polyCub.exact.Gauss: Quasi-Exact Cubature of the Bivariate Normal Density...

In bastistician/polyCub: Cubature over Polygonal Domains

Quasi-Exact Cubature of the Bivariate Normal Density (DEFUNCT)

Description

This cubature method is defunct as of polyCub version 0.9.0. It relied on tristrip() from package gpclib for polygon triangulation, but that package did not have a FOSS license and was no longer maintained on a mainstream repository.
Contributions to resurrect this cubature method are welcome: an alternative implementation for constrained polygon triangulation is needed, see https://github.com/bastistician/polyCub/issues/2.

Usage

polyCub.exact.Gauss(polyregion, mean = c(0, 0), Sigma = diag(2),
  plot = FALSE)

Arguments

polyregion	a "gpc.poly" polygon or something that can be coerced to this class, e.g., an "owin" polygon (via owin2gpc), or an "sfg" polygon (via sfg2gpc).
mean, Sigma	mean and covariance matrix of the bivariate normal density to be integrated.
plot	logical indicating if an illustrative plot of the numerical integration should be produced. Note that the polyregion will be transformed (shifted and scaled).

Details

The bivariate Gaussian density can be integrated based on a triangulation of the (transformed) polygonal domain, using formulae from the Abramowitz and Stegun (1972) handbook (Section 26.9, Example 9, pp. 956f.). This method is quite cumbersome because the A&S formula is only for triangles where one vertex is the origin (0,0). For each triangle we have to check in which of the 6 outer regions of the triangle the origin (0,0) lies and adapt the signs in the formula appropriately: (AOB+BOC-AOC) or (AOB-AOC-BOC) or (AOB+AOC-BOC) or (AOC+BOC-AOB) or .... However, the most time consuming step is the evaluation of pmvnorm.

Value

The integral of the bivariate normal density over polyregion. Two attributes are appended to the integral value:

nEval	number of triangles over which the standard bivariate normal density had to be integrated, i.e. number of calls to pmvnorm and pnorm, the former of which being the most time-consuming operation.
error	Approximate absolute integration error stemming from the error introduced by the nEval pmvnorm evaluations. For this reason, the cubature method is in fact only quasi-exact (as is the pmvnorm function).

References

Abramowitz, M. and Stegun, I. A. (1972). Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables. New York: Dover Publications.

See Also

circleCub.Gauss for quasi-exact cubature of the isotropic Gaussian density over a circular domain.

Other polyCub-methods: polyCub.SV(), polyCub.iso(), polyCub.midpoint(), polyCub()

Examples

## a function to integrate (here: isotropic zero-mean Gaussian density)
f <- function (s, sigma = 5)
    exp(-rowSums(s^2)/2/sigma^2) / (2*pi*sigma^2)

## a simple polygon as integration domain
hexagon <- list(
    list(x = c(7.33, 7.33, 3, -1.33, -1.33, 3),
         y = c(-0.5, 4.5, 7, 4.5, -0.5, -3))
)

## quasi-exact integration based on gpclib::tristrip() and mvtnorm::pmvnorm()
## Not run: ## (this example requires gpclib)
hexagon.gpc <- new("gpc.poly", pts = lapply(hexagon, c, list(hole = FALSE)))
plotpolyf(hexagon.gpc, f, xlim = c(-8,8), ylim = c(-8,8))
print(polyCub.exact.Gauss(hexagon.gpc, mean = c(0,0), Sigma = 5^2*diag(2),
                          plot = TRUE), digits = 16)

## End(Not run)

bastistician/polyCub documentation built on Feb. 3, 2024, 5:41 p.m.