pylibstats 0.3.3

Python bindings for libstats — SIMD-accelerated statistical distributions via NumPy

pylibstats

Python bindings for libstats — a C++20 statistical distributions library with SIMD batch operations.

Features

• 19 distributions: Gaussian, Exponential, Uniform, Poisson, Discrete Uniform, Gamma, Chi-Squared, Log-Normal, Student's t, Beta, Weibull, Rayleigh, Pareto, Von Mises, Binomial, Negative Binomial, Geometric, Laplace, Cauchy
• NumPy integration: pass arrays directly to pdf(), cdf(), log_pdf() — the SIMD/parallel batch path runs automatically
• GIL-releasing: batch operations release the Python GIL for concurrent workloads
• SciPy-compatible naming: pdf, cdf, ppf, fit, sample
• Input validation: all constructor, setter, and fit() parameters are validated in Python with clear ValueError messages

Quick start

import numpy as np
import pylibstats

dist = pylibstats.Gaussian(mu=0.0, sigma=1.0)

# Scalar
dist.pdf(1.0)
dist.cdf(0.0)        # 0.5
dist.ppf(0.975)      # ~1.96

# Batch (SIMD-accelerated)
x = np.linspace(-4, 4, 100_000)
densities = dist.pdf(x)

# Sampling
samples = dist.sample(n=10_000, seed=42)

# Fitting
dist.fit(samples)

Building from source

Requires Python ≥3.11, CMake ≥3.20, and a C++20 compiler.

pip install .

This fetches libstats v2.0.3 via CMake FetchContent if not already installed.

Building against a local libstats

To link against a locally built libstats (e.g. a development branch), install libstats to a prefix and point pip at it:

# In the libstats repo
cmake --install build --prefix /path/to/libstats/install

# In this repo — use libstats_DIR, not CMAKE_PREFIX_PATH
# (overriding CMAKE_PREFIX_PATH breaks nanobind discovery)
pip install --no-build-isolation -ve . \
    -Ccmake.define.libstats_DIR=/path/to/libstats/install/lib/cmake/libstats

--no-build-isolation requires build deps in the active environment:

pip install "scikit-build-core>=0.10" "nanobind>=2.0"

Running tests

pip install ".[test]"
pytest

Examples

See the examples/ directory:

• basic_usage.py — scalar/batch operations, sampling, and fitting
• benchmark.py — wall-clock comparison against SciPy (PDF and CDF)
• scipy_comparison.py — numerical accuracy verification across all 9 distributions

Known limitations

• Beta CDF performance: the regularised incomplete beta function in libstats is slower than SciPy's implementation (~0.5× speedup). All other distribution/operation combinations are faster.

Contributing

macOS ABI note

On macOS, libstats may be compiled with Homebrew LLVM while Python extensions use Apple clang. These ship different libc++ versions whose exception-handling ABIs are incompatible — C++ exceptions thrown from libstats segfault during stack unwinding instead of propagating normally.

pylibstats works around this by validating all parameters in pure Python (in __init__.py) before calling into the C++ layer, so the error path never crosses the ABI boundary. If you add new parameters or distribution classes, follow the same pattern: validate in Python, then delegate to _core.

See libstats/include/core/error_handling.h for the upstream discussion.

License

MIT