Skip to contents


In this vignette we will present the inlabru implementation of the covariance-based rational SPDE approach. For further technical details on the covariance-based approach, see the Rational approximation with the rSPDE package vignette and Xiong, Simas, and Bolin (2022).

We begin by providing a step-by-step illustration on how to use our implementation. To this end we will consider a real world data set that consists of precipitation measurements from the Paraná region in Brazil.

After the initial model fitting, we will show how to change some parameters of the model. In the end, we will also provide an example in which we have replicates.

The examples in this vignette are the same as those in the R-INLA implementation of the rational SPDE approach vignette. As in that case, it is important to mention that one can improve the performance by using the PARDISO solver. Please, go to to apply for a license. Also, use inla.pardiso() for instructions on how to enable the PARDISO sparse library.

An example with real data

To illustrate our implementation of rSPDE in inlabru we will consider a dataset available in R-INLA. This data has also been used to illustrate the SPDE approach, see for instance the book Advanced Spatial Modeling with Stochastic Partial Differential Equations Using R and INLA and also the vignette Spatial Statistics using R-INLA and Gaussian Markov random fields. See also Lindgren, Rue, and Lindström (2011) for theoretical details on the standard SPDE approach.

The data consist of precipitation measurements from the Paraná region in Brazil and were provided by the Brazilian National Water Agency. The data were collected at 616 gauge stations in Paraná state, south of Brazil, for each day in 2011.

An rSPDE model for precipitation

We will follow the vignette Spatial Statistics using R-INLA and Gaussian Markov random fields. As precipitation data are always positive, we will assume it is Gamma distributed. R-INLA uses the following parameterization of the Gamma distribution, \[\Gamma(\mu, \phi): \pi (y) = \frac{1}{\Gamma(\phi)} \left(\frac{\phi}{\mu}\right)^{\phi} y^{\phi - 1} \exp\left(-\frac{\phi y}{\mu}\right) .\] In this parameterization, the distribution has expected value \(E(x) = \mu\) and variance \(V(x) = \mu^2/\phi\), where \(1/\phi\) is a dispersion parameter.

In this example \(\mu\) will be modelled using a stochastic model that includes both covariates and spatial structure, resulting in the latent Gaussian model for the precipitation measurements \[\begin{align} y_i\mid \mu(s_i), \theta &\sim \Gamma(\mu(s_i),c\phi)\\ \log (\mu(s)) &= \eta(s) = \sum_k f_k(c_k(s))+u(s)\\ \theta &\sim \pi(\theta) \end{align},\]

where \(y_i\) denotes the measurement taken at location \(s_i\), \(c_k(s)\) are covariates, \(u(s)\) is a mean-zero Gaussian Matérn field, and \(\theta\) is a vector containing all parameters of the model, including smoothness of the field. That is, by using the rSPDE model we will also be able to estimate the smoothness of the latent field.

Examining the data

We will be using inlabru. The inlabru package is available on CRAN and also on GitHub.

We begin by loading some libraries we need to get the data and build the plots.

Let us load the data and the border of the region


The data frame contains daily measurements at 616 stations for the year 2011, as well as coordinates and altitude information for the measurement stations. We will not analyze the full spatio-temporal data set, but instead look at the total precipitation in January, which we calculate as

Y <- rowMeans(PRprec[, 3 + 1:31])

In the next snippet of code, we extract the coordinates and altitudes and remove the locations with missing values.

ind <- !
Y <- Y[ind]
coords <- as.matrix(PRprec[ind, 1:2])
alt <- PRprec$Altitude[ind]

Let us build a plot for the precipitations:

p <- ggplot() +
    x = coords[, 1], y = coords[, 2],
    colour = Y
  ), size = 2, alpha = 1) +
  scale_colour_gradientn(colours = tim.colors(100)) +
  geom_path(aes(x = PRborder[, 1], y = PRborder[, 2])) +
  geom_path(aes(x = PRborder[1034:1078, 1], y = PRborder[
  ]), colour = "red")
  error = function(e) {
    print("Unable to build the plot")

The red line in the figure shows the coast line, and we expect the distance to the coast to be a good covariate for precipitation.

This covariate is not available, so let us calculate it for each observation location:

seaDist <- apply(spDists(coords, PRborder[1034:1078, ],
  longlat = TRUE
), 1, min)

Now, let us plot the precipitation as a function of the possible covariates:

par(mfrow = c(2, 2))
plot(coords[, 1], Y, cex = 0.5, xlab = "Longitude")
plot(coords[, 2], Y, cex = 0.5, xlab = "Latitude")
plot(seaDist, Y, cex = 0.5, xlab = "Distance to sea")
plot(alt, Y, cex = 0.5, xlab = "Altitude")

par(mfrow = c(1, 1))

Creating the rSPDE model

To use the inlabru implementation of the rSPDE model we need to load the functions:

To create a rSPDE model, one would the rspde.matern() function in a similar fashion as one would use the inla.spde2.matern() function.


We can use R-INLA for creating the mesh. Let us create a mesh which is based on a non-convex hull to avoid adding many small triangles outside the domain of interest:

prdomain <- inla.nonconvex.hull(coords, -0.03, -0.05, resolution = c(100, 100))
prmesh <- inla.mesh.2d(boundary = prdomain, max.edge = c(0.45, 1), cutoff = 0.2)
plot(prmesh, asp = 1, main = "")
lines(PRborder, col = 3)
points(coords[, 1], coords[, 2], pch = 19, cex = 0.5, col = "red")

Setting up the data frame

In place of a inla.stack, we can set up a data.frame() to use inlabru. We refer the reader to vignettes in for further details.

prdata <- data.frame(long = coords[,1], lat = coords[,2], 
                        seaDist =, y = Y)
coordinates(prdata) <- c("long","lat")

Setting up the rSPDE model

To set up an rSPDEmodel, all we need is the mesh. By default it will assume that we want to estimate the smoothness parameter \(\nu\) and to do a covariance-based rational approximation of order 2.

Later in this vignette we will also see other options for setting up rSPDE models such as keeping the smoothness parameter fixed and/or increasing the order of the covariance-based rational approximation.

Therefore, to set up a model all we have to do is use the rspde.matern() function:

rspde_model <- rspde.matern(mesh = prmesh)

Notice that this function is very reminiscent of R-INLA’s inla.spde2.matern() function.

We will assume the following linkage between model components and observations \[\eta(s) \sim A x(s) + A \text{ Intercept} + \text{seaDist}.\] \(\eta(s)\) will then be used in the observation-likelihood, \[y_i\mid \eta(s_i),\theta \sim \Gamma(\exp(\eta (s_i)), c\phi).\]

Model fitting

We will build a model using the distance to the sea \(x_i\) as a covariate through an improper CAR(1) model with \(\beta_{ij}=1(i\sim j)\), which R-INLA calls a random walk of order 1. We will fit it in inlabru’s style:

cmp <- y ~ Intercept(1) + distSea(seaDist, model="rw1") +
field(coordinates, model = rspde_model)

To fit the model we simply use the bru() function:

rspde_fit <- bru(cmp, data = prdata,
  family = "Gamma",
  options = list(
    inla.mode = "experimental",
    control.inla = list(int.strategy = "eb"),
    verbose = FALSE)

inlabru results

We can look at some summaries of the posterior distributions for the parameters, for example the fixed effects (i.e. the intercept) and the hyper-parameters (i.e. dispersion in the gamma likelihood, the precision of the RW1, and the parameters of the spatial field):

## inlabru version: 2.7.0
## INLA version: 22.12.03
## Components:
## Intercept: main = linear(1)
## distSea: main = rw1(seaDist)
## field: main = cgeneric(coordinates)
## Likelihoods:
##   Family: 'Gamma'
##     Data class: 'SpatialPointsDataFrame'
##     Predictor: y ~ .
## Time used:
##     Pre = 3.58, Running = 14.9, Post = 0.0995, Total = 18.6 
## Fixed effects:
##            mean   sd 0.025quant 0.5quant 0.975quant  mode kld
## Intercept 1.943 0.05      1.844    1.943      2.041 1.943   0
## Random effects:
##   Name     Model
##     distSea RW1 model
##    field CGeneric
## Model hyperparameters:
##                                                    mean       sd 0.025quant
## Precision parameter for the Gamma observations   13.117    0.890     11.433
## Precision for distSea                          9435.082 6586.454   2373.725
## Theta1 for field                                 -1.516    0.139     -1.795
## Theta2 for field                                 -0.079    0.254     -0.573
## Theta3 for field                                 -0.185    0.921     -1.902
##                                                0.5quant 0.975quant     mode
## Precision parameter for the Gamma observations   13.094     14.937   13.060
## Precision for distSea                          7704.659  26830.797 5261.489
## Theta1 for field                                 -1.514     -1.248   -1.507
## Theta2 for field                                 -0.082      0.428   -0.092
## Theta3 for field                                 -0.219      1.740   -0.367
## Deviance Information Criterion (DIC) ...............: 2503.27
## Deviance Information Criterion (DIC, saturated) ....: 693.40
## Effective number of parameters .....................: 80.51
## Watanabe-Akaike information criterion (WAIC) ...: 2505.52
## Effective number of parameters .................: 73.32
## Marginal log-Likelihood:  -1262.55 
##  is computed 
## Posterior summaries for the linear predictor and the fitted values are computed
## (Posterior marginals needs also 'control.compute=list(return.marginals.predictor=TRUE)')

Let \(\theta_1 = \textrm{Theta1}\), \(\theta_2=\textrm{Theta2}\) and \(\theta_3=\textrm{Theta3}\). In terms of the SPDE \[(\kappa^2 I - \Delta)^{\alpha/2}(\tau u) = \mathcal{W},\] where \(\alpha = \nu + d/2\), we have that \[\tau = \exp(\theta_1),\quad \kappa = \exp(\theta_2), \] and by default \[\nu = 4\Big(\frac{\exp(\theta_3)}{1+\exp(\theta_3)}\Big).\] The number 4 comes from the upper bound for \(\nu\), which is discussed in R-INLA implementation of the rational SPDE approach vignette.

In general, we have \[\nu = \nu_{UB}\Big(\frac{\exp(\theta_3)}{1+\exp(\theta_3)}\Big),\] where \(\nu_{UB}\) is the value of the upper bound for the smoothness parameter \(\nu\).

Another choice for prior for \(\nu\) is a truncated lognormal distribution and is also discussed in R-INLA implementation of the rational SPDE approach vignette.

inlabru results in the original scale

We can obtain outputs with respect to parameters in the original scale by using the function rspde.result():

result_fit <- rspde.result(rspde_fit, "field", rspde_model)
##             mean        sd 0.025quant 0.5quant 0.975quant     mode
## 0.221596 0.0305435   0.166559 0.219968    0.28629 0.217261
## range   0.953517 0.2455040   0.566817 0.920834    1.52564 0.857939
## nu      1.834760 0.7814330   0.530797 1.771390    3.38732 1.453760

We can also plot the posterior densities. To this end we will use the gg_df() function, which creates ggplot2 user-friendly data frames:

posterior_df_fit <- gg_df(result_fit)

ggplot(posterior_df_fit) + geom_line(aes(x = x, y = y)) + 
facet_wrap(~parameter, scales = "free") + labs(y = "Density")


Let us now obtain predictions (i.e. do kriging) of the expected precipitation on a dense grid in the region.

We begin by creating the grid in which we want to do the predictions. To this end, we can use the inla.mesh.projector() function:

nxy <- c(150, 100)
projgrid <- inla.mesh.projector(prmesh,
  xlim = range(PRborder[, 1]),
  ylim = range(PRborder[, 2]), dims = nxy

This lattice contains 150 × 100 locations. One can easily change the resolution of the kriging prediction by changing nxy. Let us find the cells that are outside the region of interest so that we do not plot the estimates there. <- inout(projgrid$lattice$loc, cbind(PRborder[, 1], PRborder[, 2]))

Let us plot the locations that we will do prediction:

coord.prd <- projgrid$lattice$loc[, ]
plot(coord.prd, type = "p", cex = 0.1)
points(coords[, 1], coords[, 2], pch = 19, cex = 0.5, col = "red")

Let us now create a data.frame() of the coordinates:

coord.prd.df <- data.frame(x1 = coord.prd[,1],
                            x2 = coord.prd[,2])
coordinates(coord.prd.df) <- c("x1", "x2")

Since we are using distance to the sea as a covariate, we also have to calculate this covariate for the prediction locations. Finally, we add the prediction location to our prediction data.frame(), namely, coord.prd.df:

seaDist.prd <- apply(spDists(coord.prd,
  PRborder[1034:1078, ],
  longlat = TRUE
), 1, min)
coord.prd.df$seaDist <- seaDist.prd
pred_obs <- predict(rspde_fit, coord.prd.df, 
        ~exp(Intercept + field + distSea))

Let us now build the data frame with the results:

pred_df <- pred_obs@data
pred_df <- cbind(pred_df, pred_obs@coords)

Finally, we plot the results. First the predicted mean:

p <- ggplot(pred_df, aes(x = x1, y = x2, fill = mean)) +
  geom_raster() +
  scale_fill_gradient(low = "yellow", high = "red")

Then, the std. deviations:

p <- ggplot(pred_df, aes(x = x1, y = x2, fill = sd)) +

An example with replicates

For this example we will simulate a data with replicates. We will use the same example considered in the Rational approximation with the rSPDE package vignette (the only difference is the way the data is organized). We also refer the reader to this vignette for a description of the function matern.operators(), along with its methods (for instance, the simulate() method).

Simulating the data

Let us consider a simple Gaussian linear model with 30 independent replicates of a latent spatial field \(x(\mathbf{s})\), observed at the same \(m\) locations, \(\{\mathbf{s}_1 , \ldots , \mathbf{s}_m \}\), for each replicate. For each \(i = 1,\ldots,m,\) we have

\[\begin{align} y_i &= x_1(\mathbf{s}_i)+\varepsilon_i,\\ \vdots &= \vdots\\ y_{i+29m} &= x_{30}(\mathbf{s}_i) + \varepsilon_{i+29m}, \end{align}\]

where \(\varepsilon_1,\ldots,\varepsilon_{30m}\) are iid normally distributed with mean 0 and standard deviation 0.1.

We use the basis function representation of \(x(\cdot)\) to define the \(A\) matrix linking the point locations to the mesh. We also need to account for the fact that we have 30 replicates at the same locations. To this end, the \(A\) matrix we need can be generated by inla.spde.make.A() function. The reason being that we are sampling \(x(\cdot)\) directly and not the latent vector described in the introduction of the Rational approximation with the rSPDE package vignette.

We begin by creating the mesh:

m <- 200
loc_2d_mesh <- matrix(runif(m * 2), m, 2)
mesh_2d <- inla.mesh.2d(
  loc = loc_2d_mesh,
  cutoff = 0.05,
  offset = c(0.1, 0.4),
  max.edge = c(0.05, 0.5)
plot(mesh_2d, main = "")
points(loc_2d_mesh[, 1], loc_2d_mesh[, 2])

We then compute the \(A\) matrix, which is needed for simulation, and connects the observation locations to the mesh:

n.rep <- 30
A <- inla.spde.make.A(
  mesh = mesh_2d,
  loc = loc_2d_mesh,
  index = rep(1:m, times = n.rep),
  repl = rep(1:n.rep, each = m)

Notice that for the simulated data, we should use the \(A\) matrix from inla.spde.make.A() function.

We will now simulate a latent process with standard deviation \(\sigma=1\) and range \(0.1\). We will use \(\nu=0.5\) so that the model has an exponential covariance function. To this end we create a model object with the matern.operators() function:

nu <- 0.5
sigma <- 1
range <- 0.1
kappa <- sqrt(8 * nu) / range
tau <- sqrt(gamma(nu) / (sigma^2 * kappa^(2 * nu) * (4 * pi) * gamma(nu + 1)))
d <- 2
operator_information <- matern.operators(
  mesh = mesh_2d,
  nu = nu,
  kappa = kappa,
  sigma = sigma,
  m = 2

More details on this function can be found at the Rational approximation with the rSPDE package vignette.

To simulate the latent process all we need to do is to use the simulate() method on the operator_information object. We then obtain the simulated data \(y\) by connecting with the \(A\) matrix and adding the gaussian noise.

u <- simulate(operator_information, nsim = n.rep)
y <- as.vector(A %*% as.vector(u)) +
  rnorm(m * n.rep) * 0.1

The first replicate of the simulated random field as well as the observation locations are shown in the following figure.

opar <- par(mfrow = c(1, 2), mgp = c(1.2, 0.5, 0),
   mar = c(2, 2, 0.5, 0.5) + 0.1)
proj <- inla.mesh.projector(mesh_2d, dims = c(100, 100))
image(inla.mesh.project(proj, field = as.vector(u[, 1])),
  xlab = "", ylab = "",
  cex.main = 0.8, cex.axis = 0.8, cex.lab = 0.8
plot(loc_2d_mesh[, 1], loc_2d_mesh[, 2],
  cex = 0.2, pch = 16, xlab = "", ylab = "",
  cex.main = 0.8, cex.axis = 0.8, cex.lab = 0.8

Fitting the inlabru rSPDE model

Let us then use the rational SPDE approach to fit the data.

We begin by creating the model object.

rspde_model.rep <- rspde.matern(mesh = mesh_2d,
          parameterization = "spde") 

Let us now create the data.frame() and the vector with the replicates indexes:

rep.df <- data.frame(y = y, x1 = rep(loc_2d_mesh[,1], 30),
                      x2 = rep(loc_2d_mesh[,2], 30))
coordinates(rep.df) <- c("x1", "x2")
repl <- rep(1:30, each=200)

Let us create the component and fit. It is extremely important not to forget the replicate when fitting model with the bru() function. It will not produce warning and might fit some meaningless model.

cmp.rep <-
  y ~ -1 + field(coordinates,
    model = rspde_model.rep,
    replicate = repl

rspde_fit.rep <-
    data = rep.df,
    family = "gaussian",
    options = list(
      list(inla.mode = "experimental")

We can get the summary:

## inlabru version: 2.7.0
## INLA version: 22.12.03
## Components:
## field: main = cgeneric(coordinates), replicate = iid(repl)
## Likelihoods:
##   Family: 'gaussian'
##     Data class: 'SpatialPointsDataFrame'
##     Predictor: y ~ .
## Time used:
##     Pre = 3.68, Running = 149, Post = 11.7, Total = 164 
## Random effects:
##   Name     Model
##     field CGeneric
## Model hyperparameters:
##                                          mean    sd 0.025quant 0.5quant
## Precision for the Gaussian observations 89.59 0.635      87.63    89.45
## Theta1 for field                        -2.99 0.048      -3.18    -2.96
## Theta2 for field                         3.14 0.008       3.11     3.14
## Theta3 for field                        -1.67 0.022      -1.75    -1.69
##                                         0.975quant  mode
## Precision for the Gaussian observations      91.55 89.59
## Theta1 for field                             -2.81 -2.99
## Theta2 for field                              3.17  3.14
## Theta3 for field                             -1.58 -1.67
## Deviance Information Criterion (DIC) ...............: -5791.83
## Deviance Information Criterion (DIC, saturated) ....: 10434.14
## Effective number of parameters .....................: 4453.81
## Watanabe-Akaike information criterion (WAIC) ...: -6621.16
## Effective number of parameters .................: 2664.93
## Marginal log-Likelihood:  -4403.36 
##  is computed 
## Posterior summaries for the linear predictor and the fitted values are computed
## (Posterior marginals needs also 'control.compute=list(return.marginals.predictor=TRUE)')

and the summary in the user’s scale:

result_fit_rep <- rspde.result(rspde_fit.rep, "field", rspde_model.rep)
##             mean         sd 0.025quant   0.5quant 0.975quant       mode
## tau    0.0501975 0.00240572  0.0456527  0.0501296  0.0551105  0.0499959
## kappa 23.1549000 0.19215300 22.7782000 23.1548000 23.5327000 23.1585000
## nu     0.6340720 0.01170800  0.6112340  0.6340190  0.6572110  0.6341000
result_df <- data.frame(
  parameter = c("tau", "kappa", "nu"),
  true = c(tau, kappa, nu),
  mean = c(
  mode = c(
##   parameter        true        mean        mode
## 1       tau  0.08920621  0.05019749  0.04999588
## 2     kappa 20.00000000 23.15494333 23.15848041
## 3        nu  0.50000000  0.63407170  0.63409962

Further options of the inlabru implementation

There are several additional options that are available. For instance, it is possible to change the order of the rational approximation, the upper bound for the smoothness parameter (which may speed up the fit), change the priors, change the type of the rational approximation, among others. These options are described in the “Further options of the rSPDE-INLA implementation” section of the R-INLA implementation of the rational SPDE approach vignette. Observe that all these options are passed to the model through the rspde.matern() function, and therefore the resulting model object can directly be used in the bru() function, in an identical manner to the examples above.


Lindgren, Finn, Håvard Rue, and Johan Lindström. 2011. “An Explicit Link Between Gaussian Fields and Gaussian Markov Random Fields: The Stochastic Partial Differential Equation Approach.” Journal of the Royal Statistical Society. Series B. Statistical Methodology 73 (4): 423–98.
Xiong, Zhen, Alexandre B. Simas, and David Bolin. 2022. “Covariance-Based Rational Approximations of Fractional SPDEs for Computationally Efficient Bayesian Inference.” arXiv:2209.04670.