This was an early draft. It has been given up in favour of DensityInterface.jl.
LogDensity.jl
The simple purpose of this package is to provide the function logdensity to other packages, just
like StatsBase.jl provides loglikelihood, and several packages have logpdf. The intent is for
logdensity to allow more general semantics than the former, such as densities of measures, or
density evaluation for trace-based probabilistic programming systems.
The primary function is
logdensity(μ, x)where μ is “measure-like”, i.e., associated with volume calculation via integration:
μ(A) = ∫_A exp(logdensity(μ, x)) dx
This covers MeasureTheory.jl AbstractMeasures,
Distributions.jl distributions (which are
proability measures), and various implementations of measures through probabilistic programs, for
which logdensity might be defined over a model and a trace data type or a similar specification of
an instantionation of some random variables.
Exported functions
logdensity(μ, x)density(args..) = exp(logdensity(args...))
The latter is defaulted, since in probabilistic context, the log-space implementation is usually preferred, and the “normal-space” variant derived from it. You can of course overload the latter, too, if that is more efficient.