# NAG Library Routine Document

## 1Purpose

g05zsf produces realizations of a stationary Gaussian random field in two dimensions, using the circulant embedding method. The square roots of the eigenvalues of the extended covariance matrix (or embedding matrix) need to be input, and can be calculated using g05zqf or g05zrf.

## 2Specification

Fortran Interface
 Subroutine g05zsf ( ns, s, m, lam, rho, z,
 Integer, Intent (In) :: ns(2), s, m(2) Integer, Intent (Inout) :: state(*), ifail Real (Kind=nag_wp), Intent (In) :: lam(M(1)*M(2)), rho Real (Kind=nag_wp), Intent (Out) :: z(NS(1)*NS(2),s)
#include nagmk26.h
 void g05zsf_ ( const Integer ns[], const Integer *s, const Integer m[], const double lam[], const double *rho, Integer state[], double z[], Integer *ifail)

## 3Description

A two-dimensional random field $Z\left(\mathbf{x}\right)$ in ${ℝ}^{2}$ is a function which is random at every point $\mathbf{x}\in {ℝ}^{2}$, so $Z\left(\mathbf{x}\right)$ is a random variable for each $\mathbf{x}$. The random field has a mean function $\mu \left(\mathbf{x}\right)=𝔼\left[Z\left(\mathbf{x}\right)\right]$ and a symmetric positive semidefinite covariance function $C\left(\mathbf{x},\mathbf{y}\right)=𝔼\left[\left(Z\left(\mathbf{x}\right)-\mu \left(\mathbf{x}\right)\right)\left(Z\left(\mathbf{y}\right)-\mu \left(\mathbf{y}\right)\right)\right]$. $Z\left(\mathbf{x}\right)$ is a Gaussian random field if for any choice of $n\in ℕ$ and ${\mathbf{x}}_{1},\dots ,{\mathbf{x}}_{n}\in {ℝ}^{2}$, the random vector ${\left[Z\left({\mathbf{x}}_{1}\right),\dots ,Z\left({\mathbf{x}}_{n}\right)\right]}^{\mathrm{T}}$ follows a multivariate Normal distribution, which would have a mean vector $\stackrel{~}{\mathbf{\mu }}$ with entries ${\stackrel{~}{\mu }}_{i}=\mu \left({\mathbf{x}}_{i}\right)$ and a covariance matrix $\stackrel{~}{C}$ with entries ${\stackrel{~}{C}}_{ij}=C\left({\mathbf{x}}_{i},{\mathbf{x}}_{j}\right)$. A Gaussian random field $Z\left(\mathbf{x}\right)$ is stationary if $\mu \left(\mathbf{x}\right)$ is constant for all $\mathbf{x}\in {ℝ}^{2}$ and $C\left(\mathbf{x},\mathbf{y}\right)=C\left(\mathbf{x}+\mathbf{a},\mathbf{y}+\mathbf{a}\right)$ for all $\mathbf{x},\mathbf{y},\mathbf{a}\in {ℝ}^{2}$ and hence we can express the covariance function $C\left(\mathbf{x},\mathbf{y}\right)$ as a function $\gamma$ of one variable: $C\left(\mathbf{x},\mathbf{y}\right)=\gamma \left(\mathbf{x}-\mathbf{y}\right)$. $\gamma$ is known as a variogram (or more correctly, a semivariogram) and includes the multiplicative factor ${\sigma }^{2}$ representing the variance such that $\gamma \left(0\right)={\sigma }^{2}$.
The routines g05zqf or g05zrf along with g05zsf are used to simulate a two-dimensional stationary Gaussian random field, with mean function zero and variogram $\gamma \left(\mathbf{x}\right)$, over a domain $\left[{x}_{\mathrm{min}},{x}_{\mathrm{max}}\right]×\left[{y}_{\mathrm{min}},{y}_{\mathrm{max}}\right]$, using an equally spaced set of ${N}_{1}×{N}_{2}$ points; ${N}_{1}$ points in the $x$-direction and ${N}_{2}$ points in the $y$-direction. The problem reduces to sampling a Gaussian random vector $\mathbf{X}$ of size ${N}_{1}×{N}_{2}$, with mean vector zero and a symmetric covariance matrix $A$, which is an ${N}_{2}$ by ${N}_{2}$ block Toeplitz matrix with Toeplitz blocks of size ${N}_{1}$ by ${N}_{1}$. Since $A$ is in general expensive to factorize, a technique known as the circulant embedding method is used. $A$ is embedded into a larger, symmetric matrix $B$, which is an ${M}_{2}$ by ${M}_{2}$ block circulant matrix with circulant bocks of size ${M}_{1}$ by ${M}_{1}$, where ${M}_{1}\ge 2\left({N}_{1}-1\right)$ and ${M}_{2}\ge 2\left({N}_{2}-1\right)$. $B$ can now be factorized as $B=W\Lambda {W}^{*}={R}^{*}R$, where $W$ is the two-dimensional Fourier matrix (${W}^{*}$ is the complex conjugate of $W$), $\Lambda$ is the diagonal matrix containing the eigenvalues of $B$ and $R={\Lambda }^{\frac{1}{2}}{W}^{*}$. $B$ is known as the embedding matrix. The eigenvalues can be calculated by performing a discrete Fourier transform of the first row (or column) of $B$ and multiplying by ${M}_{1}×{M}_{2}$, and so only the first row (or column) of $B$ is needed – the whole matrix does not need to be formed.
The symmetry of $A$ as a block matrix, and the symmetry of each block of $A$, depends on whether the covariance function $\gamma$ is even or not. $\gamma$ is even if $\gamma \left(\mathbf{x}\right)=\gamma \left(-\mathbf{x}\right)$ for all $\mathbf{x}\in {ℝ}^{2}$, and uneven otherwise (in higher dimensions, $\gamma$ can be even in some coordinates and uneven in others, but in two dimensions $\gamma$ is either even in both coordinates or uneven in both coordinates). If $\gamma$ is even then $A$ is a symmetric block matrix and has symmetric blocks; if $\gamma$ is uneven then $A$ is not a symmetric block matrix and has non-symmetric blocks. In the uneven case, ${M}_{1}$ and ${M}_{2}$ are set to be odd in order to guarantee symmetry in $B$.
As long as all of the values of $\Lambda$ are non-negative (i.e., $B$ is positive semidefinite), $B$ is a covariance matrix for a random vector $\mathbf{Y}$ which has ${M}_{2}$ ‘blocks’ of size ${M}_{1}$. Two samples of $\mathbf{Y}$ can now be simulated from the real and imaginary parts of ${R}^{*}\left(\mathbf{U}+i\mathbf{V}\right)$, where $\mathbf{U}$ and $\mathbf{V}$ have elements from the standard Normal distribution. Since ${R}^{*}\left(\mathbf{U}+i\mathbf{V}\right)=W{\Lambda }^{\frac{1}{2}}\left(\mathbf{U}+i\mathbf{V}\right)$, this calculation can be done using a discrete Fourier transform of the vector ${\Lambda }^{\frac{1}{2}}\left(\mathbf{U}+i\mathbf{V}\right)$. Two samples of the random vector $\mathbf{X}$ can now be recovered by taking the first ${N}_{1}$ elements of the first ${N}_{2}$ blocks of each sample of $Y$ – because the original covariance matrix $A$ is embedded in $B$, $\mathbf{X}$ will have the correct distribution.
If $B$ is not positive semidefinite, larger embedding matrices $B$ can be tried; however if the size of the matrix would have to be larger than maxm, an approximation procedure is used. See the documentation of g05zqf or g05zrf for details of the approximation procedure.
g05zsf takes the square roots of the eigenvalues of the embedding matrix $B$, and its size vector $M$, as input and outputs $S$ realizations of the random field in $Z$.
One of the initialization routines g05kff (for a repeatable sequence if computed sequentially) or g05kgf (for a non-repeatable sequence) must be called prior to the first call to g05zsf.

## 4References

Dietrich C R and Newsam G N (1997) Fast and exact simulation of stationary Gaussian processes through circulant embedding of the covariance matrix SIAM J. Sci. Comput. 18 1088–1107
Schlather M (1999) Introduction to positive definite functions and to unconditional simulation of random fields Technical Report ST 99–10 Lancaster University
Wood A T A and Chan G (1994) Simulation of stationary Gaussian processes in ${\left[0,1\right]}^{d}$ Journal of Computational and Graphical Statistics 3(4) 409–432

## 5Arguments

1:     $\mathbf{ns}\left(2\right)$ – Integer arrayInput
On entry: the number of sample points to use in each direction, with ${\mathbf{ns}}\left(1\right)$ sample points in the $x$-direction and ${\mathbf{ns}}\left(2\right)$ sample points in the $y$-direction. The total number of sample points on the grid is therefore ${\mathbf{ns}}\left(1\right)×{\mathbf{ns}}\left(2\right)$. This must be the same value as supplied to g05zqf or g05zrf when calculating the eigenvalues of the embedding matrix.
Constraints:
• ${\mathbf{ns}}\left(1\right)\ge 1$;
• ${\mathbf{ns}}\left(2\right)\ge 1$.
2:     $\mathbf{s}$ – IntegerInput
On entry: $S$, the number of realizations of the random field to simulate.
Constraint: ${\mathbf{s}}\ge 1$.
3:     $\mathbf{m}\left(2\right)$ – Integer arrayInput
On entry: indicates the size, $M$, of the embedding matrix as returned by g05zqf or g05zrf. The embedding matrix is a block circulant matrix with circulant blocks. ${\mathbf{m}}\left(1\right)$ is the size of each block, and ${\mathbf{m}}\left(2\right)$ is the number of blocks.
Constraints:
• ${\mathbf{m}}\left(1\right)\ge \mathrm{max}\phantom{\rule{0.125em}{0ex}}\left(1,2\left({\mathbf{ns}}\left(1\right)-1\right)\right)$;
• ${\mathbf{m}}\left(2\right)\ge \mathrm{max}\phantom{\rule{0.125em}{0ex}}\left(1,2\left({\mathbf{ns}}\left(2\right)-1\right)\right)$.
4:     $\mathbf{lam}\left({\mathbf{m}}\left(1\right)×{\mathbf{m}}\left(2\right)\right)$ – Real (Kind=nag_wp) arrayInput
On entry: contains the square roots of the eigenvalues of the embedding matrix, as returned by g05zqf or g05zrf.
Constraint: ${\mathbf{lam}}\left(\mathit{i}\right)\ge 0$, for $\mathit{i}=1,2,\dots ,{\mathbf{m}}\left(1\right)×{\mathbf{m}}\left(2\right)$.
5:     $\mathbf{rho}$ – Real (Kind=nag_wp)Input
On entry: indicates the scaling of the covariance matrix, as returned by g05zqf or g05zrf.
Constraint: $0.0<{\mathbf{rho}}\le 1.0$.
6:     $\mathbf{state}\left(*\right)$ – Integer arrayCommunication Array
Note: the actual argument supplied must be the array state supplied to the initialization routines g05kff or g05kgf.
On entry: contains information on the selected base generator and its current state.
On exit: contains updated information on the state of the generator.
7:     $\mathbf{z}\left({\mathbf{ns}}\left(1\right)×{\mathbf{ns}}\left(2\right),{\mathbf{s}}\right)$ – Real (Kind=nag_wp) arrayOutput
On exit: contains the realizations of the random field. The $k$th realization (where $k=1,2,\dots ,{\mathbf{s}}$) of the random field on the two-dimensional grid $\left({x}_{i},{y}_{j}\right)$ is stored in ${\mathbf{z}}\left(\left(j-1\right)×{\mathbf{ns}}\left(1\right)+i,k\right)$, for $i=1,2,\dots ,{\mathbf{ns}}\left(1\right)$ and for $j=1,2,\dots ,{\mathbf{ns}}\left(2\right)$. The points are returned in xx and yy by g05zqf or g05zrf.
8:     $\mathbf{ifail}$ – IntegerInput/Output
On entry: ifail must be set to $0$, $-1\text{​ or ​}1$. If you are unfamiliar with this argument you should refer to Section 3.4 in How to Use the NAG Library and its Documentation for details.
For environments where it might be inappropriate to halt program execution when an error is detected, the value $-1\text{​ or ​}1$ is recommended. If the output of error messages is undesirable, then the value $1$ is recommended. Otherwise, if you are not familiar with this argument, the recommended value is $0$. When the value $-\mathbf{1}\text{​ or ​}\mathbf{1}$ is used it is essential to test the value of ifail on exit.
On exit: ${\mathbf{ifail}}={\mathbf{0}}$ unless the routine detects an error or a warning has been flagged (see Section 6).

## 6Error Indicators and Warnings

If on entry ${\mathbf{ifail}}=0$ or $-1$, explanatory error messages are output on the current error message unit (as defined by x04aaf).
Errors or warnings detected by the routine:
${\mathbf{ifail}}=1$
On entry, ${\mathbf{ns}}=\left[〈\mathit{\text{value}}〉,〈\mathit{\text{value}}〉\right]$.
Constraint: ${\mathbf{ns}}\left(1\right)\ge 1$, ${\mathbf{ns}}\left(2\right)\ge 1$.
${\mathbf{ifail}}=2$
On entry, ${\mathbf{s}}=〈\mathit{\text{value}}〉$.
Constraint: ${\mathbf{s}}\ge 1$.
${\mathbf{ifail}}=3$
On entry, ${\mathbf{m}}=\left[〈\mathit{\text{value}}〉,〈\mathit{\text{value}}〉\right]$, and ${\mathbf{ns}}=\left[〈\mathit{\text{value}}〉,〈\mathit{\text{value}}〉\right]$.
Constraints: ${\mathbf{m}}\left(i\right)\ge \mathrm{max}\phantom{\rule{0.125em}{0ex}}\left(1,2\left({\mathbf{ns}}\left(i\right)\right)-1\right)$, for $i=1,2$.
${\mathbf{ifail}}=4$
On entry, at least one element of lam was negative.
Constraint: all elements of lam must be non-negative.
${\mathbf{ifail}}=5$
On entry, ${\mathbf{rho}}=〈\mathit{\text{value}}〉$.
Constraint: $0.0<{\mathbf{rho}}\le 1.0$.
${\mathbf{ifail}}=6$
On entry, state vector has been corrupted or not initialized.
${\mathbf{ifail}}=-99$
See Section 3.9 in How to Use the NAG Library and its Documentation for further information.
${\mathbf{ifail}}=-399$
Your licence key may have expired or may not have been installed correctly.
See Section 3.8 in How to Use the NAG Library and its Documentation for further information.
${\mathbf{ifail}}=-999$
Dynamic memory allocation failed.
See Section 3.7 in How to Use the NAG Library and its Documentation for further information.

Not applicable.

## 8Parallelism and Performance

g05zsf is threaded by NAG for parallel execution in multithreaded implementations of the NAG Library.
g05zsf makes calls to BLAS and/or LAPACK routines, which may be threaded within the vendor library used by this implementation. Consult the documentation for the vendor library for further information.
Please consult the X06 Chapter Introduction for information on how to control and interrogate the OpenMP environment used within this routine. Please also consult the Users' Note for your implementation for any additional implementation-specific information.

Because samples are generated in pairs, calling this routine $k$ times, with ${\mathbf{s}}=s$, say, will generate a different sequence of numbers than calling the routine once with ${\mathbf{s}}=ks$, unless $s$ is even.

## 10Example

This example calls g05zsf to generate $5$ realizations of a two-dimensional random field on a $5$ by $5$ grid. This uses eigenvalues of the embedding covariance matrix for a symmetric stable variogram as calculated by g05zrf with ${\mathbf{icov2}}=1$.

### 10.1Program Text

Program Text (g05zsfe.f90)

### 10.2Program Data

Program Data (g05zsfe.d)

### 10.3Program Results

Program Results (g05zsfe.r)

© The Numerical Algorithms Group Ltd, Oxford, UK. 2017