NUTS.jl
NUTS.jl takes No-U-Turn Sampler transitions for a log density and gradient you provide, using a step size and metric you choose.
Install from a terminal; this command changes your Julia environment and is not run by the documentation build.
sh
julia -e 'using Pkg; Pkg.add(url="https://github.com/nsiccha/NUTS.jl")'Take one transition for a standard-normal target:
julia
using NUTS, Random, LinearAlgebra
struct StandardNormal end
function NUTS.log_density_gradient!(::StandardNormal, q, gradient)
gradient .= -q
return -sum(abs2, q) / 2
end
state = (rng=Xoshiro(42), posterior=StandardNormal(),
position=zeros(2), stepsize=0.5, scale=I, squared_scale=I)
state = nuts!!(state)
(; position=round.(state.position; digits=3), state.divergent)(position = [0.69, -0.77], divergent = false)
The returned position is a draw after one transition. Reassign the returned state on each call; its arrays are reused.
Reading
| Task | Page |
|---|---|
| Define and check a target | Log density |
| Configure transitions and retain draws | Sampler state |
| Run and inspect a complete chain | Gaussian coordinates |
| Look up the interface | API |
The package supplies the transition kernel. It does not provide automatic differentiation, constrained-parameter transformations or warmup adaptation.