SwiG: Nested Sampling with Slice-within-Gibbs

SwiG implements NS-SwiG (Nested Sampling with Slice-within-Gibbs), an algorithm for efficient nested sampling of hierarchical Bayesian models.

SwiG reduces the per-replacement cost of nested sampling for hierarchical models from O(J²) to O(J), with O(1) likelihood evaluation per local block update via budget constraint decomposition. It is implemented in JAX with GPU acceleration, built on the BlackJAX nested sampling infrastructure.

Installation

uv sync

To run the examples (adds matplotlib and distrax):

uv sync --extra examples

Examples

Script	Description
uv run python examples/funnel.py	10D funnel with analytic logZ validation
uv run python examples/glm.py	Hierarchical Gaussian with analytic evidence
uv run python examples/radon.py	Radon contextual effects (Minnesota, 85 counties)
uv run python examples/sv.py	Stochastic volatility (S&P 500, Markov variant)

Citation

If you use SwiG in your research, please cite:

@misc{yallup2026swig,
  title   = {Nested Sampling with Slice-within-Gibbs: Efficient Evidence
             Calculation for Hierarchical Bayesian Models},
  author  = {David Yallup},
  year    = {2026},
  eprint  = {2602.17414},
  archivePrefix = {arXiv},
  primaryClass  = {stat.CO},
  url     = {https://arxiv.org/abs/2602.17414},
}

SwiG builds on the following works:

@misc{yallup2026nss,
  title   = {Nested Slice Sampling: Vectorized Nested Sampling for
             GPU-Accelerated Inference},
  author  = {David Yallup and Namu Kroupa and Will Handley},
  year    = {2026},
  eprint  = {2601.23252},
  archivePrefix = {arXiv},
  primaryClass  = {stat.CO},
  url     = {https://arxiv.org/abs/2601.23252},
}

@misc{cabezas2024blackjax,
  title   = {BlackJAX: Composable {B}ayesian inference in {JAX}},
  author  = {Alberto Cabezas and Adrien Corenflos and Junpeng Lao
             and R\'{e}mi Louf},
  year    = {2024},
  eprint  = {2402.10797},
  archivePrefix = {arXiv},
  primaryClass  = {cs.MS},
}

License

Apache 2.0