dhyperdirichlet: Probability density function for, and random sampling from,...
In hyperdirichlet: A Generalization of the Dirichlet Distribution

Description Usage Arguments Details Value Note Author(s) See Also Examples

Probability density function for the hyperdirichlet distribution in terms of either p or e; and random sampling using Metropolis-Hastings

1
2
3

dhyperdirichlet_e(e, HD, include.Jacobian = TRUE)
dhyperdirichlet(p, HD, include.NC = FALSE, TINY = 1e-10, log = FALSE)
rhyperdirichlet(n, HD, start=NULL, sigma=NULL)

`HD`	Object of class `hyperdirichlet`, or coerced thereto
`p`	Vector of length `dim(HD)`, notionally summing to one
`e`	Vector of length `dim(HD)` giving the point in `e`-space
`include.Jacobian`	In function `dhyperdiriclet_e()`, Boolean with default `TRUE` meaning to include the Jabobian of the transform from `e` to `p`
`include.NC`	In function `dhyperdirichlet_e()`, Boolean with `TRUE` meaning to include the normalization factor and default `FALSE` meaning not to include it (it is expensive to calculate). Note that if the normalizing factor is not known, the function will return `NA`
`TINY`	In function `dhyperdirichlet_p()`, numeric, specifying minimum size for elements of `p` via `p <- pmax(p , TINY)`
`log`	In function `dhyperdirichlet_p()`, Boolean with default `FALSE` meaning to return the probability density and `TRUE` meaning to return its logarithm
`n,start,sigma`	In function `rhyperdirichlet()`, `n` is the number of observations to take, `start` is the start-point for the random walk (with default `NULL` meaning to use the neutral point), and `sigma`is the standard deviation for the (Gaussian) kernel, with default `NULL` meaning to use 1/d

Function dhyperdirichlet() gives the density as a function of the p_1, ..., p_d.

Function dhyperdirichlet_e() gives the density as a function of the e_i. This is useful when integrating as the simplex (in p-space) transforms to a hypercube in e-space.

Functions dhyperdirichlet() and dhyperdirichlet_e() return a scalar; function rhyperdirichlet() returns a matrix whose rows are k-tuples

Function rhyperdirichlet() uses a Metropolis-Hastings algorithm to construct a Markov chain. Note that successive observations are not independent of one another. The details of this nonindependence are poorly understood (by me, at any rate). The examples section below shows how to generate a single random observation from the hyperdirichlet distribution.

Function dhyperdirichlet() silently normalizes p by p <- p/sum(p). Currently, no check for the elements being positive is made (negative elements are nullified by TINY).

The relationship between e and p is given in e_to_p.Rd.

Robin K. S. Hankin

maximum.likelihood,e_to_p

dhyperdirichlet(c(1,4,3,2)/10, dirichlet(1:4))

rhyperdirichlet(20, dirichlet(1:3))

diff(c(0,sort(runif(9)),1))  # random sample drawn from dirichlet(rep(1,10))

# how to draw a single observation from the hyperdirichlet:
f <- function(HD,n,...){rhyperdirichlet(n,HD=HD,...)[n,,drop=TRUE]}
f(dirichlet(1:3),n=100)
f(dirichlet(1:3),n=100)
f(dirichlet(1:3),n=100)
# Note, n=100 might not be enough burn-in.