WATER RESOURCES RESEARCH, VOL. 20, NO.

2, PAGES 251-265, FEBRUARY 1984

On the Problem of Permissible Covariance and Variogram Models

GEORGE CHRISTAKOS

Centre de Gdostatistique et de Morphologie Mathdmatique, Ecole des Mines de Paris

The covariance and variogram models (ordinary or generalized) are important statistical tools used
in various estimation and simulation techniques which have been recently applied to diverse
hydrologic problems. For example, the efficacy of kriging, a method for interpolating, filtering, or
averaging spatial phenomena, depends, to a large extent, on the covariance or variogram model
chosen.The aim of this article is to provide the usersof thesetechniqueswith convenientcriteria that
may help them to judge whether a function which arises in a particular problem, and is not included
among the known covariance or variogram models, is permissibleas such a model. This is done by
investigatingthe properties of the candidate model in both the spaceand frequency domains. In the
present article this investigation covers stationary random functions as well as intrinsic random
functions (i.e., nonstationary functions for which increments of some order are stationary). Then,
basedon the theoreticalresultsobtained, a procedureis outlined and successfullyapplied to a number
of candidate models. In order to give to this procedure a more practical context, we employ
"stereological" equations that essentially transfer the investigations to one-dimensional space,
together with approximations in terms of polygonal functions and Fourier-Bessel series expansions.
There are many benefitsand applicationsof sucha procedure.Polygonalmodelscan be fit arbitrarily
closely to the data. Also, the approximation of a particular model in the frequency domain by a
Fourier-Bessel seriesexpansioncan be very effective. This is shown by theory and by example.

INTRODUCTION [1975])and intrinsicrandomfunctions(intrinsicalness

A recurring problem of geostatistical investigations in
hydrosciencesas well as in many other geophysicalsciences
is to judge whether a particular function can be used as a
covariance or variogram model, ordinary or generalized
[e.g., Matern, 1960;Matheron, 1965, 1973].The practitioner
frequently faces data which are not satisfactorily approxi-
mated by any known covariance or variogram model. In-
stead, some other function not included in the known models
may offer a much better fit. Also, the practitioner may
modify a known model so that a better fit to the data is
achieved,but it is possiblethat the model so "improved" is
not a covariance or a variogram any more.
The use of the function best fitted to the data is very
importantin the context of problemssuchas estimationand
simulationof hydrologicrandom functions. The reasonsfor
this are related to both the understandingof the underlying
hydrologic processes [e.g., Mejia and Rodriguez-Iturbe,
1974; Chiles, 1977] and to the effectiveness of the estimation
procedures [e.g., Delfther and Delhomme, 1973;Delhomme,
1978].
Consequently,criteria capableof testing, in an analytical-
ly and computationallytractable way, whether a function is
permissibleas a covariance or variogram model may be of
practical importance. This article will try to give an answer
to the problem, based on spectral analysis [Bochner, 1959;
Jenkins and Watts, 1968]. More specifically, we will first
establishthe theoretical supportto the criteria testing the
permissibility, and then we will develop procedures to
perform this testing in a more practical context. Within this
framework we will consider models for ordinary random
functionsdescribedby Yaglom [1962]as opposedto general-
ized random functions in the senseof Gelland and Vilenkin
[1964].
We will examine models for both stationary random
functions(in the weak sense,see Yaglom [1962]and Bartlett
Copyright 1984 by the American Geophysical Union.
Paper number 3W1855.
0043-1397/84/003 W- 1855505.00
251
is
used herein as an alternative term for random functions with
stationary increments in the sense of Kolmogoroff [1941];
see also Yaglom and Pinsker [1953], Yaglom [1957], and
Matheron [ 1973]). The latter class of random functions could
be viewed as a generalization of the former, and this view
involves significantadvantagesconcerningstatisticalinfer-
ences. Actually, while searchingfor permissibility, we will
keep the advantageslinked to the intuitive idea of stationar-
ity, but at the sametime we will enlargethe applicationrange
of the theory to cover practical caseswhere the stationarity
is a physically inadmissibleassumption.
The difficulty of examiningthe permissibilityof candidate
covariance and variogram models is in general greater in
more than one dimension, and this is due to obstacles of
mathematicaland technical nature. Then the operation of
space transformations may be very helpful in calculations
involving isotropic models.
Before this presentation, a brief review of the mathemati-
cal conceptsand terminologiesused in the article is given, in
order to make it as self-containedas possible. In addition,
numerousreferences are given for the reader who wishes a
more detailed study.
It must be understood that in this article we will not deal
with the problem of fitting a function (candidatemodel) to
the data available. This is a problem related to the structural
analysisof the phenomenonunder study(familiarity with the
nature of the phenomenon,the data, etc.; see Matern [1960]
or Whittle [1954]) and also to the fitting techniques [e.g.,
Daniel et al., 1971].
REVIEW OF GEOSTATISTICAL INFERENCES
The main tools of geostatistical inferences for random
functions (RF) associatedwith spatiotemporalphenomena
are either the traditional covariance function
C(Xi,Xj) --' [2[Z(xi)Z(xj)
] - [2[Z(xi)][2[Z(xj)] (1)
or the variogram function
•xi, x•)= « Var[Z(xi)- Z(x•)] (2)
252 CHRISTAKOS: PERMISSIBLE COVARIANCE AND VARIOGRAM MODELS

where Xi = {Xi,1, ''', Xi,n) representsa point in the n-
dimensionalspaceR n, Z(xi) is the associatedRF, E[ ] is the
expectationoperator, and Vat[ ] is the variance operator.
The function (2) is a second-order statistical moment intro-
duced by Kolmogoroff[1941], under the name of structural
function. Since then, (2) has appearedin the literature as the
serial variation function [Jowett, 1955] or as the variogram
Z(xi). For the case of first-order increments, i.e., IRF-0, we
simply have k(h) - -T(h). It is very importantthat in the
context of nonstationary estimation and simulation, the
variance of any increment
• )kiX(xi)
i

function [Matheron, 1965, 1971]. For many workers in the

area the variogram function is the preferable tool for statisti- dependsonly on the GC-t(, k(h). Consequently,we only need
cal inferences [e.g., Delhomme, 1976, 1978; Chiles, 1977]. to knowk(h), whichis consideredasthe ordinarycovariance
of IRF-t(.
This happensbecauseit has a number of advantagesover the
covariancefunction (1). Some of them, especiallyimportant Under the above circumstancesthe covariance, vario-
in practical applications, are (1) it is mean free, (2) its gram, and generalizedcovariancefunctions should satisfy
empirical calculation is subject to smaller errors, and (3) it severalpreliminaryproperties, suchas the following.
may offer a better characterizationof the spatialvariability. 1. They belongto the classof real, even, and continuous
In practice, one may work with stationary covariance or (except possiblyat the origin) functions, and one can write
variogram functions in one or more dimensions, such that c(-h) = c(h) (6a)
c(xi, x•) = c(xi- x;) = c(h) (3a) •-h) = •h) (6b)
OF
k(-h) = k(h) (6c)
'¾(Xi,
Xj) : '¾(X
i --Xj): '¾(h) (3b)
for every h in their domains.Hence, we simplydeal with
respectively,where h = xi - x• is the vector difference positive only values of h, and then several interesting
betweenthe two pointsxi, x;. It is interestingto note that theorems of the class of even functions are utilized.
while (3a) requires the assumption that the RF Z(xi) is 2. The covariancec(h) always has an upper bound,
stationary [see Yaglom, 1962], (3b) requires a weaker as-
sumption,the so-calledintrinsic hypothesis,(i.e., that Z(xi) Ic(h) c(0) (7)
is an RF with stationaryincrementsZ(xi + h) - Z(xi) but not while only in special cases are the variogram •h) and
necessarily stationary itself; see Yaglom [1957] and Math- generalizedcovariancek(h) so bounded(e.g., see proposi-
eron [1971]. Obviously, a stationary RF is also an intrinsic tion 5, presentedlater in the paper).
one, but the reverse is not always true. In the case of a 3. They behave at infinity accordingto the laws
stationary RF it follows from (1) and (2) that
c(h)
T(h) = c(O)- c(h) (4)
limh(l-n/2
=0 h-->
o• (8a)
where c(0) is the variance, also called sill in geostatistics.
Equation (4) does not hold in general for an intrinsic RF, ?(h)
wherec(0) may not exist(e.g., •h) = Ihl).
Many times in hydrosciences, one faces the problem of
lim
I--•-=0 hi-•o• (8b)
modelingor estimatinga nonstationaryspatialphenomenon
with a complex trend (e.g., hydraulic heads in a hilly
lim =0 Ih -->o• (8c)
aquifer). Then one may generalize the intrinsic hypothesis h 2•c+2
and employ the more sophisticated model of the trth-order
where
intrinsic random functions (IRF-t(; see Yaglom and Pinsker
[1953] and Matheron [1973]). An IRF-t( is an RF which
requires a t( + l-th order increment to achieve stationarity. h= ß
hi2 , hi
More specifically, a linear combination
are the coordinates of vector h in R n. Equations (8) are
Y= • kiZ(xi)
i relatedto the requirementsthat c(h), T(h), and k(h) can be
expandedin a Fourier integral, and they are proven in
is a so-calledgeneralized trth order increment if proposition6 (presentedlater in the paper). The equations
(6a) through(8c), while necessary,are not sufficientcondi-
bl b2
• )ki Xi,I Xi,2 '''Xi, . = 0 (5) tions for a function to be a covariance, variogram, or
i generalizedcovariance.For example, considerthe function
•h) = 1 - exp d > 2. Despitethefactthatit satisfies
for all integers b•, b2, ß ß ß , bn -> 0 such that
all the aboveconditions,we will provein a followingsection
//
thatit is nota variogram
in R•. Throughout
thispresentation
we will discussadditionalnecessaryconditions,whichmay
i=1
be very useful in deletingmodels that cannot be considered
as covariance,variogram, or generalizedcovariancefunc-
and (xi,1, xi,2, ß ß ß , Xi,n) are the coordinatesof the point xi. tions, rather than in identifyingmodelsthat couldbe so.
Next it is possible to define the trth-order generalized In general, for a continuous function chosen as a covari-
covariance k(h) (or GC-t() as the stationary part of the ance, variogram, or generalized covariance model, it is
nonstationary covariance c(xi,x;) associated
withthe IRF-t(, necessaryand sufficientto satisfy the nonnegativedefinite-
 CHRISTAKOS' PERMISSIBLE COVARIANCE AND VARIOGRAM MODELS 253

ness properties which ensure that the variance of certain •h) = (Ihl) -- •r) (1lb)
linear combinations,
n k(h) = k(Ihl)-- k(r) (1l c)
Y= • •.iZ(xi) (9) for the isotropic covariance, variogram, and generalized
i=1
covariance, respectively.
is not negative. More specifically, and assuming herein Matern [1960], presents a number of ways to produce
continuous functions, permissible isotropic covariances, which, when properly
1. The covariance c(h) of a stationary RF Z(xi) must be modified, lead to permissible variograms as well. For in-
nonnegative definite, i.e., stance, one may utilize the theorem that if/z(a) is a measure
n n
in a space U and c(r; a) is a covariance integrable over the
subspaceV of U for every r, then the function
E E •.i •.j c(h)-> 0 h: Xi -- Xj (10a)
i=1 j=l

for all n and any complex number hi.

2. The variogram •(h), since it exists when the RF Z(Xi)
c(r)
=fvc(r;
a)dla(a) (12)

is only intrinsic, must be such that -•(h) is conditionally is also a covariance. The measure/x(a) may be an arbitrary
nonnegative definite, i.e., distribution
functionin R 1,andin manycasesc(r; a) belongs
n n to the family exp (-r/a). A well-known group of covariances
-- •] •] Xih,jT(h)• O (ob) derived as above is cn(r) = const rv Kv(r/a). When using (4),
i=1 /=1 the correspondingvariogram is •(r) = const [1 - const r•
K•(r/a)], where a, v > 0, and K• is the modified Bessel
when
function of the second kind [see Gradshteyn and Ryzhik,
n 1965]. For v = 0.5, we obtain the exponential model, while
E •-i--- 0 ¾n for v = 1, we find %(r) = const [1 - (r/a) Kl(r/a)], which is a
i=1
model widely used in geosciences[see Veneziano, 1980].
Matheron [ 1973] suggestedan effective way to consruct an
3. The generalized covariance k(h) must be conditionally
isotropic covariance in R n. First, assume that Zl(xi) is a
nonnegative definite, i.e.,
stationaryRF in R 1 with a covariancecl(r). Then, by the
n n
turning bands operator, the covariance of R n will be
E E •.i •.jk(h) -> 0 (10c)
i=1 j=l

c,(r)=2G 7r-1/2 G
for all combinations of (5). 2
If, and only if, a function satisfies(10a), (10b), or (10c) it
will be called a permissiblecovariance, variogram, or gener-
alized covariance model, respectively (some authors, in- ' CI(UF)[1-- U2](n-3)/2du (13)
stead of the term "permissible," prefer the term "admissi-
ble" which, however, is not used in this article because of its
where G is the gamma function. Formula (13) can be also
connotations in game theory). The validity of (10) immedi-
used for both the cases of variogram (after taking (4) in to
ately implies other properties of the covariance and vario-
account) and generalized covariance functions. For exam-
gram models (ordinary or generalized). For example, from
ple, startingfrom cl(r) = (1 - r/a) exp (-r/a), one derives a
(10a)for n = 2, hi = c(h), h2 = -Ic(h)l, it is trivial to show
covariance useful in hydrology:
the validity of (7). Permissibility in R n implies permissibility
in R m for m < n. The reverse is not true. For example, the
variogram
T(h)= 1 - cos(h)ispermissible
inR1butnotinR2
or R 3.
Practically speaking, it is difficult to apply (10) in order to
test if a proposedmodel is nonnegativedefinite and therefore
a permissible model. What is actually done in current
practice is to fit models which are known to be permissible
or are derived from permissibleonesto the data. Essentially, inR2,whereI0,Ii aremodified
Bessel
functions
ofthefirst
one utilizes the closure properties of each class of statistical kind and orders zero, one, respectively, and L0, Ll are
moments in the R n and uses mainly linear combinations of modified Struve functions of orders zero, one, respectively
permissible models [Matern, 1960; Journel and Huijbregts, [see Mantoglou and Wilson, 1982; Gradshteyn and Ryzhik,
1978]. For most of the applicationswe may favour isotropic 1965]. For the one-dimensionalgeneralized covariance kl(r)
models which depend only on the vector length = r2C+l,c > O,thecorresponding generalizedcovariance in
R2 is foundto be k2(r)= (cl) rr-l/2[G(c+ 3/2)]-1 r 2c+l.
Difference or differential equations can lead to several
Ih=r=(•/
hi2)
1/2 classesof isotropicmodels. Considerthe differenceequation
in R2 (usefulfor soilpatterns)
We will therefore concentrate on such models, and then we
can write Z(xi, Xj) -- b[Z(xi+1, xj) q- Z(Xi_l, xj)
c(h): c(Ihl)- c(r) (11a) q- Z(xi, Xj+l) q- Z(xi, Xj-l)] q- a(xi, xj) (14)
254 CHRISTAKOS: PERMISSIBLE COVARIANCE AND VARIOGRAM MODELS

wherea(xi,xj) is usuallywhitenoiseandb is a deterministic kin, 1964]. Usually, for C(w) to exist, it is sufficientto have a
coefficient. Equation (14) leads to variograms of the form c(h)tendingto zero fast enoughas Ihl- •.
If c(h) is an impulse or a periodic function, then C(w) still
exists in the sense of generalized functions. To demonstrate
y(r) = const 1
a this, considerin R • the impulsefunctiondefinedas

For more discussion and examples in soil sciences, see

Whittle [1954], Bartlett [1975], and Christakos [1982; also a
manuscript in preparation, 1983]. Another method of con-
structingisotropicmodelswill be discussedat the end of this
article.
However, it is obvious that all the above methods, while
providing rich sources of isotropic models, are not totally
satisfactory from the practical viewpoint discussedin the
introduction. In fact, in most of the cases encountered in
practice, the functions fitted to the experimental data are
neither included among the known permissible models nor
derived from them by a method like the ones described
before. Such situations deserve to be studied, especially
because of their interest regarding applications. This is the
task carried out by the remaining sections.
ANALYSIS IN FREQUENCY DOMAINS
The covariance and variogram models presented in the
introduction are real and even functions. So are their n-fold
Fourier transforms (FT) when they exist. The transform
correspondingto the covariance c(h) is usually called spec-
tral density function, and we may establishthe FT pair
fR•h- ho)
1
f(h)
dh
=f(ho)
where f(ho) is an arbitrary function continuous at h0. By
usingthis definition, the spectralfunction of the covariance
c(h) = rS(h)is easily found to be C(w) = 1/2vt.Sucha spectral
function correspondsto the so-called white noise process
(which is equivalent to the so-calledpure nugget effect in
geostatistics).A random function with a covarianceas above
experienceszero correlation for h valuesdifferentthan zero.
Another sound example is the covariance c(h) = c(0) cos
(ah) In Rl, and after the introductionof the impulse
= 1 cos
(wx)
clx
the "extended" spectralfunctionwill be C(w) = const{tS(w
+ a) + 8(w - a)}.
As we will see later, the intrinsic variograms and the
generalized covariances may also exist in the sense of
generalized
functions(e.g., in R•, 7(h) = Ihl,r(w) - const
W-2).
The propositionthat follows is fundamentalas it relates
the spectralfunctionsC(w) and F(w) in the caseof stationary

c(h)
=fRcos
(wh)
C(w)
dw (15)
n
RF.

Proposition I

C(w) = • cos (wh) c(h) dh (16) For a stationary RF with a variance c(0), the relationships
(2rr)n , below are valid:

where wh is the inner product Ziwihi, and r(w) = c(0) a(w) - C(w) (19)

f• F(w)
n
dw
=•0)=0 (20)
holds for the n-fold integration [see Sneddon, 1972]. where •w) is the n-dimensional impulse function such that
Note that, in general, instead of the ordinary Fourier
integral (15), the covariance c(h) may be expressed in the
form of the Fourier-Stieltjes integral •w•:O)
=0 f• •w)
dw
=1 n
(21)
Proof

c(h)
=fRcos
(wh)
dQ(w)n

To prove (19), take the FT of (4). Next, using(19) together

where the distribution function Q(w) is not necessarily with (21) and (15) for h = 0, i.e.,
differentiable.However, the presentation(15) is very conve-
nient and also more than adequate for the purposesof this
study. The matter will be again consideredbelow.
Similar representations and comments hold for the FT
c(0)
=f• C(w)
dw
n
(22)
we derive (20). Note that the area under C(w) gives the
pairs
variance c(0), while (22) imposesa limitation in the increase
•h) <->F(w) (17) of C(w)withthatof Iwl.TheF(w)itselfcontains
animpulseat
w - 0 whoseareais equalto c(0),whilefromw ->0+ it starts
k(h) <->K(w) (18) following the relationship F(w) - - C(w). In practice, one
may work with w -> 0+, so as to avoid the impulseof (19).
where F(w) K(w) are the spectral functions of the variogram An immediate result of proposition 1 is that the existence
7(h) and the generalized covariance k(h), respectively. of F(w) may be derived from the existenceof C(w) and vice
The spectral functions may exist in the senseof ordinary versa. Moreover, another interesting result is presented by
or in the senseof generalizedfunctions [Gelfand and Vilen- the following corollary.
 CHRISTAKOS: PERMISSIBLE COVARIANCE AND VARIOGRAM MODELS 255

Corollary 1 wherei = V'-1 and wh is the innerproduct•iwihi. For

For a stationary RF it is valid that engineeringpractice, •bis usually differentiable, so that (26)
will take the more convenient form

3,(h)
=fR[1- cos
(wh)]
n
C(w)
dw (23)
f(h)
=f• exp
(iwh)
F(w)
dw
n

where the C(w) satisfies (19).

(see also comments following (15)).
Proof The first criterion concerns stationary and isotropic RF
and is a straightforward application of the Bochner's theo-
This is a simple consequenceof combining(4), (15), and rem for the case of real-valued and isotropic functions f(r)
(22). and F(w).
It is important to see that (23) may exist when (15) does
First Criterion of Permissibility (COP-l)
not. To show this, consider the one-dimensional case of
expressions(15) and (23): the 7(h) convergeswhen C(w) has A continuous function c(r) is a permissible covariance
a singularity at zero of the form w -•, e < 3, while at the same model in R n, i.e., satisfies condition (10a), if and only if it is
time, c(h) does not converge. Also, one may define (23) in the FT of a nonnegative bounded measure Cn(w), usually
the caseof intrinsic only RF, which is characterizedfrom the called spectral density function (SDF); i.e., we must have
nonexistenceof c(h)(e.g.,inRn,•(h) = Ihl d > O,c(h)does
Cn(t.o
) • 0 on Rn (27)
not existbut C(w) = const Theseare significant
advantages, contributing to the use of the variogram for where, because of isotropy, the n-fold FT is reduced to a
many practical situations. Hankel transform (see Sheddon [1972] and next section). In
For an IRF-0 (i.e., first-order stationary increments) the this case we have a finite variance (sill), i.e., the correspond-
relationshipk(h) = - y(h) immediately implies ing variogram is of transitive type [see Journel and Huij-
bregts, 1978]. Of course, the permissibility of c(r) immedi-
K(w) = - F(w) (24)
ately implies the permissibility of the variogram 3,(r) (see
and consequentlythe •(h) and k(h) are equivalent statistical corollary 1). Furthermore, with the aid of proposition 1, (27)
moments(note that in this case there is no impulse involved gives
in the presentation (24)). c(O) •w) - Fn(rO)-> 0 V' (28a)
Isotropy of c(r), y(r), or k(r) (see (11)) implies isotropy for
or
the correspondingspectral functions
c(0) •(ro) - Fn(rO)-> 0 •o-> 0+ (28b)
C(w) = C(IwI) = C(w) (25a)
where Fn(rO
) is the isotropic spectral function of the vario-
r(w)- r(Iwl)= (25b) gram T(r).
g(w)- g(Iwl)- (25c) The representation (28b) may be preferable sometimes,
because it avoids the impulses at the origin. However, in
respectively, where most of the situations it may be easier to check the permissi-
bility of the covariance c(r) for several reasons (e.g., it
vanishesafter some distance). Consider the model c(r) = exp
(-r2/a 2) in Rn, where
n

In the present study we will concentrate on isotropic

I'2'-- • Fi2
covariance and variogram models, which are continuous i

everywhere. The study of the so-callednuggeteffect (discon-

tinuity at the origin) is not included in this presentation. It can be written as the product

CRITERIA OF PERMISSIBILITY
c(r) = II exp(-ri2/a2) -- II c(ri)
i=1 i=l
In this section we will discuss some necessary and suffi-
cient criteria concerningthe permissibility of covariance and and, consequently,its spectral function will be
variogrammodels (ordinary and generalized). These criteria n n

are equivalent to the nonnegative definite conditions (10a), Cn(W)

= II C•(wi)- I-[ a (2X•) -• exp(-oh2a2/4)
(10b), and (10c), but they have the advantage of being i=1 i=1

convenient in their application. Furthermore, a sufficient

only criterion, but with a particularly simple enunciation, is = an(2¾/,r)
-n exp(-ro2a2/4)
-->0
also discussed.
We can now safely conclude that the above model may be
Bochner [1959] proved that a continuousfunction f(h) is consideredas a permissible covariance. So does the corre-
nonnegativedefinite if and only if it can be expressedas the
sponding
variogram
T(r) = 1 - exp(-r2/a2).Theassociated
FT of a nonnegativebounded measure •b, i.e., RF is mean square continuous and differentiable of any
order, and one can predict deterministically its realization,

f(h)
=IRexp
(iwh)
dd)(w)
n
(26) everywhere, from knowledge of it only over regions of any
finite size. The criterion which follows, despite the fact that
256 CHRISTAKOS: PERMISSIBLE COVARIANCE AND VARIOGRAM MODELS

it is only sufficient ancl deals with rather specific classesof corresponding to (29) and (30b) are
models, is very convenient because it imposes conditions
-y'(0) < 0
directly on the covariance or the variogram and not on their
spectral functions.

Second Criterion of Permissibility (COP-2)

__
f;//(//2
__
/.2)-1/2
dy"(u)
>-
0
respectively. Both of them are not valid for the present
A continuous function c(r) is a permissible covariance variogram. Instead, the first one equals zero, while the
model in R n, i.e., it satisfiesthe condition (10a), if (1) it has a second one is equal to
negative derivative at zero:
const{r[Ko2(r)+ K12(r)]- K0(r)Kl(r)}
c'(0) < 0 (29)
which takes negative values for small r (K is the modified
while, at infinity, (Sa) holds, and (2) for all positive r it Bessel function; see Gradshteyn and Ryzhik [1965]).
satisfiesthe inequalitiesbelow: We now pass to the case of intrinsic only RF, where the
in R 1 variance c(0) may not exist. Then we can use only the
variogram y(r), and the correspondingcriterion is as follows
c"(r) -> 0 (30a) (necessary and sufficient criterion).
in R 2 Third Criterion of Permissibility (COP-3)
A continuous function 3,(r) is a permissible variogram

/,/
(//2
__
/2)-1/2
dc"(u)
->
0 (30b) model in R n, i.e., it satisfiesthe condition (10b) if and only if
(1) its spectral function Fn(a,) exists in the senseof general-
in R 3 ized functions; it does not contain any impulses 8a,), while
the measure-d Fn(a,)is a nonnegative
one,i.e.,
c"(r) - r c'"(r) > 0 (30c)
--002 Fn(W)>- 0 on Rn (33)
where c"(r) and c"'(r) are the second and third derivatives
of c(r). The expressionscorrespondingto the variogram y(r) and(2) it increases
moreslowlythanta as r -• •, i.e.,
are derived easily, with the aid of (4). lim y(r)/ta = 0 r-• • (34)
To prove the COP-2, we use some results concerning
moving average models with stochastic weight functions, The justification of this COP is based on Bochner's
furnishingrealizationsof isotropicRF, as have been present- theorem properly extended to take into account the special
ed by Matern [1960]. A wide class of covariance models features of the intrinsic variogram 3/(r). More specifically,
(e.g., any completely monotonic covariance) can be ex- from the definition of the variogram and in accordancewith
pressed as the Gelfand-Vilenkin theory [Gelfand and Vilenkin, 1964;
Yaglom, 1957; Matheron, 1973], a function -3/(r) is condi-
tionally nonnegativedefinitein the senseof (10b) if and only
c(r)
=const
f;unH(r/u)
dP(u) (31) if there existsa measurex(dw) -> 0 without atom at the origin
(i.e., includes no 8(w)) and such that
where

H(r/u) = (1 - v2)(n-1)/2B dv
-•h) = f• cos
(wh)
w2- 1x(dw)
n (35)
/, 2 '2
B is the beta function [see Gradshteyn and Ryzhik, 1965],
and P(u) is a nondecreasingmeasure. This last fact gives the
fnx(dw)/(1
n
+14
;2)
<c• (36)
Equation (36) is the existencecondition of integral (35) and is
idea of solving(31) with respect to dP(u)/du, requiringit to
equivalent to (34) above. To show this, assume that (34) is
be nonnegative.Doing so, we derive (30a), (30b), and (30c)
valid. Then, if e is a unit vector in R n we find
for n = 1, 2, and 3, respectively. Consequencesof (31) are
(29) and also the recursive equation

--Cn_2'(r)= constr-1 Cn"(r) (32) •,(eh)

=fR[1- cos
(ewh)]
n
x(dw)/w
2<_
ah
2 a<•
Taking the Laplace transform of both sides of the above
which is satisfiedby the classesof models for which COP-2
holds. inequality and summingover all axes of w = (w•, w2, '' ',
Wn),we find, in a trivial way,
Let us next examine the application of the COP-2 to a few
well-known covariance and variogram models. A trivial
example is the model c(r) = exp (-r/a). For this model it
holdsthat c'(0) = -1/a < 0 and c"(r) - r c'"(r)= (r + a)a-3 f t x(dw)/(t
2+w
2)-<2na
exp (-r/a) > O, and accordingto the COP-2, it is a permissi- (notethatthe Laplacetransform of h2 is 2/t3).Furthermore,
ble covariancein R3. It is clearthatthe sameappliesin R1, lim t2/(t2 + w2) = 1 whent -• c•,andthisimplies(monotonic
R2. An interesting counterexample in R2 is the variogram continuity)
?(r) = 1 - const r Kl(r), which does not satisfy COP-2 but
which is, however, a permissible model, as we saw in a
previous section. Indeed, the conditions for variograms f x(dw)
<2na
<o•
 CHRISTAKOS.' PERMISSIBLE COVARIANCE AND VARIOGRAM MODELS 257

Clearly, x(dw) is a bounded measure and (36) follows. For function of k(r) and includes no impulses •to). This com-
the caseof isotropicvariogram'/(h) = (Ihl) - the pletes the justification of COP-4.
isotropic
measure associated
withx(dw)is _o)2Fn(o)),
where Consider the model
Fn(o))is the FT of '/(r), accordingto well-known properties
k(r) = (- 1)l+1r2l+1 (42)
of FT and assuming that F•(o)) exists and contains no
impulses•o)). Consequently,conditionx(dw) -> 0 is equiva- In R1, we find [seeGelfandand Shilov,1964]
lent to (33), and the COP-3 is fully justified. A typical
example is the variogram widely used in application, 7(r) = Kl(o))= 7r-21-22-1-1 (1!)[(2• + 1)!!](o)/27r)
-21-2
r, with Fl(o))= -2/o)2 in R1. Anotherimportantcounterex- Consequently, x(do))= o)2•+2
Kl(o))or, aftersomemanipula-
ample is the function '/(r) = r{1 + cos (ar)}, which is not a tions,x(do))= 2(2/+ 1)!o)2(•-l).
Themeasure x(do))contains
permissible variogramin R 1, because no •to) and is nonnegative, satisfying all the conditions of
O)2 COP-4 if I -< •. For I = 0, (42) is GC-0; if I = 1, it is a GC-1,
_o)2Fl(o))=2 +•+ etc. Model (42) belongs to the very important family of
(to-- a)2 (to+ a)2 polynomial models
is not a (finite)measureon R 1(in the senseof Kolmogoroff m

and Fomin [1970]). To verify this fact, assume that there k(r) = • (- 1)1+1atr2/+1 (43)
exists an intrinsic RF ZCx3 with a varioeram as above. Then. /=0

for a = 1, '/(w) = 0 and '/(2w) = 4w. But in expressingthe

Applying COP 4 to the model (43), we find that I -< K = m,
variogram in terms of norms (in the sense, again, of Kolmo-
and also the restrictions below, imposed on the coefficients
goroffandFomin),we have,by definition,IIz(2,r>- z<0>ll-< al.'
IIz(2 - Z(r)11+ IIZ(r) - Z(0)ll,i.e., ,/(2w) < 2 ,/(w). Thus
,/(r) is not a permissible variogram model. if•=O
Some of the results concerningthe permissibility of ordi-
ao -> 0 (44a)
nary covariance and variogram models discussedup to now
can be properly extended to the case of IRF-K. This is a task ifK=l
carried out by the following criterion.
a l, ao > 0 (44b)
Fourth Criterion of Permissibility (COP-4) ff•:=2

A continuous function k(r) is a permissible GC-K in R ",

i.e., it satisfies the condition (10c) if and only if (1) its a2,ao-> 0 al-> -2 •- + 1 a0 a2 (44c)

spectral function K,(o)) exists in the sense of generalized

functionsand does not contain any impulses•o)), while the where n = 1, 2, 3.
measure
o)2•+2
K• (to)is a nonnegative
one,i.e., In practice, we need the GC-•, k(r) for the kriging equa-
tions, but the so-called generalized variogram g(r) (i.e., GV-
o)2•+2
K,(o)) -> 0 on R" (37) •; see Chiles [1979]) may be computed experimentally from
the available data. When the data are along lines and
and(2) it increasesslowerthanr2•2 as r --> o•,i.e., sampledat regular intervals, g(r) is defined as

lim k(r)/(r2•+2)= 0 whenr--> o• (38)

To justify the COP-4, much of the same procedure as with
COP-3 is followed. If an RF Z(xi) is • + 1 times differentiable
g(r)
=(2K
+2)-1
(in the meansquaresense),•thenits GC-• is 2(• + 1) times
differentiable and such that
ßVar{•l(-1)/(
K+
1)Z(j+
(K+
1-/)r)}
l=0 (45) 1
where
k(h)= fRcos
(wh)
-2PK(wh)
w2
•.+
n
x(dw) (39)

IRx(dw)/(1
n
+w2)
•-1<oo (40) (;)=(n!)/[(n
-m)!]m!
Definition (45) is a generalizationof the ordinary variogram
where •r) given by (2). The g(r) is well determined as a linear
function of k(r), i.e., we can find, by using (45),
P•(w)= • (- 1Y'w2V/(2p)! (41)
p=0 g(r)= •] (- 1)1 k(lr) (46)
•+ 1 /=-g-I K+ 1 + I
and x(dw) is a positive symmetric measurewithout atom at
origin. Equation (38) is a necessaryconditionfor a permissi- The reverse is not true, except for the case of IRF - 0,
ble k(r) (the proof is given in proposition6) and is equivalent where g(r) = 7(r) = -k(r). Therefore while we can determine
to the condition (40). This can be shown in exactly the same 'the correspondingpermissible g(r) from a permissible k(r),
way as before, with (34) and (36). For an isotropic k(h) = we cannot, in general, do the reverse. Nevertheless, we can
k(Ihl) - k(r), the spectralmeasureassociatedwith the occasionallyoperatein a rather indirectway to test the
measure x(dw)is o)2•2Kn(o)),whereKn(o))is the spectral permissibility
of g(r). If the fittedg(r) is of a polynomial
258 CHRISTAKOS' PERMISSIBLE COVARIANCE AND VARIOGRAM MODELS

form, we may start by determining the order K of the Changing into polar coordinates, we get
intrinsicalness(i.e., the order of the generalized increment
Fn-k(•o)
= fFn(•o
2 + y2)myk-1d& dy
that stationarizesthe data). Then we can find the form of k(r)
from (43), but the coefficients al will still be unknown. To where w and y are the radius vectors in the subspaces
evaluate them, we employ the formula (46), and if the so- and W• (with coordinates wn-•+l, wn-•+2, ß ß ß , wn), respec-
estimatedal satisfythe restrictions(44), then the experimen- tively, and Sk is the surface of the unit sphere in the k-
tal g(r) is permissible. Let us say that we fitted to our data, dimensional
subspace,
i.e., S• = 2•/2: G(k/2).Afterintegrat-
for K = 1, the candidate variogram model ingoverthewholespaceW•andputting& = (w2 + y2)m,we
find
g(r) = -•bor+ -•bir3 (47)
where b0, bl are known coefficients.From (46) we find

g(r) = --•k(r) + «k(2r)

rn-k(w) 2Yrk/2
ff
= G(k/2) Fn(t•)(•2__ 602)k/2-1
6•d&
(48) Moreover, for proper choice of n, k, and solvingfor F2(w),
But k(r) is given from (43) for g = 1, whereas F3(w) we obtain (50) and (51), respectively.
Let us illustratethe above propositionwith two examples.
k(r) = -aor + air3 (49) First,we consider
thevariogram
3,(r)= 1 - exp(-r2/a2).In
R 1 we have
andthen(48), becauseof (49), givesg(r) = 7ao
2 r 4 3
+ 3alr.
Comparingthis last equationwith (47), we obtain ao = bo,al a

= bl. ThustheGC modelis k(r) = -bor + bir3, whichmust -Fl(w) = exp(-oo2a2/4)

-• 0
be next checked to see if it is permissible.Using (44b), we
realize that for this we only need to have bo, b• > 0. whilein R2, R3 we find, using(50) and(51), respectively,
METHODOLOGICAL AND COMPUTATIONAL ASPECTS a2
-F2(w) = •exp (-w2a2/4)-• 0
After the theoreticalbackgroundprovided by the proceed-
ing COP's, we now pass to some practical aspectsconcern- a3
ing their effective application in practice. First, there are - F3(w)8,rr3/2
exp
( oo2a2/4)
->0
severalpropertiesof a stereologicalnature satisfiedby the
isotropic spectral functions which deserve to be studied, (i.e., the same result as in the previous section). Next, for
especiallybecauseof their interest regardingapplications. the generalized covariance k(r) = G(-h/2) r x, the spectral
functionin R l is (seeGelfandand Shilov,1964]
StereologicalProperties of Spectral Functions

A very important questionin geostatisticalpracticeis, if a

statistical
model(c(r), 3,(r),or k(r))is permissible
in Rl, is it Kl(oo)=7r_X_l/2
G[(h
+1)/2]
G(-k/2)(•_•)-x-,
anisotropic permissible
modelin R2 or R3. 9 Theproposition Using (51), we immediately obtain
below gives an answer to this question by presenting the
transformation equations between the one- and two- or
three-dimensionalspectral functions of the variogram •r).
Similar equationsare valid for the ordinary and generalized
covariancesc(r) and k(r), respectively, and this is shown in
K3(t.o)
=7r_h_3/2
G((h
+3)/2)(5)
-h-3
G(-k/2)

the examples to follow.

This expressionfor K3(w) is the samewith the one obtained
from Gelfand and Shilov [1964] for n = 3.
Proposition 2. The transformation equations between
the one- and two- or three-dimensionalspectralfunctionsof It is obvious that the last proposition allows us to work in
an isotropic variogram •r) are as follows'
R l, where we calculatethe spectralfunctionusingthe
available transform tables, as did Gradshteyn and Ryzhik
From R 1 to R 2 [1965]or Gelfand and Shilov [ 1964]. Then we can employthe
(50)and(51)to calculate
thespectral
functionsinR2andR3,
respectively.
Note thatgiventhe spectralfunctionin R•, it
may be more convenient to check the permissibilityof the
From R • to R 3 statistical
modelsin R3, evenif thefieldof interestis theR2,
because thepermissibility
in R3immediately impliestheone
1 dFl(w) in R 2.
F3(w) = (51) Because
of isotropy,the FT in R2 andR3 is reduced
to a
2 vroo d oo
Hankel transform [see Sneddon, 1972]. Then the following
For the c(r) and k(r) we simply replace F with C or K, corollariesare immediate applicationsof proposition2.
respectively. Corollary 2. The spectralfunction Fn(w)is given, when it
Proof. If Fn(w)is the spectralfunction in the spaceWnof exists:
n dimensions(w•, w2, ß ß ß , ton)and Fn-k is its projection in
the subspaceWn-•, of n - k dimensions,we have in R 2

Fn_k(001,092,ß ß ' , OOn_k)

= fFn(Wl, w2,''', (On)dtOn-k+l''' dwn r2(w)

=• if: Jo(wr)
r T(r)dr (52)
 CHRISTAKOS:PERMISSIBLECOVARIANCEAND VARIOGRAMMODELS 259

in R 3

I'3(w)
= 2w,tr2 sin(or)r •r) dr (53) c(o}

where J0 is the Bessel function of first kind and zero order.

Proof. Starting from

rl(W)
=• COS
(wr)T(r)dr ylr!

we apply (50) and (51) to get, in a trivial way, the formulas

(52) and (53), respectively.
By using Hankel transforms, we find, from (52) and (53), //
respectively: /

in R 2

•r)=2•r
f:Jo(tor)
(.0
F2(to)
do) (54)

in R 3
ri

•r) = 4•r
sin
•
(tor)
0)2 I'3(to
tor
)dto (55) Fig. 1. Variogram model of polygonal form.

Similar expressionsare, of course, valid for the covariance

functions, and the matter will be consideredagain, later on in Corollary 4. The spectral function F2(to)of a variogram
the presentation. ,/(r)in R2is equalto theone-dimensional
FT of thefunction
Moreover, when taking into consideration corollary 1,
proposition2 may offer additional expressionsfor the vario-
gram functions. More specifically, it may take the following
=•1In
•b(hl) . 1 y[r 2+h22)
= (hi 1/2]
dh2
forms:
Proof. By thedefinition
of theFT in R2,we havew2= 0,
in R 2
1-'2((.
0 = WI) = •
(2•r)2 COS
(Wlhl)
1

•r)=2z-
f:[1- Jo(tor)]
(.0
C2(to)
do) (56a)

in R 3
'{fR'
T(h12-t-h2
dhl
•r) = 4•r
f:[ sinl
(tor)
1 --
tor
0)2 C3(to)do) (56b) and corollary 4 followsß
Corollary 4 may offer an alternative, and sometimes
where C2(to)and C3(to)satisfy (19) in the isotropic case. easier, way for the calculation of I'2((.0)to the one suggested
Corollary 3. For a stationary and isotropic RF with a by proposition 2. In addition, it can be made valid for the
transitive-typevariogram •r), a sutficientconditionfor •r) spectral function C2(to), by simply replacing ,/(r) with c(r).
to be permissible
in R3 is that the functionr [c(0) - •r)] is To demonstrate this, take the covariance model
nonincreasing.
Proof. From (28a),
c[r
=(h•
2+h22)
1/2]
=exp
(-h•
2+h22) a2

C(0)
(•}(to)
--I'(to
) = 2,n2to[c(0)
- T(r)]
r Then the corresponding•b(hl)function will be equal to

ß sin (tot) dr =
2-tr2to
m•0 (--1)
m
whoseFT inR1is(a2/4rr)exp(-to2a2/4).Thelatterisalsothe
spectralfunctionC2(to)in R2 of the covariancec(r).
U+ + --' sin (uto) d/•
Polygonal Models
The summation consists of a series with alternatively posi- When applyingbasic geostatistics,we frequently deal with
tive and negative terms, decreasing in absolute value, and transitive variogram modelswhich have a polygonalform, as
since the first term is positive, it follows that c(0) t$(to)- is shown in Figure 1. The proposition that follows offers an
F3(to)> 0, i.e., (28a) is satisfied. easy method to calculate the spectral functions Fn(to).
260 CHRISTAKOS:PERMISSIBLECOVARIANCEAND VARIOGRAM MODELS

Proposition3. If the variogramhas a polygonalform, its Proposition4. The Fourier-Besselapproximationsof the

spectralfunctionin R• is givenby (0) > 0+) spectral functions 1`n(0))of a variogram model •r) are as
follows:
n
1`l(W)= • {s(ri)-- s(ri-1)}COS
(wri) (57) in R •
71'O)2
i=0
FI(0))= 0)c
-1 • T(krc/0)c)
cos(krc0)/0)c) (59a)
where s(ri) is the slopeof 7(r) at the point ri (see Figure 1). k=0

For the spectralfunctionsF2(0))and I'3(0)),one uses(50) and

in R 2
(51) together with (57) above.
Proof. If 7(r) is of polygonalform, its secondderivative
is 1`2(0))
= (fro)c2)
-1 • ' y(i#0)c)
Jo(it•0)/0)c)
{J12(it0)-• (59b)
n
in R 3
T"(r)= • {s(ri)-- s(ri-1)}•(r -- ri)
i=0 F3(0))= (20)c
2 0))-• • k T(krd0)c)'
sin(krc0)/0)c)(59c)
k=l

SincetheFT of •"(r) is - 0)21`•(0))andtheFT of theimpulse

•(r - ri) is known, we find (57). whereJo is the Besselfunctionof vth order (v = 0, 1) and ik
The frequency0)cup to which we needto compute1`,(0))is are the zeros of J0 [see Gradshteyn and Ryzhik, 1965].
given a priori or determinedin someother way. For exam- Proof. The proof of this propositionis based on the
ple, in R•, usingthe boundsof - 1`•(0)), theorydiscussed
by Watson[1966,chapterXVIII]. TheR•
caseis a trivial expansionof F•(0)) in Fourier series.For the
R2,we expandthefunction2rrF2(0))assuming F2(0)> 0)c)=
0, v = 0, andwe get (59b).In R3, we expandthe function
b2= const0)-2 • {Is(ri)- s(ri-•)} 4rrX/•'r
F3(0))
andwe find,for v = 0.5, theexpansion
(59c).
i=0 Alternatively, in order to prove (59b) and (59c) one may use
proposition2 togetherwith expansion(59a).
0)½can be found solvingthe equationb2(0)c)- e b•, where e
<1.
In practice,in orderto removethe impulsesinherentin the
above equations,we may replace ,/(krd0)c)with ,/(krd0)c)-
To illustratethe applicationof proposition3, considerthe
c(0), for 0) > 0. Then the value of the later functionat origin
model
hasto be halved (i.e., at k = 0 of (59a)). Moreover, insteadof
T(r) = ar r -< 1/a 5;k
©,we usethe summation
•;•N, whereN is the numberof
T(r)= 1 r _>1/a (58) samplestaken.The choiceof N as well as the sampling
interval of the discretization are very important when pro-
-• sin2 (0)/2a)-> 0, which gramingthe equationsin a digitalcomputer.A better approx-
In R •, we find - F•(0))= 2a (fro)2)
asserts thepermissibility of •r). In R2, by applying(50)we imationmay be obtainedin somecasesby multiplyingeach
find (0) > 0) term of the summationby (1 - k/N). Such a "modification"
expressesthe approximatedspectralfunctionas a weighted
averageof the actualspectralfunctionwith a positivekernel,
- F2(0))
=• 1f•/a
Jo(0)r)
(1- ar)r dr and consequentlyit retains the true sign of the spectral
function [see also Lanczos, 1966]. These aspectsconcerning
which takes negativevalues and consequentlythe •r) is not the application in practice of the proposition4 are to be
permissible
in R2. It is interesting
to notethat (57), without emphasizedin the following section.However, becauseof
the negative sign, gives the spectralfunction of a polygonal spacelimitations,more detailsand casestudieswill be given
covariance c(r).
in a forthcoming paper.
Approximation of Spectral Functions Necessary Conditions

Proposition3 can be usedto evaluatethe spectralfunc- As we already discussedin the introduction,a numberof
tions of a curve variogram or covariancemodel, properly restrictionsare placed on the possiblebehaviour of the
approximatedby a polygonal,keepingin mind that we are covariance,variogram,and generalizedcovariancemodels.
interestedin the signof the spectralfunction rather than its We now give short proofsfor the most important of them,
exact value. The degreeof accuracyof suchan approxima- and we also present some more in the form of simple
tion can be very satisfactoryby proper discretization. inequalities.
Anotherway to evaluatespectralfunctionsapproximately Proposition5. For a stationaryand isotropic RF the
is to expand them in Fourier-Bessel series. This is an associatedvariogramshouldsatisfythe inequalitiesbelow:
excellent approximationfor the common case where they in R 1
vanish outside the finite frequency 0)c, discussedbefore.
However, even when this does not happen, such an expan- •r) -< 2 c(0) (60a)
sion is still usablebecause, as we already mentioned, we are in R 2
interested about the sign of the spectral functions and not
exact values. The propositionbelow provides the explicit 7(r) < 1.403c(0) (60b)
expressionsfor the spectralfunctions1`n(0))of a variogram in R 3
model •r). The expressionscorresponding to a covariance
model c(r) can be derived in exactly the same way. •r) -< 1.218c(0) (60c)
 CHRISTAKOS:PERMISSIBLECOVARIANCE AND VARIOGRAM MODELS 261

The usefulness of all the above restrictions regarding

permissibility consists of the preliminary rejection of non-
permissiblemodels before we proceed to the sophisticated
analysis described by the COP's.
Procedure

Based on the above results, the procedure below may be

useful in judging whether a function fitted to the data is a
permissible model:
Step 1. This is a preliminary step. Check if the necessary
conditions(propositions5 and 6) are satisfied.
Step 2.
o lO 20 1. If the candidate model is of a "convenient," closed
r i,.i form (i.e., one can find its FT in R • by usingthe existing
Fig. 2. Calculated and fitted variogram models for some water transform tables), calculate the particular spectral function
content data. The fitted model is permissible in a one-dimensional in R• andcheckits permissibility
by usingthe appropriate
but not in a three-dimensionaldescriptionof the spatial variability. COP. If the answer is positive, calculate the spectral func-
tionin the fieldof interestR2 or R3, by usingProposition
2
and the associatedcorollaries, and apply the COP's again.
Also, in R 1, 2. If the candidate model is of a transitive, polygonal
T(2mr)--<4m •r) m nonnegative integer (61) form, use proposition3.
3. Finally, if the candidate model is of some arbitrary
•(r)- •(r + r') < [2 •(r')]m (62) form, we may have to evaluate the spectral functions ap-
proximately by using propositions3 or 4.
Proof. For (60a), we simply use (4) and (7). For (60b),
Many practitioners used to fit polygonal or curve func-
starting from
tions empirically "by eye." Such a techniqueallows them to
include information and engineeringjudgement which are
•r)=2rr
f•[1- J0(•or)]
•o
C2(6o)
d•o not easily quantified (this is an interesting advantage over
the statistical fitting techniques). Steps 2.2 and 2.3 would
we see that •r)/c(O) cannot be larger than help the workers in the area to check if the fitted model is a
permissible one or not and take appropriate action. The
1- Inf Jo(•or)= 1- (-0.403)= 1.403
applications of the section that follows will illustrate this
methodology.
In an analogousway, (56b) impliesrestriction(60c). For (61), APPLICATIONS
from the valid inequality 1 - cos (2•or)-< 411 - cos (mr)], we
easily derive •(2r) _<4 •(r). Then, by induction, we get (61). Testing Candidate Models for Permissibility
Finally, for (62), use the inequality of Schwarz [see Kolmo- We will now examine the application of the results of the
goroff and Fomin, 1970]. It is not difficult to derive similar previous sectionsto further candidate variogram and covari-
restrictions for the covariance c(r), using the above results ance models. Some of them are already in use in water
together with (4). resources, while some others are functions occasionally
Let us next examine a few examples. First, the variogram fitted to the available data.
•(r) = 1 - cos (r) cannot be a permissiblemodel in more than First, consider the model
one dimension,as we see by using(60). The model •(r) = 1 •(r) = 1 - exp (-r/a) cos (br) a, b > 0 (63)
- exp (-r a) is not permissible
in R1 if d > 2. This results
from the application of restriction (61), together with the It offers a very good fit for many soil properties [see, Lumb,
expansions 1975]. However, caution is needed when using (63) because
it is not a permissible variogram for arbitrary values of the
7(ar)•:l• ( 1)k+
= •(ar)ka
k!
- a = 1 and 2. coefficients a and b. By applying the methodology of the
precedingsectionwe can avoid suchpitfalls:In R•, the
spectral function is
Proposition 6. An ordinary covariance c(r) and a gener-
alized one k(r) should behave at infinity, as in expressions a2 + b2 + o)2
-Fl(•O) = > 0
(Sa) and (8c), respectively. ' [(•o- b)2 + a2][(•o+ b)2 + a2]
Proof. Equation (8a) is a straightforward application of
the lemma of section 14.41 of Watson [1966], assuming that But in R3 we find,by usingproposition
2, that the spectral
function

•C(oo)
(.o
(l-n)/2
&o - F3(to)=
604q_2•2 602q-(2a -- •2) •2
0•7r2(0)
4 q-2c•ro
2 + •34)2
exists.For (8c),from(39)andthewell-knowninequalitycos a = (1 - a2b2/a
2) 13= (1 + a2b2)/a
2
(x) - Pk(x)l< x2K+2/(2t•
+ 2)!,wehavek(r) _<al + a2r2K+2,
when al, a2 are constants.Then (8c) follows by the dominat- is positiveonly if ab < 0.577. To illustratethe importanceof
ed convergencetheorem. For K - 0, we find (8b), keeping in these restrictions, the model (63) is fitted to the water
mind that in this case k(r) = -•r). contentdata of Figure2, for a - 6.3 m and b - 0.3 m-1.
262 CHRISTAKOS.' PERMISSIBLE COVARIANCE AND VARIOGRAM MODELS

o Data ble one,evenin Rl, wherewe get,by usingproposition 3,

1.4
-- Fotted Model -F•(03) = const03-2{1 + 15 cos(403)- 16cos(503)}
which takes negativevalues. On the other hand, the model of
1.0 o Figure 3 is immediately rejected from the application of the
o
necessarycondition (61). For instance, for rn = 1, r = 0.5 we
vlr)
get 342 x 0.5) - 341) > 4 340.5). The use of proposition3 in
0.6 practice may offer some interesting hints concerning the
permissibility of polygonal models or models approximated
by polygonal functions. For example, a large initial slope
0,2 o
seriouslycontributesto the permissibilityin R 1 (in the
present model the initial slope is only 0.25, see Figure 3).
0.0
• o The same appliesto a nonincreasingslope of the variogram
0.0 0.2 0.6 1.0 1.4
(in the present model we have a slope increasingfrom r =
Fig. 3. The fitted model is not a permissiblevariogram in one or 0.8- to r = 0.8+).
more dimensions.
In the application to follow, we will use the methodology
of proposition 4 to investigate the permissibility of the
According to our discussion,the fitted model is permissible variogram model
in R l, but it cannotbe usedfor a three-dimensional
descrip- ?(r) = r2/(1+ r2) (66)
tion of the spatial variability, since ab = 1.89 > 0.577,
leading to erroneous results (negative variances, etc). Note fitted to the data, as shown in Figure 4. Its spectralfunction
that the model (63) is continuous at r = 0 and thus for all r > in R3 is calculated
employing
a computerprogrambasedon
0. However, (59c), after replacing ?(krd03c)with ?(krd03c)- 1 (see discus-
sion following proposition4). A samplinginterval of Ar =
dy(r)
0.50 and N - 16 samples, with a highest frequency of 03c=
r=0 2rr has been used. Such a 03cvalue insures no aliasing. In
Figure 5 are shown both the calculated and the theoretical
and therefore we conclude that the associated RF Z(Xi) ,
spectralfunctionsfor comparison. Obviously, the --F3(03)is
while continuous (in the mean square sense), is not mean
square differentiable.
always nonnegativeand hence the model (66) is permissible
For the extensively used exponential variogram
in R3 (andalso,of course,in R2 andR1).Whenprograming
proposition4 to the computation of spectralfunctions, care
7(r) = 1 - exp (-Ha) a > 0 (64) is needed in some details, such as the choice of Ar and N (a
large Ar or a smallN may lead to poor approximationof the
the spectralfunctionin R 1is suchthat
spectral function at higher frequencies 03) and also the
correct interpretation of the results (the approximatedspec-
--F1(03) -- > 0 tral function may be symmetrical about N/2, related to
'rr(1+ 032a2)
negative frequencies, and this follows because it is an even
and the COP-1 is satisfied. Next, by using proposition 2 we function).
find The variogram

in R 2 7(r) = r m rn _>0 (67)

is a nonbounded one, and for the associated RF only the
-['2(03) = > 0 intrinsichypothesis
is valid.In R3it satisfies
theCOP-3only
2•ra[(a203
2 + 1)/a2]
3/2
if0 -< m < 2, when
in R 3

a3 - 092
F3(03)
= G[(m+ 3)/2]
G(-m/2)
•r -(m+3/2)
03-m-1
>0 03>0
--['3(03) = > 0
'tr2(1+ 032a2)2
lim 'y(r)/r2= 0 r--> •
andaccording to COP-l, model(64)is permissiblein R •, R2, Consequently,
it is permissible
in R3, andit is notdifficultto
or R3. Like model(63), it corresponds
to a continuous but show that the same applies in R n, n > 3. The model (67) is
not differentiable RF.
continuous but not always differentiable, depending on the
A covariance
modelpermissible
onlyin R • andR2 is the value of rn.
one below:
The GC-1 spline model
c(r) = exp (-r2/a 2)Jo(br) (65) k(r) = rm log (r) rn > 0 (68)
In R2, its spectralfunctionis suchthat is a very useful model for nonstationary spatial phenomena,
and it is derived from the family (43). When applyingCOP-4,
we find
C2(03)
= zi7.
rIo(a2b03/2)
exp- 4 ' >0
dCm
and the COP-1 is valid. This is a model for rather strong K(to)
:• (oo/2
•r)-m-n+ Cm
(00/2
•r)-m-nlog(0o/2
•r)
correlations and tends rapidly to zero, where it oscillates.
The underlying RF is mean square differentiable. whereCm= 2m+nYrn/2G[(rn+ n)/2]/G(-m/2)[seeGelland
The polygonal model shown in Figure 3 is not a permissi- and$hilov, 1964].For 03
4K(03)• 0, we need0 < rn < 2. The
 CHRISTAKOS: PERMISSIBLE COVARIANCE AND VARIOGRAM MODELS 263

1.0
o u wherep = (n - 2)/2. Equation (70) impliesthat the "normal-
ized" spectral function C,(oo)/c(O)can be viewed as the
probabilityfunction Po(v) of an n-dimensionalrandom vec-
tor, i.e., Pdv) - C,(oo)/c(O).Furthermore, P•(v) is isotropic
0.8
(i.e., it is fully definedby the probabilitydensityfunctionof
Iv[, say,P•(w), whereIv is in w units).The necessaryand
sufficientconditionsfor c,(r) to be a permissiblecovariance
0.6
model is that it can be expressedby (70) where C,(w) is
nonnegativeand summable(see COP's). Basedon this fact,
and since it is easier to construct a probability density
function P• than a nonnegativedefinite covariance c,(r), a
convenientway to generate covariance and (consequently)
0,4
variogram functions in R" may be as follows:
1. Construct a probability density function P• for a
random vector v.
0.2 2. SubstitutePo into (70) to obtain the corresponding
Data
covariance c,(r).
F•tted Model
3. Substitute c,(r) into (4) to find the corresponding
0,,0 i i i
variogram ¾,(r).
0 2 4 6 Let us examinea few examples. Supposethat the random
vector v is uniformly distributed on the surface of an n-
Fig. 4. Calculatedandfittedvariogrammodelsin R3. dimensionalspherewith radius u. The probability density
function of v will be

value m = 2 is also included becausethe k(r) may be derived

as the limit of the first-order generalizedcovariance G(n/2)

r 2+c - r 2 0< c< 2

Pv(v
= w)=2,trn/2
(.on_l
(•((.0-
/,/)
and the spectralfunction C,(o) = Po(o) c(0) is nonnegative
for all w. Hence we can substitute it into (70), and then by
r2+C _ r 2 using (4), derive the variogram model below:
k(r) = lim
c--•O

¾.(r)= c(O){1 - 2p G(n/2)(ur)-p Jp(ur)} (71)

andthe limit of a generalizedcovarianceis still a generalized
covariance. On the other hand, condition (38) is satisfiedfor The model(71) for n = 1, gives ¾•(r)= c(0) [1 - cos (ur)]; for
m < 4. Hence the splinemodel (68) is permissibleas a GC-1 n = 2, gives ¾2(r)= c(0) {1 - Jo(ur)};and finally, for n = 3;
inR", for0 < rn < 2. gives ¾3(r)= c(0){1 - [sin(ur)/ur]}.
Finally, a remarkable generalized covariance model, re-
sultingfrom the combinationof the models(43) and (68), is
1.0
k(r) = -ao r + a]r 2 log(r) + a2r 3 (69) Approximation
Theoretical Model
In R3 the COP-4 leadsto the followingconstrainsfor the
parametersa0, a 1, and a2:
0.8

a0,a2 -> 0 a] -> -1.47(a0a2)v2

For R2 and R • the constraintson a• are less severe,being
- 1.50 and - 1.56, respectively.
0.6

A Spectral Method for Generating Covariance

and Variogram Models
So far we have concentrated on the problem of testing a
0.4
given function to see if it is a permissiblestatisticalmoment
(covariance,variogram, or generalizedcovariance). Next,
we will discusshow the sametheory as above can be applied
to create severalfamilies of isotropicmodels. The idea is to 0.2
work in the oppositedirectionand try to constructvariogram
or covariancemodels from properly chosen spectralfunc-
tions. If p, = c,(r)/c(O) is the correlationfunction, similar to
(54) and (55), expressionsmay be established,i.e., we can 0.0 i i i •
write 0.0 0,25 0,5 0.75 1.0

=c,(r)
p.(r) -_(2
c(O) •r)
rp"/2
fR "2•C,(w)
+l Jp(oor)
00 c(O) doo (70)
Fig. 5. Calculatedand theoretical spectralfunctionsfor the vario-
gram model of Figure 4.
264 CHRISTAKOS'. PERMISSIBLE COVARIANCE AND VARIOGRAM MODELS

Following the same procedure, if we take the Pv to be of (74) for n = 5, we obtain the model
exponentialform, the correspondingvariogramwill be
73(r)= c(0){2.5cr- 2.5c3r
3 + cSr5} r- < 1/c

¾3(r)= c(0) r > 1/c

7n(r) = c(0) 1 - 1 q- (72)
which is a model for stronger correlations.
which may be a useful model in geotechnicalapplications. To conclude, the methods discussedin this section may
Also, if we take provide a number of variogramand covariancefunctions
usefulfor a wide range of practicalapplications.From this
pointof view, they couldbe consideredas toolsof potential
Pv(oo)= 1+ l • n/2 usefulnessfor the geostatisticalinvestigations.
CONCLUSIONS
it is not difcult to prove that the correspondingvariogram The important problem of judging the permissibilityof
will be
candidate covariance and variogram models (ordinary or
%(r) = const(1 - constr* K,(r/a)) generalized) has been discussed in this article. Several
criteria of permissibility, covering the case of stationary
where v = (1 - n)/2. A convenient class of models is random functions as well as the case of Kth-order intrinsic
generatedby assumingthat random functions, have been introduced for this purpose.
A procedurerecommendedfor testingthe permissibilityin
one or more dimensionsmay be suitablefor a wide rangeof
P•(w)
= (2/w)
• • J•/2(w) (73) practical applications. To maintain the proper balance be-
tween theory and facts, we have presentednumerousexam-
In this case we may derive the following models: ples that have been computed in detail.
forn= 1 The author believes that the problem of identifying per-
missiblemodelsis by no means simple. However, he hopes
3q(r) = constr r -< 1/const that the reader of this article will be helpedto test efficiently
whethera particularfunctionfitted to the data is permissible
yl(r) = c(0) r > 1/const
or not, and also that he will find some new sources of
forn = 2 potential models in addition to the ones already used.

c(0) Acknowledgments. The writer acknowledgeswith appreciation

¾2(r)= 2 {cr(1- c2/'2)
1/2q-arcsin(cr)} r _< 1/c the commentsand criticisms made by G. Matheron, Director of the
centre de Geostatistiqueet de Morphologie Mathematique,Ecole
des Mines de Paris. Thanks are also owed to K. Potter and P.
y2(r) = c(0) r> 1/c Holland for their comments,which helpedimprovethe quality of the
article.
forn= 3
REFERENCES

¾3(r)= c(0)

¾3(r)= c(0)
'}
cr -- --
2

r > 1/c
C3r
3 r --<1/C Bartlett, M. S., The Statistical Analysis of Spatial Pattern, Chap-
man and Hall, London, 1975.
Bochner, S., Lectures on Fourier Integrals, Princeton University
Press, Princeton, N.J., 1959.
Chiles,J.P., G•ostatistiquedes ph•nom•nesnon stationaires,th•se
Such models have the advantagethat data points can be de Docteur-Ing•nieur,Univ. de Nancy, Nancy, France, 1977.
ignored after the distance 1/c and so the computationsare Chiles, J.P., Le variogrammeg•n•ralis•, Res. Rep. N-612, Centre
de G•ostat., Fontainebleau, France, 1979.
reduced.
Christakos,G., Mathematical modelling of advancedestimation
We can apply the samemethodas aboveto find modelsfor approaches for geotechnicalsystems,M.S. thesis,Mass.Inst. of
n > 3. For example, a commonfourth dimensionmay be the Technol., Cambridge, Mass., 1982.
time. The models obtained for n > 3 are also valid for n -< 3, Daniel, C., F. S. Wood, and J. W. Gorman, Fitting Equations to
Data, Wiley-Interscience, New York, 1971.
a fact that further enrichesthe classesof available isotropic Delfiner, P., and J.P. Delhomme, Optimum interpolationby krig-
modelsfor the one-, two-, and three-dimensionalspaces.For ing, in Displayand Analysisof SpatialData, editedby J. C. Davis
instance,(73) may be used for n = 5 (say) to give and M. J. McCullagh,pp. 96-114, JohnWiley, New York, 1973.
Delhomme,J.P., Applicationde la th•orie des variablesr•gionali-
¾s(r)= 0.25c(0) ß[7.5cr- 5c3r3 q- 1.5c5r
5] r < 1/c s•es dans les sciences de 1' eau, th•se de Docteur-Ing•nieur,
Univ. Pierre et Marie Curie, Paris, 1976.
¾5(r)= c(0) r > 1/c Delhomme, J.P., Kriging in the hydrosciences,Advan. Water
Resour., 1(5), 251-266, 1978.
which may be a convenientmodel for n < 5, also. Delhomme,J.P., Spatialvariability and uncertaintyin groundwater
An alternativemethod, to producepermissiblemodelsfor flow parameters:A geostatisticalapproach, Water Resour. Res.,
n < 3 from known permissiblemodelsin n > 3, is introduced 15(2), 269-280, 1979.
by the recursive formula below: Gelfand, I. M., and G. E. Shilov, GeneralizedFunctions1, Academ-
ic, New York, 1964.
r ¾n'(r) dYn(r) Gelfand, I. M., and N. Vilenkin, Generalized Functions 4, Academ-
Tn-2(r) = 'yn(r)+ 'yn'(r)= • (74) ic, New York, 1964.
n-2 dr Gradshteyn,T. S., and I. M. Ryzhik, Tablesoflntegrals, Seriesand
Products, Academic, New York, 1965.
Formula (74) is derived from the combination of (4) and (32). Jenkins, G. M., and D.G. Watts, Spectral Analysis and its
Consideras an example the model ¾5(r),above. When using Applications, Holden-Day, San Francisco, 1968.
 CHRISTAKOS:PERMISSIBLECOVARIANCEAND VARIOGRAM MODELS
Journel, A. G., and C. J. Huijbregts, Mining Geostatistics, Academ-
ic, New York, 1978.
Jowett, G. H., Sampling properties of local statisticsin stationary
stochastic series, Biometrika, 42, 160-169, 1955.
Kolmogoroff, A. N., The distribution of energy in locally isotropic
turbulence, Dokl. Akad. Nauk SSSR, 32, 19-21, 1941.
Kolmogoroff, A. N., and S. V. Fomin, Introductory Real Analysis,
Dover, New York, 1970.
Lanczos, C., Discourse on Fourier Series, Hafner, New York, 1966.
Lumb, P., Spatial variability of soil properties, paper presentedat
the Second International Conference on Application of Statistics
and Probability to Soil and Structural Engineering, Tech. Hochs-
chule Aachen, Aachen, Federal Republic of Germany, 1975.
Mantoglou, A., and J. Wilson, The turning bands method for
simulation of random fields using line generation by a spectral
method, Water Resour. Res., 5, 1379-1394, 1982.
Matern, B., Spatial variation, Medd. Statens Skogsforskniginst.
Swed., 5, 144 pp., 1960.
Matheron, G., Les Variables regionalisees et leur Estimation,
Masson et Cie, Paris, 1965.
Matheron, G., The Theory of Regionalised Variables and its Appli-
cations, Ecole des Mines, Fontainebleau, France, 1971.
Matheron, G., The intrinsic random functions and their applica-
tions, Advan. Appl. Prob., 5,439-468, 1973.
Mejia, J. M., and I. Rndriguez-Iturbe, On the synthesisof random
field samplingfrom the spectrum:An applicationto the generation
265
of hydrologic spatial processes, Water Resour. Res., •0(4), 705-
712, 1974.
Sneddon, I. N., The Use of Integral Transforms, McGraw-Hill,
New York, 1972.
Veneziano, D. A., Random processesfor engineeringapplications,
course notes, Dep. of Civ. Eng., Mass Inst. of Technol., Cam-
bridge, Mass., 1980.
Watson, G. N., Theory of BesselFunctions, CambridgeUniversity
Press, London, 1966.
Whittle, P., On stationary processesin the plane, Biometrika, 41,
434-449, 1954.
Yaglom, A.M., Some classes of random fields in n-dimensional
space,related to stationary random processes,Theory Probab. Its
Appl. Engl. Transl., 3, 273-320, 1957.
Yaglom, A.M., An Introduction to the Theory of Stationary
Random Functions, Prentice-Hall, Englewood Cliffs, N.J., 1962.
Yaglom, A.M., and M. S. Pinsker, Random processeswith station-
ary increments of order n, Dokl. Akad. Nauk SSSR, 90, 731-734,
1953.
G. Christakos, 4 Promitheos Street, Koliatsou Square, Athens
(817), Greece.
(Received March 9, 1983;
revised November 2, 1983;
accepted November 16, 1983.)