Rusting Faster: Simulation using Rcpp

The rust package implements the multivariate generalized ratio-of-uniforms method of simulating random variates from a d-dimensional continuous distribution. The user specifies (the log of) a positive target function f(x) proportional to the density function of the distribution. For an introduction to rust see the vignette Introducing rust.

This vignette describes a new feature of rust: the option for the user to provide a C++ function to evaluate the target log-density, rather than an R function. The Rcpp Eddelbuettel (2013) and RcppArmadillo (Eddelbuettel and Sanderson 2014) packages are used to speed up simulation from the target density. The improvement results from faster function evaluations and (in particular) from performing using C++ the looping in the ratio-of-uniforms algorithm. The new function ru_rcpp requires the target log-density to be specified using (externals pointers to) C++ functions, whereas the existing ru requires input R functions. Otherwise, the functionality of these two functions is the same. There are also Rcpp-based versions of functions for setting Box-Cox transformation parameters: find_lambda_rcpp and find_lambda_one_d_rcpp

In this vignette we describe in general terms the general setup of the Rcpp-based functions and use examples to illustrate their use. For more information about these examples see the vignette Introducing rust

Providing a C++ function to `ru_rcpp`

The general way that rust enables users to provide their own C++ functions uses external pointers and is based on the Rcpp Gallery article Passing user-supplied C++ functions by Dirk Eddelbuettel. For a detailed case study of the general approach see the RcppDE package (Eddelbuettel 2016) vignette at the RcppDE page on CRAN.

The user writes a C++ function to calculate log f(x). The current implementation in rust requires this function to have a particular structure: it must take a constant reference to an Rcpp::NumericVector, say x, a constant reference to an Rcpp::List, say pars, and return a double precision scalar. x is the argument x of f(x). pars is a list containing the values of parameters whose values are not specified inside the function. This allows the user to change the values of any parameters in the target density without editing the function. If there are no such parameters then the user must still include the argument pars in their function, even though the list provided to the function when it is called will be empty.

A simple way for the user to provide their C++ functions to create them in a file, say user_fns.cpp. Example content is provided below. The full file is available on the rust Github page. The functions in this file are compiled and made available to R, either using the Rcpp::sourceCpp function (e.g. Rcpp::sourceCpp("user_fns.cpp")) or using RStudio’s Source button on the editor toolbar. The example content below also includes the function create_xptr, which creates an external pointer to a C++ function. See Passing user-supplied C++ functions. It is this external pointer that is passed to ru_rcpp to perform ratio-of-uniforms sampling. If the user has written a C++ function, say new_name, then they need to add to create_xptr two lines of code:

else if (fstr == "new_name")  
  return(Rcpp::XPtr<funcPtr>(new funcPtr(&new_name))) ;

to create an external pointer for new_name using create_xptr. The following excerpt from the example user_fns.cpp file contains code for a standard normal density. Note that for this particular example we don’t need RcppArmadillo: we could replace #include <RcppArmadillo.h> with #include <Rcpp.h> and delete using namespace arma;. However, RcppArmadillo is used in the the multivariate normal example below and will be useful in many examples.

// [[Rcpp::depends(RcppArmadillo)]]

#include <RcppArmadillo.h>

using namespace arma;
using namespace Rcpp;

// [[Rcpp::interfaces(r, cpp)]]

// User-supplied C++ functions for logf.

// Note that currently the only interface available in rust is
// double fun(const Rcpp::NumericVector& x, const Rcpp::List& pars).
// However, as shown in the function logdmvnorm below RcppArmadillo
// functions can be used inside the function.

// Each function must be prefaced by the line: // [[Rcpp::export]]

// One-dimensional standard normal.

// [[Rcpp::export]]  
double logdN01(const Rcpp::NumericVector& x, const Rcpp::List& pars) {  
  return (-pow(x[0], 2.0) / 2.0) ;
}

// A function to create external pointers for any of the functions above.  
// See http://gallery.rcpp.org/articles/passing-cpp-function-pointers/  
// If you write a new function above called new_name then add the following
//
// else if (fstr == "new_name")  
//   return(Rcpp::XPtr<funcPtr>(new funcPtr(&new_name))) ;  

// [[Rcpp::export]]  
SEXP create_xptr(std::string fstr) {  
  typedef double (*funcPtr)(const Rcpp::NumericVector& x,  
                  const Rcpp::List& pars) ;  
  if (fstr == "logdN01")  
    return(Rcpp::XPtr<funcPtr>(new funcPtr(&logdN01))) ;  
  else  
    return(Rcpp::XPtr<funcPtr>(R_NilValue)) ;  
}  

// We could create the external pointers when this file is sourced using   
// the embedded R code below and/or (re)create them using create_xptr() in 
// an R session or R package..

/*** R
ptr_N01 <- create_xptr("logdN01")
*/

Examples : `ru_rcpp`

All the examples in the documentation for ru are replicated in the documentation for ru_rcpp. Here we consider a subset of the examples from the Introducing rust vignette, to illustrate how to provide user-supplied C++ functions to ru_rcpp and to compare the performances of ru and ru_rcpp. We use the microbenchmark package (Mersmann 2015) to make the comparison.

library(rust)
library(Rcpp)
# Is microbenchmark available?
got_microbenchmark <- requireNamespace("microbenchmark", quietly = TRUE)
if (got_microbenchmark) {
  library(microbenchmark)
}  
# Set the size of the simulated sample
n <- 1000

It is assumed that the user has already compiled their C++ functions and made them available to their R session, either using the Rcpp::sourceCpp function (e.g. Rcpp::sourceCpp("user_fns.cpp")) or using RStudio’s Source button on the editor toolbar.

Standard normal density

We start with a simple example: the (1-dimensional) standard normal density, based on the C++ function logdN01 in the example user_fns.cpp file above.

# Normal density ===================

# Create a pointer to the logdN01 C++ function
# (not necessary if this was created when the file of C++ functions was sourced)
ptr_N01 <- create_xptr("logdN01")

# Use ru and ru_rcpp starting from the same random number seed and check
# that the simulated values are the same.
set.seed(47)
x_old <- ru(logf = function(x) -x ^ 2 / 2, d = 1, n = n, init = 0.1)
head(x_old$sim_vals)
#>            [,1]
#> [1,]  0.7764728
#> [2,]  0.5310434
#> [3,] -0.1046049
#> [4,]  1.2111509
#> [5,]  1.1391379
#> [6,]  0.5180914
set.seed(47)
x_new <- ru_rcpp(logf = ptr_N01, d = 1, n = n, init = 0.1)
head(x_new$sim_vals)
#>            [,1]
#> [1,]  0.7764728
#> [2,]  0.5310434
#> [3,] -0.1046049
#> [4,]  1.2111509
#> [5,]  1.1391379
#> [6,]  0.5180914

# Compare performances of ru and ru_rcpp
if (got_microbenchmark) {
  res <- microbenchmark(
    old = ru(logf = function(x) -x ^ 2 / 2, d = 1, n = n, init = 0.1),
    new = ru_rcpp(logf = ptr_N01, d = 1, n = n, init = 0.1)
  )
  print(res, signif = 4)
}
#> Unit: milliseconds
#>  expr   min     lq   mean median     uq    max neval
#>   old 12.52 12.890 13.460  13.07 13.280 16.200   100
#>   new  1.54  1.617  1.704   1.64  1.677  4.392   100

As we would hope, ru_rcpp is faster than ru. If we start from the same random number seed we get the same simulated values from ru and ru_rcpp.

Multivariate normal density

To execute this example we add the following function to user_fns.cpp

// d-dimensional normal with zero-mean and covariance matrix sigma.

// [[Rcpp::export]]
double logdmvnorm(const Rcpp::NumericVector& x, const Rcpp::List& pars) {
  arma::mat sigma = as<arma::mat>(pars["sigma"]) ;
  arma::vec y = Rcpp::as<arma::vec>(x) ;
  double qform = arma::as_scalar(y.t() * arma::inv(sigma) * y) ;
  return -qform / 2.0  ;
}

and add

else if (fstr == "logdmvnorm")
  return(Rcpp::XPtr<funcPtr>(new funcPtr(&logdmvnorm))) ;

to the function create_xptr in user_fns.cpp.

# Three-dimensional normal with positive association ----------------
rho <- 0.9
covmat <- matrix(rho, 3, 3) + diag(1 - rho, 3)
# R function
log_dmvnorm <- function(x, mean = rep(0, d), sigma = diag(d)) {
  x <- matrix(x, ncol = length(x))
  d <- ncol(x)
  - 0.5 * (x - mean) %*% solve(sigma) %*% t(x - mean)
}
# Create a pointer to the logdmvnorm C++ function
ptr_mvn <- create_xptr("logdmvnorm")

if (got_microbenchmark) {
  res <- microbenchmark(
    old = ru(logf = log_dmvnorm, sigma = covmat, d = 3, n = n,
             init = c(0, 0, 0)), 
    new = ru_rcpp(logf = ptr_mvn, sigma = covmat, d = 3, n = n,
                  init = c(0, 0, 0))
  )  
  print(res, signif = 4)
}  
#> Unit: milliseconds
#>  expr     min      lq    mean  median      uq    max neval
#>   old 126.500 131.700 135.300 134.400 136.700 200.40   100
#>   new   6.704   6.944   7.328   7.072   7.204  12.22   100

Again, the improvement in speed obtained using Rcpp is clear.

Log-normal density after Box-Cox transformation

In this example we use a log transform (Box-Cox parameter λ = 0) so that the ratio-of-uniforms sampling is based on a normal distribution. The C++ function to calculate the log-density of a lognormal distribution is:

// Lognormal(mu, sigma).

// [[Rcpp::export]]
double logdlnorm(const Rcpp::NumericVector& x, const Rcpp::List& pars) {
  double mu = pars["mu"] ;
  double sigma = pars["sigma"] ;
  if (x[0] > 0)
    return -log(x[0]) - pow(log(x[0]) - mu, 2.0) / (2.0 * pow(sigma, 2.0)) ;
  else
    return R_NegInf ;
}

ptr_lnorm <- create_xptr("logdlnorm")
if (got_microbenchmark) {
  res <- microbenchmark(
   old = ru(logf = dlnorm, log = TRUE, d = 1, n = n, lower = 0, init = 0.1,
            trans = "BC", lambda = 0),
   new = ru_rcpp(logf = ptr_lnorm, mu = 0, sigma = 1, d = 1, n = n,
                 lower = 0, init = 0.1, trans = "BC", lambda = 0)
  )
  print(res, signif = 4)
}  
#> Unit: milliseconds
#>  expr    min     lq   mean median    uq     max neval
#>   old 26.450 27.350 29.390 28.000 30.13 100.000   100
#>   new  3.655  3.744  3.963  3.797  3.92   7.004   100

Generalized Pareto posterior density

The C++ function to calculate the log-posterior density is:

// Generalized Pareto posterior based on an MDI prior truncated to
// shape parameter xi >= -1.

// [[Rcpp::export]]
double loggp(const Rcpp::NumericVector& x, const Rcpp::List& ss) {
  Rcpp::NumericVector gpd_data = ss["gpd_data"] ;
  int m = ss["m"] ;
  double xm = ss["xm"] ;
  double sum_gp = ss["sum_gp"] ;
  if (x[0] <= 0 || x[1] <= -x[0] / xm)
    return R_NegInf ;
  double loglik ;
  Rcpp::NumericVector sdat = gpd_data / x[0] ;
  Rcpp::NumericVector zz = 1 + x[1] * sdat ;
  if (std::abs(x[1]) > 1e-6) {
    loglik = -m * log(x[0]) - (1.0 + 1.0 / x[1]) * sum(log(zz)) ;
  } else {
    double t1, t2, sdatj ;
    double total = 0;
    for(int j = 0; j < m; ++j) {
      sdatj = sdat[j] ;
      for(int i = 1; i < 5; ++i) {
        t1 = pow(sdatj, i) ;
        t2 = (i * sdatj - i - 1) ;
        total += pow(-1.0, i) * t1 * t2 * pow(x[1], i) / i / (i + 1) ;
      }
    }
    loglik = -m * log(x[0]) - sum_gp / x[0] - total ;
  }
  // MDI prior.
  if (x[1] < -1)
    return R_NegInf ;
  double logprior = -log(x[0]) - x[1] - 1 ;
  return (logprior + loglik) ;
}

We simulate some data from a Generalized Pareto distribution, calculate summary statistics involved in the likelihood and calculate an initial value in the search for the posterior mode.

set.seed(46)
# Sample data from a GP(sigma, xi) distribution
gpd_data <- rgpd(m = 100, xi = -0.5, sigma = 1)
# Calculate summary statistics for use in the log-likelihood
ss <- gpd_sum_stats(gpd_data)
# Calculate an initial estimate
init <- c(mean(gpd_data), 0)

Again we see that ru_rcpp is substantially faster than ru.

# Arguments for ru_rcpp
ptr_gp <- create_xptr("loggp")
for_ru_rcpp <- c(list(logf = ptr_gp, init = init, d = 2, n = n,
                 lower = c(0, -Inf)), ss)

if (got_microbenchmark) {
  res <- microbenchmark(
   old = ru(logf = gpd_logpost, ss = ss, d = 2, n = n, init = init,
            lower = c(0, -Inf)),
   new = do.call(ru_rcpp, for_ru_rcpp)
  )
  print(res, signif = 4)
}  
#> Unit: milliseconds
#>  expr    min     lq   mean median     uq   max neval
#>   old 47.240 50.720 52.600 52.870 54.030 57.02   100
#>   new  7.802  8.128  8.681  8.273  8.509 12.98   100

Examples : `find_lambda_one_d_rcpp` and `find_lambda_rcpp`

We repeat two examples from the Introducing rust vignette.

Gamma density: example for `find_lambda_one_d_rcpp`

We make use of the Rcpp sugar function dgamma.

// Gamma(alpha, 1).

// [[Rcpp::export]]
double logdgamma(const Rcpp::NumericVector& x, const Rcpp::List& pars) {
  double shp = pars["alpha"] ;
  return Rcpp::dgamma(x, shp, 1.0, 1)[0] ;
}

alpha <- 1
max_phi <- qgamma(0.999, shape = alpha)
ptr_gam <- create_xptr("logdgamma")
lambda <- find_lambda_one_d_rcpp(logf = ptr_gam, alpha = alpha,
                                 max_phi = max_phi)
lambda
#> $lambda
#> [1] 0.2727968
#> 
#> $gm
#> [1] 0.5689906
#> 
#> $init_psi
#> [1] -0.2016904
#> 
#> $sd_psi
#> [1] 0.7835109
#> 
#> $user_args
#> list()

Generalized Pareto posterior density: example for `find_lambda_rcpp`

In this example we supply an external pointer to a C++ function phi_to_theta that ensures that both parameters of the model are strictly positive, a requirement for the Box-Cox transformation to be applicable. The function phi_to_theta must have the same structure as the function used to calculate log f. See Providing a C++ function to ru_rcpp for details. See the Introducing rust vignette for the form of the transformation.

temp <- do.call(gpd_init, ss)
min_phi <- pmax(0, temp$init_phi - 2 * temp$se_phi)
max_phi <- pmax(0, temp$init_phi + 2 * temp$se_phi)

# Create external pointers
ptr_gp <- create_xptr("loggp")
ptr_phi_to_theta_gp <- create_phi_to_theta_xptr("gp")
# Note: log_j is set to zero by default inside find_lambda_rcpp()
lambda <- find_lambda_rcpp(logf = ptr_gp, ss = ss, d = 2, min_phi = min_phi,
                           max_phi = max_phi, user_args = list(xm = ss$xm),
                           phi_to_theta = ptr_phi_to_theta_gp)
lambda
#> $lambda
#> [1] 0.1624226 0.3678549
#> 
#> $gm
#> [1] 1.10542493 0.03225836
#> 
#> $init_psi
#> [1]  0.1054021 -0.2184344
#> 
#> $sd_psi
#>       Var1       Var2 
#> 0.12670792 0.02477219 
#> 
#> $phi_to_theta
#> <pointer: 0x55fe8dc9b0d0>
#> 
#> $log_j
#> <pointer: 0x55fe935fda10>
#> 
#> $user_args
#> $user_args$xm
#> [1] 1.846219

References

Eddelbuettel, D. 2013. Seamless R and C++ Integration with Rcpp. New York: Springer.

Eddelbuettel, D. 2016. RcppDE: Global Optimization by Differential Evolution in C++. https://CRAN.R-project.org/package=RcppDE.

Eddelbuettel, D., and R. Francois. 2011. “Rcpp: Seamless R and C++ Integration.” Journal of Statistical Software 40 (8): 1–18. doi:10.18637/jss.v040.i08.

Eddelbuettel, D., and C. Sanderson. 2014. “RcppArmadillo: Accelerating R with High-Performance C++ Linear Algebra.” Computational Statistics and Data Analysis 71: 1054–1063. doi:10.1016/j.csda.2013.02.005.

Mersmann, Olaf. 2015. microbenchmark: Accurate Timing Functions. https://CRAN.R-project.org/package=microbenchmark.

Rusting Faster: Simulation using Rcpp

Providing a C++ function to ru_rcpp

Examples : ru_rcpp