]\gt ; Mr. E. Cunningham .
The nctions , [ May 29 , The author has pleasure in his indebtedness to Dr. T. J. Baker information from which research inated ; to Mr. .
F. Hudson for the preparation of the micro-photographs ; and to Mr. J. L. Dixon , Bowen Research Scholar in the Metallurgical Department of the University of crham , for the care with which he has repeated and extended many of the earlier observations .
The unctions , ; of Normnl ctions occurring in Statistics .
By , JM .
A. , Fellow of St. John 's , Cambridge , Lecturer in Applied Mathematics , University , London .
( Communicated by Professol ' Pearson , .
Received , \mdash ; Read Jull , 1908 .
) I. ltctor , The resent paper erinated in an to discover the ynificance of certain functions developed ] ) rofessor Karl Pearson in a memoir entitled " " of audom tion , \ldquo ; and by him called to the of normal functions , are applied in the memoir named to obtain an expansion to represent a distribution ical about a point in a plane .
The distribution is not fixed but pends on a paralneter function being a function being distance the centre of distribution .
The fundamental In the coul se of the ] ) ration it soon appeared that the same function led olutiollS of ) of conduction of heat in bions f. the of sylnnlctr round the , the time taking the place of Tn fact , it that if sohution of that equation is in follD is ) function nrultiplied ) a ) olynoUlial , is to sclution of the blenl of the of infinite , the a ( distance ) vays zero .
in the to ; and all the of the equation fonnd which of the form is a lh ' co-ordinates the point 1908 .
] of occurring in Stafistics .
311 alone .
The result is that , as in other problems , is a spherical harmonic , or circular function , in the case of three and two dimensions respectively , that is of the form , while satisfies the generalised equation , where or ) , if This equation is solved by the Laplace transformation , and the ] ution , , taken round a loop contour from infinity round , is denoted by .
The asymptotic expansion of is which terminates if either , or is a positive integer .
Thus the solutions of in the cases of one , two , and three dimensions respectively are , , being a spherical harmonic of order , which reduce for certain values of the constants to simple well-known ions , .
:\mdash ; ( i ) To for , ( ii ) to for , ( iii ) to .
The one-dimensional solution obtained is not new , it practically coincides with the function of the parabolic cylinder , or the functions considered by in connection with the conduction of heat , by Thiele and Charlier in reference to probability , and by Hermit The two-dimensional functions seem to be new , but they arc shown to be the equivalent in polar co-ordinates of the functions considered ) Hermit and the relations connecting the two types of function are It is shown in the paper that a.ny function of and can be anded in the form , whele and a proof of the ence and validity of the is iven in the case where is independent of , with all indication of the extension to the general case .
Hence a series is obtained solution of the * Liouville , I and II .
' Reudus , ' vol. ) Mr. E. Cunningham .
-Functions , [ May , which reduces for to a function iven all over the } ) viz. :\mdash ; , the coeHicient being obtained as above .
This series is shown , by direct expansion , to be equivalent to Laplace 's definite integral solutio of the cooling of an infinite plate , Various other expansions are iven a to Neumann 's expansions of in a series of Bessel functions .
The last part of the paper considers the equivalent functions in Cartesian co-ordinates , in terms of Hermite has expanded any function. .
It is shown that an equivalent method of Hermite 's expansion is that of equating the successive moments of the series to those of the given , l'unction , with special reference to the best surface for a given frequency distri } ) ution in two dimensions .
The series of approximations so obtained is compared with that of which is ultimately the same , and is shown to converge rapidly under the same conditions .
An example is then given } coutour lines of a surface obtained by terms up to the fourth order .
These contour lines show much similarity with those actually ) rved in statistical work , such as those given by Perozzo in his memoir .
The possibility of a fairly close approximation to a iven distribution ) means of a knowledge of the product .
moments up to the fourth order is thus blished .
II.\mdash ; The 1 .
\mdash ; In older to obtain the types of function suitable to the prol ) lelllS mentioned , the most creneral solution of the equation will be found in the form T. , where is a tiull of alone , lnction of , and of the ular inates .
begin with , cases and t dimensions are considered in the , it becolJlcs If , and functions of and respectively , we have ' so that both sides lnnst be independent of and ' Camb .
Phil. Trans tlsl , 1904 .
'Annali di Statistica , ' , 5 , 188.3 .
1908 .
] of Functions in Statistics .
313 Hence and ) where is any constant .
Thus ( b ) \mdash ; The equation in polar co-ordinates is satisfies this equation , a function of only , and being independent of , the functions and satisfy a constant .
Thus , and yain , putting , and , the same procedure as above gives and .
( 2 ) ( c ) Thrce Dimensions.\mdash ; Exactly similar procedure in this case leads to the conclusion that , if .
Z. a function of and must be a spherical harmonic\mdash ; say of order must be , and that then satisfies the equation .
( 3 ) ( d ) .\mdash ; In this system of co-ordinates eneral solution is sinh , where satisfies the equation \amp ; .
( 4 ) The equations ( 1 ) to ( 4 ) can be brought under the same form .
Putting in the four cases respectively , and also the equations become , using dashes to denote differential coefficients , , ( 6 ) , ( 7 ) .
( 8 ) as the standard form , the particular solution dealt with in this paper will be denoted by he corresponding solutions of ( 5 ) , ( 7 ) , and ( 8 ) are therefore Mr. E. Cunningham .
The -Functions , [ May 29 , and the complete solutions of the original equations are ( 9 ) , ( 10 ) , ( ll ) , ( 12 ) The simplest solutions are given in the four cases as follows:\mdash ; ( i ) , ( ii ) , ( iii ) , ( iv ) 2 .
of and Laplace transformation of the equation , we find that must tisfy the / ?
while the must be such that the same at the miti to nd positive , this will be attained by a colltoul .
of a leal axis from to then a circle of Iius ftlld a line the real axis to inhnity .
A similar contour the point instead of 1 ) ) ill ) ) 1 and we have for the cxpression ( ) } The the lur ( 13 ) which if either or is a positive The portant c is that in .
is a ) osil i : in that ( 14 ) , il this \ldquo ; ' omCb l. This es on is .
1908 .
] Class of Functions in tistics .
3 .
jcting Successive lctions .
eqnations\mdash ; Thus .
( 16 ) this twice and simplifying by means of the fundamental differential equation , we find .
( 17 ) ( b ) of \mdash ; Since the derivative of any solution of with respect to is also a solution , and since } , we deduce at once , after adjusting the numerical coefficients , the following relations in the cases of one , two , and three dimensions respectively , ( 10 ) , ( 11 ) ) ( 18 ) III.\mdash ; The } 1 .
liestrictin the work in this section to functions arising in the one-dimensional case , which are briefly called the ] , it will be convenient to write for and for , and in place of From ( 18 ) we then have ' assuming that is an integer .
If is an have , similarly , Thus in either Thus , as a function of is identified the function } .
by Conl p Mr. .
Cunningham .
The -Functions , [ May 29 2 .
s.\mdash ; The expansion is , which , if an odd and if is an even Hence , coefficient of in { and thereSore ( 22 and ( 23 the former mation being for even values of and the latter for odd , ether , ( 24 of Tl oing properties will now be applied to obtain an expansion of Laplace 's solution of the problem of the an finite l. We have , , for as as is a function when f to . .
in , the subject of as above , we obtain 1908 .
] Class of Functions occurring in Statistics .
317 the following series to represent a the same equation reducing when , to : where In accordance with a remark made above , it may be noted that , when is real , is imaginary , so that the expansion is wholly real .
A different method of establishing the same expansion is given below .
IV.\mdash ; The cncral co-Function\mdash ; In fcgral Let The asymptotic expansion of the -function shows that the integration in regard to to the infinite upper limit is valid , and the older of integration may be inverted .
Thus For every value of the subject of integration is zero to the second order at least when is infinite if is positive , and if ' is negatiye .
positive for the sake of argument , , if , or if .
Thus but .
( 27 ) , considel the moment VOL. LXXXI.\mdash ; A. Mr. E. Cunningham .
The -Functions , [ May , if is a positive integer , is a polynomial subject of integration has no residue at ? .
But if is zero a negative integer , the residue is Thus , ( 28 according as , or .
In this is not necessarily an and are assumed to be integers .
If these results , ( 26 ) and ( 27 ) , be applied in the one-dimensional cas considered above , assuming that , we obtain at one .
Consequently , if and is a function satisfying , and reducin ] when to the iven function .
Thus the expansion is obtained \ldquo ; nnctions ia to the special values of and suitable to the deyelopment lnnctions in plane , integer .
is an relation ( 19 ) , we have .
In particular , if we start the function independent of .
put ?
, so ticular set of functions ulay conveniently be called the -0i -fnnctions .
If be , a p , so that is polynomial , .
the functions found essed in terms of the fttives ol the with respect to , and Since ivatives of aIly solution of with respect tu co-ordinates ' : and are also solutions , we may expect all ] utious to expressible in ) of tho derivatives with respect to of the solution .
1908 .
] a of Normal ctions occurring Statistics .
319 The following equations may be shown to hold for the planar -functions , , and , therefore , being an , 0 ( 32 ) , we have and giving the values of in terms of and .
Thus the connection is established between the planar -functions and Hermite 's functions , to which reference is made below .
of Planar 1 .
Suppose that a function of and may be expanded in the form where being a parameter .
Then , the Fourier method for all values of , we have ' and In order to find the coefficients , we now use the rals developed above , and obtain , ( 33 ) with a similar expression for .
Thus , in order to obtain the coefficient of , we make nse of the allied function , which bears to it the relation .
The values of and to be taken do not appear , but it will be shown presently that takes all integer values and ?
takes those values for which is zero 01 ' a positive .
The expansion is therefore one in a series of polynomials .
responding expansion in rtesian co-ordinates is iven by 'Comptes Rendus , ' vol. 68 .
Mr. E. Cunningham .
The -Functions , [ May 2 !
2 .
Th(siolb of the Dcfinite Integrat in \mdash ; An expa1 sion analogous to ( 26 ) will first be obtained .
standing , as everywhere , , let ' stand for and for .
Then , taking all valu ' and for which are both positive integers , ! .
coefficient of ) in ; or , 'taking all half-integer values from to infinity .
Again , taking the same range of values for Expanding the last expression by Taylor 's theorem , we have Thus , putting and This gives the proof of the relations given above ( 32 ) .
and in ( 35 ) , that expression , where ( uv ) are Cartesian co-ordinates .
Hence ' by , where .
Now put and .
Then the above becomes The sunmlalion oughout for all values of and such that and are ) ositive ihere thus correspond to terl in sulllmation an equal but opposite value of and , 1908 .
] a Class of Normal Functions occurring in Statistics .
321 jhese terms will only differ by being } to ering also that ?
is an integer , .
Hence .
( 36 ) This expansion is gestive of Neumann 's expansion in the theory of the 3essel metion .
Another theorem is given below , which , in some respects , is nearer analogue .
The development of the Laplace definite integral follows once from ( 36 ) and is identical with that in the last paragraph ( 33 ) .
The nethod of this paragraph shows that the values of and ?
are as there tated .
3 .
The Expoension of in a of -Functions.\mdash ; As in the last section , ( the summation being taken for the same values of and as above ) \ldquo ; where the summation extends only to values of such that is a ositive i. Thus is expanded in a of -functions in the variable , and from this may be derived , by neans of relations ( 32 ) , the expansions of lsin , differentiations with to , and 4 .
The ergence and Validity of the of a Function in a of -functions.\mdash ; The proof of the validity of expansions in a series of normal functions is always a difficult matter , and a general theorem to cover them has yet to be established .
A discussion is here given of the expansion series of zonal -functions of a function symmetrical round a point , which seems capable of extension to the general -functions in a plane without much increase of difficulty , though with considerable addition to the analysis .
The multiplication of different discussions for each particular set of normal Mr. E. Cunningham .
The ctions , [ May 29 , functions seems undesirable , however , and the present investigation will therefore be limited to the single case of the functions .
A erent method appears possible in the discussion of the validity of the processes in the development of the Laplace integral given above .
In the , the basis of the work is the definite expression for the -function , The expansion to be justified is , where Consider the sum of of the sel.ies A necessary condition is at once obvious for the valiLlity of the processes involved .
For , if is very large , since the -contour includes points for which real part of is negative , is not everywhele finite on that contoul But if be such that , no matter reat y may be , , where is some finite ative constant , we ma ] the -contour such that the real part of ?
is always oo.reater than , and then the rations lequired are justified .
Let Then , ?
is finite , is , since tho ?
-colltour includes no point at anishes for any value of contour .
Also , by the condition to ) satisfied , the of from values of reater than finite nitude can bc tade c slnali as we ) lease by taking ] arge .
Thus it to be ememljered that of the is sought when / tends uity , nd that since the contour encilcles the point the contours .
have ) assumed such that does not vnnish for a ) of within these ) orcatel D 1908 .
] a Class ] Functions occurring in 323 than umity on the major portions of the con tours , so that the subject of therefore becomes indefinitely great as increases .
Consider now the integral divide into two pol .
Cions for which ?
less than respectively , them and Let be the same integral with the limits and .
( say ) .
Considel ' the latter part ; it is absolutely and uniformly and thelefore be written ( say ) .
Within the contours considered there is no pair of points which .
If the ?
-contour be extended so as to include all the which yanishes for points on the 1-contour , we add to the contoul , but not , and this addition since the contour does not include the point , and since is positive .
Thus the -contour in may be supposed extended so to include the point at which vanishes for of or Let the 1-contour be now specified more precisely as that part of a ol radius , with centre at , that lies on positive side uf straight line on the real part of ether with rt of line which lies within that circle .
This contour is considered in , as follows : , the straight portion ; ) of the curved ljacent to and a small at the centre of the , the remainder .
For values of on , the -contour , extended to include to a con tour just surrounding the complete -contour .
For on and this contour is extended in the negative direction Mr. E. Cunningham .
The -Functions , [ May 29 , to be a short distance on the negative side of .
The -contouris considered in three portions similar to .
: , the straight portion ; , the arcs contiguous to and subtending an angle at the middle point of , the remainder .
Thus the whole of the double integral is divided into nine portions obtained by } one of the partial contours with one of the partial contouls Now remains always finite as and become , except for , a value which lies outside the under consideration .
Thus and .
Thus , if .
Hence , if therefore , on the arc where , and this quantity tends to zero as diminishes , provided remains finite .
On the arc , the expression certainly remains finite .
On the straight takes its reatest yalue at the point where it adjoins , i.e. is finite on , but along a length of measured from the real axis it becomes comparable with with .
with , where is large and finite .
Similar considerations apply to the various portions of the -contour , and them together we that each of the portions of the double tends to zero as tends to infinity .
Thus for all values of and also tcnds to zero as ) ecomes infinite , for any value of less than ) , in particular if Thus nd is lependent of the lower limit The , and -contours the same , this last will be unaltered if and be interchanged .
Hence 1908 .
] of Normal Functions Statistics .
325 where where tends to zero as becomes large ; since , by reasoning similar to that used above , the double with the -contour extended to include the point vanishes as tends to infinity .
Thus and Finally , thel.efore , , whatever the value of Apply now the same treatment to the integral , viz. :\mdash ; The coefficient of in the exponential in the second integral being now greater than that of , we carry out first the integration in regard to .
The first part of is , as above , equal to .
Thus In extending the -contour of this last integral , instead of the ?
/ , -contour as in , we add to it ; , as above , standing for a contour encircling , and not zero .
As before , the extended integral is zero , and therefore the unextended integral equals .
Thus for all values of We have .
E. -Fnmctions , [ May 29 , and ; , therefore , finally and if the function is continuous , It will ) noticed that success of ) method depended on the s1ubject of in the a eoInetrical progression .
If the general expansion for the unsymmetrical case is treated similarly , the subject of integration is a double progression and may be treated , but the analysis will not bc carried out th th 1 .
was lneutined at the outset , the linear -functions have been nised for some time as suitable fnnctions purpose of a given frequency distributioJl to any degree of } ) proxinlation .
The method commonly ) in fitting the coefficients is to make the btccesve Dloments of the equal to those of the iven distribution .
* But it does not ) to ] ) been noticed that this nlethod i exactly equivalent to of ) the coefficients .
, is { immediate the fact that the lnctioIt : is a ( .
Thus the ) thod by luoments , comnlon]y as an artificial of obtiig a series to repl.esent a distl iution of a particular ) ' ) , is leally the and unique of by a of -fUnctions .
anothcl thod of a series of successive approximations to a tril ) in the } ] if be developed , leads to series of -functions .
The lelld the cociticicnls is that of the molnents , so tcly the series obtained by must ansio in -functions .
lt vever , that , for statistical ( ions , of meant u by give .
cit. .
Phil. ) 1908 .
] Class of Norr Functions occurring Statistics .
327 more rapid approximation .
The argument as to the nitudes of the coefficients , however , appears equally applicable to the present series .
It is shown in the memoir quoted that the coefficients rapidly only if the standard deviation be small .
Let this be assumed in the series , where .
Then , if be the moment of the distribution , and , we have and in pal'ticular and hence , since is the standard deviation , is zero .
ument , is of the order of magnitude itnd hence , by taking the above equations for different values of , we see that the coefficient is of the order \mdash ; the word order being used in the same sense as in the work cited .
Thus the coefficients will diminish with about the same rapidity as the quantities , while the of the equations much pler .
Pass now to the case of two dimensions .
eference has been made to Hel.mite 's note on the expansion of a function in a series of the form , vilich l been shown to be equivalent to an expansion -functions .
It is a problem of some importance to obtain a series to represent a distribution in correlated characteristics , such , for example , is considered by Perozzo in his memoir on the of wife The lines of equal by erozzo are far from being concentric ellipses , as they be if the frequency surface were a Gaussian normal surface .
It is worth .
whether a few terms of the expansion in -functions cau be made to an approximation that is tolerably near .
The theol.etical possibility of repre- the distribution by a is not considered , the conditions of validity of the expansion are lnlost certainly satisfied .
Hermit gives a method of obtaining the coefficients .
Using the notation , ?
and are both polynomials of degrees in and .
If be exlJanded in the form , it is proved that 'Annali di bistica , ' 1883 .
Mr. E. Cunningham .
The -Functions , [ May 29 , where is a numerical constant depending on and alone .
Now it will be much more convenient in practice to obtain the coefficients by means of the successive moments to the corresponding moments of the series .
This is equivalent to Hermite 's process , a polynomial ; so that above is a linear function of the moments .
Putting ?
can be reduced by successive integrations by parts .
If , we have whatever the values of and , and , similarly , the is zero if ?
whatever the values of and .
But if , and !
and if and and this last is zero if is an odd the notation , and , if we multiply the equation by , and integrate , we have there terms on the right , of which those { zero in which the sum of the subscripts of is odd . .
the first fe equations , by choosing the origin at the mean of the distribution , and the constants so that the constants all become zero .
By this means the normal face that best fits the distribution is obtained , viz. , But the fact the contour lines in the example quoted are by no means concentric ellipses with centre at the mean shows that the fit is not yet sufficiently good .
The next approximation to the distribution is 1908 .
] Class of Functions occurring in tistics .
3 The equations the coefficients of the fourth order terms are : , etc. Thus in terms of the moments the coefficients are quickly obtained , and 's argument as to the rapidity of approximation applies equally well here if the standard deviations are sufficiently small .
But the labour of obtaining the moments is very great , and a suitable example for testing the closeness of approximation is not available .
What will be done here , therefore , is to show more or less generally how , in a simple case , contour lines can be obtained which are approximately of the type found in the example given above ( see diagrams , Perozzo , loc. 2 .
An of Contour Lincs Two uensions.\mdash ; The example chosen here is that of .
and it will be assumed that there is symmetry about the line The most general first correction is , where , , .
In ular , if , and is negative , vanishes on the straight line , and the ellipse ; on the positive side of it is positive outside the ellipse , and }ative inside , and vice on the negative side of Thus the effect of the correction , while not altering the mean or second moments , is to introduce skewness about , the maximm ordinate being displaced in the negative direction along .
Outside the ellipse the surface is depressed on the ative side of , and raised on the positive side .
The zero-line of is .
shown in the figure ( curve A ) for The contour lines on the poSiGiyG part of the surface will now be oval curves .
ounding the point of maximum ordinate with their reatest width parallel to , and crowding closely together in the neighbourhood of the zero curve .
These have not been shown in the figure to plication .
variation in their form could be produced by introducing a term , but the effect of skewness is again the most noticeable .
Consider now a fourth order correction , which for the end in view has been conveniently taken to be The which this is is drawn in the figure ( curve B ) , and ortions of the plane are indicated in which it is positive and ative respectively .
Thus the effect of this COl.rection on the contour lines in the hbourhood of the maximum ordinate is to contract them in the direction The , and to in direction .
A sel.ies of contoul lines is drawn for ; so ) the lines shown are curves of equa ] frequency for the function The curves arc distinctly similar in ) to Perozzo 's , though no attempt ] been made at exactly what the coefficients should be .
Probably a better nation would be found with a rather largel coefficient for the order corrcctio1l , as skewness is not so marked it should be .
Li of A\mdash ; the B\mdash ; the cve tectic R Similar curves can clearly be obtained there is correlation to be taken into account .
For by making the change of variable is turned into a normal surface with correlation , while are linear functions of and Eutectic Research .
No. 1.\mdash ; The Alloys of Lead and Tin .
By WALTER , B.C.E. , with .
A. TUCKEr , .
( Communicated Dr. R. T. Glazcbrook , .
Received June , 1908 .
( From Ph atory .
) ( Abstract .
) Attempts to prepare pure eutectic alloys of known constitution led to the discovery of discrepancies between the authors ' experiments and the data on lead-tin alloys published by Roberts-Austen .
A complete redeCel'mination of the equilibria of the lead-tin system was therefore undertaken , by both pyrometric microscopical methods .
-curves of the taken by both inyerse-rate and differential methods are given ; these , with microscopic data , lead to the equilibrium diagram shown in the .
This diHers pcipally from that given by in eutectic point is placed at a concentration of 63 per cent. of tin instead of 69 , that the eutectic line towards the lead end of the series terminates at a concentration close to 16 per cent. of tin , and that a series of transformations the line ) have been found in the solid alloys near the lead end of the series .
The discrepancies of these results , as regards the eutectic composition , arise from the more delicate lnethod employed in the presentl.esearch .
While alloys within cent. on either side of the true eutectic composition show no detectable diffelence in or -point , the presence of small amounts of excess of either constituent can be detected by the microscope , and in this manner the composition of the pure eutectic has been ascel.tained .
As solnbility of tin in solid lead , it was found that the of the eutectic arrest-point in alloys containing less than 16 per cent. of tin -Austen , ' Fifth Report to Rese rell Committee of the Institutio of Mechanicnl iueers , '1897 .

