python-KernSmooth

An AI-assisted port of the R package KernSmooth (v. 2.23-26) to Python, together with the complete analysis, conversion, and validation pipeline used to produce it.

KernSmooth implements the kernel-smoothing methods of Wand, M.P. and Jones, M.C. (1995), Kernel Smoothing, Chapman and Hall, and ships as a recommended package with every standard R installation. This repository documents and executes a systematic, dependency-order translation of KernSmooth into an installable Python package, r2py_kernsmooth, which compiles the original Fortran 77 computational core verbatim via numpy.f2py and replaces the R wrapper layer with semantically equivalent, type-annotated Python functions built on NumPy and SciPy.

A full account of the project's methodology, findings, and validation results is given in the technical report at docs/report/main.tex (compiled PDF: docs/report/main.pdf) and in the accompanying paper at docs/paper/main.tex.

Repository Structure

Path	Contents
KernSmooth/	Unmodified R package source (v. 2.23-26).
r2py_kernsmooth/	The installable Python package; also mirrored as a standalone repository at github.com/r2py-project/r2py_kernsmooth via git subtree.
structural_analysis/	Dependency-graph artefacts and per-function JSON analysis of the R source.
language_dependency_analysis/	Per-file R-to-Python language-dependency tables and the generated conversion guides.
conversion_results/	Per-function JSON translation artefacts produced by the conversion pipeline.
docs/	Architecture document, planning materials, per-phase summaries, the technical report, and the paper.
.claude/agents/, .claude/commands/	Sub-agent and skill specifications for the Claude Code-based conversion workflow.
git_pull.sh, git_push.sh	Cluster batch scripts synchronising this repository with the r2py_kernsmooth subtree remote.
install_environments.sh	Cluster batch script provisioning the r-to-python conda environment.

Conversion Methodology

The port was carried out in seven phases, each documented in detail in the technical report:

1. Static analysis of the R package's function-level dependency structure and Fortran integration conventions.
2. Python build infrastructure: a meson-python build backend compiling the original Fortran sources via f2py, with cross-platform BLAS discovery and cibuildwheel-based PyPI distribution.
3. Language-dependency cataloguing: systematic extraction of every R standard-library call site and generation of a dedicated Python translation guide for each, to avoid silent numerical discrepancies (e.g. FFT normalisation, sample-variance denominators, integer width).
4. Automated function conversion, in dependency order, from R to Python, followed by package assembly and correction of R-versus-f2py return-value semantics.
5. Code-quality audit: type-annotation completion, namespace hygiene, dead-code removal, and alignment of error and warning messages with the R source.
6. Regression test infrastructure: R test scripts ported to pytest/rpy2-based assertions that compare live output against the R reference implementation.
7. Comprehensive test suite construction: positive, negative, and edge-case tests generated for all seven public functions, closing out with 518 passing tests and zero failures against R KernSmooth 2.23 via rpy2.

Installation

The Python package can be installed independently:

pip install r2py_kernsmooth

For local development from this repository (requires a Fortran compiler and a BLAS implementation such as OpenBLAS):

cd r2py_kernsmooth
pip install --no-build-isolation .

See r2py_kernsmooth/README.md for package usage and the list of public functions.

Testing

python -m pytest r2py_kernsmooth/tests/ -q

The test suite requires rpy2 and an R installation with the KernSmooth and carData packages, since tests assert numerical agreement against live R reference values.

License

The original content of this repository is distributed under an "Unlimited" license, matching the terms of upstream KernSmooth. See LICENSE for details. The vendored KernSmooth/ subdirectory retains its own upstream license (see KernSmooth/DESCRIPTION and KernSmooth/LICENCE.note), and r2py_kernsmooth/ carries its own LICENSE/NOTICE pair for standalone distribution.

Attribution

KernSmooth was originally authored by Matt Wand, with LINPACK Fortran routines contributed by Cleve Moler and the R-language port maintained by Brian Ripley. The Python port and this project's conversion pipeline are authored by Yufei Cai (ycai9@nd.edu) and Jun Li (jun.li@nd.edu).