code-numerically-robust-fixed-point-smoother

This repository contains the code for the paper

Krämer, Nicholas. "Numerically Robust Fixed-Point Smoothing Without State Augmentation." Transactions on Machine Learning Research (2025).

Here is a bibtex entry:

@article{
  kramer2025numerically,
  title={Numerically Robust Fixed-Point Smoothing Without State Augmentation},
  author={Nicholas Kr{\"a}mer},
  journal={Transactions on Machine Learning Research},
  issn={2835-8856},
  year={2025},
  url={https://openreview.net/forum?id=LVQ8BEL5n3},
  note={}
}```


## Warning 
This is experiment code.
But if you want to work with this repository, proceed as follows.


## Installation

We use Python 3.10 for all experiments.
Other versions might also work.

First, ensure that JAX is installed.
Then, run
```commandline
pip install .

which installs the source code plus all dependencies.

Experiments

To run the experiments, execute (for instance)

python experiments/estimate_parameters.py

or run all experiments via

make run-experiments

To turn the results into the tables from the Paper, execute the scripts in from_results_to_paper/*. The scripts' names match the experiments' names, for example,

python from_results_to_paper/measure_robustness.py

The parameter estimation experiment plots result in the experiment script.

Using the code

Everything is contained in a single module. To use it, and after installation, import

from fpx import fpx

print(help(fpx))

and access all code via fpx.* ("fpx" stands for "fixed-point smoothing in JAX"). Consult the test file in tests/test_fpx.py for examples.

You may also run mkdocs serve to get a list of all types and functions.