Biometrika Trust

The Generalised Product Moment Distribution in Samples from a Normal Multivariate Population Author(s): John Wishart Reviewed work(s): Source: Biometrika, Vol. 20A, No. 1/2 (Jul., 1928), pp. 32-52 Published by: Biometrika Trust Stable URL: http://www.jstor.org/stable/2331939 . Accessed: 20/03/2012 03:42
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Biometrika Trust is collaborating with JSTOR to digitize, preserve and extend access to Biometrika.

http://www.jstor.org

THE GENERALISED PRODUCT MOMENT DISTRIBUTION IN SAMPLES FROM A NORMAL MULTIVARIATE POPULATION.
By JOHN WISHART, M.A., B.Sc. Statistical Department, Rothamsted
Experimental Station. 1. Introduction. For some years prior to 1915, various writers struggled with the problems that arise when samples are taken from uni-variate and bi-variate populations, assumed in most cases for simplicity to be normal. Thus "Student," in 1908 *, by considering the first four moments, was led by K. Pearson's methods to infer the distribution of standard deviations, in samples from a normal population. His results, for comparison with others to be deduced later, will be stated in the form
1 w-1
N-.

dp =
V

-A

da .....................

(1),

2J
N
A=2
2 s2, 2

where N is the size of the sample, and ~a= a- being the standard deviation of the sampled population, and s that estimated fromthe sample. Thus, if xl, x2, ... xN are the sample values,
N

and

Ns2

N E (X_ -c)2 1

When bi-variate popuilations were considered, other problems arose, such as the distribution of the correlation coefficient and of the regression coefficient in samples. These problems, taklen by themselves, were found to be difficult, and only approximative results had been reached, when, in 1915, R. A. Fishert gave a formula for the simultaneous distribution of the three quadratic statistical derivatives, namely the two variances (squared standard deviations) and the product coefficient. Thus, let x1, x2, ... $N represent the sample values of the momnent x-variate, and Y1, Y2, ... y the corresponding values for the y-variate, let a1 and c2 be the standard deviations of the sampled population and p the correlation between x and y. We then calculate the following statistical derivatives from the sample: Nac-= Y.($) Ny = Y,(y)
1 N 1 Nsi2E(

Ns2
N

1 N 2 )( 1

- y_)2

Nrs,82

= E ($Y-1).

* Biometrika,

Vol. VI. 1908, pp. 4-6.

t Biometrika, Vol. x. 1915, p. 510.

JOHNWISHART If If weut weput

A

33

N =2(rJ2 (I _

p2)

B-

=2o-22 (I _

p2)

H=- -2lo-2N -P p (l 2of

h = s1s2r, then Fisher's result, for the simultaneous distribution of a, b and h, may be put in the symmetrical form* 1 H ~~~A I,22 ebAa-Bb-2Hh. a h ?v- dadbdh dp = h b 2 H B N2

a = sl2,

b = s22,

......... (2).
The distribution of the correlation coefficient was deduced by direct integration from this result. Further, K. Pearson and V. Roinanovsky, starting from this fundamental formula, were able to deal with the regression coefficients. Pearson, in 1925 t, gave the mean value and standard deviation of the regression coefficient, while Romanovsky and Pearson", in the following year, published the actual distribution. New problems arise when we are dealing with three or more variates. There are, for example, the distributions of the partial and multiple correlations, and of the partial regression coefficients. In this domain our knowledge is still far from complete. The partial correlation coefficient has been shown to be distributed exactly as a total coefficient, when the size of the sample has been reduced by the number of variates eliminated?. The actual distribution of the multiple correlation coefficient, for the particular case where no real correlation exists in the sampled population, was shown in 192311to be of the form

dp

r 2 )

2-f3

N-n-2

(Nn)

r(n_ l)

(R2

(1-R2)

d(R2) .

(3),

where n is the total number of variates. A recent experimental investigation by the present writer?Tmay be regarded as confirming this result, as well as was to be expected on the basis of the inadequate numbers used. On the other hand, L. Isserlis**, in 1917, found approximations to the mean value and standard deviation of the multiple correlation coefficient, in the case of three variates, while recently P. Halltt, for an n-fold system, determined (a) the mean value, and (b) the standard deviation, of the nmultiple and partial correlation coefficients, as far as terms of the order of N
* See Romanovsky, ConzptesRendus, Tome 180, 1925, p. 1897. t Biometrika, Vol. xvii. 1925, pp. 195-196. 646. Proc. Roy. Soc. A. 112,1926, p. 1. * Bulletin de l'Acadfmie des Sciences de l'U.R.S.S., 1926, p. 1-2. ? R. A. Fisher, Metron, Vol. iII. Nos. 8-4, 1924, pp. B. 11 A. Fisher, Phil. Trans. B. Vol. 213, 1923, p. 91. Memoirs, IT Memoirs of the Royal Meteorological Society, Vol. ii. pp. 29-37, 1928. Rothamsted

Vol. xiv.
** Phil. Mag. Vol. xxxiv. 1917, pp. 205-220. ++ Biometrika. Vol. xix. 1927, pp. 100-109.
Biometrika XA

34 ThteGeneralisedProduct Moment Distribution in Samples

What is now asserted is that all such problems depend, in the first instance, on the determination of a fundamental frequency distribution, which will be a generalisation of equation (2). It will, in fact, be the simultaneous distribution in samples of the n variances (squared standard deviations) and the n (n1) product

moment coefficients. It is the purpose of the present paper to give this generalised distribution, and to calculate its moments up to the fourth order. The case of three variates will first be considered in detail, and thereafter a proof for the general n-fold systemnwill be given.

Distribution. 2. Tri-variateProduct Moment

Let the frequency distribution of the population sampled be

dp =
xe1L(X
1

I
(27r)i a-,a, a,
- Ml)2 2

Ai (yr2)2 +Y -M2

22

(Z --

-A3 rn3)2a_+2+2(Y(

2) Z-ma )A23+ 2'(z '' +zr

'-'

ml)3,A

2'x -,nt)Y x-

2)Al

dxdydz ...
where A is the determinant IPet s, t = 1, 2, 3, and A8, is the minor of Pst in A. Now let x,, x2, ... xN represent the sample values of the x-variate, and
Yl,

(4),

Y2

...

YN,

Z1, Z2,

...

ZN

be the corresponding values for the y- and z-variates. Then the chance that
X1, yl, Zx, X2, Y2, Z2,
.

..-

XN, YN, ZN

should fall within the elementary ranges dx,, dy1, dzl, ... dXN, dy.N, dzN is

= a dpd= --3; N 3N (27r)(212~2~A

xe

-11 iA 1

NIix -rm1)2
(rl2

IL

Ali +

(y-

(z -rnt)2 (Y-rnMO 2+ (Z8)

2)2

+ 2 (Y

77z2)

(Z - MS)
A2

X + 2 (z m3) (X ml) 4131+ 2A(

0'30'1

l (Y- in2) ~ 2 12] (

O

O12

'T2

....... dx,dyldzldx2dy2dz2 dXNdyNdzN ............

The following statistical derivatives are now to be calculated from the sample and substituited in (5):
N N

(5).

,N-x= (x)
N

N-y =YX (y)

Ns2 Nr31S,S1=
X 1

N z = Xw (z)
N Ns32=X(z_z)l
1

= Nsl2 !, ( _ z).s
1

(y_ y)2 (Z-Z) (X-&)

N (y -

Nr23S28 =

Y) (z-2)

Nr12S1S2
1

(x -)

(Y- Y)

In order to transform the element of voluime,we require to extend somewhat the geometrical reasoning employed in the two variate case*. The N values of x may be regarded geometrically as specifying a point P in an N-dimensional space,
* See Biometrika, Vol. x. pp. 509-510.

JOHN WISHART

3"

whose co-ordinates are x1 - $, x2- ., ... xN -. Similarly the N values of y and the N values of z specify points Q, R in the same space. When Y and s8 are fixed, as when a particular sample is chosen, P is constrained to move so that its perpendicular distances from the line x, = 2 ... = xN and from the "plane"
XI+X2+

...+XN=nX

remain constant. It must therefore lie on the surface of an N- 1-dimensional sphere which is everywhere at right angles to the radius vector X1-=2 = *. = XN. The element of volume is then proportional to (VNV8)N-2 ds8dy. For the factor of proportionality we require the entire area of the surface of a sphere in N - 1 dimensions. If of radius r this is, for N = 3, 4, 5, 6, 7, 27rr, 47rr2, 27r2r3,?7r2r', 7rsr6,
N2N

the general result being

2.

ri-2

Thus we have a contribution to the transformed element of volume of

NS-1 N-2

2. rs2

2 1N-2ds8jd

By similar reasoning Q and R must lie on concentric spheres in the same space, and there will be corresponding contributions to the transformed element of volume. Let the radii vectores OP, OQ, OR be cut by the unit sphere whose
p

N
r23

R centre is at 0 in the points L, M, N. Then LMN is a spherical triangle, specified by the nature of the sample. To find the chance that this particular triangle should be chosen we note that, P being fixed, the chance that Q (or M) should fall within the elementary range dO1 being the angle 2- LOM) is equal to (01
cosNlOidGi .fcosN3Oid9
3-2

36

The Generalised Product Moment Distributtion in Samples T zN-1>

2

i.e. to

-CcosN301d0j.

Similarly the chance that R (or N) should fall within dO2 (LON= 2-- 0) is equal to

N/s

-2) cos

02 N-3

dO2.

But the points M and N do T(IAT2 independently. They are connected by the not vary relation that, if ( be the angle between the planes LOMand LON,
-2r2
csl-r2

-rI2r3

/1 - r132

Now LM being fixed, the chance that LN should fall between the angles ( and ( + do, measured from LM,is equal to
sinN4(pd(4

? f Sin' N)4
0

=
V/-rT(IV 2J)

sin N-4(

do.

The transformned voluine element will consist of the product of all the above The exponential term in (5) is easily expressed in terms of s1, S2, S3 probabilities. and the r's, and we have

(N
?T3T(

-1)
2
F-

(N2 -N2N-2A-

3 (N-2) 2

r1 (
2 _-M_

o 22 ) T (N3)~~-1 2 ) (8s2 -2 n3 )2
0
-__)n___2 2 2 _323

xe2AL

o~~-12

+ (
311

A12+

-3 + 263 +

+ r2

+ IA33 2 23233

S2J3

(g -n2)

A23

+ 82 I2 r.1s3s1 + (f - M3)(P - MI) A32 2 'r,281 + (I - mI)(Y - i2) A2]

x SI2 S22N -2s3N-2 cos N-3 91 cos N-3 02 sin N-440 d$d,d- ds1ds2ds3d01d02d(
............ We can integrate out immediately 8 for , y anid z. The integration
N- 1

(6). produces

(V /27r)3N'o1 o3 VA, and we get ao2

_

dp

7r2

LITpN

(NI2) 1

Dr

F-(N2- )rtNF~1lA _ 2A L0i2

3)8.2o.

22T2

+C-~-A +a-A
03' 2
03

2S23A 23+ 23sA+28Is2A] +2232

2 3 0'303 oI

xe

X S81N28S9N-283N-2

cos

N-3

COS N-'02 ......... .(7).

sin N-40 ds,ds2ds3dO1 dO2do

JOHNWISHART
Following Romanovsky * in the method of expression, let us put

37

s,2= a
S22 =

SIS2

sin61

= h,
=g, cos O.COS43+ sin 63 sin 62) = f;

sis3sin62
S2.8 (cos 61

S32 = C

(for

r12=

= sin 6,, r,3= sin 62, and r5#.1 cos=

r32

For the Jacobian of this final transformation we have

? a@f ah9
a? a(a,b,c,3-f,g,@h)o=8818s283 , a2,
ag2

= 8 {bc -h2) (ac (ab

abc denotes the determninant where I a g N-2{(ab-h2)(ac- g2)} 2 V ~~(abcY~ ~~~-3 aN (bc) 2
N-3
*

8sS3 S.3S33

61 cos2 62 sin
-

g2)

4 IbC }t

h g'

h bf

f
a

ci
V-4

The element of voluime then becomes on transformation

AT- 4 2

IabcIF2

(
2

abel-

{(ab- h)(ac- g2)} a h g N-5 =ih b f

2)(cl{(ab

h2) g2(ac

g f c a very neat result. Finally we have for the simultaneous distribution of the three variances and the three product moment coefficients the symmetrical expression

dp =

~ N 1)
7r
2

I
-__N3__2

L N3\V-1~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

)rN

2)
N a b
2 "II {2^11+ 1

2(
2 A22 +

__).2-3

ch C
A33+2 (732

x xe

f &X+2-LA.+2 0':; 73(71

0'1

A12

a h g x h b f g f c
*

N-2.

dadbdedfdgdh.

ComptesRendus, loc. cit. See also Mletron,Vol. v. No. 4, 1925, p. 31.

38 The GeneralisedProduct Moment Distribution in Samples

We may simplify this expression by writing

k All
2o.12A'

Af

2o~2Al

2N
2Ff2f 0.I 2o2

A2

T\>

'

=23r

AGN ,A 2a,ao1A' X
i

N AS 2ar'A' N A12
H=2

2a0o2'* A 2

when it becomes

=1
dp__ N__2N_

A l Nl1 G
H B F

,-Aa-Bb-Ce2Ff-2Gg-2Hh

a h g x h b f g f c
It is to be noted that i abcI is equal to
s12s22S32 1rp, 1, p, q=1,

-5

dadbdcdfdgdh
............ .(8).

2, 3.

This is the fundamental frequency distribution for the three variate case, and in a later section the calculation of its moment coefficients will be dealt with. 3. Multi-variate Distribution. Use of Quadratic co-ordinates. A comparison of equation (8) with the corresponding results (1) and (2) for uni-variate and bi-variate sampling, respectively, indicates the form the general result may be expected to take. In fact, we have for the simultaneous distribution in random samples of the n variances (squared standard deviations) and the n (n2 1) produictmoment coefficients the following expression: A1l A12... Ain A21 A,2... A2n dpAn,
A21,. An N-i

in (n - 1

- 1)(f

N
-

2).
...

N(f

)
-

x e- Aliai
al,
X an,

A22a22

A,nat,n-2A2al2
2

2A18a

- ... -2A"lnan-la

a12 ...

a,,
a2.,

N-n-2

a2.

dallda,2... da,,n .

(9),

an, ... ann

A,

where ap, = spsqrpq, and A p = 2N

aPpqdq,tPemqioo1,f2,

A being the determinant

3i...nn,

in the minor of ppq A. and Apq

JOHN WISHART

39

This result can be proved by the aid of the following general geometrical considerations. We shall begin by defining the quadratic co-ordinates of a set of points, and thereafter develop the argument by the use of these co-ordinates.

of (a) If XPI,IX2, I Ipn are the rectangularco-ordinates a point (p) in space of n dimensions, then the conifiguration of a set of n points relative to the origin may be specified by the co-ordinates
GPq =

k=1

(XpkXqk).

These co-ordinates will be unchanged by any rotation of the whole system about the origin, the configuration of the set of (n + 1) points including the origin being unchanged. (b) The determinant Itpq1,p, q = 1, 2, 3, ... n, on expanding the several terms, may be recognised as the square of the determinant Ixpq 1, p, q = 1, 2, 3, ... n, and is therefore equal to the square of the volume enclosed by completing the parallelfaced figure having one corner at the origin and edges running to the n points of the figure. This volume will be represented by vn. the plane space passing (c) The perpendicular distance of the point (n) fromn through the origin and the points (1) to (n-1) is v I

(d) If the points (1) to (n - 1) are fixed, and that of the point (n) is specified

by the co-ordinates nll

tna,

...

{nn)

then the eleinent of volume in the neighbourd1,

dn

hood of the point (n) inay be represented by 2v,, For the Jacobian
-v

.2 d.n.

X11
tn ** $5nl)

Sa*
X12

1
.
XlnCI

(tX

iil

2n

(Xcni,X

n2l

...

$n)

:
2xln

:
2x., 2Xnn

(e) If in a space of N dimensions the points (1) to (n - 1) are fixed, and that of the point (n) is specified by the co-ordinates tm, fna, ... {nn within ranges dtnl X d n2, .... dotnnl then the point (n) is free to move on the surface of a sphere in N-n + 1 dimensions, of which the centre is the foot of the perpendicular from (n) upon the space The radius of this sphere is therefore v)t containing the points (1) to (n-1). and the area of its surface is 2
4 (N1

+n1

Vn)Nn

(see above, p. 35).

40 The GeneralisedProduct MomentDistribution in Samples

This surface is everywhere normal to the space passing through the origin and the points (1) to (n). But the element of volume in this space is

2v.

2v dg911X ) dtn2

... dtnnX

Hence the entire volume that may be occupied by (n) is U N-} + 7rT (N- {+1) 10 V7,XNz-1 dgM, ...d &dtni,n2 N r n + lj Vn
IP

nn

................

(10).

(f) The quadratic co-ordinates of n points in space of N dimensions may be also regarded as co-ordinates of a single point in space of n sets, each of N dimensions. If the projections of the point upon the first n - 1 spaces are fixed, the Will volume element corresponding to variations of the co-ordinates Eni, n2, ... be that found above.
enl`

(g) We now require the volume element corresponding to variations of all the co-ordinates of a point so specified in Nn dimensions. The component spaces of N dimensions each are defined by rectangular co-ordinates, hence the entire volume may be found by multiplying the volume element defined by the variations of
'11)
e12)
.''

-1,i-,

by that defined by the variations of

tnl,

2-

...

1nn.

Conse-

quently the volume element for variations of all the co-ordinates is found by multiplying together the volumes (equation (10)) got by putting n successively equal to 1, 2, ... n. The comnponentvolumes vnar appear in these expressions successively in the numerator and the denominator with equal indices, and therefore disappear from the product, thus we have finally for the volume element
7r'4i (21V- it + 1)

IV

(X

!) ,, F (N

1 V)

-n-ldell

d2,

1)

(h) Now write N - 1 for N and let epq Napq, to correspond with our earlier work. The above expression then becomes all a12 ... aln N--n- 2
,7fl (2Nn-I)
.

NP?n(N-l )

a21

a22 ... daa2,

da,2

... dann,

r(

2 l)r

(f2

r)

2
fl(

).|a,
1)

an2 ...nn

and this multiplied by a density factor

(N

(27r) -2 becomes
1)rN (21r 0I r(f

1()

1 (
) I -2

a,,

a12 ... a22

an

IN-n-2

2 1)(2)F

N-n)

a2

a2Ub

~~~~~~~~~~~~~~~~2 da12 ......dann

dall

)2
atnl an2.

ann

............ (12),

whence our general result (9) follows.

JOHN WISHART

41

4. MomentCoefficients the Distribution. of

It is now possible to find the moments, up to any desired order, of the simultaneous distribution given by (9). A number of these moments will be identical with those found previously for the cases of one and two variates, and published by different writers. To illustrate for the case of three variates, let

F (abcfgh)

1
_ __

A HG
N
___

IV2

N=
7r

H B C, 'G F F
a h
N-5

x e, Aa-Bb-Cc2-Ff-2Gg-2Hh

h b f

..*...

(13).

g f c Then the generalised product moment coefficient of this distribution is given by

= Mjklmpq ffaibkclf

mgphq F(abcfgh) da dbdc dfdgdh,

where the limits of integration for a, b and c are 0 to oo, and for f g and h - Vb to + Ibc, - lac to + lac and - lab to + lab respectively. Now following the method of Romanovsky *, we define the function b(a,8uyXpv) =

da 7db

eaa bh +hv dcffF (abcfgh) + +cy+fx+g/h dfdgdh

as the generating function of the product moments, which are given by (af3,yXov)j mj+k+q+m+p+q = = D Mjklmpq is-. aa@8ka7 @ XMapPaV Since the total frequency volume in (8) is unity, we have
e-Aa-Bb-Cc-2Ff-2Gg-2Hh u=M=O
N-1

(1)

A H G dv = const. H B F

G F C where the integration is taken throughout the volume. A parallel result is obtained for the integration in (14) simply by replacing A by A - a, B by B -,, H by H - v, etc., in the determninarntABC . What is then left to deal with is I the fundamental frequency volume and we have H-lv A-a -y1 G-4i, (a,y,v) = A H G N-1
H B G; F F C x H-tv

2G-t

F-aX F-j_X C-fy

B-fl

a convenient expression from which to calculate the moments. It is of the same form when generalised for the case of n variates.

We shalluse the notation

corresponding to the determinant

1-

k m a h g
h b f (I f c

* Metron, loc. cit. pp. 27, 32, 35-6.

in Product MomentDistributtion Samples 42 The Gaeneralised

to represent the general product moment coefficient of the jth power of s82, kth power of 822, Ith power of 332, mth power of s2s,r2., pth power of s1s3rl, and qth power- of s,s2r12. These moment coefficients are in all cases except the first calculated about the mean of the sample, and are expressed in terms of the following moment coefficients of the general population:
/2Ow =i, /20 = 02 , How/-= 32,

4110 = 12pl2P 1 T

P101 0=3P13,

/1011= U2 3p23W

In most cases there will be a number of similar results deduced from one another merely by a cyclical interchange of the suffixes, but one representative result for each class only will be stated. The following pages list every independent result up to the fourth order, and as far as eight variates. They therefore embody certain results which are already known. " Student" * first gave the one-variate moments up to the fourth order, while Sopert, to his degree of approximation, gave all the twovariate results, to the same order. His second and third order moments agree with 1 1 the following, except for the terms in N, which he makes N and -2, in place of N-i
N2

and

N- I
N3

. The divergence of the fourth order results is wider, for the same

reason. Isserlist later gave the general 2nd order-4 variate, 3rd order-6 variate and 4th order-8 variate moments, sampling from a limited population. For an infinite normal population his results can be expressed in terms of the p's, and, of course, all the moments up to the fourth order can be deduced from his three results by associating the variates in all the possible ways. But his results were calculated about the general population value, and not, as in the following, about the mean of the sample. 5. Derived Coefficients. We are now able to deduce certain other constants of the distribution curves specified by the moments cited. In particular we have for the Betas of a12, i.e. of
81 S2r12,

(.3}
.4

2( )}
2
2

4
}

(3 P122 + p12)
6___

a12t2

2/(

)=

J.{P(

3+ N

6P12

+p14

(1+ p2)2

These values show that when pl2, the correlation coefficient of the sampled population, is zero, the distribution of a12 iS symmetrical, but not normal. For
/82 =

3 + N-i'

showing that as far as the first four moments the curve agrees with Type VII. On
* Biometrika, Vol. vi. 1908, p. 4. Vol. ix. 1913, pp. 103-104. t Biomnetrika, '- Bionetrika, Vol. xiI. 1918, p. 183.

JOHN WISHART

43

= the other hand, when P12 1 we have the Betas of s2, the square of the standard deviation of the sample, 12 8
82/81=N-1,
82,82=

3+ N

l,

the well-known Type III result*. Since a12/32 is always greater than 3, the curve of distribution is leptokurtic. We also have 2,82-3,81 - 6

2 (1 p2)2 N1- 1'(1i + p2)3'

12. 1
N-i

which is always positive, having a maximumnvalue

Also

kP
2-4(1

(3 + p2)21 +
Np2)21

1 + 6p2+ p42 f1 (1 +p2)2 J

2 + 5p2+ 8p4+p6
(I+p2)3 J

showing that the curve is of Type IV for small p, but beyond a value of about 45 (according to the size of the sample) it becomes Type VI. Certain correlation coefficients also follow at once from the moments given. It will be remembered that Pearson and Filon in Phil. Trans. Vol. 191 A, 1898, p. 242, gave r,,12= P122 and r,171, p12/V/2 approximately. They also gave approximate expressions for r,1.r23,rr,. r, and rr12.r,4(pp. 256, 259, 262). In our case the correlations we are able to deduce are those between the product moments themselves and are exact, thus
rail.
rail.a2
a22 =

~
{C2.li)}

~
+

~
{
(

~
2)}(

~~=P12,
{p2
P}

=a

={4
{PU

(ll)} ,)}

* {H

(.
2#

:}
(

t ,)}

(~;
.)})

P12 + p+p2 p

P13P23 (1 +

ra23. als

(. .

)}

.2.

. )i

+ pU3p24 PU4p23

My thanks are due to Dr R. A. Fisher, in whose laboratory this paper was written, and without whose critical help it would have been difficult to generalise the geometrical methods employed by him.
* " Student," Biometrika, Vol.
VI.

1908, p. 4.

GeneralisedProduct MomentDistributiontin Samples 44 Thae

1ST ORDER:

1 variate -ia, (1) p (1) = NN2ND ORDER: 1 variate (3) 2 variates (4) u(l). (6) u(1 3 variates 2N ) 2N -1a,2a.
p12,

2 variates

(2) .LQ'=N (

N-

1~2
l a2P

u (2) =2

N2 1a

(5)

u(.2)=N

aaT22

(1 + P122)

1 al3a2Plg.

(7) (1:.)2N i) (7) u

4 variates (9)
3RD ORDER:

1=N-1i 2 NN-2a2a3Pl2PIP3 =

(8)

p2+p3p) 2 a(la2a32(P12+P13Pa )

N2Y

ca2a3 a4 (P1lP24 + P14P23)

1 variate
(10) u (3) =8

N-1i
NN3

al a,

2 variates

11 (ii) (13)

(2i)=8N (2i)1 = 8

i
aN3

41O2P3

a2 P12
(1

(12)

(32

2 N a-1

N3

l3a2Pl2

+ P12)) (3 + P12 12(1 + P12),

(1 2) = 2 N

a4a22

+ 3P2),

(14) uL('

)4

NN3 a1

(15) U (21) = 8 NN
3 variates (16) (17) (18)
u(i)

1 Oa2P12a

=8 N
=2

1,4a2a3

p12 p13 + 2P12 P23) P13

(:

N-3 1

a1222a32 (pl22 + pf32

u
1

2 = 2-N-J= 8 N3

aa22a33

+ (p13 + 2P12P23 P13P23)

(19) l (20) ,u (2 1) (22) u

(23)

1al a2 a3 P12p13PM3
a10a2a3 (p23 + 3P12P13),

('V)=
( i)

2 N3

) = N N3-10a,2a232 = 4'-N
= 2

(1

+ P122 + P232 +

p238 + 4Pl2P1SP23)

1a 2a0123a3P12 (P13+ P12P23))

a13a2a32 (P12 + P13PO + 2 P12P13).

,u (1

N-

JOHN WISHART
4 variates (24) u (ti

45

4 N= 2

a122aa

ajpn (PaPp +

Pi4P23),

(25) u (1 i)
(26) (27) , ( (28) u ( (29) ..
)

- a12a2a32a4 [PlspIi

+ Pia (Pi2Ps4 + PiaP24+ P14P23)]I

=2

+ + a,3aa3a4 (P12P34 PiaPM

+ PNP23 PiaPiaPio),

i) = 2N= 2-Y3
')

+ + al3a2a3a4 (P13PMP1uP24 2PP2P12P4)

a32aa2a3a4
[pl3p2

+ pa3 + P12(P1SPM + PisP).)]i pm

,u(
(

= N 1 a12a22a3a4

[P34 + pP4

+ + 24 + P12 P2a P (PiPaP33PpP2a)], + P1SPM

5 variates

(30) (31)

=2
=

1 a2a + + P13 + 2aa3a4a5 (P14P3MP1MP34) (P14PM5P15P24)] [P12

+ 1a22a3a4a [P14P3
pp34
+

( lt

2P1sPaPp2 +

(Pm P12 PU+ P231P34)

6 variates
/.

+ P23 + (P14P25 P15PM4)].

1..-..
)

(32) u

aN

+ + P26 , [(a2a3a4a6 [P13(PM + P26P45)P14(PMP3P PA) P46

+ P15 + + + (P2aP46 PS4PaP) P18 (PS3P4. P24PaM)1.

4TH ORDER:

1 variate (33) u(4) 2 variates

(34)
(3)

12( 12(N

N4 N+

)a,
22

1) + 3) (N
1 U4 al4a24

(35)

&

(2.)

=4

[N

1+

8P122

+ 2 (N + 1) P124],
2P12),

(36) p (13)

N4 4(N-1) (N + 3) 3)

51 U2P12

(1 +

(37)y()=6(N--1(4N+ (38)
(3 1) =

a, a2P12(1 + p12)

12 (N-1)

(N + 3) U7a2p12,
6

(39)
(40)
(41)

2 (N-1) (N + 3)
N
424 1

2 (1

+ 5P12),

,u (12)

=2

[2 + (5N + 11) P122 (N + 5) P124], + 2 (N + 5) pl52].

, (p

4)=3

a 4a2 N +'p24)+ 4[(N + 1) (1

46

The Generalised Pr oduct Moment Distribution in Samrples

(3
i)

3 variates (42)
t. -i

12 (N-1)

N4

(N + 3) a1 %
U

as P12 P1 231P3 P232 + 8P12P13P23 +

(43) fu( i i) (44)

(45)
y

/2 . i ii)

= 4 N=

4 NN4 u1 a4a22a32 [(N-1) N -i 1 I N4

al-4a23u3

2 (N + 1) P122p132], 2 (N + 1) P123P13] ,
+ 8P12P13P23

[(N

1) p23 + 4p12(P13 +
P232)

P121P23)+

(2

i)

1 4a22a32[(N - 1) (1 +

+ 4 (p122+
+

p,32)

4 (N

(46)

21 =4(21.)

-2

(N - i) (N
N,6a4 +

+ 1) P12P13],

3
)UU2U3(P23

+ 5P12P13))

(47) (48)

(/4k(i)

a)=2(Ni(N+ = 44
=

3)

5a2aU32(P12 +

pisp2a+ 4P12P132)
P132)

a2a23a33 [2p12p13 +

(N

+ 1) P23 (P122 +
+ 2Pia2 +

+ (N + 5) P12P13P3a])

( (49) uR i 2)

2 uN2a24a32 NN4 N 1
N4
4120

[(N + 3) P122

4 (N + 2) P12P13P23

(N

+ 5) P122P232]

(50)

N)

+ 2 (N a23a3 [2p23 +2(Np +

ia )P233] + + 2) P12P13 (3N ++7) P122P23 (N + 5) 1312PP1],

(51) ,a (t i 1)

2N
6

32 [2 u1 a3u23 (P12+ Np13 p23)+ 4P12 (P132

+ p232)

(N + 1) P,a3

+ (3N + 7) P12P13P23],

(52) H ( u (52)

3)

N4

uo 12a23a33

[(N

+ 1)

P12P13 (1 + p232) + 2p23 (p122

+ pJ32)],

(53) u ( * ) -2
2_

(N - i) (N

aN(4 +3)

3
35ac2a32 (P12

+ 2P.13P23 3P12P132)) +
p232)

(54) , (

2)

2 N-N4-

a1 3a22a33

[(N + 1) P13(1 +

4p12P23

2P133

+ 2

(N + 2) P122P13

+ 2 (N + 3) P12P123],

(55)
(56)

=2

1N4 1al 422a32

[i

+ p232 +

(N + 2) (P122 +
P13 p23) +

P132)

+ (3N + 7) P12P13P23

+ (N + 5) P12P13],

/4(.) (57)1 (

= 3N-

a, a23a34 + 1) (P12+ [(N

1 a,2
2q4

(N + 5) P23(P13+
2p232)

P12P23)]b

2) _2

[(N + 1) (1 +

2PJ22 +

+ (N + 5) (P132 + (58) u(.)

4 variates
=

p232 + 4P12P13P23)i

N -1i 2o,2333 [(N + 5) P23+ (3N + 7) P12P13 (N + 1) p233 +

+ 2 (N + 3) P23(P122 +
(59)
HR

P132) +

(3N +

II) P12P13P232]

(2i

i) =
-

---N1-a,

u3 Or2 a4 [(N -)

P23P24 +

+ 4P12(P13P24 P14P23)
+

2(N

+ 1)P122P13P14],

(60) _ /4 (60) a(

2N (-1)

(N +
N4

3)a2

a3

(P13P24+ P14P23 +

4P12P13P14)

JOHN WrISHLART (6 1) ) 2 -A_-N_ 4

-

47
+ 4p12 (PI2PM + P13P24 + P14P23)
P122}],

al4a22aa4

[(N

1) (pH +

p23p24)

+ 4P13P14 + {I

(N +1)

(62)

4 NN- 1

2a 2a2a42

[(N -1)

(P122p342 + P132P242+ P14P232)

+ 4(P12P13P24P34 P12P14P23P34 P13P14P23P24)], + +

(63)

4(
(

(3r22 a3

4 [(PI2P34+ P13P24) {2p23 + (N + 1) P12P13}

+ p14p23 {(N -

1) p23 +

4P12P13}I,

(64)

=2

g a,3a23a3a4[2(N +1)P12(P13P14 + P23P24)

+ + (P13P21 P14P23) + (N + 5) P122}], {2

(65)

,u

('i

2-

1 al4a22 a3aO [2p23p24+ 2 (N + 2) P12 (P13P24 +

P14P23)

+ P122 {(N + 3) p34+ (N + 5) P13P14}],

(66) u (ti

2 ) = 2 3N 1 u13 22cr2cr [2P23 (P12P34 +

P14P23) + 2P13p24 (1
P122 (P14

+ 2pI3 ) + 2NP13P23P24

(N +

1)

+ P13P34) 2 (N + 3) P12P13P14P23], +

(67)

(ii)

a,3a23a3a4

+ [2{p12P34+ P13P24 (N -1) plp23} + 4P12P13P14 + (N + 1) P122 (P12P34 + 3pi3P24) + 4P12P23 (P24 + P12P14)],

(68)

(ui H

2)N 2IYA=1
i)

U12U22U32 42

[(N

1) P122 (1

+ p342)

+ 2 (P132P242 +

P142P232)

+ 4P12{P13P23 P14P24 P34(P13P24 + + + P14P23)}+ 4NP13PI4P23P24],

(69)

(
j 2)

1 cr4u22a3c4[(N + 3) p13P14 +

2p23p24

+ 2 (N + 2) P12(P13P24 P14P23) +

+ (N + 5) P12P13P14],
(70)
,u
=

2 N-4

alr3a22a32a4 [(N -

1) p14 (1 +

p232) + 2 (PI2P24 + P13P34) + 2P14 (P122 + P132) 4P12P13P14P23],

+ 2 (P12P34 + P13P24) {p23 + (N + 1) P12P13} + (71)

(21)
=

2N

1012a23a32a4 a

(P122 + P,e2) + 4 (P13P14P23 + [2p24

+ p2a2)

P12 P23P34)

+ (N + 1) P12 (1 {P14

2P13 (P34

P23P24)}]'

(72)
(73)

(
=

_(N - 1) (N +3 2 NN (4 + 3) C02v30(P12PS4

+ P13P24 + P14P23 3PI2P13P14)2 +

2 N-NS1 al4a22a3a4[p34+

p23p2p +

(N + 2) {P13 ( P14 +
P122P13P14

P12P24) + P122Ps4}
P12P14P231,

+ (N + 5) (74)

+ (2N + 5)

,LK)

2 N-N-

al a2 a3a4 [(N + 1) P12(P34+ P23P24)+ 2 (N + 2) P12P13P14 + + 2 (P13P24 P14P23 P123P34) (N + 3) pIU2 (P13P2 + P14P23)], + +

(
(75)

1" \) =2-N
=2

(N+

1 03a2204

[P2

(P

+ 3P23P24) +

p14{I +

2P122 + p232 +

(N + 1) pl32

+ 2 (N + 2) P12P13P2} + NP12(P24

PM P3p4) + P12Pl3

P34 {2P12 + (Nl+ 3) p.3p2}],

48

The Generalised Product Moment Distrib2ition in Samples

('~~N-i12
2 N
a2a2a32a42 [P22

(76)

+ pl32 + p14 + P122ps42 + p32Pm2 + p

23

+ 1) p23 + (N + 3) P24 + P12P14 {(N + 1) P24 + (N + 3) P23Ps4} p34 + P13P14 + 1) p34+ (N + 3) p23p24}], {(N
+ P12P13 {(N

(77)

i)

3 = 33 N-

a a2a3a4

+ [4P12(P13P14 P23P24)+ + 1) (p34 2p2p + +

(N + 1) (1
3P12P13P14)

+ P122) (P13P24 + P14P23)],

(78) ,u (

~)

N4~ 1

14 22a3a4 [(N

+ (N + 5) (79)
=

(P13P14 + P122P24 2P12P3P24 + 2PL2P23P14)] +

i33

a4 [(N +

+ 5) P12P24 4P13P24

+ (N + 1) (3P14P23+ 2P122P13P24 + P123P34) + 2 (N + 3) P12 (PlaPi4 + P23P24)+ (N + 9)

P122P4P23],

(80)

a,3 + a1322a32ar4[4P12(P24 + P23P34) 4P132Pp4

+ (N + ]) {(1 + P12) (P14 + P13P34) 2 (p.14 + + p232 + 2 (N + 3) p13p23P24 + 2 (N + 5) P12P13P14P23],
P12P13P24)}

(81) 1 (~ 2) =

NN4 al2222[aa3 2 -1)2(1

+ P12)

(1

+ p342) 2 +

(P32+

pu42+ P232 + p242)

+ P1423) + 4{P12P13P23 P12P14P24P13P14P34 + 2N (PL31P242 + + + P23P24P34 + P12P34 (P13P24 + P14P23) + (N + 2) P13P14PM3P24}], (82)

( *

a NN o13a22a324 [p14{2(1 +

p232)+ (N + 1) (p122 + P132)}

2P12P13P14P23)

+ + + 2 (N + 2) p23(P12P34 + P13P24) (N + 3) (P12P24+ P13P34 + (N + 5) P12P13 (p12P34 + P13P24)],

(83)

* 1)

NN41 al2a22a32a42 + [1

P122 + pl32 + p242+ p342+

N (Pu2 +

p22)

+ P122P342 + P13P242 + (N + 2) p142p232 + + + + (N + 3) (P12P13P23 P12P14P24P13P14P34PO23P2434) + (N + 7) P14P23 (P12P34 + PlsP24) + (3N + 1) P12P13P24P34]-

5 variates

(84)

Z (2

+ + 2 NN4 a14a2a3a4a, - 1) (P24PM4 P2sP34) [(N

+ 4P13(Pl4P25 + P15P24) +

4P12 (PI4P35+ p15 p34)

4 (N + 1) P12P13P14P15]I
-

(85) I (;i

4
+

[(N 1 2a22a32a4a3

1) (P122P24P + P132P24P23 P23P14pl6) 3 + P33+ P15P34)

2P12PI3 (P24P35 + P25P34) + 2P12P23 (P4

+ 2p13 (p14 + P1 P24)], P23 P25

(86)

(ii

2N-

1 a,2a23a3a4a5 [2p12(p14 p35 +

P14P15P23 +

Pu5P43) +

2 (P13+

2P22P23) (P14P25 + p15p24)

+ 2 (N -1)

(N + 1) P12 {P12 (P24 + P23 + 2p13p24P2p}], P3 P34)

JOHN WISHART
(87) 2( . 1) 2 aN
a12a22a32a4 a5 [(N

49

-1) Pls2(p4g+ pH p36)+ 2 (p1a2p24 + pga2pu4P15) p2s

+ + 2PnP213 (p24p36 + P2Pp34) 2P12P2a (2pwp45+ p14 + Pi5P34) 3p + 2 (P12+ N p13p23) (P14P25 + P15p24)], (88) 2, (

1222

(1 a [(N - 1) P14P15 + P22) + 2 (P122PaPa +

P132P24P2)

+ 2Np12 p13 (P24 P.(5 + P2 P34)+ 2P12 (Pn6 + P23 P34) P15 P14 P35)+ 2P12 (P24 + P23 + 2pj3P14 (P35 + P23P25) + 2pi3P15 (P3 + PM )] pM)

(89)

=2

a14a2a3a4a5 pa + p25pu + 2P12 P13P14P15 [pu + (N + 2) {P12 (P14P35 + P15P34)+ + (N + 3) P14P15 + p12PI3)]S (P23
+ P13 (P14P25 P15P24)}

(90)

*11

= 2 N - 1a3a2a32a4a5[P14P25+

+ P14 P15PM 2P12 P15 +

(N - 1) P12(P45+

Pa4P35)

+ + P13(P23P45 + P24P35 P256P34)+ (p2 + 2p,2p13) (P13P46 + + (N + 1) {P13 (P14P25 + P16P24)+ 2p13P14P15 P23}] (91) (

P14P35 + P15p34)

)=

2Y

1a4 ala2aa2a4a5[(1

+ 2pta2) (pltp2 + P16P24)

+ + + 2P12P34P35 + P23(P14PP3 P15P34)+ Np13 (P24P35 P25 P34) + + (N + 3) P12P13 (P14P35 P15PM4)+ (N + 1) P14P15 (P12 + P13P23)1,

(92) u (

=2N

. 1 a3a2a32a4a5 [(N

+ 1) P12(p45+ pHPM)+ (N + 1) {P132(P14P25 P15P24)

+ + + + 2P13P14P15P23} P14P25 P15P24 2P12P14P15 2P12P13 (P14P35 + P15P31) + + + + 2P13P45 (p2 + pi2p13) + PM(PI3P25+ P15P23) P35(P13P24 P14P23)],

(93) ,u (1:

=2 N

1 a 2a2a32a4a52

[Pl2

P4

(1

+ p352) + P12P15P45 + P24 (pG32+ P152)

+ P13P23(P14+ P15 P35) P34 P23 P25 P45 P45) + P13 (P13 + 3p14P35) + P15 (P15 + 3P12

+ N {P12P13 (P34+ P36P45)+ P14P15 + P23P35)}+ (N + 1) P13P15P24P35 (P25

+ (N + 2) P13P15P25P34],

(94)

=N
+

13

, 2
1)

+ a4a [4p12 (p24pP3 P2 P34)+ 4p3P24 + 4p,4P15(P13+ 2pI2P23) P26

(N +

{( 1+

P122)

+ + (PI4P25+ P15P24)}]M (P14P35 P15P34) 2 (p23 + P12P13)

+P1S )(pg6+p34

(95)

=N

4 1a2a22a2a4a[(N-1)(1

p3) + 2NP(p152p24pz+p232pupis)

+ + + (p432 + p 2) + 2 (N + 2) P13P23 P25) 2P46 (Pu4p2s + P15P24) 2 (P14P15 P24 + + 2P12(P14P25 P15P24)+ 2pi3(Pu4p + P15P34)+ 2p2a (P24 P3 + P26P3) + + + + + 2P12P13 (P23P45 P24P35 P25Pa4) 2PI2p23(P13P45 + P14P36 P15PM)]

(96) A(

a-

a,a4oa23oa5[(N + 1) (pap4s +

p24p35 +

+ P25PH 3P12P13P14P15)

+ (N + 5)
Biometrika xXA

+ + + + {P12 (PI3P45 P14P35) P14(P13Pa + P15P23) Pi (P(122P34P13P2P)}] 4

50 The Generaliaed Product Mom,ent Diwtributionin Samples

(97) (3)eN
a l3eaast

+ [2p, (p. + p3 PM)+ 2p'o (p.s + PaPam) 4P132PlP24

+ (N + 1) {P132 (P12P45 + P14P25) + 2ps ps PM + (P15+ PI3P35)(P24 + P1SP14)} + (N + 3) pIs(pa pw + PSSPm)+ (N + 5) Pisap1 (P1SPm + p1ip2s)]

(98)

a,(

.::)

a,2a2 3a4q5 [2 {Pi4P1s (1 +

p22)

+ P24Psa (1 + P13)

paPa (1 + P122)}

+ (pi4 p2 + P5sPmp?) {(N + 1) p12 + (N + 3) PiaPe} + + (Pu4PMU P1s PU) ((N + 1) P13+ (N + 3) P12P23} + (p24pa + p2sp3) {(N + 1) pm + (N + 3) pllp3}],
(99)

''I)

iYj7

a 12a22as2a4as

[p46(1

+ P122 + p,32 + Np22) + P14P15{N

+ (N + 2) P232}

+ P24 P25(1 + p2) + p34p35( + P12) + + (P25 + P23P35) 3P13P15 (PM4+ + 3P12P14 + N {P12P13 (P23 P45+ P24P35 + P2sP34) + + (N + 1) P12P34 (PlSP2s + Pl5P23) + (N (100)

p24 p23 p + 2P13P14P23P25 (P13P45 + P15P35) P23P24)+ 3P12P23 P12 P24+ P13P14PM} P15 + 2) p23p235 (p + P13P14)]

a 1Ta2Ua3Ua4(a6

[(1 + P122) (p35 + P34P45)+ (PI3 + 3PI2P23)(Pls + P14P45)

+ + P242 (P35 + Pl3PI5) + 3P24(P25P34 P13P14P25) + + N {P142P35 (P23+ Pl2P13)(P25+ PMP45)}+ (2N + 1) P12P15P24PM + (- + 2) P14(P15P3 + P14 P25) + (NS+ 3) P12 P24P35 P14 P23 + (N-^ 4) P14(P12 P34+ P15P23P24)] + P25

6 variates (101)
(ii:a

=a

al
2P13P14P25P26

as [(Gv-1) {P122 P46+ P36 (P35 P45)

2P135P1G6P23P24} + 2 (P13P24 + P14P23) (P1sP26 + P1,8P25) 2P12P14 (P25P36 + P26P35)

+ 2P12PI3 (P25P46 + p26P4s) +

2P12 P15 (P3 P4e

P24P36) + 2P12 P16 (P23 P35)], P45 + P24 [(N

-

(1(2)

Al4

a132a3a4a35a6

1) P12(P3sP46+ P36P4s)+ P13(P2sP46+ P26P45)

+ P14(P2sP3s+ P26P3s) P15(P23P46 P24P36)+ P16(P23P4s+ P24P35) P35) P45) + 2P12P14 (Pl5P36 ? P1e (PI P46e+ P1e -2P12P13 (N + 1) (Pl3PIPs(PI P26+ P16P24) + PI4Pie (P13P25 + P15P23)}], )=2 2 y
a12a32a4Uasa6

(103)

a(

1 *

[P12 (P15Pe46+ P16Ps) + P14 (P1P26 + Pl6P2s)

+ 2P12 P14P35P36 2PI3P15 P24 + 2P13 P1eP24P3s + P132 (P25 P4e + P26P45) P3P -{pP12P3 P4e P,P P4s) + P13 (P25 + P26 P14 P3e (P35 P46 P35)}+ P13 (P23 + P2eP34) P1s + P13P1e p45sr p2p34) t- (N - 1) P1sP1e(P24 + p p3) (P3 + + (P12P34 P14P23) (PIsP36+ P16P35)],

(104)

-(

2 1;')=2-Y< -(N

,~~~~N-1

+ [P12 a-a2ua3a42uaa6 (P15P36 Pi6P3) + P13(P15 + P6 P25) P26

- 1) P12 (p5e + P4e P45) + NP14 {P16(P26P34+ P24P3s) P13 + P1e(P24P35-r P25 P46 P14 P34)} + P12 (P35 + P36 P45) + P13 (P2ePs5 + P2sP46) P14 + P142 (P25PM+ P2ep3) + 2 P1s1P2ePM + + + (P12P34 P13P24) (P15P46 P1eP45+ 2pI4P56e)j,

JOHN WISnIART

51

(105) u(

1)

NN-1
+

2a22a3a4a5a [(N -

1) (1

+ P122)(p35p4 + P36P45)

2N (P13P24+

+ 2 (P13 + + 2 (p23 +

P12P23) (P15P46 + P18P45) +

+ + (P15P2M P18P25) + 4 (P13P14P25P26 P15P18P23P24) P14P23) 2 (P14 + P12P24) (P15PM+ P18P35) + P35)], (P25PM+ PM P12P13) (P25P46 + P28P45) 2 (p24 + P12P14)

(106) u (

a=N'

a a6

[2P12

+ (P.sP4e+ P3e,P45) 2pI3(P25P46+ P2 P45)

+ 2p14 (P25P38 + P26P35) + 4P15P18 (P12P34 + P13P24 + P14P23) + (N.+ 1) {(P23 + P12P13) (P15P4e + P16P45) + (P24 + P12PA4)(PI5P36 + P16P35)

+ (P34+ P1sPu4) P2 + P1eP25)}] (P15

. 11
N

(107)

[(1 N ) 1 u1i el2u22u3eoea6

+ + + +

P45) + P12) (P36 + P36 P46

+ (PI5P46+ P16P45)+ (P24+ 3Pl2P14) (P25P36 P26P35) (P13+ 3P12P23) + + 2p34 (PI5P16 + P25P26) 2 (P13P14P25P26 P15P16P23P24) {P13P24+ (N + 1) P12P34+ (N + 2) P14P23} (Pl5P26+ P16P25) + + + (P15P36 P16P35)}], N {(p23 + P12P13) (P2sP4ff P26fP45) (PI4 + P12P24)

u (1(08)) (

*1 +N

2 a3'a4 a5 e [ (N -1)

P46) P23)(P56+ P45 (pI2 + P13

+ + (P24P35 P25P34)} + (P24P36 P26P34)+ (PA4P36 P16P34) {(p,4p3p + P15P34) + + (P35+ P34P45) (P35P46 P36fP45) P13P26 + 2P12PP36 P45 + P13P24

+ + + + P35)+ 2P24P34(P15P36 P16P35) 2p34P56(P12P34 P13P?A) + 2pl4 p34(p25pP3f P26 + (PIs + P14P45) (P26 + P23P36)+ (PIe + Pi4P4e)(P25+ P23P35) (2PI2P3s+ P13P25)] + P34)(2P14P56+ P15P46 Pie P45) + (P36+ P34P46) + (P24+ P23 7 variates

(109)

(1

)= 2 2-N--

u3a a 6u7

P46) P15 P66) + P12 (P36P47 + P37 P57+ P37 [PI2 P14(P36

Psf) P57 P14 P46) + P13 (P26 + P27 P17 P57 P16 + P12 (P34 + P35P47) + P12 (P34P56+ P35 + (P24P56 P25P46) P25P47) + P13P17 P34) P-7 P35) + P1sP1e(P24 + P27 P17 + P14 (P25 + P27 P35) + P14 (Ps5P36+ P26 P37 P1e + + + P13P15 (P24P57 P46) + P13P16 (P26P47 P27 (P46 + P1sP17 (P24P3e+ P26P34)+ (N - 1) {P12P13 Ps7 + P47P56) P34)}]I P17 + P14 (P2eP37+ P27P36)+ P16 (P24P35+ P25 P15

(110) *

/.~~~~ 2...
\ /

-4

a, a2a3a4a5a6 7 [(P24 +

Ps5) P14) Ps7 + P37 (P3 P12

(p25 ~~~+ +r

P57 + P35P47) P46) + (PeS+ P12P16) (P34 P12P15) (P36P47 + P37 + P14P23) (P1eP57 + P17P50) + (P27 + P12P17) (P34P56 + P3sP46) + (P13P24
+ (P13P25 + P15P23) (P1e P47 + P17P46) + (pI3 P2e + P16P23) (P14 P57 + P15P47) + (P13P27 + P17P23) (PA4P6e + P15P46) + (PA4P2e + P16P24) (PI5P37 + P17P35) + (P14P27 + P17P24) (P15P3e + P1eP35) + (P15P2e + P16P25) (PA4P37 + P17P34) + (P15P27 + P17P25) (PA4P3e + P16P34) + (N - 1) {(P23 + P12P13) (P46fP57+ P47P5B) + (P14P2s + P15P24) (PI6 P37+ P17P36) + (P14P35 + P15P34) (PIS P27 + P1?P26)}]e

4-2

;52 The GeneralisedProduct MomentDistributiot iut Samples

8 varia*es
/.~~~~~1
(111)
C)_ N-Noi

a&

a7 a[(Ps

+ PISPS)(P47PO P4SPr) +

(Pn)P(.

+ (P PO+ P16PS) (AP PM + PM + + PS) (P4SPOs P46 P61) + (PUPO7 P17 PM) + PnP23)(PupsP+ PNSP7)+ (Pu Ps + PiP20) (Pa7 + P P.,) + Po

+ (PiaP2 + P1e (Pa7Ps+ P38P5)+ (PlBPC P17 (PAPS + PMPMB) + P2) PM) + (Pi4Psi+ PnePsi) Po + P3BPS7) (Pi P. + P17 (Pup.A+ PaeP3B) + Ps) (Pu + (Pn Ps + PisPr.)(PePu7+ Pa37P) + (Pus + P17uPu) Pe + PMPu) P, (Pu + (Pie pa + pB p.) (ps p47+ pa, p0) + (N- 1) {(pu pum P14 + p) (pa, p, + p.sp.7) + (pISPS+ PIePS) (PWp.A P3sP47) (Pu7PS P1PO) (P85PA PeePM)}]. + + + +