LogPSplinePSD 0.1.0

LogPsplines in JAX

LogPSplinePSD estimates power spectral densities (PSDs) with Bayesian log-P-splines. It supports univariate and multivariate time series, fits smooth spectral matrices with NumPyro/JAX, and returns ArviZ-compatible xarray.DataTree outputs for diagnostics and plotting.

Highlights

• Log-domain P-spline models for positive PSDs.

• Multivariate Wishart likelihoods for spectral matrices.

• VI warm starts and factorised multivariate NUTS.

• Optional frequency-domain coarse graining.

• Posterior PSD quantiles, coherence summaries, and diagnostic plots.

Install

For development, use the repository virtual environment:

source .venv/bin/activate
python -m pip install -e '.[dev]'

For package use:

python -m pip install LogPSplinePSD

Quick Example

from log_psplines.example_datasets.varma_data import VARMAData
from log_psplines.mcmc import run_mcmc
from log_psplines.pipeline.config import PipelineConfig

data = VARMAData(n_samples=256, fs=64.0, seed=7)

idata = run_mcmc(
    data.ts,
    PipelineConfig(
        n_knots=6,
        n_warmup=50,
        n_samples=100,
        vi_steps=200,
        outdir="runs/varma_quickstart",
    ),
)

Documentation

Build the docs locally with:

source .venv/bin/activate
.venv/bin/jupyter-book build docs

The public docs focus on package usage, configuration, outputs, API reference, and implementation notes. Domain-specific examples are intentionally kept out of the main docs for now and can be added later as separate studies.

References

Eilers, P. H. C., & Marx, B. D. (1996). Flexible smoothing with B-splines and penalties. Statistical Science, 11(2), 89-121.

Maturana-Russel, J., & Meyer, R. (2021). P-spline spectral density estimation with a discrete penalty. arXiv:1905.01832.