]\gt ; [ 429 ] IX .
Mathematical Contributions to the Theory of Evolution.\mdash ; XIX .
Second Supplement to a Memoir on Variation .
KARL PEARSON , F.R. Rcceived Febru , ary 2 , \mdash ; Read , 1916 .
[ PLATE 1 .
] ( 1 ) IN a memoir presented to the Royal Society in 1894 , I deaIt with skew variation in homogeneous material .
The object of that memoir was to obtain a series of curves such that one or other of them would agree with any observational or theoretical frequency curve of positive ordinates to the following extent:\mdash ; ( i ) The areas should be equal ; ( ii ) the mean abscissa or centroid vertical should be the same for the two curves ; ( iii ) the standard deviation ( or , what amounts to the same thing , the second moment coefficient ) about this centroid vertical should be the same , and ( iv ) to ( v ) the third and fourth moment coefficients should also be the same .
If be the moment coefficient about the mean vertical , the area , be the mean abscissa , the standard deviation , , then the equality for the tWo curves of and leads almost invariably the case of frequency to excellency of fit .
Indeed , badness of fit generally arises from either heterogeniety , or the difficulty in certain cases of accurately determining from the data provided the true values of the moment coefficients , e.g. , especially in J- and -shaped frequency distributions , or distributions without high contact at the ; here the usnal method of correcting the raw moments for sub-ranges of record fails .
Having found a which corresponded to the skew binomial in the same ma1lner as the normal curve of errors to the symmetrical binomial with finite , it occurred to me that a development of the process applied to the hypergeometrical series would achieve bhe result I was in search of , i.e. , a curve whose constants would be determined by the vational values of and The hypergeometrical series was one not only arising naturaIly in ance problems , but covering in itself a most extensive range of functions .
The advantage of the hypergeometrical series is that it abrogates the xioms o which the Gaussian frequency is based .
The equality in frequellcy of plus and errors of the same magnitude is replaced by an arbitrary ratio , the number of contributory VOL. CCXVI .
causes is no longer indefinitely large , and the contributions of these causes are no longer independent but correlated .
* Since and are by nature positive we can represent all possible values of on a chart in which and are the co-ordinates of a point in the positive quadrant .
But a little consideration shows that must be greater than , thus one-half the area of the quadrant , that above the lin is emoved from the field of possible occurrences .
Further , there is a limit to the application of the series of curves discussed when gets large , for the high moments of two of the types of curves , i. e. , Types IV .
and VI .
, or and become infinite when the order of the moment is greater than , or the probable error of the fourth moment would become indefinitely large for , i.e. , we are practically limited by the line .
The first four moments of the curve remain finite , but from the fifth onwards they can become infinite , the lines corresponding to these , however , lying outside the above line .
For curves corresponding to points below this line it is fitting to take as differential equation ' ( i ) or a slightly more general form which is related to the higher hypergeometrical as the present series of curves to the simple hypergeometrical .
The whole theory of curves of the above type has been worked out for some time past , but has remained unpublished , for we failed to find any definitely homogeneous data by which it could be effectively illustrated , and for this reason heterotypic curves have for the time being been left in abeyance .
We may , however , notice the following point .
If we take our generalised hypergeometrical to be Then and this will correspond to the ordinary form if , i.e. , .
*Just as values of the binomial with negative and very often give good fits to frequency distributions , so we have recently found that with imaginary and are of common occurrence in frequency distributions , and when applied to individual samples from real hypergeometrical populations may give better fits than the theoretical series , i.e. , in card drawings .
See RHIND , ' Biometrika , ' vol. VIL , p. 133 .
PROF .
KARL PEARSON ON SKEW VARIATION .
We have and accordingly we get the curve approximating to the hypergeometrical of the higher order by putting , ( ) where the six independent constants can be express in of the original six , .
It will be seen that a etrical of the second order will , in general , have two modes , the exception bein when ; . . . . .
( iii ) in which case ( ii ) coincides with ( i ) the general equation to the approximation of curves when and fall into the heterotypic area .
It will thus be noted that such curves approximate to hypergeometric series of second order when the special condition ( iii ) holds ; always assuming the unimodal character of homogeneous lnaterial .
It seems probable that for the most part bimodal frequencies would be those that lead to values of and lying in the heterotypic re , , and such are excluded from practical statistics .
In the original four types of curves were dealt with beside the Gaussian curve corresponding to an isolated point .
A supplementary memoir issued in ) dealt with two further types , which had been overlooked until actual experience demonstrated their existence .
I have now to confess the omission of further types , not to speak of a horizontal straight line , as sub-groups of the J-section of curves , which are themselves in practice so rare , that the region of the , plane in which they occur had not been very fully investigated .
My atteiltion was drawn to these curves while considering the frequency curves for the correlation of small samples . .
If we take a sample of four from uncorrelated material , the is equally likely to have every correlation from to In this case , , and the frequency curve is a horizontal straight line .
What would my of curves give in this case ?
I discovered that they also gave a rectangle of frequency or a horizontal straight line , and this discovery led to a closer tion of the sub-groups of curves in the of the -curve area .
The point in the 'Phil .
TraIlS vol. ( 1895 ) , pp. 'Phil .
Trans , vol. 196 ( 1901 ) , pp. 'Biometrika , ' vol. , p. 306 , and vol. X. , p. 313 .
plane for which , I term the -point and denote by R. ( See folding diagram , Plate 1 , at end of paper .
) The rectangle-point is the point of contact with the axis of of the biquadratic which bounds the area of -curves .
The novel curves are in part limiting curves which occur when the point lies on this biquadratic , i.e. , transition curves from -curves to -curves and from -curves to limited range curves , and in part a limiting curve which exists along the line which passes through the rectangular poin and never again meets the biquadratic in the loop in the positive quadrant .
It would be convenient to speak of this lin as the axis of the biquadratic loop , but unfortunately the loop is not symmetrical about i and to avoid misunderstanding I term it the -line .
Up to the present the minimum limit to the area of -curves had not been given .
Since is , half the positive quadrant was impossible , but a recent observation shows that frequency curves above the line are impossible .
This limit was suggested in the following manner .
When samples of three are taken from an indefinite population , the frequency curves for the correlation of any two variates of the three individuals sampled are -shaped frequency curves , but when samples of two are taken the correlation must be either positive or negative , and accordingly the frequency is collected into two lumps or blocks as a limiting case of a -shaped distribution .
But for two such lumps .
Iri other words , along the line , the -shaped frequency either brings all frequency to an end , or passes through a transitional case .
The former is the true state of affairs , for cannot be less than .
To demonstrate this , , and let there be quantities .
Clearly , , and .
Now by BURNSIDE and PANTON , ' Theory of Equations , ' vol. II .
, p. 35 , , * I owe this neat proof to the kindness of Mr. G. N. WATSON .
PROF .
KARL PEARSON ON SKEW VARIATION .
which must therefore be either zero or a positive quantity .
Thus we see that the whole area covered by my frequency curves is limited above by the line and below by the line .
The first line limits all frequency ; the second line limits my types .
* ( 2 ) Before proceeding further , let us examine the limit to all frequency .
Consider the line The form of the curve is Now , where therefore and also Hence and , or The form of the curve is accordingly or , apparently , -shaped .
Now and is finite .
But limit of But limit of * It is not accurately correct to say it limits my types of sksw curves .
What it actually does is to cut off an area in which the probable errors of the constants of Types .
and .
curves can be very grcat .
The curves may give a good fit , but the Ilstants ca1mot be cited as charactcristics of the frequency distribution as they are unstable .
The notation throughout is that of my original ' Phil. Trans. ' memoirs of 1895 and 1901 .
when both are to be made vanishingly small , being and respectively .
Thus the limit Hence vanishes or is zero at all points , but and where it is undetermined .
Since , we have , and the frequency really consists of two concentrated groups at and , or at If and be the distances of the centroid from the two ends of the range , where and are the concentrated at the range terminals .
But , or we have , or is the finite quantity which marks the ratio of the vanishing of and ; this , therefore , is equal to Clearly and Thus giving as verification Thus the whole problem is solved if we know the magnitude of the two frequencles and concentrated at - and As special cases the point on the -axis gives , and represents two equal concentrated frequency lumps .
The point at on the line , or represents a single frequency lump , for which .
I speak of these concentrated ency lumps lyin on the line PROF .
KARL PEARSON ON SKEW VARIATION .
as block-frequency , and represent them by the letter ; they correspond to points on the -line .
( See Diagram , Plate 1 .
) The most remarkable limiting case of this kind has been already referred to .
It will be shown in practical examples in a memoir on\ldquo ; small samples now nearly ready for press , that the correlation between two variates may be determined by sampling these populations in pairs , and merely observing , which can be usually done without measurement , whether the pair is positively or negatively correlated .
The ratio of the two frequency " " lumps\ldquo ; easily provides the ( 3 ) Let us now consider the nature of the frequency on the loop of the biquadratic .
Taking the form of the curve to be we know that and are the roots of the quadratic where , and Now provides the biquadratic ; actually Now and are by their nature essentially positive .
Hence , provided is positive , i. e. , as long as we deal with points above the line , i.e. , the Type III .
curve will be positive , if lie outside the loop of the biquadratic .
But within the loop it is egative , or one value of must be negative , or we reach an infinite ordinate at or , i.e. , -shaped curve .
The other ordinate at , or is zero , because the other must be a finite positive quantity .
If , i.e. , along biquadratic loop , one value of ? ?
is zero , the other is positive if be greater than 2 , and negative if it be less .
But Accordingly above the line , al1 above Lhe line will be negative , but these lines do not meet in the positive quadrant .
Hence all * See " " STUDENT ' Biometrika , ' vol. , p. 304 , and FISIIEP , ' Biomctrika , ' vol. X. , 1 ) .
508 .
along the upper boundary of the loop one is zero and the other negative .
Accordingly , from the -point round the upper boundary of the loop , we have the curve I call this curve Type VIII .
Since , and is zero while and are finite , it follows that , and accordingly the range of frequency is from to .
The curve is therefore a -shaped curve with infinite ordinate at one end of the range and a finite ordinate at the other .
Now consider the lower side of the loop .
Here will be positive , for this side is below the -line and will also be positive until the point in which the line meets the lower side of the loop , i.e. , the point .
Hence from the -point up to , a poin practically outside the range of the customary statistical frequencies , will be positive , or will be positive .
Further and being finite and zero , it follows that is zero , or the curve is In this case the curve has a zero ordinate at one end and a finite ordinate at the other .
I term this curve Type IX .
At the poin where the line meets the biquadratic , Type IX .
agrees with my earlier Type III .
The equation to that type where and Hence for , and .
Thus is zero and the curve becomes the range being from to But in Type IX .
, since has become infinite , is infinite and the limit to is accordingly the exponential curve as we shall see shortly must equal , where is the standard deviation .
I propose to call this exponential curve Type X. , and the point or the exponential point .
* Phil. Trans , vol. 186 , p. 373 .
ARL SON ON ARIATION .
Beyond the exponential point , our biquadratic brancb of Type VI .
curves , *and m will now be negative .
Now the equation to Type VI .
is and the range from to .
The special case of this the branch of the biquadratic occurs when , leading to or ere which is ) lines This curve .
which will be more fully considered below , has a range a certain value to .
It thus starts with a finite ordinate and asymptotes to zero .
It is a transition curve extending from the exponential point along the lower limb of the biquadratic loop .
I call this curve Type .
The biquadratic neyer cuts the cubic along which Type V. no further change occurs in Type XI .
I now pass to the consideration of the -line or The general ential equation S to type of frequency under sideration is the ol.igin being Hence if ternl iu disappears the numera can further get rid of by for it .
Making this substitution , we reach ' Phil. Traus , vol. 197 , p. 449 . .
cil .
, Equations , bottom of p. 449 .
As we pass outwards from the exponential point along the biquadratic ranges from to 5 , which it reaches : the asymptote to the iquadratic / , or when / S " " Mathematical Contri})utions to the Theory of Evolution , .
On the hcory of ' Correlatiott inear 1cgrcssion , \ldquo ; p. 6 , Conll ) ridgo lJniversity Press .
VO CCXV .\mdash ; A. 3 PRO ARSON O This leads on to I term this Type XII .
, or the -curve .
origin is the mean , the range from to .
It separates -curves\mdash ; so long as we above the line 0\mdash ; for which2 is positive from those in 2 is negative .
But Hence below the -line the positive is greater than the negative ?
) , but above this line the positive is less than the negative , i.c. , the upright of the is emphasised at the expense of the horizontal part , while below the -line this condition is reversed ulltil on the biquadratic the finite ordinate of the -upright is replaced by a finite ordinate .
I propose now to consider little detail the ture of these new types of frequency and the manner of fitting to actual data .
I have dealt above ] fully block-frequency\ldquo ; its criterion nd therefore need consider ypes VIII .
to ( 4 ) .\mdash ; Range , from to is clearly the value of , the finite ordinate at the tail .
We easily deduce if ) ] , total frequellcy taking the origin at , .
ence for about the These lead to ' Clearly could be found from the value of by solving the cubic equation *Of -hether is ) egtive will depend on the sig given to it , the direotio1l of the PROF .
KARL ON SKEW then is determined from the sign b determinable from the obseryed value , and from and the placing of the frequency curve on the ) } ervati 1 .
If , however , find and ) , we giving and thus since is to be positive , the point must be abore the line .
The line does not neet in the positive quadrant , so that a point below both these lines does not exist in real frequency .
Clearly , , and thus if these values be substituted in as above , reach the equation to the biquadratic , proving that the point associated with the above frequency curve lies on the biquadratic .
Again will always be positive , or less than unity .
For the upper of the loop of the biquadratic lies below its asymptote , or , and accordingly below the line ; thus the numerator of is always positive .
So also is the denominator , for the upper branch always ] above the line * In fact the -line the parallel to the asymptote , the limi frequency line , and the Type III .
line meet in the point ) of the negative quadrant and the upper branch of the loop lies in the angle between the first and in the positive quadrant , As is positive and less than unity the and moments of the curve are all real and finite .
When the point moves along the loop of the biquadratic towards the , the value of becomes more and more nearly unity , and ultimately at we have , or the frequency curve is a rectangle , i. e. , we reach the rectangle point .
If on the other hand we move towards infinity along the upper of the biquadratic loop , we find approaches the value and thus ultimately becomes , or .
Thus the limiting form of the frequency curve is a rectangular hyperbola , or rather the part such hyperbola from the vertical asymptote to But this is clearly only a theoretical limit , for it involves , and this means that be finite , and are infinite\mdash ; results impossible in anyactual frequency if the population be finite .
It is clear indeed that must be less than , for obviously .
Again , is 1 , and accordingly But these limits are of small service for practical statistics , where even for small samples , say , , they would scarcely ever be approached .
Thus the rectangular hyperbola can only be treated as a limiting form of Type VIII .
far beyond the region of actual statistical experience .
For practical purposes the point is that is limited to values between and 1 , or Type VIII .
ranges from the rectangle to the rectangular hyperbola .
The suggestiveness of this is that curves in the and the areas , i. e. , above and below the upper branch of the biquadratic loop , must approach these types as they approach the extremes of this branch .
enerally a -curve near biquadratic will be close to a curve resembling a curtailed hyperbola .
* Mr. G. N. has given mc a nc limit to namely , .
But , except as showing that must be finite , which is otherwise obvious , this is again of no real service .
The highest observed values that I know of for ) are those given by ( ' Biometrika , ' vol. VIII .
, p. 238 ) .
He gives ' Armzahl , ' Asieriam exigua 33 13 , There are only three groups of frequency in each , 4 , 6 and 6 , and the bulk of the observations are concentrated in 5 .
The observations do not give , as hc suggests , PEARSON'S Type .
and Type curves respectively ; the in ) cases is less than unity , corresponding to Type IV .
But both fall into the heterotypic area of Type .
The attempt to fit with heterotypic curves would hardly be profitable until there was absolute certainty that the group with 4 ' Armzahl ' was not the result of accident .
Th eoretically vory high values of anrl , can easily be found , i.e. , for samples of four , when the population sampled has , say , correlation of here the frequency curve for the correlation coefficient gives and , but it is the rapidly approaching zero of which leadS to these results .
PROF .
KARL PEARSON VA In concluding our discussion of this curve we may nots that , rhaps , the easiest way of tracing the biquadratic is to calculate and from by giving a succession of values to For to 3 we get the points on the lower branch of the loop ; for to we obtain the points on the upper branch of the loop .
It will be seen that amounts to taking the origin at , and rotating line point round it to intersect the curve .
slope this line to the axis is 3/ ( 3The cubic , it may here 1loted , which ] the Type V. curve 1lliL be Here must be given values frolll 1 to 3 .
The III .
line , which point , passes and the above nled ns ofetting i the points on the cubic corresponds to finding the points which a line passing through and rotating from the position of the Type III .
line cuts the cubic\mdash ; its slope in position being as before .
Actually if ) the angle between the aboye line frol n to ) , , , but to use equation not ) eelJ fnaInJel ( .
plotting the cub .
Range to , ; one of th to the other .
The analysis proceeds precisely as case ) curve VIII .
, except that is now opposite in .
We ( distance ) ) ) , *The parts of the ) in the other three ] ) lotted Miss B. .
B. .
Geometrically the interrelatiolts of the two cnrv .
, their asymptotic and othe ] oritical lines are of much , but until some interprotation be put on of the moment coefficients , these interrelations have no statistical bearing : PROF .
KARL PLARSON ON SKEW yARIATION .
leading to Thus would } would bo found the being found from the observed value of .
Lastly Practically it is better to determine from which value of substituted in the expression for gives the biquadratic .
Clearly since the lower branch of the biquadratic lies below the line is positive until the line is reached , and in this section of the branch , i.e. , from to , or from up to ( the exponential point , E ) occurs an interesting isolated point\mdash ; the line-point L. When , then , and Type IX .
degenerates into a sloping straight line , , or the frequency lin is .
Up to the IX .
curve rises at to the axis , of , at the line-point it makes a finite angle less than 90 degrees , and after the linepoint we start with contact at It is to note the sloping line arising as a case of these generalised frequency curves , and we observe that its locus is separated from the rectangle locus by a considerable interval along .
the biquadratic in which the curve of Type IX .
is very trapezoidal in form .
( 6 ) quency Curve of X. The Curve.\mdash ; Beyond the line-point , at , we reach as steadily mounts a series of frequency curves which culminate in the exponential curve at or ) early when is infinite , StlN ON IATit ) Further the range to , if we take the positive sign\mdash ; and from to , if we take the negative .
is thus sufficient to consider with range from to two of the about are and ) Thus and as it and The fitting of the exponential cnrve presents difficulty .
exponential tr nsition point of interest than point \mdash ; the meeting 1nany types .
At changes to at the familiar Type II1 .
passes a ordinal at the limited end of the range to a -curve with infinite ordinate .
Further , is a point at which the of Type T. ) as a limited range with dinates at its terminals , and as a limited ange one infinite ordinate a terminal ) meet .
Finally , Type , which lies veell 7 nd Typc V. cubic , is divided into two sections by Type XI .
, which lies along the lower of the biquadratic loop below the biquadratic , Type .
takes the a botb positive .
the ) , however , nd Xl .
, .
takes , loll]l haped ' with ange f lvill be uegative , since we below is since we tho cnbic or V. Thus our atic corresponds of necessity to real roots , of which one will be negative the ot , her positive .
The positive root ) be PROF .
KARL ARSON ON SKEW VARtATION .
and is therefore numerically the smaller root since is negative ; it will be less than unity , and therefore will be negative if or but this is the condition for the point lying inside the loop of the quadratic .
Thus } this case we reacl ) the -shaped of , or that the area of this and its onlents should be finite , it is clearly needful that should be less than unity .
( 7 ) \mdash ; Beyond the exponential point the lower branch .
the biquadratic is below the line , and sequently m is again negative and the curve takes the forul where The range is , however , only limited in one direction , it is from to , say .
This lower })ranch of the biquadratic loop tends to become vertical and asymptotic to the line .
Hence takes all values from down to 5 .
Clearly , for moments about these will be real and finite if or only the fourth moment would fail at the limit , which indeed cannot in practice be reached .
At the same time if we want the probable error of the fourth moment to be finite , it is needful that should be or we must have .
Thus must be where the curye passes the heterotypic-region and becomes of doubtful application .
We easily find from the above result for leading to PROF .
KARL PEARSON ON SKEW VARIATION .
Thus for we find , which satisfy the equation of the heterotypic line .
may be found from or from alone by the cubic then and , while the mean enables us to place the curve on the observations .
There is no discontinuity in the form of the curve down to , but only discontinuity after in the probable errors of its moment-coefficients .
The curve starts with a finite ordinate and meets that ordinate at a finite angle ; it asymptotes to the -axis at , and has no poin of inflexion except at infinity .
( 8 ) Frequency Curve .
Type XII.\mdash ; This -curve arises along the , or .
Its range is from to , and then its mean is the origin .
When is zero it degenerates into a rectangle ( i.e. , at the rectangle point ) .
In order to illustrate the nature of the curve more fully let us start from the general equation which arises when the denominator of the differential equation has real roots , where and the origin being the mode and the range .
Transferring to the mean as origin this becomes 'Phil .
Trans , vol. 186 , p. 369 .
Loc .
, p. 370 .
VOL. CCXVI.\mdash ; A. .3 where* ' on substitution for and as above .
Now put , or , say .
Then while It remains to find Now and are the roots of where Hence , when , we have or , the -line , and .
Whence or But , and it is well known that Thus This is the full equation to the -line -curve , the mean being origin .
It requires for its determination only a knowledge of , but we must be also certain that the .
cit. , p. 369 .
, pp. 368-9 .
Deduced at once from by putting rhesign of in must be determined from that of .PROF .
KARL PEARSON ON SKEW VARIATION .
condition is satisfied within the limits of random sampling .
Its possibilities extend from to .
When , the rectangle .
Now consider what happens for any frequency curve of the limiting character when both and become infinite , say , in the ratio .
Then and accordingly will be finite if is finite , except along the Type III .
line .
Accordingly for will be zero .
Thus the ratio of to is from their values , which agrees with the above result for : For the special case when , we have , which agrees with the ratio of along the -line .
Now when we have from or Thus from the equations on page 445 , if we change the sense of the axis of and take from to Now in order that should be finite it is needful that should be infinite when , for But if be infinite , owing to the factor , for every value of , except Hence the frequency is a concentrated lump at , and this involves of itself But if must be finite or zero , and these both again throw us back on a concentrated frequency at Accordingly , when and both become infinite , we deal with a concentrated frequency lump .
But the ratio of to will depend on the manner in which we have reached this limiting case .
For example , if we are dealing with the correlations in samples of two drawn from a population in which the correlation is , the frequency consists of two lumps , but as approaches unity , one lump ivels up , and both become infinite , but their ratio is one of equality , i.e. , we approach infinity along the line When we take samples of three from a population of correlation , the frequency curves are U-shaped , but as approaches unity the frequency concentrates in one leg of the and both become indefinitely larger , but their ultimate ratio appears to equal The -curve flattens down into an -curve , of which the horizontal limb extends to infinity and becomes indefinitely thin , while the vertical limb contains all the frequency .
( 9 ) Scheme of Skew Frequency Represented as \mdash ; We are now able to considerably enlarge our diagrammatic representation of frequency curves .
( See Diagram , Plate 1 .
) Every distribution is represented by its characteristic co-ordinates and , which must be positive , and therefore we need only deal with the positive quadrant .
No frequency distribution at all can lie above the line ; this restriction removes more than half the positive quadrant .
No frequency distribution can be adequately represented by one of the present system of skew curves , if it falls below the line .
The area below this line is therefore termed heterotypic .
Heterotypic distributions are to say the least of it very rare , if they be not extremely improbable .
We have seen that there is some reason to suppose that bimodal distributions would give rise to such heterotypic distributions , but with our present views as to frequency such distributions when they do not arise from the mere anomalies of random sampling are classed as heterogeneous , and supposed to be due to mixtures .
Having thus limited our area at top and bottom we proceed to consider the various possibilities that arise .
The axis , where , is the axis of frequency distributions .
Possibilities begin at the -line or the point , or we have two equal concentrated frequency blocks at any arbitrary distance .
This is the case of two alternative values , either of which is equally probable .
For example , heads or tails in the repeated tossings of a coin , or positive or negative perfect correlation in samples of two taken from a population of individuals bearin two uncorrelated * I use the word ' ' appears\ldquo ; advisedly , because the ratio has been obtained.by determining the value of for bigh numerical value of .
The actual ratio for depends upon approaohing a limit in rather complicated elliptic integral expressions , which I have not yet accomplished .
PROF .
KARL PEARSON ON SKEW VARIATION .
characters .
Below the point , descending the -axis , the two concentrated frequencies expand into a symmetrical -curve .
This is Type with the equation and the criterion Here* , , and When , and we reach the " " rectangle-point " " R. Here and Samples of three individuals from a population whose individuals carry two uncorrelated characters give a symmetrical -frequency for the coefficients of correlation of those characters in triplets of individuals .
In this illustration Samples of four individuals from the same population boive a , for the frequency distribution of the coefficients of correlation .
Passing still lower down the axis of symmetrical frequency the type is now Type , or the limited range frequency curve and the criterion is In this range increases from to , and , We see that the range grows greater ches i , or ) , reach the Gaussian point .
If samples of individuals be taken from an indefinitely in the individuals carry two uncorrelated characters , then if be 5 or over , all the frequency curves of the correlation coefficients of these samples of Type , only approaching the Gaussian when is very considerable indeed .
For exanlple when , and the frequency is still a good from When , it is thus fairly close to it , but is not coincident .
* It is , perhaps , worth noticing that for we obbai11 the ordinary ) } ' of frequency-curve .
After we have passed the Gaussian point we obtain curves of unlimited range of Type VII .
, of which the equation is The range of is from 3 to and , falls from infinity to ; while , Illustration of curves of Type VII .
are not infrequent in biological statistics .
We see that the Gaussian is a mere point in an infinite range of symmetrical frequency cllrves , and a single point in a doubly infinite series of general frequency distributions .
Now let us consider the asymmetrical frequency curves displayed on the Diagram .
If we approach from the " " impossible area\ldquo ; we reach on the -line the first available type of frequency\mdash ; the alternative concentrated blocks .
At one end of the -line we have two equal isolated frequencies , and at the other a single isolated frequency .
Crossing -line we reach the area of limited range -shaped curves , i.e. , Type , which has for its equation : This -area extends as far as the upper branch of the loop of the biquadratic , the asymptote of which , , is indicated by a broken line .
In -shaped frequency curves both and are necessarily less than unity , for their product is , which is less than unity and positive above the upper branch of the biquadratic .
Type is fitted as Type I. ( see ' Phil. Trans vol. 186 , p. 367 ) , and has been illustrated by me ( ' Roy .
Soc. Proc vol. 62 , p. 287 ) , by fitting curves of frequency to cloudiness .
The frequency curves for the correlation coefficients of samples of three drawn from a population whose individuals have two characters of any degree of correlation are also skew -shaped frequency curves , although their algebraic form has not the above simplicity .
*Type was discussed in my first memoir , ' Phil. Trans vol. 186 , p. 372 .
Type and Type VII .
are briefly referred to in 'Biometrika , ' vol. Iy .
, p. 174 , but , unfortunately , with some rather disturbing misprints .
They are correctly placed on ) ' diagram , 'Biometrika , ' vol. VIL , p. 131 , but the formulae for fitting are not given .
The formulae have been given for many years in lecture-notes , and the curves have been frequently used .
PROF .
PEARSON ON KEW VARIATION .
On the upper branch of the biquadratic loop we reach curves of Type VIII .
, i.e. , discussed on p. 444 of the present memoir .
Here is less than unity .
We now pass into the loop of the biquadratic between the upper branch and the -line .
Here we have -curves , Type , of the form where is less than unity , and is less than to the -line , becomes equal to and we have Type XII .
, or discussed on p. 446 of the present memoir .
Below the -line , we return to Type but is now greater than We now reach the lower branch of the biquadratic loop .
This is divided into three portions by three critical points .
The first portion is from the ) oint ( to the line-point L. In this portion we start from with the curve of Type IX .
or , for , or the rectangle , and proceed from that value to , which gives us the line ( or triangle ) ; the range is to .
Since is always , the curve rises perpendicularly at , and approximates to a trapezoidal form .
The lnethod of fitting is discussed in this memoir , p. 441 .
The of the line curve is dealt with on p. 442 .
Beyond the line-point we have Type which differs in no way from Type except that is now greater than unity , and there is contact of a rapidly increasing order at When we find Type X. the exponential curve , at the exponential point E. The fitting of this curve has been discussed on p. 443 .
example , at the point , between tho -line nnd npper branch , but at vecn the line : the ) Since is the junction of several types , we turn to consider Type III .
which is the curve found along the critical line It passes through the Gaussian point , and its equation is It is fully discussed in my first memoir ; see 'Phil .
Trans. , ' , vol. 186 , p. 373 , et seq. From to the exponential point ranges from to zero , which latter value provides the exponential curve .
After the exponential point becomes negative and we reach Type , a -curve with range limited in one direction only .
This curve separates the doubly limited curves of Type from curves of Type , which lie below the line , and above the lower branch of the biquadratic loop .
On this lower branch of the loop we have Type XI .
, or the form the range being from an arbitrary value to , and ranging from to 5 .
This type is fully discussed in the present memoir ; see p. 444 .
It continues right away along this branch of the biquadratic , but at and , the eighth moment of the theoretical curve would become infinite , and accordingly the probable error of the fourth moment coefficient would become theoretically infinite .
Thus since the fitting of the curve depends on the fourth moment its constants would cease to be reliable measures of the distribution .
We enter at this point the " " heterotypic area for this type of curve .
We have now two further areas to clear off , namely those between the Type III .
line and the lower branch of the biquadratic loop .
Above the former and below the latter we have the ran ge of double limited frequency cnrves , i.e. , Type , or This curve was fully discussed in my first memoir ( ' Phil. Trans , vol. 186 , p. 376 , et seq. ) and are both positive , and experience has shown that probably the bulk of all frequency distributions cluster into this area .
Above the biquadratic loop and below the line , we have curves of Type , or with range from to * Of course , by using the actual ghth moment of the data , instead of the eighth moment of the theoretical curve , the standard deviation of the fourth moment would be finite , but this procedure would really indicate that , as far as the high moments are concerned , curve and data were discordant , and that we should not really be finding the probable error of a constant of our theoretical frequency curve .
PROF .
KARL ]ON SKEW VARI A 45.3 They have been considered on p. 443 of the present memoir .
Their full theory is precisely that of curves of Type VI .
in general , discussed in the first supplement to my memoir on skew variation ( ' Phil. Trans. , , vol. 197 , p. 448 , et seq The only point to be emphasised is that the of Equation XIX .
of that memoir in this area is negative and less than unity .
The treatment is identical .
Below both the Type III .
line and the biquadratic , we have a space bounded by the cubic This is the of Type VI .
proper , with range from to ?
) being positive , and is fully discussed in th ' memoir just cited .
The area o .
is limited by the ) which Type V. , or , from to , describes the frequency .
Its full consideration will be found in ' Phil. Trans. , ' , vol. 197 , p. 446 , et seq. Below the Type V. cubic we reach the area of Type IV .
curve , or This has unlimited range in both directions and its treatment is fully discussed in my first memoir ( ' Phil. Trans vol. 186 , p. 376 , seq Theoretically , ypes IV .
and VI .
describe all types lying below the line .
The objection to their use lies in the increasing probable errors of their constants , vever good their general fit may be .
To warn the statisticians of this , the line , is drawn on the diagraln and the area below it is marked " " heterotypic area I use this terlu to signify that it is doubtful whether my skew-frequency curves , depending only on the first four moluents , can adequately describe distributions of types below line ; they require the use of the fiftb and nonlent c Their occurrence in practice , however , must be rare .
It will be noticed that the line ) is drawn the point .
This is the relation which must be satisfied in the case of POISSON'S exponential linnit to the binomial .
HeIlce , in the case of a distribution with , near this line , it is worth while investigating " " of small numbers " " is appropriate .
Above this line every real binomial distribution , i. e. , cases of and both positive R1ld than unity , and positive taking the binomial as ) must lie , for VOL. CCXVI.\mdash ; A. 3 and the right-hand side is clearly less than unity .
This limited area covered by the real binomial explains its relative infrequency as a descriptive series in practical statistics .
If , however , we take the negative binomial as admissible , i.e. , allow forms of the type , where we extend the possible area of a binomial down to the line Such a type of binomial is by no means of infrequent occurrence and can be more or less justified on priori grounds .
* Below Type III .
line , the values of and become in the mathematical sense unreal , i.e. , imaginary .
It is by no means certain , however , that such imaginary binomials with real moment coefficients may not , like imaginary hypergeometricals , give statistically good fits and be ultimately provided with physical interpretations .
( 10 ) Concluding Remarks.\mdash ; It is very difficult to assert finality for any scientific investigation , but I trust this second supplement to my original memoir on skew variation of 1894 has garnered the last harvest of possible types within the limits proposed in that investigation .
The object was the discovery of a system of frequency curves providing for.every possible variation of the first four moment coefficients of a distribution and provision for their rapid treatment and calculation .
Since 1894 much has been done by the provision of tables of the new functions and improved tables of old functions necessary to carry this out .
Diagrams like that accompanying this memoir , enable the statistician who has calculated the characteristic and , to select at once the appropriate type , from the position of the point in the plane .
The first diagram , prepared by Mr. A. J. RHIND at my suggestion , has been long in use .
For the present very carefully prepared and much extended diagram I have to thank my colleague , Miss ADELAIDE G. DAVIN , whose labours cannot fail to be appreciated by those having to handle practically statistical data .
Since the publication of my original memoir on skew variation , many attempts have been made to express the nature of skew distributions by other systems of curves or by expansions in series .
I have given careful attention to these competing systems and have discussed some of them elsewhere Biometrika , .
My chief objections to them arise from the fact that they either cover far less than the area ; or ( ii .
) involve constants the probab ] errors of which can be indefinitely great ; or ( iii .
) involve constants the probable errors of which have not been or possibly cannot be calculated .
In no case that I know of have they systematically been applied to extensive ranges of data , and the goodness of fit compared with that of other systems .
The existence of such competing systems is at any rate See ' Biometrika , ' vol. , p. 209 , and vol. XI .
, p. 139 .
Now collected in ' Tables Statisticians Biometricians issued by the Cambridge University Press .
' Biometrika , ' vol. , p. 131 . .
KARL PEARSON ON SKEW VARIATION .
noteworthy evidence that to attempt to describe frequency by the Gaussian curve is hopelessly inadequate .
It is strange how long it takes to uproot a prejudice of that character !
If the reader will turn again to the present diagram , he will see that the Gaussian frequency occupies a single point in an indefinitely extended area .
Those who support the Gaussian theory have to prove that no distribution occurs at a distance from the point of our diagram greater than could be accounted for by the probable errors of sampling of and .
These errors are known and have been and that position is.quite untenable .
Frequency distributions occur every day which by no manner of means can be described by Gaussian systems .
It has been said that my skew curves suddenly change their algebraic type and that the statistician is puzzled by a slight change in the constants and involving such radical changes in the equation to the type .
But if the reader examines the present diagram , he will see that the main ypes I , IV .
, VI .
and occur in areas , while the remaining types occur in the critical curved or straight lines which bound these areas .
Special cases like the Gaussian , the exponential or the rectangular distributions occur where critical lines intersect .
Now all these critical lines are really critical in the sense that a change of important physical significance occurs in this neighbourhood , and it is very unlikely that physical changes will be unaccompanied by sharp algebraical changes of form , such as are directly obvious in my curves , but are disguised by discontinuities in some of the proposed alternative expressions in series .
Any one illustration that the frequencies which occur in actual statistical data can practically cover the whole possible area of the planes , and can present frequency distributions which change abruptly in type , will suffice to confute both the argument that frequency is concentrated in or near the Gaussian point , and the argument that it is undesirable that skew-frequency curves should be so manifold in form , although how they are to change from to , to " " cocked hat to rectangle and to exponential forms without this abrupt change will be a puzzling problem to solve for the professed mathematician .
An illustration of this character has been several times referred to in the course of this paper .
Let us suppose there exists an indefinitely large population , each individual of which carries any number of characteristics which are correlated together , for simplicity we will say according to the normal law .
We may suppose that there are enough pairs of characters to all values of the correlation from to-l .
'Tables for Statisticians and Biometricians , ' pp. 68-71 .
An analogy might be given in the case of the ) of a ' cocked-hat\ldquo ; shape of finite range and a -shaped by a serics .
Here the trigonometrical expression ) the FOURIER'S series would be the if kept symbolic form , while the algebraic , form of the -curve would require two vertical asyllptotcs and its cquation would be wholly fronI that of the " " cocked-hat ' form .
The Fourier expression would o1lly disguise the real discontinuity .
In the same real discontinuity of form is disguised in the series which expl.ess skew frequency in terms * of a long series of moment coeflicients .
456 PROF .
KARL PEARSON ON SKEW yARIATION .
Now from this population we will take a large number of samples of individuals .
If in each one of these samples we calculate the correlation , , between two variates , then will not be equal to the value of in the sampled population , but the samples will give a frequency curve for , which is limited in range between and and is determined by the number of individuals in the sample and by the correlation of the characters in the indefinitely large population sampled .
We thus obtain a doubly infinite series of frequency distributions .
The general theory of such distributions has been worked out by " " STUDENT\ldquo ; ( ' Biometrika ' vol. VI .
, p. 302 , et seq. ) , Mr. H. E. SOPER ( Ibid. , vol. IX .
, p. 91 , et seq and Mr. R. A. FISHER ( ibid. , vol. X. , p. 507 , et seq The actual forms of the frequency curves are not usually expressible by simple single functions , but the ordinates and the admit of numerical determination .
The calculations are extremely laborious , but up to the present the members of my laboratory staff have calculated some 270 frequency curves with nearly 40 ordinates each for values of ranging from to 1 , and of from 2 to .
The great bulk of these curves show no approach to normality .
The values of range from points on the -line down to infinity , the distributions contain concentrated blocks , -shaped curves , -shaped curves , rectangles , trapezoid-like forms and every variety of skewness in doubly limited range curves .
Only in cases where is very considerable and is neither a positive nor a negative high correlation is there approximation to the Gaussian .
For a series of curves in which can be 5 and 9 , \mdash ; or both , if we will\mdash ; ten times these amounts , it is idle to talk about the value of the Gaussian curve in describing variation .
These frequency curves can be actually obtained by experimental sampling , although the process is laborious , and indeed were so obtained in the first place .
* They from observation and experiment .
The remarkable point about them is that they illustrate all the types we have been discussing andjustify sharp transitions in algebraic forms by showing that such transitions correspond to actual physical facts arising from experimental statistical data .
The whole illustration , details of which will shortly be published , indicates the evil of implicit reliance on a classical The Gaussian theory of error has , with great weight of authority , been applied to determine significant differences in statistical constants .
The theory of the " " probable error\ldquo ; must be justified in the case of each statistical constant to which it is applied .
Psychologists have been busy discussing the differences found in mental correlations deduced from small samples on the basis of significance judged by the Gaussian theory of probable error .
That theory has practically no application , as the " " probable error\ldquo ; has really no meaning in the case of the bulk of the samples dealt with .
Applications of the theory of probable error in other sciences than psychology to experimental results based on small samples will readily occur to the reader .
The conclusions may be correct or incorrect , but they are unquestionably based on an 'Biometrika , ' vol. PROF .
KARL PEARSON ON SKEW VARIATION .
inflation of the ussian point , , to cover all that may be happening in the whole area of possible points in our diagram .
It cannot at present be too often emphasised that such inflation is illegitimate , and that , as Dr. ISSERLTS has recently indicated , assulnption that the distribution curves of statistical constants follow the Gaussian curve is not legitimate , especially in the case of " " small samples which not only for many commercial purposes , e.g. , experimental brewing , but in numerous branches of science , e.g. , psychology , astronomy , and eveIl pbysics , are all that economy of money or time perlnits of recorded .
' Roy .
Soc. Proc , I ) .
23 .

