Biometrika Trust

On the Variation of Yield Variance with Plot Size

Author(s): P. Whittle
Source: Biometrika, Vol. 43, No. 3/4 (Dec., 1956), pp. 337-343
Published by: Oxford University Press on behalf of Biometrika Trust
Stable URL: http://www.jstor.org/stable/2332912
Accessed: 20-03-2018 14:56 UTC

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].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

Biometrika Trust, Oxford University Press are collaborating with JSTOR to digitize,
preserve and extend access to Biometrika

This content downloaded from 200.131.199.26 on Tue, 20 Mar 2018 14:56:07 UTC
All use subject to http://about.jstor.org/terms
 [ 337 ]

ON THE VARIATION OF YIELD VARIANCE WITH PLOT SIZE

BY P. WHITTLE

Applied Mathematics Laboratory, New Zealand Department of Scientific

and Industrial Research, Wellington

The problem examined is that of evaluating the spatial covariance function of yield density, from a
knowledge of the way yield variance varies with plot size and shape. Results are obtained in ? 3 for
several kinds of plot. Results are also obtained (?4) on the dependence of the yield variance on plot
geometry for very small and very large plots. Special attention is paid to the case for which the
covariance follows a power law at large distances.

1. INTRODUCTION

It is well known that, in order to explain the observed variation of yield variance with size
and shape of plot, it is necessary to allow the possibility of correlation between yield den-
sities at any two points in the plot. (We shall restrict ourselves to the stationary case, for
which the expected yield density is constant over the area.) Moreover, it appears that this
spatial correlation must often fall off relatively slowly with increasing distance between
the two points; as a power function of the distance rather than as an exponential.
The same behaviour is shown by observations on yarn diameter, flood height (Feller,
1951), and response from population samples. The calculations of this article will apply to
these cases, too, but for concreteness we shall continue to speak of plots and yields (although
we shall sonmetimes use the word 'region' instead of 'plot ', indicating that we do not confine
ourselves to two dimensions).
The type of calculation most usually made is to evaluate the yield variance for a plot of
definite size and shape, and for a given spatial covariance function p(s). However, the
inverse calculation would probably be more useful in general: to determine the covariance
function from a knowledge of yield variance as a function of plot geometry. Such a pro-
cedure would enable one to make use of experimental results to obtain at least a partial
estimate of p(s). We consider this question in ? 3.
A solution of one form of problein is most easily reached by using a Mellin transform, so
that the covariance which falls off as a power of the distance,

p(s)- C Is 1-P (s large), (1)

makes an unforced debut.
There is a good deal of evidence to show that many observed covariance functions are of
the form (1) (Fairfield Smith, 1938). This is all the more interesting, in view of the fact that
the simpler linear models of yield variation do not predict a power law, but rather a much
more quickly diminishing law of the type

p(s)-CIsj-ve-a1s1 (s large). (2)

We shall pursue this question further in ? 4.
It is easy to show that the 'yield' variance of a large one-dimensional interval is
asymptotically proportional to a power of the length of the interval if the covariance func-
tion is of the form (1), but analogous results are not obtained quite so immediately in two
or more dimensions. This topic is also covered in ?? 3 and 4.
22 BioTin. 43

This content downloaded from 200.131.199.26 on Tue, 20 Mar 2018 14:56:07 UTC
All use subject to http://about.jstor.org/terms
338 Variation of yield variance with plot stze

2. FORMULAE FOR THE YIELD VARIANCE

We shall use a vector co-ordinate x to denote points in the plane. Let us consider a plot,
or region, Q, with area A and yield y(Q). (We use the words 'plane' and 'area', but most
formulae of the paper apply to regions in n-dimensional space, n = 1, 2, 3, ....)
Consider two regions Q1 and Q2 which contract to zero in such a way that ultimately they
constitute two points a distance s apart. We shall define the covariance function as

p(s) lim cov [IY(A), Y(Q2)]

X,, ,o ->0 A1A2

and shall restrict ourselves to cases for which this limit exists.
If V(Q) is the population variance of y(Q), then it follows from definition (3) that

V(Q) = Xf P(x x2)dxldx2' (4)

If p(s) is continuous at the origin then we can draw one immediate conclusion from (4):

V(Q)#p(O)A2 (Q small), (5)

that is, for small regions (i.e. for regions which are small in all their dimensions), V is pro-
portional to the square of the area. This is to be contrasted with the case of zero spatial
correlation, when V is proportional to area under all circumstances.
For the one-dimensional case, when Q consists of a segment of a line of length a, then
(4) simplifies to
Za
V(a) = 2 (a-x) p(x) dx. (6)

For the case when Q consists of a rectangle in a plane, with sides a, b parallel to the
co-ordinate axes, we have similarly

Rb Za
V(a, b) = 4f (a - x) (b- y) p(x, y) dxdy. (7)

The integral (6) may be evaluated for the commoner choices of p(s), but (7) and other two-
dimensional formulae can generally be evaluated only for rather unlikely forms of the
function p(s).
If we introduce the spectral density function
00

F(w) = fei.x p(x)dx (8)

_00

and the areal characteristic function

G(w) = fei. xdx, (9)

then equation (4) can be rewritten as

V(A2 = (2)n G, (@) 12F(w)dw. (10)

This formula sometimes has advantages over (4), since the limits of integration are simpler,
the weights given to different wave numbers w are quite evident, and approximate methods

This content downloaded from 200.131.199.26 on Tue, 20 Mar 2018 14:56:07 UTC
All use subject to http://about.jstor.org/terms
 P. WHITTLE 339

such as the method of steepest descents cani be uise(d immediately. The function G(W) is
readily evaluated for simple plot shapes. Thuis for the rectangle with sides a, b

G(w) = 4 sin (aol) sill (bo02) (11)

and for a circle with radius a and centre at the origin

27Ta
G(w) J J( a), (12)

where J1 is the Bessel function of

is its absolute value.

3. THE INVERSE PROBLEM

From the practical point of view, it would be useful to be able to invert relation (4), and
express the covariance function p(s) in terms of the yield variance V(Q), which can pre-
sumably be observed for varying Q.
This inversion is very simple for the special cases (5) and (6), corresponding to linear
and rectangular plots. We have respectively

I la2V(x) (13)

1_ l V (x,y) (14)
P(X, Y)= 4 x 2ay2 .(4
However, there are few cases as tractable as these.
We shall now largely confine ourselves to one general type of case. We shall assume that
the covariance is isotropic, so that p(s) is a function of s only, and we may write

p(s) = p(s). (15)

Further, we shall assume that we are dealing with a plot of constant shape, which can only
vary similarly, i.e. by changing all its dimensions in the same ratio. For a given plot shape,
the size of the plot is conveniently specified by its largest dimension, x (say), and we shall
write
V(Q) = V(x). (16)

(Thus, for a circular plot x equals

etc.) Now, corresponding to each plot shape there will be a function K(s) describing the
distribution of distances between two points chosen randomly in the plot, such that

V(1) = f:K(s) p(s) ds. (17)

If we now increase all linear dimensions in the ratio x: 1, then the infinitesimal elements of
integration will be increased in the ratio Xn: 1, and we shall have quite generally

V(X) = X2n K(s) p(xs) ds. (18)

This content downloaded from 200.131.199.26 on Tue, 20 Mar 2018 14:56:07 UTC
All use subject to http://about.jstor.org/terms
340 Variation of yield variance with plot size

The form of the relation (18) between V and p invites us to introduce the Mellin transforms

V(v) = f V(s) sv-lds, (19)

-(v) = fp(s)sv-1ds, (20)

K(v) = K(s)sv-1ds, (21)

in terms of which (18) becomes
V(v-2n) = K(1-v)p(v). (22)

We shall consider in ? 4 the common range of v for which the transforms (19) and (20)
exist, and (22) is valid. Assuming for the moment that a suitable range exists, we can
immediately obtain a solution for p(s) by taking the inverse transform of the expression for
p provided by (22). In fact, if the desired inverse relation to (18) is

p(x) = x-2nJ L(s) V(xs) ds, (23)

0

then p(v) = L(1 + 2n-v) V(v-2n), (24)

where L(v) has a similar definition to K(v). Comparing (22) and (24) we see that

1
L(v) K(v-2) (25)

In order to be able to carry out the calculation explicitly we must then be able to
perform the integration (21) and then invert the L(v) yielded by (25). One of the few cases
for which this is possible is the rather unrealistic one of a circular plot (x= diameter, n = 2).
For the circle we obtain by direct methods

K(s) = i[s cos-1 (s)- S2 1( _ S2)], (26)

whence K(V)= _- __2-_ (27)

2~~~~~~~~~~(7

L(v) = 3 (28)

However, upon attempting to invert L(v), we obtain a divergent integral. This may be taken
as an indication of the fact that p(x) should be expressed not only in terms of V(x) but also
of its derivatives V(j)(x). In view of results (13) and (14) this is not surprising. If for (23)
we substitute the more general relation
r1 X
p(x) = x-28
Jo j=o
E Lj(s) [(xs)i V(j)(xs)] ds, (29)
we find that, provided that
[sv+-l Vj)(s)]0 = 0 (j = 0, 1, 2, ...), (30)
then L(v) = Y (v-1) (v-2) ... (v-j) Lj(v). (31)
j

Now, the L(v) of (28) can be written

This content downloaded from 200.131.199.26 on Tue, 20 Mar 2018 14:56:07 UTC
All use subject to http://about.jstor.org/terms
 P. WHITTLE 341

which, in view of (29) and the integral representation of the Beta-function, correspoinds
to the relation

p(x) = X [(Xs) V'(s) - (X8)2 V"(X8) + (X8)3 V"'(X8)](j -82) (33

A case of greater practical interest would be that of the rectangle. However, while K(s)
is easily calculated for this figure, the transform K(v) is not.

4. POWER-LAW COVARIANCES

Let us first return to relation (22). If a functionf(s) is 0(s8-) and 0(s-fl) for small and large
positive s respectively, then its Mellin transform f(v) will certainly exist for

a < Re (P) < ,8 (34)

and f(P) will in fact have simple poles at a and ,.

Now, p(s) is 0(1) at the origin (if we suppose it continuous there), aiid if we assuime it is
O(s-A) for large P, then p(v) exists for 0 < Re (P) < A. For small s, K(s) is p)roportional to th
surface area of an n-dimensioinal sphere of radius s, anid is heince Q(s1-1), wlile for s > 1 it is
zero. Consequently K(r)) exists for ( 1-n) < Re (P) < oo.
It thus follows that both quantities in the right-hand member of (22) exist for

0 < Re (P) < min (n, A). (35)

Since relation (22) determines V in terms of K and p, the relation is also valid for these
values of P, and V(v) exists for

-2n <Re (P) < min (- n, A - 2n), (36)

and has simple poles at these extreme values of v}.

We can thus conclude that for small plots

V(x) = 0(x2n) = O(A2) (37)

while for large plots
V(x) = Q(xe) = O(A), (38)

if p(s) falls off more rapidly than s-}, or

V(x) = O(x2n-A) = O(A2--A/n) (39)

if p(s) falls off as 8-A (A < n). In other words, it is a power law [p(s) C,Cs-'-] which marks the
transition between the two cases where V is proportional to plot area for large plots, or
increases faster thaln plot area.
Relation (22) may be used in reverse: if V(x) is observed to increase as Al- for large areas
(1 < I < 2), then we can deduce that p(s) falls off as 8'7(/-2) for large distances.
It should be emnphasized that these conditioins hold oiily for plots of constant shape,
since V will in general be a functioin of shape as well as area. However, somne results pre-
sented by P. Fairfield Smith (1938) would seenm to iindicate thlat this dependence on shape
may be weak, aand that V may be determinied alnmost eiitirely by plot area. This is presumniably
true only if the shape is not too extreme, e.g. Inot too narrow and elongated.
Fairfield Smith's results are also of in-terest in that they provide very coniviincing evidence
of the behaviour associated with equation (39). He found that to a very good approximationi

This content downloaded from 200.131.199.26 on Tue, 20 Mar 2018 14:56:07 UTC
All use subject to http://about.jstor.org/terms
342 Variation of yield variance with plot size

the variance of yield per unit area in plots of area A could be represented as a curve
const. A-0-749. Replacing the index by-3, we then have

V(A) const. A2-- = const. Al, (40)

corresponding to a covariance function

p(s) const. s-i. (41)

(These equalities cannot, of course, hold for indefinitely small A and s.)
This result, well founded in observation, provides evidence of three intriguing possi-
bilities: (a) that covariances decaying as S-A do occur in nature; (b) that the rate of decay
may be so small that yield variance increases faster than plot area; (c) that the observed
index A may have simple rational values.
None of the simple linear models hitherto proposed to represent yield variation over an
area (see Whittle, 1954; Heine, 1955) lead to covariances of the above type. For instance,
the model which relates yield density 6(x, y) symmetrically to the yield at all points around
(x, y): 2 2
+ (aa) _K2] 6(X, y) = e(x, y) (42)

leads to the covariance function

p(s) = const. sK,(KS)

const. 81 e-KS (KS large). (43)

(Here K1 denotes a modified Bessel function.)

For the one-dimensional problem of yarn diameter variation D. R. Cox (personal com-
munication regarding unpublished work) has proposed a model which does in fact predict
a power-law covariance for large s. The only multi-dimensional models known to the author
which predict a power-law covariance for any range of s are those of turbulence (see
Batchelor, 1953, p. 122). For these models dimensional arguments show that the index A
must have simple rational values.
It seems likely to the author that any model which is to provide a satisfactory explanation
of the power-law decay observed in agricultural work must embody two of the features
common to Cox's yarn model and the turbulence model: (a) it must be non-linear, and (b) it
must consider the variate (yield density) to be a function of time as well as of the spatial
co-ordinates. (In Cox's case 'time' is discrete, the number of smoothings that the yarn has
undergone.) As an example of such a model, one could suppose that fertility gradients in
the soil tend to be smoothed out in the course of time by a diffusion process, which is non-
linear to the extent that only gradients which are greater than a certain value tend to
diminish.
The value of the index A deserves some discussion. Is it fortuitous that Fairfield Smith's
estimated index b = 0*749 lies so near the rational value 4 (corresponding to A = ) ? Certaiily
little can be inferred from this single instance. However, Fairfield Smith lists the b values
estimated from uniformity data for several different kinds of crop. The histogram of the
thirty-nine estimated b values has its main peak in the interval b = 0 41-0 50, and a second
peak and upper cut-off point at b = 0*71-0*80. This provides at least a suggestion that the

This content downloaded from 200.131.199.26 on Tue, 20 Mar 2018 14:56:07 UTC
All use subject to http://about.jstor.org/terms
 P. WHITTLE 343

values b -i and i (correspondin

should say no more until further data or predictions from theory are available.

The author is grateful to Dr D. R. Cox for drawing his attention to the observations on
yarn diameter and flood height.

REFERENCES

BATCHELOR, G. K. (1953). The Theory of Homogeneous Turbulence. Cambridge University Press.

FAIRFIELD SMITH, H. (1938). J. Agric. Sci. 28, 1-23.
FELLER, W. (1951). The asymptotic distribution of the range of independent random variables. Ann.
Math. Statist. 22, 427-32.
HEINE, V. (1955). Models for two-dimensional stationary stochastic processes. Biometrika, 42, 170-8.
WHITTLE, P. (1954). On stationary processes in the plane. Biometrika, 41, 434-49.

This content downloaded from 200.131.199.26 on Tue, 20 Mar 2018 14:56:07 UTC
All use subject to http://about.jstor.org/terms