Diffusion differentiable resampling

This repository features implementation for the paper "Diffusion differential resampling". In this paper, we have introduced a new differentiable-by-construction, consistent, and informative resampling method via diffusion models. Check out our preprint for details at https://arxiv.org/abs/2512.10401.

Installation

Install the package via a standard procedure.

git clone git@github.com:zgbkdlm/diffres.git
cd diffres

python -m venv .venv
python .venv/bin/activate
pip install -e .

Example

We have provided a few examples for demonstration in folder ./demos, e.g., the Jupyter Notebook gaussian_mixture.ipynb.

Quick start

Suppose that you have weighted samples and wish to do differentiable resampling, the simplest code using the diffusion resampling would look like this:

import jax
from diffres.resampling import diffusion_resampling

key = jax.random.PRNGKey(666)

# The given weighted samples
samples = ...
log_ws = ...

# Resampling parameters
a = ...
ts = ...

# Resampling
_, resamples = diffusion_resampling(key, log_ws, samples, a, ts)

Reproduce experiments

The folder ./experiments contain all the files that we use to generate the results. Each bash file is associated with one experiment. Check the .sh files to see how we ran the experiments, and you can easily adapt them in your local machine. Running them will exactly reproduce our results in the paper.

After running the experiments, you can then run the scripts in ./experiments/summary to print and plot the results. These scripts also guarantee to exactly reproduce the tables and figures in the paper.

Citation

Please cite using the following bibtex.

@article{Andersson2025diffres, 
    author = {Andersson, Jennifer R. and Zhao, Zheng}, 
    title = {Diffusion differentiable resampling},
    journal = {arXiv preprint arXiv:2512.10401},
    year = {2025},
}

License

GPL v3 or later. For those aiming proprietarisation, please invest your own time to re-implement the algorithm by yourself.

Contact

Zheng Zhao, Linköping University, https://zz.zabemon.com. Jennifer R. Andersson, Uppsala University.