IMPORTANT: this package is now abandoned, please use TransformVariables.jl instead.
Continuous transformations (or more precisely, homeomorphisms) from ℝ (and two-point compactified version) and ℝⁿ to various open (or closed) sets used in statistics and numerical methods, such as intervals, simplexes, ordered vectors.
Work in progress, API may change without notice.
This package was born because I was tired of coding the same transformations over and over, with occasional bugs, and wanted something well-tested.
Transformations defined by the package can be
logjac(transformation, x)method for the log Jacobian determinant,
inverse(transformation, x)method for the inverse.
Log jacobian determinants and their log are useful for domain transformations in MCMC, among other things.
In addition, the package includes types to represent intervals, and some basic methods of working with them. The concept of intervals is slightly different from IntervalSet.jl and ValidatedNumerics.jl, and as a result not compatible with either.
The convenience function
bridge(dom, img) figures out the right transformation from
img. Currently implemented for intervals.
using ContinuousTransformations t = bridge(ℝ, Segment(0.0, 3.0)) # will use a real-circle transformation, stretched t(0.0) # 1.5 inverse(t, 1.5) # ≈ 0.0 logjac(t, 0) # ≈ 0.405 image(t) # Segment(0.0, 3.0)
ArrayTransformation(transformation, dimensions...) transforms a vector of numbers to an array elementwise using
TransformationTuple(transformations) can be used for heterogeneous collections of transformations.
TransformLogLikelihood wraps a log likelihood function, and
TransformDistribution transforms a distribution. Both of them take care of the log Jacobian determinant adjustment.
Special transformations, useful for Bayesian methods, are also available (WIP). Feature and pull requests are appreciated.
Stan Development Team (2017). "Modeling Language User's Guide and Reference Manual, Version 2.17.0" (pdf)
Lewandowski, Daniel, Dorota Kurowicka, and Harry Joe. "Generating random correlation matrices based on vines and extended onion method." Journal of multivariate analysis 100.9 (2009): 1989–2001.
5 months ago