\name{GaussRF}
\alias{GaussRF}
\alias{InitGaussRF}
\title{Gaussian Random Fields}
\description{
These functions simulate stationary spatial and spatio-temporal
Gaussian random fields using turning bands/layers, circulant embedding,
direct methods, and the random coin method.
}
\usage{GaussRF(x, y=NULL, z=NULL, T=NULL, grid=!missing(gridtriple), model, param,
trend=NULL, method=NULL, n=1, register=0, gridtriple,
paired=FALSE, ...)
InitGaussRF(x, y=NULL, z=NULL, T=NULL, grid=!missing(gridtriple), model,
param, trend=NULL, method=NULL, register=0, gridtriple)
}
\arguments{
\item{x}{matrix of coordinates, or vector of x coordinates}
\item{y}{vector of y coordinates}
\item{z}{vector of z coordinates}
\item{T}{vector of time coordinates, may only be given if
the random field is defined as an anisotropic random field,
i.e. if \code{model=list(list(model=,var=,k=,aniso=),...)}.
\code{T} must always be given in the \code{gridtriple} format,
independently how the spatial part is defined.
}
\item{grid}{logical; determines whether the vectors \code{x},
\code{y}, and \code{z} should be
interpreted as a grid definition, see Details. \code{grid}
does not apply for \code{T}.}
\item{model}{string or list; covariance or variogram model,
see \command{\link{CovarianceFct}}, or
type \command{\link{PrintModelList}}\code{()} to get the list of all implemented
models; see Details.}
\item{param}{vector or matrix of parameters or missing, see Details
and \command{\link{CovarianceFct}};
The simplest form is that \code{param} is vector of the form
\code{param=c(NA,variance,nugget,scale,...)}, in this order;\cr
The dots \code{...} stand for additional parameters of the
model.}
\item{trend}{trend surface: number (mean) or a vector of length
\eqn{d+1} (linear trend \eqn{a_0+a_1 x_1 + \ldots + a_d x_d}{
a_0 +a_1 x_1 + ... + a_d x_d}), or function; you have the choice
of using either x, y, z or X1, X2, X3, ... as spatial variables, as
time variable T should be chosen}
\item{method}{\code{NULL} or string; method used for simulating,
see \command{\link{RFMethods}}, or
type \command{\link{PrintMethodList}}\code{()} to get all options.
If \code{model} is given as list then \code{method} may not be
set if \code{model[[i]]$method}, \eqn{i=1,3,..} is given, and vice
versa. However, a global parameter \code{method} and
specific \code{method}s may be given, e.g.
\code{list(list(model=..., method="TBM3"), ..., method="ci")};
then the specific ones overwrite the global \code{method}.
}
\item{n}{number of realisations to generate}
\item{register}{0:9; place where intermediate calculations are stored;
the numbers are aliases for 10 internal registers}
\item{gridtriple}{logical. Only relevant if \code{grid=TRUE}.
If \code{gridtriple=TRUE}
then \code{x}, \code{y}, and \code{z} are of the
form \code{c(start,end,step)}; if
\code{gridtriple=FALSE} then \code{x}, \code{y}, and \code{z}
must be vectors of ascending values
}
\item{paired}{
logical. If \code{TRUE} then the second half of the simulations is
obtained by
only changing the signs of all the standard Gaussian random variables,
on which the first half of the
simulations is based. (\dQuote{Antithetic pairs}.)
}
\item{...}{\command{\link{RFparameters}} that are locally used only.}
}
\details{
\code{GaussRF} can use different methods for the simulation,
i.e., circulant embedding, turning bands, direct methods, and random
coin method.
If \code{method=NULL} then \code{GaussRF} searches for a
valid method. \code{GaussRF} may not find the fastest method neither the
most precise one. It just finds any method among the available
methods. (However it guesses what is a good choice.) See
\link{RFMethods} for further information.
Note that some of the methods do not work for all covariance
or variogram models.
\itemize{
\item
An isotropic random field is created by
\code{GaussRF} where
\code{model} is the covariance or variogram model and the
parameter is \code{param=c(mean,variance,nugget,scale, ...)}.
Alternatively the \code{trend} can be given; then
\code{param=c(variance,nugget,scale, ...)}.
\item
Nested models can be defined in the same way as a nested
\command{\link{CovarianceFct}}. If the \code{trend} is not given
it is set to 0.
\item
An anisotropic random field (i.e. zonal anisotropy, geometrical
anisotropy, separable models, non-separable space-time models)
and a random field based on multiplicative or nested models
is defined as in the case of
an anisotropic \command{\link{CovarianceFct}}.
If the \code{trend} is not given it is set to 0.
The \code{method} may be specified by the global \code{method}
or for each model separately, as additional parameter
\code{method} for each entry of the list;
note that methods can not be mixed within a multiplicative part.
If \code{model=list(list(model=,var=,k=,aniso=),...)} then a time
component might be given. In case of \code{model="nugget"},
\code{aniso} must still be given as a matrix. Namely if
\code{aniso} is a singular matrix then a zonal nugget effect
is obtained.
}
\code{GaussRF} calls initially \code{InitGaussRF},
which does some basic checks on the validity of the parameters. Then,
\code{InitGaussRF} performs some first calculations, like the first
Fourier transform in the circulant embedding method or the matrix
decomposition for the direct methods. Random numbers are not involved.
\code{GaussRF} then calls \command{\link{DoSimulateRF}} which uses the
intermediate results and random numbers to create a simulation.
When \code{InitGaussRF} checks the validity of the parameters, it
also checks whether the previous simulation has had the same
specification of the random field. If so (and if
\command{\link{RFparameters}}\code{()$STORING==TRUE}), the stored intermediate
results are used instead of being recalculated.
Comments on specific parameters:
\itemize{
\item \code{grid=FALSE} : the vectors \code{x}, \code{y},
and \code{z} are interpreted as vectors of coordinates
\item \code{(grid=TRUE) && (gridtriple=FALSE)} : the vectors
\code{x}, \code{y}, and \code{z}
are increasing sequences with identical lags for each sequence.
A corresponding
grid is created (as given by \code{expand.grid}).
\item \code{(grid=TRUE) && (gridtriple=TRUE)} : the vectors
\code{x}, \code{y}, and \code{z}
are triples of the form (start,end,step) defining a grid
(as given by \code{expand.grid(seq(x$start,x$end,x$step),
seq(y$start,y$end,y$step),
seq(z$start,z$end,z$step))})
%\item \code{model="nugget"} is one possibility to create independent
%Gaussian random variables. Without loss of efficiency, any
%covariance function with parameter vector
%\code{c(mean, 0, nugget, scale, ...)}
%can also be used. If \code{model="nugget"} is used
%the second component of \code{param} must be zero.
%% this has to be changed in later versions;
%\item The sum of the components variance and nugget in the argument
%\code{param} equals the sill of the variogram.
\item \code{register} is a parameter which may never be
used by most of the users (please let me know if you use it!). In
other words,
the package will work fine if you ignore this parameter.
The parameter \code{register} is of interest in the following
situation. Assume you wish to create sequentially
several realisations of two random fields \eqn{Z_1}{Z1} and
\eqn{Z_2}{Z2} that have different
specifications of the covariance/variogram models, i.e.
\eqn{Z_1}{Z1}, \eqn{Z_2}{Z2}, \eqn{Z_1}{Z1}, \eqn{Z_2}{Z2},...
Then, without using different registers, the algorithm
will not be able to profit from already calculated intermediate
results, as the specifications of the covariance/variogram model
change every time.
However, using different registers allows for profiting from
up to 10 stored intermediate results.
\item The strings for \code{model} and \code{method} may
be abbreviated as long as the abbreviations match only one
option. See also \command{\link{PrintModelList}}\code{()} and
\command{\link{PrintMethodList}}\code{()}
\item Further control parameters for the simulation are set by means of
\command{\link{RFparameters}}\code{(...)}.
}
}
\note{
The algorithms for all the simulation methods are controlled by
additional parameters, see \command{\link{RFparameters}}\code{()}. These
parameters have an influence on the speed of the algorithm
and the precision of the result.
The default parameters are chosen such that
the simulations are fine for many models and their parameters.
If in doubt modify the example in \command{\link{EmpiricalVariogram}}\code{()}
to check the precision.
}
\value{
\code{InitGaussRF} returns 0 if no error has occurred and a positive value
if failed.\cr
The object returned \code{GaussRF} and \command{\link{DoSimulateRF}}
depends on the parameters \code{n} and \code{grid}:\cr
if \code{vdim > 1} the \code{vdim}-variate vector makes the first dimension
if \code{grid=TRUE} an array of the dimension of the
random field makes the next dimensions.
Else if no time component is given, then the values are passed as a
single vector. Else if the time component is given the next 2
dimensions give space and time.
if \code{n > 1} the repetitions make the last dimension
}
\references{
See \link{RFMethods} for the references.
}
\author{Martin Schlather, \email{martin.schlather@math.uni-goettingen.de}
\url{http://www.stochastik.math.uni-goettingen.de/~schlather}
Yindeng Jiang \email{jiangyindeng@gmail.com} (circulant embedding
methods \sQuote{cutoff} and \sQuote{intrinsic})}
\seealso{
\command{\link{Covariance}},
\command{\link{CovarianceFct}},
\command{\link{DeleteRegister}},
\command{\link{DoSimulateRF}},
\command{\link{GetPracticalRange}},
\command{\link{EmpiricalVariogram}},
\command{\link{fitvario}},
\command{\link{MaxStableRF}},
\command{\link{RFMethods}},
\code{\link{RandomFields}},
\command{\link{RFparameters}},
\command{\link{ShowModels}},
% \command{\link{winddata}}.
}
\examples{
% library(RandomFields, lib="~/TMP"); RFparameters(Print=6)
#############################################################
## ##
## Examples using the symmetric stable model, also called ##
## "powered exponential model" ##
## ##
#############################################################
PrintModelList() ## the complete list of implemented models
model <- "stable"
mean <- 0
variance <- 4
nugget <- 1
scale <- 10
alpha <- 1 ## see help("CovarianceFct") for additional
## parameters of the covariance functions
step <- 1 ## nicer, but also time consuming if step <- 0.1
x <- seq(0, 20, step)
y <- seq(0, 20, step)
f <- GaussRF(x=x, y=y, model=model, grid=TRUE,
param=c(mean, variance, nugget, scale, alpha))
image(x, y, f)
#############################################################
## ... using gridtriple
step <- 1 ## nicer, but also time consuming if step <- 0.1
x <- c(0, 20, step) ## note: vectors of three values, not a
y <- c(0, 20, step) ## sequence
f <- GaussRF(grid=TRUE, gridtriple=TRUE,
x=x ,y=y, model=model,
param=c(mean, variance, nugget, scale, alpha))
image(seq(x[1],x[2],x[3]), seq(y[1],y[2],y[3]), f)
#############################################################
## arbitrary points
x <- runif(100, max=20)
y <- runif(100, max=20)
z <- runif(100, max=20) # 100 points in 3 dimensional space
(f <- GaussRF(grid=FALSE, Print=5,
x=x, y=y, z=z, model=model,
param=c(mean, variance, nugget, scale, alpha)))
#############################################################
## usage of a specific method
## -- the complete list can be obtained by PrintMethodList()
x <- runif(100, max=20) # arbitrary points
y <- runif(100, max=20)
(f <- GaussRF(method="dir", # direct matrix decomposition
x=x, y=y, model=model, grid=FALSE,
param=c(mean, variance, nugget, scale, alpha)))
#############################################################
## simulating several random fields at once
step <- 1 ## nicer, but also time consuming if step <- 0.1
x <- seq(0, 20, step) # grid
y <- seq(0, 20, step)
f <- GaussRF(n=3, # three simulations at once
x=x, y=y, model=model, grid=TRUE,
param=c(mean, variance, nugget, scale, alpha))
image(x, y, f[,,1])
image(x, y, f[,,2])
image(x, y, f[,,3])
#############################################################
## ##
## Examples using the extended definition form ##
## ##
## ##
#############################################################
#% library(RandomFields, lib="~/TMP"); RFparameters(Print=6)
x <- (0:100)/10
m <- matrix(c(1,2,3,4),ncol=2)/5
model <- list("$", aniso=m,
list("*",
list("power", k=2),
list("sph"))
)
z <- GaussRF(x=x, y=x, grid=TRUE, model=model, me="TBM3")
Print(c(mean(as.double(z)),var(as.double(z))))
image(z,zlim=c(-3,3))
## to know more what GaussRF does, use Print
## TMB can be very slow. To trace the iteration, use every
##
z <- GaussRF(x=x, y=x, grid=TRUE, model=model, me="TBM3",
Print=3, every=100)
image(z,zlim=c(-3,3))
## here, GaussRF uses `direct decomp' to simulate on the line
## and the square root of the covariance matrix is
## calculated by the Cholesky decomposition
## non-separable space-time model applied for two space dimensions
## note that tbm method works in some special cases.
#% library(RandomFields, lib="~/TMP")
x <- y <- seq(0, 7, if (interactive()) 0.05 else 0.2)
T <- c(1,32,1) * 10 ## note necessarily gridtriple definition
model <- list("$", aniso=diag(c(3, 3, 0.02)),
list("nsst", k1=2,
list("gauss"),
list("genB", k=c(1, 0.5))
))
z <- GaussRF(x=x, y=y, T=T, grid=TRUE, model=model,
method="ci", CE.strategy=1,
CE.trials=if (interactive()) 4 else 1)
rl <- function() if (interactive()) readline("Press return")
for (i in 1:dim(z)[3]) { image(z[,,i], zlim=range(z)); rl();}
for (i in 1:dim(z)[2]) { image(z[,i,], zlim=range(z)); rl();}
for (i in 1:dim(z)[1]) { image(z[i,,], zlim=range(z)); rl();}
#############################################################
## ##
## Example of a 2d random field based on ##
## covariance functions valid in 1d only ##
## ##
#############################################################
x <- seq(0, 10, 1/10)
model <- list("*",
list("$", aniso=matrix(nr=2, c(1, 0)),
list("fractgauss", k=0.5)),
list("$", aniso=matrix(nr=2, c(0, 1)),
list("fractgauss", k=0.9))
)
z <- GaussRF(x, x, grid=TRUE, gridtriple=FALSE, model=model)
image(x, x, z)
#############################################################
## ##
## Brownian motion ##
## (using Stein's method) ##
## ##
#############################################################
# 1d
kappa <- 1 # in [0,2)
z <- GaussRF(x=c(0, 10, 0.001), grid=TRUE, Print=5,
model=list("fractalB", kappa))
plot(z, type="l")
# 2d
step <- 0.3 ## nicer, but also time consuming if step = 0.1
x <- seq(0, 10, step)
kappa <- 1 # in [0,2)
z <- GaussRF(x=x, y=x, grid=TRUE, model=list("fractalB", kappa))
image(z,zlim=c(-3,3))
# 3d
x <- seq(0, 3, step)
kappa <- 1 # in [0,2)
z <- GaussRF(x=x, y=x, z=x, grid=TRUE,
model=list("fractalB", kappa))
rl <- function() if (interactive()) readline("Press return")
for (i in 1:dim(z)[1]) { image(z[i,,]); rl();}
#############################################################
## This example shows the benefits from stored, ##
## intermediate results: in case of the circulant ##
## embedding method, the speed is doubled in the second ##
## simulation. ##
#############################################################
RFparameters(Storing=TRUE)
y <- x <- seq(0, 50, 0.2)
(p <- c(runif(3), runif(1)+1))
ut <- system.time(f <- GaussRF(x=x,y=y,grid=TRUE,model="exponen",
method="circ", param=p))
image(x, y, f)
%hist(f)
%c( mean(as.vector(f)), var(as.vector(f)) )
cat("system time (first call)", format(ut,dig=3),"\n")
# second call with the same parameters can be much faster:
ut <- system.time(f <- DoSimulateRF())
image(x, y, f)
%hist(f)
%c( mean(as.vector(f)), var(as.vector(f)) )
cat("system time (second call)", format(ut,dig=3),"\n")
#############################################################
## ##
## Example how the cutoff method can be set ##
## explicitly using hypermodels ##
## ##
#############################################################
## NOTE: this feature is still in an experimental stage
## which has not been yet tested intensively;
## further: parameters and algorithms may change in
## future.
#% library(RandomFields, lib="~/TMP");source("~/R/cran/RandomFields/tests/source.R")
## simuation of the stable model using the cutoff method
#RFparameters(Print=8, Storing=FALSE)
x <- seq(0, 1, 1/24) # nicer pictures with 1/240
scale <- 1.0
model1 <- list("$", scale=scale, list("stable", alpha=1.0))
rs <- get(".Random.seed", envir=.GlobalEnv, inherits = FALSE)
z1 <- GaussRF(x, x, grid=TRUE, gridtriple=FALSE,
model=model1, n=1, meth="cutoff", Storing=TRUE)
(size <- GetRegisterInfo(meth=c("cutoff", "circ"))$S$size)
(cut.off.param <-
GetRegisterInfo(meth=c("cutoff", "circ"), modelname="cutoff")$param)
image(x, x, z1)
## simulation of the same random field using the circulant
## embedding method and defining the hypermodel explicitely
model2 <- list("$", scale = scale,
list("cutoff", diam=cut.off.param$diam, a=cut.off.param$a,
list("stable", alpha=1.0))
)
assign(".Random.seed", rs, envir=.GlobalEnv)
z2 <- GaussRF(x, x, grid=TRUE, gridtriple=FALSE, model=model2,
meth="circulant", n=1, CE.mmin=size, Storing=TRUE)
image(x, x, z2)
Print(range(z1-z2)) ## essentially no difference between the fields!
#% library(RandomFields)
#############################################################
## The cutoff method simulates on a torus and a (small) ##
## rectangle is taken as the required simulation. ##
## ##
## The following code shows a whole such torus. ##
## The main part of the code sets local.dependent=TRUE and ##
## local.mmin to multiples of the basic rectangle lengths ##
#############################################################
# definition of the realisation
RFparameters(CE.useprimes=FALSE)
x <- seq(0, 2, len=4) # better 20
y <- seq(0, 1, len=5) # better 40
grid.size <- c(length(x), length(y))
model <- list("$", var=1.1, aniso=matrix(nc=2, c(2, 1, 0.5, 1)),
list(model="exp"))
# determination of the (minimal) size of the torus
InitGaussRF(x, y, model=model, grid=TRUE, method="cu")
ce.info.size <- GetRegisterInfo(meth=c("cutoff", "circ"))$S$size
blocks <- ceiling(ce.info.size / grid.size / 4) *4
(size <- blocks * grid.size)
# simulation and plot of the torus
z <- GaussRF(x, y, model=model, grid=TRUE, method="cu", n=prod(blocks) * 2,
local.dependent=TRUE, local.mmin=size)[,,c(TRUE, FALSE)]
hei <- 8
do.call(getOption("device"),
list(hei=hei, wid=hei / blocks[2] / diff(range(y)) *
blocks[1] * diff(range(x))))
close.screen(close.screen())
sc <- matrix(nc=blocks[1], split.screen(rev(blocks)), byrow=TRUE)
sc <- as.vector(t(sc[nrow(sc):1, ]))
for (i in 1:prod(blocks)) {
screen(sc[i])
par(mar=rep(1, 4) * 0.0)
image(z[,, i], zlim=c(-3, 3), axes=FALSE, col=rainbow(100))
}
##############################################################
## Simulating with trend (as function) ##
##############################################################
x <- seq(-5,5,0.1)
z <- GaussRF(x=x, y=x, model = "exponential", param=c(1,0,1), grid=TRUE,
trend=function(x,y) 3*sin(x)*cos(y))
colors=heat.colors(1000)
image(x,x,z,col=colors)
##############################################################
## Simulating with linear trend surface ##
##############################################################
x <- seq(-5,5,0.1)
##trend surface: 3x - y
z <- GaussRF(x=x, y=x, model = "cubic", param=c(1,0,2), grid=TRUE,
trend=c(0,1,-1))
colors=heat.colors(1000)
persp(x,x,z, phi=30, theta=-3)
%##############################################################
%## for a further, more complicated example on ##
%## space-time modelling, see data(winddata) ##
%##############################################################
}
\keyword{spatial}