hitro.new: UNU.RAN generator based on Hit-and-Run sampler (HITRO)

Description

UNU.RAN random variate generator for continuous multivariate distributions with given probability density function (PDF). It is based on the Hit-and-Run algorithm in combinaton with the Ratio-of-Uniforms method (HITRO).

[Universal] -- MCMC (Markov chain sampler).

Usage

hitro.new(dim=1, pdf, ll=NULL, ur=NULL, mode=NULL, center=NULL,
          thinning=1, burnin=0, ...)

Arguments

dim

number of dimensions of the distribution. (integer)

pdf

probability density function. (Rfunction)

ll,ur

lower left and upper right vertex of a rectangular domain of the pdf. The domain is only set if both vertices are not NULL. Otherwise, the domain is unbounded by default. (numeric vectors)

mode

location of the mode. (numeric vector)

center

point in typical region of distribution, e.g. the approximate location of the mode. If omitted the mode is used. If the mode is not given either, the origin is used. (numeric vector)

thinning

thinning factor. (positive integer)

burnin

length of burnin-in phase. (positive integer)

...

(optional) arguments for pdf

Details

Beware: MCMC sampling can be dangerous! This function creates a unuran object based on the Hit-and-Run algorithm in combinaton with the Ratio-of-Uniforms method (HITRO). It can be used to draw samples of a continuous random vector with given probability density function using ur.

The algorithm works best with log-concave distributions. Other distributions work as well but convergence can be slower. The density must be provided by a function pdf which must return non-negative numbers and but need not be normalized (i.e., it can be any multiple of a density function). The center is used as starting point of the Hit-and-Run algorithm. It is thus important, that the center is contained in the (interior of the) domain. Alternatively, one could provide the location of the mode. However, this requires its exact position whereas center allows any point in the typical region of the distribution.

If the mode is given, then it is used to obtain an upper bound on the pdf and thus its location should be given sufficiently accurate. The HITRO algorithm is a MCMC samplers and thus it does not produce a sequence of independent variates. The drawn sample follows the target distribution only approximately. The dependence between consecutive vectors can be decreased when only a subsequence is returned (and the other elements are erased). This is called thinning of the Markov chain and can be controlled by the thinning factor. A thinning factor $k$ means that only every $k$-th element is returned.

Markov chains also depend on the chosen starting point (i.e., the center in this implementation of the algorithm). This dependence can be decreased by erasing the first part of the chain. This is called the burn-in of the Markov chain and its length is controlled by the argument burnin.

References

R. Karawatzki, J. Leydold, and K. P"otzelberger (2005): Automatic Markov Chain Monte Carlo Procedures for Sampling from Multivariate Distributions. Research Report Series / Department of Statistics and Mathematics, Nr. 27, December 2005 Department of Statistics and Mathematics, Wien, Wirtschaftsuniv., 2005. http://epub.wu.ac.at/dyn/virlib/wp/showentry?ID=epub-wu-01_8cb

Examples

Run this code

## Create a sample of size 100 for a 
## Gaussian distribution
mvpdf <- function (x) { exp(-sum(x^2)) }
gen <- hitro.new(dim=2, pdf=mvpdf)
x <- ur(gen,100)

## Use mode of Gaussian distribution.
## Reduce auto-correlation by thinning and burn-in.
##  mode at (0,0)
##  thinning factor 3
##    (only every 3rd vector in the sequence is returned)
##  burn-in of length 1000
##    (the first 100 vectors in the sequence are discarded)
mvpdf <- function (x) { exp(-sum(x^2)) }
gen <- hitro.new(dim=2, pdf=mvpdf, mode=c(0,0), thinning=3, burnin=1000)
x <- ur(gen,100)

## Gaussian distribution restricted to the rectangle [1,2]x[1,2]
##  (don't forget to provide a starting point using 'center')
mvpdf <- function (x) { exp(-sum(x^2)) }
gen <- hitro.new(dim=2, pdf=mvpdf, center=c(1.1,1.1), ll=c(1,1), ur=c(2,2))
x <- ur(gen,100)