SBL prior normalization — Python companion

Python implementations of prior-normalizing (KR) transport maps for sparse Bayesian learning (SBL), packaged as a small library plus worked examples that plug into TRIPS-Py deblurring test problems.

This repository is meant as a public companion to:

• TRIPS-Py — regularization and test problems for inverse problems (here: 1D/2D deblurring).
• The paper Efficient sampling for sparse Bayesian learning using hierarchical prior normalization (Glaubitz & Marzouk, 2025) and its reference Julia code.

The Julia repository remains the authoritative source for the full numerical study; this repo focuses on a minimal, reusable Python core (prior_normalization) and two scripts that combine TRIPS-Py forward models with Tikhonov/GCV and prior-normalizing MAP estimation.

Contents

Path	Description
src/prior_normalization/	Chebyshev-built Φ and Knothe–Rosenblatt inverse map utilities used by the examples.
examples/deblurring_1d/run_deblurring_1d.py	1D deblurring (Deblurring1D): whitening via first differences, four SBL hyperparameter sets from the paper.
examples/deblurring_2d/run_deblurring_2d.py	2D deblurring (Deblurring2D): pixel-domain MAP; requires TRIPS-Py demo image .mat files (see below).
examples/bootstrap.py	Resolves prior_normalization on sys.path and locates TRIPS-Py (TRIPS_PY_ROOT or common clone locations).

Reference figures shipped with the project live under each example’s figures/reference_*.png. Running an example writes figures/results_deblurring_*.png (ignored by git unless you choose to commit them).

Citation

If you use this code or the underlying method, cite the article:

@article{glaubitz2025efficient,
  title   = {Efficient sampling for sparse Bayesian learning using hierarchical prior normalization},
  author  = {Glaubitz, Jan and Marzouk, Youssef},
  journal = {arXiv preprint arXiv:2505.23753},
  year    = {2025},
  month   = {05}
}

Also cite TRIPS-Py if you use its operators and test problems; see that repository for their preferred reference.

Requirements

• Python 3.9+
• NumPy, SciPy, Matplotlib, PyLops (declared in pyproject.toml)
• A checkout of TRIPS-Py that includes trips.test_problems (clone from GitHub; PyPI packages may not match this layout).

Installation

From the repository root:

git clone https://github.com/<your-org>/SBL-PriorNormalization-Python.git
cd SBL-PriorNormalization-Python
python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e .
pip install -r requirements.txt   # optional lock-style install without extras

Point the examples at TRIPS-Py

Clone TRIPS-Py somewhere on your machine, then either:

export TRIPS_PY_ROOT=/absolute/path/to/trips-py

or place the clone at ~/Documents/trips-py or ~/trips-py, which examples/bootstrap.py discovers automatically.

The 2D script temporarily sets the process working directory to <trips-py>/demos so Deblurring2D can load ./data/image_data/*.mat; the prior directory is restored afterward.

Running the examples

Run from any working directory; use the repo’s Python so editable prior_normalization resolves:

export TRIPS_PY_ROOT=/path/to/trips-py   # if not auto-discovered
export MPLBACKEND=Agg                    # optional: non-interactive plotting

python examples/deblurring_1d/run_deblurring_1d.py
python examples/deblurring_2d/run_deblurring_2d.py

2D runtime: The full default settings (24×24 unknowns, four SBL models) involve large-scale optimization and can take a long time. For a smaller dry run (e.g. CI or laptops):

export SBL_EXAMPLE_QUICK=1
python examples/deblurring_2d/run_deblurring_2d.py

This uses a coarser grid and a single SBL model with a reduced iteration budget (see the script). Omit SBL_EXAMPLE_QUICK for results comparable to the shipped reference PNGs.

Scope and disclaimer

This repository does not reproduce every experiment from the Julia paper; it provides the prior-normalization machinery in Python and two TRIPS-Py deblurring demonstrations (Tikhonov/GCV vs prior-normalizing MAP). The code is provided as-is for research and education.

License

MIT — see LICENSE.

CMDA capstone (Virginia Tech)

This work was carried out as a Computational Modeling and Data Analytics (CMDA) capstone project at Virginia Tech.

Role	People
Project team	Daniel Miles, Ash Vadicherla, Jun Young Kwon
Project sponsor	Dr. Mirjeta Pasha

Acknowledgments

The project builds on research by Jan Glaubitz and Youssef Marzouk and on the TRIPS-Py software ecosystem.