or and are seen by ( lxu .
) and ( lxiv .
) to have small percentage probable errors , and are , therefore , good for use as characteristic physical or biological constants .
But it should be noted that the modal frequency is considerably more exact for moderate values of than the mean frequency .
For example it would be somewhat better to tabulate the modal than the mean frequency of the barometer as a physical characteristic of climate .
The probable errors of the distances from the mean to the mode and from the mean to the terminal of the range are given by ( lxxii .
) and .
( lxxiii .
) .
Since , we may write the first .
* Phil. , vol. 186 , p. 374 .
PROFESSOR K. PEARSON AND N. G. FILON This remains finite , sven if be indefinitely great .
On the other hand , the probable error of , and even its percentage probable error , becomes indefinitely great with .
It is to be noted that in this case becomes infinite .
Results ( lxv .
) to ( lxvii .
) give the probable errors of the second , third , and fourth moments .
It will be seen that , for a large , the percentage error of the fourth moment is about double that of the second .
It might thus appear , at first sight , safer to work with the second than with the fourth , but is by no means necessarily the case , for to deduce any quantity from one or the other they must be reduced to the same order .
For example , the square root of must be compared with the fourth root of , and the probable errors of and will be sensibly of the same order .
Remembering that , we may write This tends to a finite limit as increases indefinitely , and we conclude that the probable error of is always finite , and will in general be a small fraction of the cube of the standard deviation .
The above remarks are ajustification for the use of higher moments in frequency calculations .
Equations ( lxviii .
) to ( lxx .
) give the error-correlations between the first three moments .
They show that an error in the value of one of these moments will most probably lead to an error in the other two .
We see that for fairly large is a large ation , while and .
are small .
In other words a random selection of an even moment makes a far larger correlated change in another even moment than in an odd moment .
If increase indefinitely we find the ratio approaches the value 2/ 3 ; in other words , is more closely correlated to the.higher moment than to the lower moment Formulae ( lxxi .
) give the probable errors of the useful constants and .
We see that they are small and approach the value zero as is indefinitely increased .
S. Let us restate the formulae for indefinitely great , i.e. , for the normal curve of frequency In this case we have , skewness mean and mode coincide .
Several of these zero quantities , however , tend to have definite probable errors .
We have MATHEMATICAL CONTRIBUTIONS TO THE THEORY OP EVOLUTION .
, first , second , fourth and fifth of these results are old ; the roast appear to be novel and of some importance .
In the first place we notice that given a population which is really normal , we should not expect a random selection to exhibit all the signs of normality .
Its skewness will differ from zero with a probable error .
For example , in a random selection of 600 from a normal population , the skewness will be as likely to exceed as to fall short .
Hence an exhibition of skewness of less than once to twice67449 must not in itself merely be taken to indicate an absence of normality in a general population .
Again , in a random selection from a general population , the mode will differ the mean , even if the population be normal , with a probable error of .
Thus , in a population of 600 , a difference between the mean and the mode of should not be taken to indicate want of normality .
Generally , the divergence between mean and mode in a population must at least exceed once to twice , for us to be able to on this ground alone that the population has not a normal distrib tion .
Again , the third moment not being zero , but having a value of once or twice .
, is not in itself an argument for skew frequency .
The above statements are an important addition to the second memoir of this series ; they give us the criterion , there wanting , to distinguish between a skewness which is characteristic of a population and one which might arise by the random selection of a population of the given size out of a larger , but really normal , population .
We may now note the exceedingly interesting conclusions which these results have for the theory of evolution .
Suppose an to have , as so many do , skew variation , then we notice PROFESSOR K. PEARSON AND MR. L. N. G. FItON Any selection of the organ by size tends to alter its variability , but not its skewness ; this follows from and ( lviii .
) .
Further , if , as have supposed , the range be limited on the side of dwarf organs , then any increase of size means a decrease of variability , and vice ( ii .
) Any selection of val.iability is a selection of skewness ; this follows from .
If a selection be made from a general population , which has less variability , then it will tend to greater normality .
In other words , it would appear that stringent selection tends to generate normal distribution .
Thus , if out of a skewly distributed population we make a number of random selections , that with the least variability will be most normal .
Select at random again out of this latter selection , and the least variable group will again be the most normal , and so on .
Now take a problem of this kind involving group , and not individual , selection .
Let a large general break itself up at random into groups , and let us suppose these groups , not individuals among them , to carry on a struggle for existence\mdash ; an inter-group , not an intra-group , struggle .
Then , if it be an advantage to a group that its members shall be among themselves close to a type , i.e. , less variable , then the more normal groups will survive , for variability is positively correlated with skewness .
Now suppose each group to be periodically subdivided at random into new groups\mdash ; the mathematical description of some process of group reproduction\mdash ; then we see how normal distribution may be a result of a stringent inter-group selection of groups whose individuals have the closest resemblance to each other\mdash ; intra-group resemblance .
( iii .
) Any selection of the size of an organ produces by ( lxxiv .
) an alteration in the distances between the mean and the mode , and between the mean and the end of the range .
A random selection which has its mean larger than that of the general population , will , if the mode be on the dwarf side of the mean , tend to have its mode and mean nearer ether than are the mode and mean of the general population , while on the other hand , to raise the mean is to raise the dwarflimit to the range .
A considerable number of like results might be stated , but the above will be sufficient to emphasi the general principle that a random and a fortiori an artificial selection of the size of an organ , does , whenever its distribution is skew , influence in a definite manner the variability of the organ .
It is quite safe to assert that it will also influence the correlation of organs .
When we notice how wide-spread is skew variation in nature , we may assert that the general rule is that no modifica- tion can be made in any of the features\mdash ; mean sizes , variabilities and correlations\mdash ; of a group of organs without at the same time modifying all the others .
* * A paper has recently been published by Messrs. DAVENPORT and BULLARD in the 'Proceedings of the American Academy of Science ' ( see Illustration II .
below ) on ' The Variation and Correlation of the Glands in the Legs of Swine Unfortunately the authors have overlooked the markedly skew chragter MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF VOLUTION .
279 As a result of Articles ( 15 ) and ( 16 ) , it is possible to use the frequency curve of type with as much certainty as to the nature and magnitude of the errors made in the constants as has hitherto been possible in the case of the normal distribution .
The method has been exemplified numerically in twenty-three cases in a memoir on the " " Variation of Barometric Frequency\ldquo ; ( see ' Phil. Trans , vol. 190 , p. 423 ) .
It may not , however , be amiss to illus rate it further in a special case having closer bearings on the theory of evolution .
( 18 .
) Numerical Jllustration\mdash ; Incidence of Enteric Fever .
In a memoir in the ' Phil. Trans , vol. 186 , p. it is shown that the curve closely represents the distribution with age of 8,689 cases of enteric fever received into the Metropolitan Asylums Board Fever tals .
The unit of is five years , the origin is the mode at years .
The criterion is not very nearly zero , although small , but the curve is graphically a good fit ( see Plate 12 , fig. 9 ) .
The following are the numerical values of all the constants:\mdash ; Mean years .
mean-mode .
Mode ] years .
unit .
Sk skewness 462,594 .
modal frequency mean frequency standard deviation units years .
From these the numerical values of the probable el.rors and of the correlations between the errors of the constallts were found by the processes indicated and the formulae given above .
We found whence o distribution .
It is , howevel , clear their tables and pluto that no tiuu oould made of the absolute number of glands without altering the riability of the gland distribution and the correlation between different systems of glands .
Probable Error .
Percentage Probable Error .
1.7854 Correlation of errors in and 9581 9 .
1.3951 mean year .
mode year .
These are the constants which determine the position and algebraical equation to the frequency curve , and we see at once that they are all determined with a close degree of accuracy .
The largest percentage probable error is in , but this is under per cent. , and , owing to the high correlation between and much error would produce no sensib ] change in the shape of the curve .
Two tant facts may also be drawn from these results , which indeed follow from the general formulae , namely : The position of the mean is sensibly more exactly than the position of the mode .
Here about times as ( ii .
) rrhe modal frequency , on the other hand , is sensibly more accurate than the mean frequency .
Here about times as accurate .
Hence the advantage of using the mean as origin of rement for the curve is accompanied by the counterbalancing , and here relatively greater , disadvantage of the increased inaccuracy of determination of the mean frequency .
Passing to the " " physical\ldquo ; constants of the curve , we have Probable Error .
Percentage Probable ]Brror .
Sk .
022845 These fully determine the non-symmetrical nature and spread of the curve , and since the errors in the skewness and in the distance between the mean and mode are less than and cent. of the respective values of these quantities , we conclude that skewness and divergence between mode and mean are characteristic features of enteric fever distribution , and not mere anomalies due to a random selection of cases .
They are significant constants peculiar to each type of fever distribution and no description of such a distribution is sufficient unless their values are stated .
Before giving a table of the correlations between what we have termed the " " physical\ldquo ; constants , it may be well to write down some of the correlations between the errors in the physical and algebraical constants , which arise in the course of their calculation .
We find MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
281 By aid of these we find the following table of error correlations:\mdash ; Now this table enables us to draw some remarkable conclusions with regard to enteric fever .
We see at once that no random selection of a group of individuals , which has any single characteristic differing from that of the general population will , except in the case of mean of incidence and skewness , leave the other characteristics unmodified .
Thus the most probable result of any selection which alters the nature of the distribution of enteric fever can be predicted .
The reader will possibly appreciate this better , if we replace the above table by another giving the ressions in years , number of cases per thousand , \amp ; c. PROGRESSION TABLE .
* The frequency oi incidence in the modal year since the unit is five years To make this 1000 we must multiply by error in incidence of mod year .
Thus we have to replace by ( orror per thousand iu modal year of incidence ) .
VOL , We see at once from the above table that if the mean age of incidence of enteric fever in any group were raised , the disease would be concentrated in fewer years , the modal mean incidence would be brought closer together , and the incidence in the modal year of frequency would be heavier .
The changes here are very sensible .
Thus , if we raised the mean age of attack to that of phthisis , or about nine years , the moda ] frequency would be about 41 per cent. , the concentration of the incidence of the fever increased about 40 per cent. , while the distance between mode and mean would be reduced to nearly 2/ 5 of its original value .
The skewness would not be changed .
Much less ed effects would arise from a selection of modal frequency .
Any increase of modal frequency tends to slightly raise the mean age of attack , to increase slightly the concentration , to draw the mode towards the mean and reduce the skewness .
The changes produced by closer concentration of the attacks of the disease , i.e. , the nitation of its incidence to fewer years , wouId be of a more marked character , they would raise the mean age of attack and the modal frequency , they would ecrease the interval between mode and mean , and reduce the skewness .
ration of the disease would thus tend to render its distribution more normal .
To increase the interval between mean and mode lowers the mean age of attack , reduces the modal frequency , increases the period of liability to incidence , and much increases the skewness .
Finally , increase of skewness decreases the modal frequency , increases the period of liability and the interval between mean and mode .
These statements with regard to the manner in which enteric fever would affect different groups selected at random from the general population seem of considerable interest , for there is reason to believe that what is thus stated for enteric fever in different groups may be applied to different fevers in one and the same group .
For example , the lower the mean age of attack of any fever , the greater its concentration ; the less the concentration , the more nearly normal is its distribution , \amp ; c. , \amp ; c. ( 19 .
) Probable Errors of the Constants of the Generalised Probability Curve of the Type \mdash ; Transfer the origin to one end of the ran , and the equation to the curve becomes ( lxxv .
) , where is the number of observations is the range .
The following values are given in 'Phil .
Trans , 186 , where : MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF .
283 Mean ( lxxvi .
) , Mode ( lxxvii .
) , , , Skewness ( lxxx .
) .
For the moments we have , , . .
Lastly , for the mean and modal frequencies of Sx and Sx , we have expt .
( lxxxiv .
) , expt .
\mdash ; @ ( lxxxv .
) .
where \mdash ; \mdash ; etc. .
( lxxxvi .
) .
Let I where then I , and we easily find , by the fundamental property of funcbions , that I I ( lxxxvii .
) .
I Fron ] we have ( Jxxxviii .
) , where will be needful to find , and ( lxxxix .
) .
Similarly . . .
( xc .
) , where ( xci ) , ( xcii .
) .
, and can now be readily expressed in semi-convergent series admitting of calculation .
( xciii .
) , where and .
( xciv .
) .
MATHEMATICAL CONTRIBUTIONS TO THEORY OF EVOLUTION .
285 It is clear that if and .
be at all large , which they frequently will be , we may omit the series , or even reduce to , and , respectively .
Making use of ( lxxxvii .
) and ( lxxxviii .
) easily find .
( xcvi .
) . . .
( xcvii .
) .
( xcviii .
) .
( xcix .
) . .
( ci .
) . . . .
( ciii .
) . .
The next stage is to calculate the determinant and the minors \amp ; c. , of , \amp ; c. We shall then determine , PROFESSOR K. ARSON AND MR. L. N. G. FILON and correlations .
The algebraic expressions .
the expanded determinant and its minors are very lengthy , and it will be found easiest in any numerical case to calculate the numerical values of the .
, and then find the values of the and its minors numerically .
So soon as the above four standard deviations and six error correlations been calculated , the determination of the probable errors in the constants of the frequency distribution becomes easy .
We have for the mean organ , if now defines the origin of coordinates , , ( cvii .
) .
Whence \mdash ; .
Similarly , the modal value of the organ can be found ( cix .
) .
* The easiest numerical method in this , as in the previous case , is to proceed as follows :Write where the , are umbral symbols , and let be the numerical values of the coefflcients on the right .
Then put Now square this equation , and , whenevsr a product , , multiply it by the corresponding error correlation , , already calculated , putting if .
Then , actually , BARLOW'S Tables rapidly give the squares and CRELLE'S Tables , or a Brunsviga , the products .
MATHEMATICAL TO THE THEORY OF EVOLUTION .
287 In like manner from ( lxxviii .
) say , where the numerical values of and can easily be found in any actnnl case .
Hence .
Again , from ( lxxix .
) .
Or .
( cxi .
) .
Further from ) , say .
Hence ( cxii .
) .
From the results given above we can deduce the effects on size , ran ge , variability , or skewness of a selection at random of any one of these four .
Writing ( cvii .
) we find ( cxiii .
) .
( cxiv .
) . .
These results show us that it is , in the general of skew , impossible to select one of the quantities\mdash ; mean size , range , variability , or skewness of an organ , without at the same time in all probability modifying all the others .
For example , the frequency of the incidence of certain types of diseases at different ages follows a distribution of this character .
Hence , if any special class of the community had a mean age of incidence differing from that of the general population , we should expect correlated changes in such other characteristics of the disease as its first appearance ; ( ii .
) its appearance ; ( iii , ) its tendence to heavier incidence above or below the mean age of incidence ; ( iv .
) the concentration of its incidence about the mean age of incidence for this selected class .
Precisely similar series of changes would arise in bhe case of a random selection of individuals having the variability of a certain organ greater or less than that of the general population , there would be correlated changes in the size , range , and skewness of the distribution of this organ .
Turning to the mean and modal frequencies , we have , say . . . . .
( cxvi .
) .
Similarly , , say . . . .
( cxvii .
) .
Here can be easily calculated , if we note that . . .
( cxviii .
) .
, where is the same as on p. 272 .
MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
289 and can be foumd in the usual manner by squaring after the insertion of the numerical values .
From ( lxxxi.)-(lxxiii .
) the probable errors and correlations of the moments can be found if required , the calculation being numerically somewhat laborious , but presenting nothing of novelty .
The probable error of the criterion where and , may be found as follows : Put , ; then we find ( cxix .
) , and accordingly . . . .
( cxx .
) , where and admit of easy calculation .
Hence , . . . . . .
The value of can thus be found , and the steadiness of the curve to its type ascertained .
Illustra from.\mdash ; Glands of the Fore-legs of Swine .
In the 'Proceedings of the American Academy of Arts and Sciences , ' vol. p. 87 , 1896 , is a memoir by C. B. DAVENPORT and C. BULLARD , on the variation in number of the Mullerian glands in the fore-legs of 4,000 swine .
The paper especially attracted our attention , because the authors are content to describe the frequency distribution of these glands by means of a normal curve .
They write , after discussing the ling of the normal curve on their diagram ( pp. 90-91):\mdash ; These and other characters of the ' probability ' curve are indica ed in that shown in dotted in the accompanying diagram .
The diagram also shows curve of * The authors a esent the ourve by -sided polygon .
VOL. CXCI.\mdash ; A. 2 distribution of the various numbers of glands occurring on a leg from 1 to 10* .
This curve is drawn from the right female leg only ; the curve for the other legs would be very similar .
We shall speak in a moment of the method of construction of these curves ; but we want now to caII attention to the fairly close similarity of the two curves\mdash ; that gained by observation and the theoretical one\mdash ; a similanty so that we are justified in concluding that the law of distribution of the variants in the leg glands of swine is the same as that of accidental errors Now , in our opinion , the curve was markedly skew , and it seemed to us that most interesting properties bearing on the action of selection on the Mulerian glands in swine actually depended on this skewness .
We have taken the distribution of for 2,000 9 swine .
To illustrate the difficulty of applying the normal curve we may l.emark that it gives about 6 swine per mill with gland , and about 15 with glands , while authols have forgotteh that thele a sensible perc.entage of Ztro.glands .
MATHEiMATICAL CONTRIBUTIONS TO THE THEORy OF EVOLUTION .
it gives 30 per mill instead of 10 per mill with no glands .
These difficulties are entirely met by the skew curve , which gives no frequency whatever of negative glands ( see figure ) .
Taking the number of glands in the right fore-leg of female swine , we have the frequency series : We have worked with the frequency per mill for convenience of reduction , although the actual number of observed cases , 2,000 , is used , of course , in the determination of the probable errors .
Using the method of the paper in the ' Phil. Trans A. , vol. 186 , p. 367 , we found mean glands .
Thus the criterion is greater than zero , or the frequency is of Type I. , or has a limited range .
Proceeding we found 3.79623 Mode 14.24837 Sk .
Start of curve gland .
Thus it would appear that both the distance ( d ) from mean to mode and the skewness are very sensible , and that , unless their probable be very large , it is quite impossible to represent the results by a normal curve .
We may note that the range starts from gland and runs to glands , so since it gives zero at \mdash ; 1 gland , we see that it sensibly confines the possible number of glands between and 17 , but we should have to examine considerably more than 2,000 swine to have a probability of more than 10 glands occurring .
The total range given is thus both in magnitude and position extremely satisfactory ; supposing the frequency , not the actually measured quantity , i.e. ; number of glands , to be known , the theory would have given a very accurate determination of the limits of possibility , especially the start of possibility with the whole number of glands .
In order to work out easily the determinant , , and its minors , found it desirable to bring out certain factors and reduce the formulae given above slightl different forms , which , as they are likely to be of general service , are here repeated .
Let , where have the values given on p. 284 , then we found , 1 , , , , 2 \mdash ; 1 , , 1 , MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
293 , 1 , , , In our particular case we found With the aid of the values for and given above , the determinantal parts of the 's were then calculated .
If these be 8 , we have @ 153,7969 , , , , , , From these the standard deviations and correlations of errors in the algebraical constants are at once found .
We have , Then as a step towards the determination of other probable errors , the standard deviation and umbral equation*for were found Antl .
Antl .
This led to By aid of these auxiliary results the probable errors of all the algebraical and " " physical\ldquo ; constants were determined .
PROBABLE Error Table .
Now it will be clear from an examination of these results that all the " " physical\ldquo ; constants are determined with great accuracy .
The mean is subject to less probable error than the mode , the modal frequency has a slightly less probable error than the mean , and as it is less than per cent. in the former case , either are closely known .
The skewness and distance from mean to mode are known respectively with less than and with probable errors .
Thus they are both significant constants .
In other words , the curve differs significantly from a normal curve , and it is erroneous to represent the frequency by such a normal curve .
The range which ought to be such that there is no frequency at gland , gives no frequency at * See footnote , p. 286 , and Iater , p. 305 .
It may be as well to remind the reader that , as in the other illustrations , logarithms of the full , not the oited values , were nsed in the calculations .
The percentage are high , but we have several times pointed out , is of small importance , as , owing to their high correlalion , the actual shape of the curve is not changed sensibly by large changes in and NATHENATICAL CONTRIBUTIONS TO THE THEORY EVOLUTION .
295 818 gland with a probable error of .
It is , therefore , clear that our method gives the start of the range with very considerable accuracy .
The whole length of the range runs to glands , with a probable error of .
We may , accordingly , conclude that the maximum possible number of glands is hardly likely to be less than 16 or more than 20 .
We consider that this example is a good illustration of the accuracy with which the principal " " physical\ldquo ; characteristics of a distribution may be obtained by aid of skew curves , and how they provide much information which is not given by the use of the normal curve .
The next ) oint is the determination of the umbral equations the error correlations of the\ldquo ; physical\ldquo ; constants .
They , if Antl .
stands for antilogarithm : Antl .
\mdash ; Antl .
\mdash ; Antl .
AntL 1 Antl .
Antl .
Antl .
} Antl .
Antl .
\mdash ; Antl .
Antl .
Antl .
\mdash ; Antl .
Antl .
Antl .
Multiplying these out pair and pair , we found RROR Correlation able .
Hence , proceeding to multiply rows and divide columns by the corresponding standard deviations , we have , after nltering the units , the PROGRESSION Table .
An tion of this table brings out several interesting features of the frequency distribution of Mullerian glands in the fore-legs of swine .
If a group of swine were isolated , and found to have a higher mean number of glands , then this group would most probably have an increased possible range , but at the same time a decreased variability and a marked increase of skewness .
increase of the possible range with a decreased variability is especially notable , since the rough-andready class of statistician is very apt to treat the range observed as a measure of variability ; we have here a case in which the same cause , raising of the mean , produces opposite effects on range and variability .
Increase of range , it ] next be observed , produces very little effect on any of the physical constants , but such effect as there is , is an increase of them all .
To increase the modal is to increase the range and to reduce both the variability and the skewness .
Thus the more mediocre swine there exist in any group , the more nearly their distribution will be normal .
Change in the variability is the cause which on the whole produces most effect .
Increased variability mean - lowered mean less mediocrity , but much increased skewness .
Finally increased skewness denotes probable increase of range , variability , and mean .
As we have suggested in a previous illustration the principles of multiple correlation easily enable us to predict the probable change in a random selection in which two or more of the characters differ from those of the general population .
TO THE THEORY OF EVOLUTION .
297 ( 20 .
) Probable of the of the Probability Curve of Type .
This curve is discussed at length , 'Phil .
Trans , vol. 186 , pp. 376-80 .
The chief constants are given as follows , if l@ ( r2 ) , , and denote the origin :\mdash ; Moments\mdash ; ( cxxiii .
) , cxxiv .
( cxxv .
) .
Distance of centroid from origin . . .
, Size of mean organ ( cxxvii .
) , Size of modal organ ( cxxviii .
) , Distance from mean to mode ( cxxix .
) , Skewness ) ( cxxx .
) , Standard deviation ) .
( cxxxi .
) , .
( cxxxii .
) , where and ( cxxxiii .
) , is the formula of reduction .
VOL. CXCL\mdash ; A. 2 PROFESSOR K. PEARSON AND MR. L. N. FIhON Further , we have the following BERNOULLI number series for , ( cxxxiv .
) .
To find Sx , the mean freqnency , we have only to put in , and we have where stands for the summation in ( cxxxiv .
) .
Hence ( cxxxv .
) , or , ( cxxxvi .
) .
As typical constants we require the probable errors of the mean , the standard deviation , the skewness , and the mean frequency .
It is clear that these will require us first to find , 2 , and We shall only indicate briefly the steps towards finding the integrals of the second differentials of log , ' whence , remembering that , and integrating the first integral by we find NATHEMATICAD IONS TO THEORY OF- EVOLUTION .
299 or ; .
Precisely similar reductions lead us to ( cxxxviii .
) , ( cxxxix .
) , , ( cxli .
) , , ( cxliii .
) , ( cxliv .
) , ( cxly .
) , ( cxlvi .
) .
It will now be needful to find easily calculable series for the second differentials of .
These can be obtained from ( cxxxiv . .
We find ( cxlvii .
) , .
( cxlviii .
) .
Hence ; , .
These allow of the fairly rapid calculation of .
The values of the standard deviations of the errors , and of the error correlations , can then all be calculated from the determinant its minors .
Let , and let , then if , be the minor corresponding to in , we must work out for any special numerical case ' ' ' ' ' ' * As in the , these vrere all developed , but the tremakngth of th resulting formula givos them no over in auy cirl c vritl the rioaT determinants .
MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
301 In general none of these correlations vanish , and their values must be found before the errors and correlations of the chief characteristics of the frequency can be found .
The following results , easily obtained by aid of the relation , will of service . . .
( clii .
) , ( cliii .
) , ( cliv .
) , ( clv .
) , ( clvi .
) .
By ( cxxvii .
) and ( cxxxi .
) if be the mean size of organ , Hence ( clvii .
) , , ( clviii , ) , and .
( clix .
) .
From ( cxxx .
) we have , if skewness , .
Or , taking logarithmic differentials , whence , .
( clx .
} .
In a similar manner and can be found if desired .
None of these quantities will , as a rule , vanish , and as many measurements on animals give curves of the tangent type , we conclude that in general all selection of the.size of organ alters its variabitity the skezvness of its distribution , selection of variability connotes alteration of the size and skewness of the selected The probable errors of , and , as well as the error-correlations of these quantities , can all be found from the differentials of ( cxxiii .
) , ( cxxiv .
) , and ( cxxv .
) ; the calculation is laborious , but presents no novelty .
Lastly , the probable errors of the mean and modal frequencies may be deduced .
For the mean frequency we start from ( cxxxv .
) and use ( cxxxiv .
) .
This requires us to know , where .
We have , as in ( cxlvii .
) and ( cxlviii .
) , \mdash ; / where and admit of fairly easy calculation .
Hence , by ( oxxxv .
) , we find , - .
( clxii .
) .
If the problem be to find the modal frequency sx , we easily deduce by putting MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
303 in the equation to the curve .
Writing and the same function of that is of we have ( dxiii .
) .
Since is greater than , and and are less than and respectively .
it follows that is greater than , as it should be .
Further we find ( clxiv .
) .
Here and are the same functions of and that and are of and The usual process of squaring and introducing the standard deviations into the square terms and the product of standard deviations and correlations into the product terms will give us ( 21 .
) In order to illustrate the difficulties which may arise in determining the probable errors of the constants and the error correlations , we have selected for this illustration not a curve markedly skew , but one which is extremely nearly normal .
The problem in this case is accordingly the following one : Are the values of the constants obtained for the distribution and distinguishing it from a normal distribution really significant The difficulties which arise in the course of the arithmetical work depend upon the fact that , as the distribution is nearly normal , its constants approach the values at which the type of the skew curve passes over into the normal curve , and consequently not only will their probable errors be large , but , as ) cases of approach to limits , they will depend upon tending to become indeterminate .
Thus in the evaluation of the determinant and its minors , we at once found our results depended on the ratio of the differences of very small quantities .
We were accordingly in this case obliged to calculate our constituents to a degree of accuracy which will , in general , be quite unnecessary , and which was only possible and straightforward owing to the ready help of a large sized Brunsviga .
That the method , even in a critical case of this kind , gives correct results is evidenced by the agreement of our values of the constants with those ( probable errors of mean and standard deviation ) which can be readily calculated by other processes .
The example we have selected is that given for the stature of 2192 St. Louis school girls of 8 years of age in ' Phil. Trnns , vol. 186 , p. 386 .
The equation to the frequency curve is the axis of being positive towards dwarfs and the origin on the positive side of the mean .
The unit of is 2 .
of height , and all the constants the mean height are given in two-centimetre units .
We the following values of the constants:\mdash ; , Mean height centims. , modal frequency mean frequency 185,606 , Sk .
skewness 04885 , 4.56967 , It will be seen at once that the skewness is small , that the mean and mode are close together , and the mean and modal frequencies are almost identical .
Our roblem is : Are these differences significant or not Let , the total population ; then the values of given above were assumed to be absolutely exact , and calculated with its constituents to 9 places of figures , as it depends on the differences of very small quantities .
We shall indicate one or two stages in the arithmetical work 131 , 017 , 063,438,906 Thoe evaluation of this deterlninant and its minors was then carried out by of the Brunsviga , found MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
305 We can now , remembering that , and writ , down the standard deviations and error correlations of the algebraical constants 1.7078 , 4.4773 , 1898 , 4.5840 , 6835 , 9980 , 8340 .
Here marks the position of the origin of the curve , and the numerical values are only retained to four places of figures , although , of course , in the further calculations the logarithms of the full values of the 's and 's have been used .
It will be noticed at once that though , and have very probable errors , the correlation between them is very high .
In other words , as the curve approaches its limiting shape , , and may vary very considerably , but owing to their close correlation this will not sensibly affect the ical shape of the curve .
The next stage was determine the standard deviations and error correlations of certain idiary constants .
Here , as in the determination later of the like quantities for the " " physical\ldquo ; characters , we found the umbral notation of great service .
It consists , as we have seen , in writing down a difference equation betweell any constants , and then replacing the differences Su by , @v by \amp ; c. , where \amp ; c. , are quantities which obey the relations .
Thus , if we find for the umbral equation .
Whence , putting in the numerical values , we have Antl .
\mdash ; Antl .
where Antl .
stands for antilogarithm , in which form we found it easiest to keep the umbral coefficients .
The square of this result gave at once .
087,926 and dividing out by its logarithm , we have the pure umbral equation Antl .
\mdash ; Antl . .
CXCI.\mdash ; A Our object was then to find such pure umbral equations connecting the " " physical\ldquo ; constants with the algebraic constants .
Their products will then give the error correlations of all the " " physical\ldquo ; constants in terms of the correhtions already known between the a]gebraical constants .
For example , multiplying the above equation for by we have , ce , \amp ; c. , are ready known , , actually It was these logarithms , of course , which were used in the further calculations .
Since is measured negatively towards dwarfs , positive ) , we must write for transferring origin to the mean where ban is the distance between the old origin and the mean , or if be used to represent the mean we have Hence we find the umbral equation Antl .
Antl .
Hence we determine and the pure umbral equation Antl .
Antl .
Antl .
In ecisely the same way all the other " " physical\ldquo ; constants , i.e. , the standard deviation , , the mean frequency , , the distance between mean and mode , , and the skewness , Sk .
, were found , and the umbral equations investigated .
It is only necessary here to give the results .
MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
307 Now it will be seen at once that the probable errors in the algebraic constants are large , but that the probable errors in the position of the mean , of the mode , and in the magnitudes of the mean frequency and standard deviation are small .
The position of the mean is sensibly more correct than that of the mode .
On the other hand , the distance of the mean from the mode and the skewness have large probable errors , not , however , so large but what these quantities are probably significant .
The frequency distribution probably differs significantly from the normal distribution , but the difference is small and would require a very large number of observations to determine it with extreme accuracy .
That there is a significant divergence from normality is also indicated by the sensible difference between the percentage errors in and which would be equal for a normal distribution .
we taken a normal distribution , the probable error of the mean would have been0400 , and of the standard deviation , 02831 .
In fact , the standard deviation of the standard deviation , if calculated for the normal curve , if culated by our present method , and if calculated b.v a modified form of the fourth moment formula given by CZIJBER* .
This shows that the arithmetic of our process has been substantially correct .
We now place together the umbral equations for the correlations of the errors in the " " physical\ldquo ; constants .
They are Antl .
Antl .
Antl .
\mdash ; Antl .
Antl .
Antl .
\mdash ; Antl .
Antl .
Antl .
Antl .
\mdash ; Antl .
\mdash ; Antl .
From these resulcs any correlation between pairs of errors , " " physical\ldquo ; or algebraic , can be found at once .
The following table gives the chief results:\mdash ; CORRELATION Coefficients between Errors in Constants .
Now it is clear that although the curve is nearly normal , there is still correlation between quantities\mdash ; e.g. , mean and or d\mdash ; which would have no relation between them if the curve were absolutely normaL This will be clearer as in the previous illustration , we replace this table by a table.of regression coefficients .
This table has now finally to be thrown into suitable units and attention paid to the fact that increases towards .
We have , after the proper changes , the following results:\mdash ; PROGRESSION Table .
This table is extremely suggestive .
It shows us that a random selection of girls of eight which had an increase of stature would have a less standard deviation , less distance between the mode and mean and less skewness .
In other words , a selection giving baller children would be less variable and more nearly normal .
Now as children grow older their stature increases , is less variable , and is more normal in its distribution .
T.hus , a selection of taller children from among children of eight would broadly tend to reproduce the characters of the stature distribution of old .
MATHEMATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
309 children .
In the same manner a selection of shorter children is more variable and normal than the distribution of the general population of eight years of age , i. e. , tends to reproduce the characteristics of a younger population .
Generally , a random selection , which increases variability , very sensibly increases skewness and decreases stature .
What , perhaps , would hardly be expected , is that increase of skewness as well as increase ofinterval from mean to mode , i.e. , greater divergence from normality , increases the frequency of the mean stature .
It will be clear that by aid of this table we are able to predict the probable changes in all the other physical characters of distribution when any sub-class has been selected at random from the general population with a difference of one character .
If two or characters differ in the sub-class , the probable changes in the other characters can be found by the principles of multiple correlation from the correlation table on page 307 .
( 22 .
) Conclusion .
This study of the probable errors and error correlations shows us that these quantities can be determined for the most complex system of organs in the case of normal correlation , and in the case of either normal or skew variation with considerable ease .
It is only in the case of skew variation that the arithmetic becomes at all laborious .
But numerical examples suffice to show that the errors here made are of the same order as in the case of normal variation , confine our attention to the characteristic features of the frequency , e.g. , the mean or modal frequency , the standard ation , the skewness , \amp ; c. Certain constants of the algebraic form of the frequency curves have large probable errors , but these errors are so highly correlated , that their existence does not suffice to substantially modify either the form of the curve , or the " " physical\ldquo ; characteristics of the distribution calculated f'rom such values .
For the theory of evolution certain very important principles flow , beyond the mere advantage of knowing the probable errors made in the measurement of racial or organic characters .
Above all we note the importance of a random selection in altering in a systematic manner all racial constants .
In most cases even size cannot be altered without alteration of the -ize , variation and correlation of all correlated organs .
This principle is developed more at length in a memoir , nearly completed , on the fluence of directed selection , which covers as a special case that of random selection .
Later , we hope to apply the general theorem from which our memoir starts to determine the probable errors in the constants of the components into which a heterogeneous frequency distribution may be resolved by the method of the first memoir of this series .
* It applies equally to such an investigation .
* The importance of such a determination was emphasized by Professol .
] OBGE DARWIN in the soUssiou which took place at the reading of thal memoir .
[ NOTE.\mdash ; Added May 25 , 1898 .
One point ought to been mooe ffilly dealt .
in the above memoir , namely , the probable error of the criterion upon which the selection of the type of the frequency depends .
Clearly , if probable error of this criterion is as large as the criterion itself , there can be stability of type , or the frequency may change over from one type to another .
On page 289 we have found the standard deviation of the oriterion in terms of known quantities for the curve It is in fact given by the umbral equation ( clxv .
) where and are functions of and given in ( cxx .
) and , is known from .
The standard deviation of the criterion for the curve of type may be found by taking differentials of a value readily obtainable from 'Phil .
Trans , vol. 186 , p. 377 .
We thus find the umbral equation ( clxvi .
) , say , where 2 and are given by ( clii .
) and ( cliv .
) .
Applying these results to the numerical examples , we find : For the glands of swine whence the probable error of ; or , For the stature of children whence the probable error of ; or , MATBENATICAL CONTRIBUTIONS TO THE THEORY OF EVOLUTION .
In both cases , therefore , we may consider that the sign of is beyond question , or that the type selected is really a significant character of the frequency .
With regard to the probable error made in estimating a criterion to be zero , and using a curve of type we must remark that , the criterion bein assumed zero is equivalent to assuming that its probable error is zero .
Accordingly the only satisfactory method of testing whether a curve really falls under this type is to work out the probable error of its criterion on the hypothesis that it belongs to one or other of the two types , with positive or negative criterion as the case may be .
If the probable error of the criterion thus calculated is sensibly as large as the criterion itself , then we may assume that the frequency distribution is of the type

