ars (version 0.8)

ars: Adaptive Rejection Sampling

Description

Adaptive Rejection Sampling from log-concave density functions

Usage

ars(n=1,f,fprima,x=c(-4,1,4),ns=100,m=3,emax=64,lb=FALSE,ub=FALSE,xlb=0,xub=0,...)

Value

a sampled value from density

Arguments

n

sample size

f

function that computes log(f(u,...)), for given u, where f(u) is proportional to the density we want to sample from

fprima

d/du log(f(u,...))

x

some starting points in wich log(f(u,...) is defined

ns

maximum number of points defining the hulls

m

number of starting points

emax

large value for which it is possible to compute an exponential

lb

boolean indicating if there is a lower bound to the domain

xlb

value of the lower bound

ub

boolean indicating if there is a upper bound to the domain

xub

value of the upper bound bound

...

arguments to be passed to f and fprima

Author

Paulino Perez Rodriguez, original C++ code from Arnost Komarek based on ars.f written by P. Wild and W. R. Gilks

Details

ifault codes, subroutine initial

0:

successful initialisation

1:

not enough starting points

2:

ns is less than m

3:

no abscissae to left of mode (if lb = false)

4:

no abscissae to right of mode (if ub = false)

5:

non-log-concavity detect

ifault codes, subroutine sample

0:

successful sampling

5:

non-concavity detected

6:

random number generator generated zero

7:

numerical instability

References

Gilks, W.R., P. Wild. (1992) Adaptive Rejection Sampling for Gibbs Sampling, Applied Statistics 41:337--348.

Examples

library(ars)

#Example 1: sample 20 values from the normal distribution N(2,3)
f<-function(x,mu=0,sigma=1){-1/(2*sigma^2)*(x-mu)^2}
fprima<-function(x,mu=0,sigma=1){-1/sigma^2*(x-mu)}
mysample<-ars(20,f,fprima,mu=2,sigma=3)
mysample
hist(mysample)

#Example 2: sample 20 values from a gamma(2,0.5)
f1<-function(x,shape,scale=1){(shape-1)*log(x)-x/scale}
f1prima<-function(x,shape,scale=1) {(shape-1)/x-1/scale}
mysample1<-ars(20,f1,f1prima,x=4.5,m=1,lb=TRUE,xlb=0,shape=2,scale=0.5)
mysample1
hist(mysample1)

#Example 3: sample 20 values from a beta(1.3,2.7) distribution
f2<-function(x,a,b){(a-1)*log(x)+(b-1)*log(1-x)}
f2prima<-function(x,a,b){(a-1)/x-(b-1)/(1-x)}
mysample2<-ars(20,f2,f2prima,x=c(0.3,0.6),m=2,lb=TRUE,xlb=0,ub=TRUE,xub=1,a=1.3,b=2.7)
mysample2
hist(mysample2)