ARTICLE IN PRESS

Computers & Geosciences 32 (2006) 1615–1628

www.elsevier.com/locate/cageo

TBSIM: A computer program for conditional simulation

of three-dimensional Gaussian random fields
via the turning bands method$
Xavier Emerya,, Christian Lantuéjoulb
a
Department of Mining Engineering, University of Chile, Avenida Tupper 2069, Santiago 837 0451, Chile
b
Centre de Géostatistique, Ecole des Mines de Paris, 35 rue Saint-Honoré, 77 300 Fontainebleau, France
Received 4 July 2005; received in revised form 4 March 2006; accepted 5 March 2006

Abstract

The simulation of spatially correlated Gaussian random fields is widespread in geologic, hydrologic and environmental
applications for characterizing the uncertainty about the unsampled values of regionalized attributes. In this respect, the
turning bands method has received attention among practitioners, for it allows multidimensional simulations to be
generated at the CPU cost of one-dimensional simulations.
This work provides and documents a set of computer programs for (i) constructing three-dimensional realizations of
stationary and intrinsic Gaussian random fields, (ii) conditioning these realizations to a set of data and (iii) back-
transforming the Gaussian values to the original attribute units. Such programs can deal with simulations over large
domains and handle anisotropic and nested covariance models.
The quality of the proposed programs is examined through an example consisting of a non-conditional simulation of a
spherical covariance model. The artifact banding in the simulated maps is shown to be negligible when thousands of lines
are used. The main parameters of the univariate and bivariate distributions, as well as their expected ergodic fluctuations,
also prove to be accurately reproduced.
r 2006 Elsevier Ltd. All rights reserved.

Keywords: Geostatistics; Stochastic simulation; Spectral method; Ergodic fluctuations; Multivariate normal distribution

1. Introduction attribute, either at a single location (local uncer-

tainty) or jointly over several locations (spatial
Geostatistical simulation is increasingly used in uncertainty). For instance, it is helpful to estimate
earth and environmental sciences for assessing the the probability that the average value of the
uncertainty about the unsampled values of a spatial attribute over a given area exceeds a critical level,
a problem that arises in ore reserve evaluation,
$
agricultural land management, groundwater con-
Code available from server at http://www.iamg.org/ tamination or polluted soil management.
CGEditor/index.htm.
Corresponding author. Tel.: +56 2 978 4498; Continuous attributes are often modeled by
fax: +56 2 672 3504. nonlinear transforms of Gaussian random fields,
E-mail address: [email protected] (X. Emery). for which the spatial distribution is multivariate

0098-3004/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cageo.2006.03.001
 ARTICLE IN PRESS
1616 X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628

Gaussian (in short, multigaussian). The simulation turning bands method is still scarcely used in
algorithms developed in this respect can be classified geostatistical applications and the existing free
in two families: packages are often limited in their implementation
or applicability. As an example, the first GSLIB
(1) Exact algorithms, for which the statistical version (Deutsch and Journel, 1992) offers a basic
properties of the simulated field match that of program for simulating a regular grid in R3 , based
the desired model. This family comprises the LU on an earlier code by Journel and Huijbregts (1978,
decomposition of the covariance matrix (Davis, pp. 538–545). However, apart from generating non-
1987), convolution methods such as the auto- conditional realizations, this program uses a re-
regressive and moving average models (Black stricted number of lines (15) and allows only
and Freyberg, 1990; Chilès and Delfiner, 1999), isotropic spherical and exponential covariance
and spectral approaches based on discrete and models to be reproduced.
fast Fourier transforms (Pardo-Igúzquiza and A more complete free library is the R package
Chica-Olmo, 1993; Dietrich and Newsam, 1993, RandomFields by Schlather (2001). It uses either a
1996; Wood and Chan, 1994; Chilès and continuous spectral or a circulant-embedding algo-
Delfiner, 1997). The LU algorithm is applicable rithm for line simulation. Also, a large number of
to any configuration of the simulated locations covariance models are available and the realizations
and any covariance model, but CPU time can be made conditional to a set of data, which
becomes prohibitive when more than a few makes the package appealing. However its limita-
thousands locations are considered. Although tions are twofold. The first one concerns the line
computationally efficient, the other aforemen- simulation process: the continuous spectral method
tioned methods can simulate only gridded is not efficient for simulating covariances that are
locations and are not applicable to any covar- not smooth at the origin (Lantuéjoul, 2002, p. 192),
iance model (in particular, the presence of large- while the circulant-embedding algorithm (Dietrich
scale structures may be problematical); also, and Newsam, 1993) requires discretizing each line
memory storage requirements become important into regular intervals and therefore loses accuracy in
when millions of locations have to be simulated. the covariance reproduction (see discussion in
(2) Approximate algorithms, which produce realiza- Section 2.2). Second, several facilities that are
tions of random fields whose spatial distribution helpful for practical applications are not implemen-
is close to multigaussian. In practice, they are ted, such as a back-transformation from normal
often the only alternative when dealing with scores to original variable, the definition of a
huge conditioning datasets or with simulation moving neighborhood for conditioning kriging, or
domains containing millions of nodes. Approx- the use of dual kriging to quicken calculations when
imations take place because of algorithmic working in a unique neighborhood.
simplifications (e.g. neighborhood restrictions) Such limitations have motivated the present
or because the algorithm relies on the central work, which aims at supplying and documenting
limit theorem and convergence to multinorm- computer programs for simulating three-dimen-
ality is only asymptotic. In general, this is a sional Gaussian random fields via the turning bands
minor concern for practical applications, but it method. A second objective is to illustrate the
may be a crucial issue in studies that aim to efficiency and reliability of this method, through an
assess statistical procedures. To date, the most analysis of the bivariate distributions and statistical
widespread approximate algorithms are the fluctuations of the simulated values. The proposed
sequential Gaussian (Deutsch and Journel, programs are based on a former code published by
1992) and continuous spectral methods (Shino- Lantuéjoul (1994, pp. 169–177), which has been
zuka and Jan, 1972). Another option is the completed to include the following features:
turning bands method (Matheron, 1973), which
achieves 2D or 3D simulations via a series of
possibility to handle nested and anisotropic
one-dimensional simulations along lines that covariance models,
sweep R2 or R3 .
possibility to simulate either gridded or scattered
locations on large domains,
Although it is one of the oldest techniques for
accurate reproduction of the desired covariance
simulating multidimensional random fields, the model (without approximation),
 ARTICLE IN PRESS
X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628 1617

availability of the most commonly used covar- that of continuous covariances in R. This property
iance models, allows one to replace the simulation of a multi-

conditioning to a set of existing data, dimensional random field with covariance CY by the

definition of either a unique or a moving simulation of a one-dimensional field with covar-
neighborhood for conditioning kriging, iance CX. The expression of CX as a function of CY

back-transformation from normal values to is often complicated when d ¼ 2 (planar simulation)
original units, (Brooker, 1985; Gneiting, 1998), but it has a simple

change of support (upscaling) of the realizations, solution in three dimensions, which is the frame-

improvements in the simulation code to quicken work considered in this work. By using spherical
calculations. coordinates, formula (2) can be inverted as
d
In the following, the simulation of stationary or C X ðrÞ ¼ ½rC Y ðrÞ, (3)
dr
intrinsic random fields is considered. In the intrinsic
case, the traditional covariance function does not which determines the covariance CX associated with
exist and one works with a generalized covariance; a given isotropic covariance CY in R3 .
also, the multivariate Gaussian assumption does not
concern the random field itself, but its increments or 2.2. Comments on the line simulation process
generalized increments (Dimitrakopoulos, 1990;
Pardo-Igúzquiza and Dowd, 2003). Henceforth the There exists a wide choice of methods for
term ‘covariance’ will stand for either the traditional simulating a one-dimensional random field with
or the generalized covariance; this function is prescribed covariance CX. Several options have been
assumed invariant under translation of the coordi- proposed by Mantoglou and Wilson (1982), Man-
nates, i.e. it only depends on the separation vector toglou (1987), Christakos (1987), Tompson et al.
between the current locations. (1989), Lantuéjoul (1994) and Dietrich (1995),
among others. Exact Gaussian simulation algo-
2. The turning bands algorithm rithms such as the moving average and the circulant
embedding methods are suited to the simulation of
2.1. General presentation regularly spaced locations, which is usually not the
case for the projection of the nodes to be simulated
The principle of the turning bands method is to onto a line with random orientation, even if these
simplify the simulation problem in Rd (with, in nodes are on a regular grid in R3 . A discretization of
general, d ¼ 2 or 3) into a problem in R, which the line at a given mesh and a migration of the
allows one to perform multidimensional simulations projected nodes are therefore required to perform
at the CPU cost of one-dimensional simulations. the one-dimensional simulation. This procedure
Let us consider a random field fX ðtÞ; t 2 Rg with provokes a bias in the reproduction of the
zero mean and continuous covariance function CX, covariance model, as well as difficulties in memory
and a random vector U with a uniform distribution management (and increase in computing time) if the
over the unit sphere Sd of Rd . For any location discretization mesh is very small with respect to the
x 2 Rd , let us put: size of the projected three-dimentional domain.
These impediments can be avoided by resorting to
8x 2 Rd ; Y ðxÞ ¼ X ðox; U4Þ, (1) continuous simulation algorithms along the line. A
whereo,4is the standard inner product in Rd . common example is the continuous spectral meth-
Eq. (1) defines a random field fY ðxÞ; x 2 Rd g with od, which provides an accurate simulation of the
zero mean and isotropic covariance CY such that target covariance without need for regularly spaced
Z locations. This method has been implemented in
C Y ðrÞ ¼ C X ðoh; u4Þ$d ðduÞ; (2) program TBSIM.M for simulating covariance
Sd models that are smooth at the origin. Regarding
where h is a vector of Rd , r its modulus and $d the other models, ad hoc algorithms have been pre-
uniform distribution over Sd . ferred, allowing an efficient and accurate simulation
Matheron (1973, p. 462) proved that relation (2) of the one-dimensional random field at any set of
is a one-to-one mapping between the set of irregularly spaced locations. Note that the use of
continuous and isotropic covariances in Rd and multigaussian simulation algorithms is needless
 ARTICLE IN PRESS
1618 X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628

since multivariate normality is lost anyway in the Let us put

spreading process from R to R3 [Eq. (1)]. Z þ1

3 r 1 r3
C Y ðrÞ ¼ C 1 þ 3 f ðuÞ du (6)
r 2u 2u
2.3. Available covariance models
for a probability density function f(.) in Rþ to be
determined. By differentiating with respect to r and
Fifteen models are implemented in program
after simplification, one finds
TBSIM.M. Their definition and the way to obtain  unu o
the one-dimensional simulations are described here- u
8u40; f ðuÞ ¼ exp  þ 1 . (7)
after. 3 a2 a a
(1) Spherical covariance with sill C and scale One recognizes the probability density function of
factor (range) a: a gamma random variable G with parameters (3,a)

 with probability 1/3 and (2,a) with probability 13.
3 r 1 r3 The simulation of a random field with an exponen-
C Y ðrÞ ¼ C 1  þ 1rpa . (4)
2 a 2 a3 tial covariance can therefore be performed in the
The associated one-dimensional random field following fashion:
fX ðtÞ; t 2 Rg is simulated in the following fashion
(Lantuéjoul, 1994, p. 157):
simulate the random variable G;

simulate a random vector U with a uniform
distribution over the sphere S3 ;

choose an offset on the real axis at random

divide the line with orientation U into intervals
(uniformly in [0,a[) and divide this axis in
with the same length G. Within each interval,
intervals with length a;
draw a linear function with positive or negative

within each interval, draw a linear function with
slope with equal probability (Fig. 1).
an equal probability of being increasing or
decreasing; the slope sign is independent from
(3) Gamma covariance with sill C, scale factor a
one interval to another one (Fig. 1).
and parameter b40:
 r b
(2) Exponential covariance with sill C and scale C Y ðrÞ ¼ C 1 þ . (8)
factor a (practical range 3a): a
 r This model is simulated as a scale mixture of
C Y ðrÞ ¼ C exp  . (5) exponential covariances [Eq. (5)]. The scale factors
a
are the reciprocals of gamma random variables
The one-dimensional random field can be simu- with shape parameter b (Chilès and Delfiner, 1999,
lated thanks to a Poisson point process in R p. 70).
(Lantuéjoul, 1994, p. 156). Here, we use a faster (4) Stable covariance with sill C, scale factor a and
method, consisting in viewing the exponential parameter b A ]0,2]:
covariance as a scale mixture of spherical models.    
r b
C Y ðrÞ ¼ C exp  . (9)
a
When b41, this model is simulated as a scale
mixture of Gaussian covariances (see model type no
6 hereafter); the scale factors are the reciprocals of
square-rooted stable random variables with para-
meter b. Mixing Gaussian covariances is also valid
when bp1, but this approach is not efficient because
the covariance becomes convex at the origin. In this
case, it is preferable to see the stable model as a
mixture of exponential covariances with scale
factors inversely proportional to stable random
variables with parameter b. Note that the case b ¼ 1
Fig. 1. Realization of a 1D random field associated with a 3D corresponds to the exponential model and b ¼ 2 to
spherical model the Gaussian model.
 ARTICLE IN PRESS
X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628 1619

(5) Cubic covariance with sill C and scale factor (9) K-Bessel covariance with sill C, scale factor a
(range) a: and parameter b40:

 C 1b  r b  r 
r2 35 r3 7 r5 3 r7 C Y ðrÞ ¼ 2 Kb . (16)
C Y ðrÞ ¼ C 1  7 2 þ  þ 1rpa . GðbÞ a a
a 4 a3 2 a5 4 a7
(10) This covariance is simulated as a mixture of
Gaussian covariances whose scale factors are
This model is obtained in a way similar to the
square-rooted gamma random variables with shape
spherical covariance, but with a cubic function
parameter b. Indeed, by using the Sonine–Schläfli
instead of a linear function on each interval (Chilès
representation of the K-Bessel function (Matheron,
and Delfiner, 1999, pp. 488, 649).
1965, p. 290), Eq. (16) becomes
(6) Gaussian covariancepffiffiwith
ffi sill C and scale (

Z þ1 2 )
factor a (practical range a 3): C r
    C Y ðrÞ ¼ exp  pffiffiffi expðuÞ ub1 du.
GðbÞ 0 2a u
r 2
C Y ðrÞ ¼ C exp  . (11) (17)
a
This covariance can be simulated by applying the The previous approach is not efficient when the
continuous spectral method, which constitutes a parameter b is less than 1/2, as the covariance is not
particular case of the turning bands method. It smooth at the origin. In this case, it is preferable to
consists in putting view the K-Bessel covariance as a mixture of
pffiffiffiffiffiffiffi exponential models whose scale factors are square-
8x 2 R3 ; Y ðxÞ ¼ 2C cos ð2pox; T4 þ FÞ, (12) rooted beta random variables with parameters
(b,1/2b) (Matheron, 1965, p. 47):
where F is a uniform phase in [0,2p[ and T a
random frequency distributed according to the pffiffiffi Z 1  r
2C p
spectral measure of CY. In this case, T is a Gaussian C Y ðrÞ ¼ exp  u2b1 ð1  u2 Þb1=2 du.
GðbÞGð1=2  bÞ 0 au
vector in R3 .
(18)
(7) Cardinal sine covariance with sill C and scale
factor a:
Also note that the case b ¼ 0:5 corresponds to the
a  r  exponential model.
C Y ðrÞ ¼ C sin . (13) (10) Generalized Cauchy covariance with sill C,
r a
scale factor a and parameter b40:
Again, the continuous spectral method is used for   r 2 b
simulating this covariance. Here, the spectral C Y ðrÞ ¼ C 1 þ . (19)
measure in R3 is uniform on the sphere centered a
on 0 with radius 1/a.
(8) J-Bessel covariance with sill C, scale factor a This model is a scale mixture of Gaussian
and parameter bX0.5: covariances. The scale factors are the reciprocals
 r b  r  of square-rooted gamma random variables with
C Y ðrÞ ¼ C2b Gðb þ 1Þ Jb . (14) shape parameter b.
a a (11) Exponential-sine covariance with sill C and
In R3 , the spectral density of CY is scale factor a:
np  r o
Gðb þ 1Þa3 ð1  a2 jtj2 Þb3=2 C Y ðrÞ ¼ C sin exp  . (20)
8t 2 R3 ; f ðtÞ ¼ 1jtjo1=a . 2 a
Gðb  1=2Þp3=2
By truncating a Gaussian random field with such
(15)
a covariance to the median threshold (y ¼ 0), one
If T is a random vector with this density, then a2 obtains the indicator of a random set with
|T|2 is a beta random variable with parameters proportion 0.5 and exponential covariance (Lantué-
(3/2,b1/2). Note that the cardinal sine model joul, 2002, p. 216). For the one-dimensional
[Eq. (13)] belongs to the J-Bessel family and simulation, the continuous spectral method is used.
corresponds to the particular case b ¼ 1=2. The The spectral density of the exponential-sine covar-
case bo1/2 is not allowed in three dimensions. iance is a mixture of Gaussian densities with
 ARTICLE IN PRESS
1620 X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628

random variances: When b is greater than 1, an adaptation of

Z þ1
2 3=2 the continuous spectral approach is considered,
3 2a u with a one-dimensional random field fX ðtÞ; t 2 Rg of
8t 2 R ; f ðtÞ ¼ expða2 jtj2 uÞhðuÞ du,
0 2p the form
(21) 8t 2 R; X ðtÞ ¼ yðRÞ cos ð2pRt þ FÞ, (26)
where h(.) is a probability density function given by
where F is a uniform phase in [0,2p[ and R is the
a series expansion:
ratio between two independent gamma random
1 X þ1
ð1Þn p2nþ1 variables with shape parameter 0.5. Function y(.) is
8u 2 Rþ ; hðuÞ ¼ pffiffiffiffiffiffi expfð2n þ 1Þ2 ug:
pu n¼0 ð2nÞ! 2 defined by
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
(22) Cp0:5b Gðb þ 2ÞGðbþ3 2  kÞ 1 þ r
8r 2 Rþ ; yðrÞ ¼ 2k3
.
Random variables with density h(.) are simulated 2 Gðk  2ÞGðb þ 2  2kÞ rbþ0:5
b
via an acceptance–rejection method (Lantuéjoul,
(27)
2002, p. 63), by noting that h(.) is an alternate series,
hence less than the term corresponding to n ¼ 0 in Note that Eq. (27) is undefined if b is an even
Eq. (22), which is p/2 times the standard gamma integer. In this particular case, the random field
density with shape parameter 1/2. fX ðtÞ; t 2 Rg is simply a monomial of degree k ¼ b=2.
(12) Linear generalized covariance with slope C (14) Mixed power generalized covariance with
and scale factor a: slope C, scale factor a and parameter bA]0,2]:
  r b 
r C
C Y ðrÞ ¼ C . (23) C Y ðrÞ ¼ 1 . (28)
a lnðr=aÞ a
This covariance is simulated along the real axis by This model is a mixture of power covariances
drawing a random field that increases or decreases whose exponents are random variables uniformly
by unit jumps at regularly spaced discontinuity distributed in ]0,b[.
points (Chilès and Delfiner, 1999, p. 507). Another (15) Spline generalized covariance with slope C,
option is to use a power covariance model (next scale factor a and parameter b ¼ 2k with k 2 N:
type) with exponent b ¼ 1.  r b  r 
(13) Power generalized covariance with slope C, C Y ðrÞ ¼ ð1Þkþ1 C ln . (29)
scale factor a and parameter b40: a a
 r b The associated one-dimensional random field
C Y ðrÞ ¼ ð1Þkþ1 C , (24) fX ðtÞ; t 2 Rg is still given by Eq. (26), except that
a
function y(.) is now defined as
where k is an integer (order of the intrinsic random sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
field) such that 2kobp2k+2. This model is Cp1b Gðb þ 2Þ 1 þ r
invariant by the turning bands operator [Eq. (3)], 8r 2 Rþ ; yðrÞ ¼ . (30)
2b1 rbþ0:5
i.e. the corresponding covariance CX is also a power
model with the same exponent, but not the same This expression can be obtained by considering
slope. When b is less than or equal to 1, a locally the spline covariance with parameter b as the
stationary representation of the power covariance derivative of the power covariance given in
can be simulated continuously along the real axis, Eq. (24) with respect to the exponent b (Chilès
by means of a mixture of stationary triangular and Delfiner, 1999, p. 264).
covariances:
 r b Z 1  r 2.4. Ergodicity and multivariate normality
8r 2 ½0; a; 1  ¼ 1 F ðduÞ, (25)
a r=a au
Formula (1) is not fully satisfactory as it leads to
with 8u 2 ½0; 1½; F ðuÞ ¼ ð1  bÞub and F ð1Þ ¼ 1. The non-ergodic three-dimensional random fields: each
triangular model is simulated by dividing the real realization exhibits a zonal anisotropy oriented
axis into regular intervals and assigning to each along vector U, so that its experimental covariance
interval a constant value drawn at random from a does not match the theoretical isotropic model
given distribution, independently from one interval (Chilès and Delfiner, 1999, p. 474). To solve this
to another one. issue, one has to repeat the spreading process many
 ARTICLE IN PRESS
X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628 1621

times, say N, and put 2.6. How should the lines be distributed on the
sphere?
1 X N
Y ðxÞ ¼ pffiffiffiffiffi X i ðox; Ui 4Þ; (31)
N i¼1 So far, the presentation assumed that the lines
had a uniform distribution over the sphere S3 .
where {Ui, i ¼ 1... N} are independent directions
However, Eq. (32) ensures that the covariance is
randomly distributed over the sphere S3 and {Xi,
also honored if the line directions form an
i ¼ 1... N} are independent one-dimensional ran-
equidistributed (not necessarily uniform) sequence
dom fields, each of them with covariance CX(.)
over S3 , i.e. if the proportion of directions falling in
fulfilling Eq. (3). Given the realization {Ui ¼ ui,
a given sub-domain B  S3 converges to the
i ¼ 1... N}, the covariance of fY ðxÞ; x 2 R3 g for a
fraction of S3 covered by B when the number of
separation vector h with modulus r is
directions becomes infinite. In practice, the conver-
1X N
gence is faster when the lines are regularly
C Y ðrÞ ¼ C X ðoh; ui 4Þ. (32) distributed, as it amounts to calculating the integral
N i¼1
in Eq. (2) by a regular discretization of the sphere
This constitutes a discrete approximation of (Chilès and Delfiner, 1999, p. 476).
Eq. (2): the theoretical covariance model (an The maximal number of regular directions in the
integral over the unit sphere S3 ) is reproduced three-dimensional space is equal to 15. This number
experimentally by Monte Carlo integration. Eq. (31) relies on the icosahedron approximation (Journel
explains the name of the method, as the simulated and Huijbregts, 1978, p. 503) and has led several
random field is the sum of several random fields authors to propose the use of 15 regular lines for the
generated along lines whose spatial orientation turning bands algorithm. Actually, the lines do not
scans the unit sphere of R3 . need to be regularly distributed (the previous
Provided that the simulation algorithm used presentation assumes that U is a uniform random
along the lines is accurate, the turning bands vector over the sphere), so a number greater than 15
method reproduces the theoretical covariance model is by no means forbidden.
without bias, irrespective of the number N of lines Freulon and de Fouquet (1991) examined several
chosen by the user. Even so, it belongs to the class sequences of line directions and showed that the
of approximate methods because the spatial dis- ergodic properties of the simulation substantially
tribution of the simulated random field (or, in the improved by choosing the directions from a van der
intrinsic case, of its increments) is neither ergodic
nor multigaussian. Such properties are only ensured
asymptotically, i.e. if N is infinitely large, by virtue
of the central limit theorem.

2.5. Simulation of nested and anisotropic models

A Gaussian random field with nested covariance

model can be simulated as the sum of independent
Gaussian random fields, each of them associated
with one basic structure of the nested model.
Program TBSIM.M also allows the simulation of
covariances with a geometric anisotropy, which can
be transformed into isotropic covariances via a
rotation and rescaling of the coordinate axes. For
the one-dimensional simulation, a different scale
factor is assigned to each line, according to its
orientation on the sphere S3 . Extensions of the
turning bands method to the simulation of
more general anisotropic models (not implemented
in TBSIM.M) can be found in Lantuéjoul (2002, Fig. 2. A set of 300 equidistributed directions over the unit
p. 192). sphere of R3 (van der Corput sequence).
 ARTICLE IN PRESS
1622 X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628

Corput sequence (Fig. 2) instead of uniform dimensions, the 15 regular lines are clearly insuffi-
directions. The construction of such a sequence is cient and several hundreds of lines should be used,
explained in Lantuéjoul (1994, p. 164) and has been say N ¼ 1000, as confirmed by the experience of
implemented in subroutine VDC.M. The directions several authors (Tompson et al., 1989; Freulon
are defined up to a random rotation in order to and de Fouquet, 1991). Lantuéjoul (2002, p. 198)
avoid repeating the same set of directions for all the showed that hundreds of lines are necessary to make
realizations. sure that the moments up to order 4 of any weighted
average of the simulated values are close enough to
2.7. How many lines should be used? the expected Gaussian moments.
Actually, the number of lines is not the limiting
Several factors should be considered when factor of the turning bands method, as the simula-
choosing the number of lines to use for simulating tion along the lines is very fast. In general, a greater
each structure of the nested model: CPU time is devoted to post-process the realiza-
tions, in particular concerning the conditioning to a
(1) The distribution of the lines over the sphere set of existing data (see next section).
(almost regular or purely random, see previous
discussion). 3. Conditioning the realizations to Gaussian data
(2) The relative importance of the basic structure in
the covariance model: the greater the contribu- To facilitate practical applications, program
tion of the structure (fraction of the total sill), TBSIM.M allows conditioning of the realizations
the more lines. to a set of normal data. Non-conditional realiza-
(3) The criterion used to decide whether the multi- tions are converted into conditional realizations
gaussian model is well reproduced or not. through a kriging step (Journel and Huijbregts,
Lantuéjoul (1994, p. 159) gives clues on how 1978, p. 495). Let fY S ðxÞ; x 2 R3 g be a non-
to quantify the closeness of the simulated conditional simulation. In the stationary case (finite
spatial distribution to a multivariate Gaussian variance models), the random field defined by
distribution.
(4) The type of covariance and the associated 8x 2 R3 ; Y CS ðxÞ ¼ Y S ðxÞ þ ½Y ðxÞ  Y S ðxÞSK (33)
simulation technique along the lines

The spectral method is suitable for simulating constitutes a conditional simulation, i.e. it repro-
covariance models whose spectral densities duces the distribution of fY ðxÞ; x 2 R3 g conditional
rapidly vanish at infinity (e.g. a Gaussian to the Y-data. In formula (33), the superscript SK
covariance); few lines are then needed to stands for a simple kriging (with mean zero and
sample correctly such densities. covariance CY) from the values at the data

Concerning the simulation methods based on locations. To summarize, the conditional simulation
a division of the lines into intervals (e.g. for is constructed as follows:
simulating the spherical and exponential
covariances), more lines should be used if (1) draw a non-conditional realization at the target
the scale factor increases. Indeed, the average location x and at the data locations, via the
number of discontinuities (Fig. 1) within a turning bands method;
given segment [x,x0 ] of R3 is proportional to (2) compute the deviations (residuals) between the
the number of lines and inversely propor- data values and simulated values at the data
tional to the scale factor. If they are not locations;
numerous enough, these discontinuities may (3) perform a simple kriging of the residual from its
not compensate and provoke artifacts in the values at the data locations
simulated maps. (4) add the result to the non-conditional realization,
in accordance with Eq. (33).
In two dimensions, Mantoglou and Wilson (1982)
suggest using 4 to 16 lines, Chilès (1977) recom- Although this approach is less straightforward
mends the use of 180 lines (one per degree of angle), than the direct conditioning performed by the
while Gneiting (1999) has found that 64 lines lead to sequential Gaussian method, its advantages are
an excellent covariance reproduction. In three threefold (Emery, 2004, p. 413). First, the kriging
 ARTICLE IN PRESS
X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628 1623

weights needed in step (3) are the same when several and zmax of the original attribute:
realizations are required, hence only one kriging is
necessary to make all the realizations honor the (
8y 2  1; y1 ; z ¼ zmin þ b0 expðlyÞ;
data. Second, the kriging matrix only involves the (35)
8y 2 ½yp ; þ1½; z ¼ zmax þ bp expðl0 yÞ:
original conditioning data, not the simulated values,
therefore it is smaller than in the sequential
algorithm and its inversion is faster. The dual (
b0 ¼ ðz1  zmin Þ expðly1 Þ
formulation of kriging advantageously reduces the with .
amount of calculations when a unique neighbor- bp ¼ ðzp  zmax Þ expðl0 yp Þ
hood is used. Third, if the simulation is performed To quicken the algorithm, grid simulation is
on a fine grid, numerical instabilities in the kriging performed row-wise. At each step, one or several
systems are avoided. rows of the 3D grid are projected onto the lines and
In the intrinsic case (infinite variance models), their values are simulated, conditioned to the data,
conditioning is ensured by substituting intrinsic back-transformed and written in the output file; the
kriging for simple kriging in Eq. (33) (Dimitrako- maximum number of nodes to consider at each step
poulos, 1990, p. 369). The other steps of the is defined by the user. This way, the 3D grid is not
conditioning process are unchanged. loaded in the Matlab workspace and the program is
able to deal with large simulation domains. Two
options are available for the conditioning kriging
4. Program description step: if the search radius is set to 1, a dual kriging
with a unique neighborhood is used, otherwise a
The proposed programs are ASCII files written in conventional kriging is performed in a moving
MATLAB language (version 5.0 or later). The main neighborhood, in association with a super-block
routine, called TBSIM.M, allows one to draw search strategy (Deutsch and Journel, 1992, p. 31).
several realizations of a Gaussian random field over The reader is referred to the header of the
a regular grid or over scattered locations in R3 program file for a description of the input para-
condition these realizations to Gaussian data, back- meters of TBSIM.M. No output variable is gener-
transform the values from the Gaussian scale to that ated in the Matlab workspace. Instead, an external
of the original attribute, and regularize the simu- ASCII file is created, which contains one realization
lated values to a block support. The covariance per column. For grid simulations, the ordering of
model can consist of several nested structures (see the nodes follows the GSLIB conventions (Deutsch
Section 2.3 for a list of the available structure and Journel, 1992): point by point to the east, then
types), each of them with a possible geometric or row by row to the north, and finally level by level
zonal anisotropy. The nugget effect is treated apart upward.
as it amounts to adding a white noise to the Alternatively, program TBSIM.M can be used
simulated random field. with a parameter file: in this case, no input
Regarding the back-transformation from Gaus- argument is required and the user is prompted for
sian values (y) to original values (z), a piecewise the name of the parameter file. If no name is
interpolation is performed according to a conver- entered, a default file (tbsim.par) is assumed. If the
sion table fðzi ; yi Þ; i ¼ 1 . . . pg specified by the user: file does not exist, a blank file tbsim.par is created.
When resorting to the parameter file mode, the
8i 2 f1; :::p  1g 8y 2 ½yi ; yiþ1 ; z ¼ ai þ bi y, (34) coordinates of the locations to be simulated (if these
locations are not gridded), the data coordinates and
data values must be in external ASCII files without
with headers. TBSIM.M can also be compiled and
converted to a stand-alone executable that works
ziþ1  zi
bi ¼ and ai ¼ z i  bi yi . like a GSLIB program.
yiþ1  yi The generation of random variables requires the
availability of the Matlab Statistics Toolbox (routines
For tail extrapolation, an exponential function is for simulating uniform, normal, gamma and beta
used, which requires the definition of two positive random variables). TBSIM.M uses twelve additional
scalars l and l0 , as well as the extreme values zmin subroutines, the description of which can be found in
 ARTICLE IN PRESS
1624 X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628

the corresponding program files: BACKTR, COVA, As stated earlier, the use of 15 lines is clearly
CREATE_PARAMFILE, DUAL, KRIGE, PICK- insufficient and leads to artifacts (striping) in the
SUPR, SEARCH, SETDUAL, SETROT, SU- simulated map. However, such artifacts are no
PERBLK, TBMAIN and VDC. longer perceptible when using 1000 lines. In
practice, to choose a suitable number of lines, users
5. Investigating the quality of the simulation can construct a few non-conditional realizations
program: example of a spherical covariance model and check whether their ‘texture’ is acceptable or
not. Based on the current example and on the
This section examines the quality of the realiza- comments in Section 2.7, the following heuristic rule
tions obtained by program TBSIM.M, through a is deemed appropriate for simulating a spherical
particular example consisting of a non-conditional covariance model with relative sill C (sill of the
simulation over a domain of size 1024  1024  1. spherical model divided by the total sill):
The covariance model is composed of an isotropic  
au
spherical model with range 50 and sill 0.9 plus a N  20C max , (36)
u2S3 meshðuÞ
nugget effect with sill 0.1. In the following, three
kinds of checks are performed: where au is the range along the direction spanned by
vector u and mesh(u) the minimal distance between

a visual appreciation of the artifact banding due consecutive nodes (or locations to be simulated)
to the use of a finite number of lines; along that direction. This formula can also be

an analysis of the statistics of the simulated applied to the exponential and cubic models, the
univariate and bivariate distributions; construction of which is close to that of the

an inspection of the fluctuations of the first spherical model.
moments over a set of realizations.
5.2. Reproduction of the univariate and bivariate
distributions
5.1. Artifact banding
To validate the quality of the algorithm beyond a
Fig. 3 presents two realizations in grayscale visual check of the simulated maps, a set of 100 non-
representations, the first one obtained by using 15 conditional realizations is drawn with the same
lines, the second one corresponding to 1000 lines. model and same domain as in the previous

Fig. 3. Two realizations of random fields simulated with the turning bands algorithm by using 15 and 1000 lines, respectively. The
covariance consists of an isotropic spherical model plus a nugget effect.
 ARTICLE IN PRESS
X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628 1625

Fig. 4. Validation of the reproduction of the univariate and bivariate distributions, from 100 non-conditional realizations obtained by
using 1000 lines. The covariance consists of an isotropic spherical model plus a nugget effect.

subsection, by using 1000 lines for each realization.
median indicator variogram:
The simulated univariate distributions are then pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1=p arcsin 1  rðhÞ=2.
compared to the expected standard Gaussian
distribution via a quantile–quantile plot. Also, the
bivariate distributions are examined through three In each case, one observes that the average
parameters: the simulated variograms, madograms simulated statistics match the theoretical model
and median indicator variograms (Fig. 4). For a almost perfectly. These tests are easy to perform and
standard Gaussian random field with correlogram helpful for investigating the quality of a simulation
r(h), one has the following relationships (Lantué- algorithm, see for instance Emery (2004) for an
joul, 2002, p. 208; Emery, 2005a, p. 166): application to the sequential Gaussian algorithm.

variogram: 1r(h); 5.3. Ergodic fluctuations

madogram (first-order variogram):
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Another way to validate a simulation algorithm is
1  rðhÞ=p; to examine the ergodic fluctuations in the simulated
 ARTICLE IN PRESS
1626 X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628

statistics. The amplitude of such fluctuations In log–log coordinates, the points plotting the
depends on the size of the simulated domain and dispersion variance of W within V as a function of
on the spatial distribution of the random field (in P1 should be asymptotically aligned with slope 1
the present case, a multigaussian distribution). In when P decreases. In practice the dispersion
the following subsections, we focus on the fluctua- variance is estimated as the average of the experi-
tions of the spatial averages and of the simulated mental variances of the sub-domains over the 100
variograms. realizations. In Fig. 5A, the test is applied to two
types of sub-domains: squares and stripes with
length 1024. In both cases, the points are almost
5.3.1. Fluctuations of the spatial averages over a set
aligned along the theoretical asymptote for small
of increasing sub-domains
values of P, which corroborates the consistency of
Let V be the entire simulated domain. Suppose
the observed fluctuations with the multigaussian
that it can be partitioned into P sub-domains
model.
fW p ; p ¼ 1 . . . Pg with the same shape, size and
orientation as a reference domain W. Let us
consider the dispersion variance D2(W|V) of the 5.3.2. Fluctuations of the regional variogram
average value of the simulated random field over Let GV(h) be the probabilistic version of the
these sub-domains. By applying Krige’s relation- regional variogram over domain V and g(h) the
ship, the following approximate identity is estab- theoretical variogram model. Then one has (Math-
lished (Matheron, 1989, p. 85): eron, 1989, p. 81):


2 1 1 var½GV ðhÞ
D ðW jV Þ  A  , (37) Z
jW j jV j 1
¼ ½gðu þ hÞ þ gðu  hÞ  2gðuÞ2 K h ðuÞdu,
where |V| and |W| are the volumes of V and W, and 2K 2h ð0Þ
A is the integral range of the correlogram model (in ð39Þ
the present case, A ¼ 1413:7). Eq. (37) holds
provided that A is finite and non-zero and that where Kh(.) is the geometric covariogram of the
|W| is much greater than A (Lantuéjoul, 1991). It intersection of V and V shifted by h.
can be re-written as follows: The regional variogram GV(h) is the average of
many identically distributed random variables (the
A squared differences between pairs of simulated
D2 ðW jV Þ  ðP  1Þ. (38)
jV j values). For lag distances |h| much smaller than

Fig. 5. Validation of the fluctuations for (A) the spatial averages over a set of increasing sub-domains, and (B) the simulated regional
variograms.
 ARTICLE IN PRESS
X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628 1627

the diameter of the simulated domain V (say, of continuous attributes represented by Gaussian-
|h|o100 in the case under study), these variables related models (Lantuéjoul, 2002; Emery, 2005b).
are not independent but are located in an area that
is several hundreds times the integral range of the Acknowledgments
covariance model. By assuming a mixing property
in that area, the central limit theorem states that the The authors are grateful to two anonymous
distribution of GV(h) at small distances is almost reviewers for their valuable comments on an earlier
Gaussian, with mean g(h) and variance given by version of this paper.
Eq. (39). Consequently, the distribution of the
dispersion of the regional variogram over a set of
Appendix A. Supplementary materials
n realizations is proportional to a w2 random
variable with n1 degrees of freedom (in the present
Supplementary data associated with this article
case n  1 ¼ 99). This result allows one to define a
can be found in the online version at doi:10.1016/
95% probability interval around the expected
j.cageo.2006.03.001
variance [Eq. (39)] and check that the observed
dispersion of the simulated regional variograms lies
within such an interval (Fig. 5B).
Lantuéjoul (1994) showed that the turning bands
References
method produces fluctuations that depart from the
theoretical expression given in Eq. (39). However, Black, T.C., Freyberg, D.L., 1990. Simulation of one-dimen-
the departure tends to zero when the number of sional correlated fields using a matrix-factorization moving
lines increases, so it can be controlled in order to be average approach. Mathematical Geology 22 (1), 39–62.
smaller than a given accuracy level. When using one Brooker, P.I., 1985. Two-dimensional simulation by turning
thousand of lines as in this example, the bias is bands. Mathematical Geology 17 (1), 81–90.
Chilès, J.P., 1977. Géostatistique des phénomènes non station-
negligible. naires (dans le plan). Doctoral Thesis, Université de Nancy I,
Nancy, 152pp.
Chilès, J.P., Delfiner, P., 1997. Discrete exact simulation by the
6. Conclusions Fourier method. In: Baafi, E.Y., Schofield, N.A. (Eds.),
Geostatistics Wollongong ‘96. Kluwer Academic, Dordrecht,
pp. 258–269.
Thirty years after being developed, the turning Chilès, J.P., Delfiner, P., 1999. Geostatistics: Modeling Spatial
bands approach is still scarcely used in geostatistical Uncertainty. Wiley, New York 695pp.
applications, although it is more efficient than most Christakos, G., 1987. Stochastic simulation of spatially correlated
other methods for simulating Gaussian random geo-processes. Mathematical Geology 19 (8), 807–831.
Davis, M.W., 1987. Production of conditional simulations via the
fields. The non-conditional simulation step is gen-
LU triangular decomposition of the covariance matrix.
erally very fast as it achieves the three-dimensional Mathematical Geology 19 (2), 91–98.
simulation through a series of one-dimensional Deutsch, C.V., Journel, A.G., 1992. GSLIB: Geostatistical
simulations. Furthermore, a single kriging run is Software Library and User’s Guide. Oxford University Press,
required to make all the realizations conditional to a New York 340pp.
set of data. By setting adequately the parameters of Dietrich, C.R., 1995. A simple and efficient space domain
implementation of the turning bands method. Water Re-
the algorithm (in particular, by using several sources Research 31 (1), 147–156.
hundreds or thousands of lines), the artifact Dietrich, C.R., Newsam, G.N., 1993. A fast and exact method for
banding is almost imperceptible in the realization multidimensional Gaussian stochastic simulations. Water
maps, whereas the simulated statistics of the Resources Research 29 (8), 2861–2869.
univariate and bivariate distributions and the Dietrich, C.R., Newsam, G.N., 1996. A fast and exact method for
multidimensional Gaussian stochastic simulations: extension
associated statistical fluctuations are consistent with to realizations conditioned on direct and indirect measure-
the theoretical multigaussian model. ments. Water Resources Research 32 (6), 1643–1652.
The proposed computer programs allow one to Dimitrakopoulos, R., 1990. Conditional simulation of intrinsic
simulate continuous attributes over large domains random functions of order k. Mathematical Geology 22 (3),
361–380.
and can handle nested and anisotropic covariance
Emery, X., 2004. Testing the correctness of the sequential
models. They may also be a basis for the simulation algorithm for simulating Gaussian random fields. Stochastic
of categorical attributes through the truncated Environmental Research and Risk Assessment 18 (6),
Gaussian and plurigaussian algorithms, or for that 401–413.
 ARTICLE IN PRESS
1628 X. Emery, C. Lantuéjoul / Computers & Geosciences 32 (2006) 1615–1628

Emery, X., 2005a. Variograms of order o: a tool to validate a Mantoglou, A., Wilson, J.L., 1982. The turning bands method
bivariate distribution model. Mathematical Geology 37 (2), for simulation of random fields using line generation by a
163–181. spectral method. Water Resources Research 18 (5),
Emery, X., 2005b. Conditional simulation of random fields with 1379–1394.
bivariate gamma isofactorial distributions. Mathematical Matheron, G., 1965. Les variables régionalisées et leur estima-
Geology 37 (4), 419–445. tion. Masson, Paris 305pp.
Freulon, X., de Fouquet, C., 1991. Remarques sur la pratique des Matheron, G., 1973. The intrinsic random functions and
bandes tournantes à trois dimensions. In: Cahiers de their applications. Advances in Applied Probability 5,
géostatistique, Fascicule 1, Centre de Géostatistique, Ecole 439–468.
des Mines de Paris, Fontainebleau, pp. 101–117. Matheron, G., 1989. Estimating and Choosing. Springer, Berlin
Gneiting, T., 1998. Closed form solutions of the two-dimensional 141pp.
turning bands equation. Mathematical Geology 30 (4), 379–390. Pardo-Igúzquiza, E., Chica-Olmo, M., 1993. The Fourier integral
Gneiting, T., 1999. The correlation bias for two-dimensional method—an efficient spectral method for simulation of
simulations by turning bands. Mathematical Geology 31 (2), random fields. Mathematical Geology 25 (2), 177–217.
195–211. Pardo-Igúzquiza, E., Dowd, P.A., 2003. IRFK2D: a computer
Journel, A.G., Huijbregts, C.J., 1978. Mining Geostatistics. program for simulating intrinsic random functions of order k.
Academic Press, London 600pp. Computers & Geosciences 29 (6), 753–759.
Lantuéjoul, C., 1991. Ergodicity and integral range. Journal of Schlather, M., 2001. Simulation and analysis of random fields. R
Microscopy 161 (3), 387–404. News 1 (2), 18–20.
Lantuéjoul, C., 1994. Non conditional simulation of stationary Shinozuka, M., Jan, C.M., 1972. Digital simulation of random
isotropic multigaussian random functions. In: Armstrong, processes and its applications. Journal of Sound and
M., Dowd, P.A. (Eds.), Geostatistical Simulations. Kluwer Vibrations 25 (1), 111–128.
Academic, Dordrecht, pp. 147–177. Tompson, A.F.B., Ababou, R., Gelhar, L.W., 1989. Implementa-
Lantuéjoul, C., 2002. Geostatistical Simulation, Models and tion of the three-dimensional turning bands random field
Algorithms. Springer, Berlin 256pp. generator. Water Resources Research 25 (8), 2227–2243.
Mantoglou, A., 1987. Digital simulation of multivariate two- and Wood, A.T.A., Chan, G., 1994. Simulation of stationary
three-dimensional stochastic processes with a spectral turning Gaussian processes in [0,1]d. Journal of Computational and
bands method. Mathematical Geology 19 (2), 129–149. Graphical Statistics 3 (4), 409–432.