GWFrames

Manipulate gravitational waveforms — changing frames, and so on.

The code in this project extends code written for the paper "Angular velocity of gravitational radiation from precessing binaries and the corotating frame", giving an explicit implementation of the methods discussed in it. The first commit of this project is the last commit of the paper's project.

The license for using this software is basically open (see the LICENSE file in this directory), though citations to the original paper are appreciated, where relevant. Also, if your work depends on features found in this module that have not been described in a separate publication of mine, I would appreciate the opportunity to be a coauthor.

Note that this code is not optimized. In many cases, obvious optimizations were rejected in favor of clearer or simpler code, or code that simply reflected the discussion in the paper more directly. The code is generally more than fast enough for interactive use on individual waveforms, but there is plenty of room for improvement. In particular, rotations may be painfully slow for large data sets.

Installation

GWFrames requires Python 3.11 or newer, on Linux or macOS. Binary wheels for Linux and macOS (x86_64 and arm64) are published on PyPI.

The HDF5 functions (ReadFromH5, OutputToH5, ReadFromNRAR, and OutputToNRAR) and the GWFrames.Extrapolation module need h5py. PyPI has no h5py wheels for Python 3.15, and there may never be any, so on Python 3.15 and newer GWFrames is installed without h5py; everything else works without it. If you need those functions, either use Python 3.14 or older with uv or pip, or get h5py from conda-forge with pixi or mamba, which do build it for Python 3.15.

With uv or pip

uv add GWFrames          # in a uv project
uv pip install GWFrames  # or in any virtual environment
pip install GWFrames     # or with pip

With uv, adding --python 3.14 to uv init or uv venv selects a Python for which h5py is available.

To install the latest code from GitHub, or on any platform without a wheel, GWFrames is compiled from source. That needs a C++ compiler, GSL, and FFTW, which are available from most package managers:

sudo apt-get install libgsl-dev libfftw3-dev   # Debian or Ubuntu
brew install gsl fftw                          # macOS with Homebrew

after which

uv pip install "git+https://github.com/moble/GWFrames"

builds and installs it. (The submodules are fetched automatically.) If GSL or FFTW is installed somewhere unusual, set CMAKE_PREFIX_PATH to its prefix.

With pixi

pixi can take NumPy, SciPy, and h5py from conda-forge and GWFrames itself from PyPI:

pixi init my-project
cd my-project
pixi add python numpy scipy h5py
pixi add --pypi GWFrames
pixi run python -c 'import GWFrames; print(GWFrames.__version__)'

With mamba or conda

Similarly, the dependencies can come from conda-forge and GWFrames from PyPI:

mamba create -n gwframes -c conda-forge python numpy scipy h5py
mamba activate gwframes
python -m pip install GWFrames

To build the latest code from GitHub instead, all of the build requirements are available from conda-forge:

mamba create -n gwframes -c conda-forge python numpy scipy h5py gsl fftw \
    cxx-compiler cmake ninja pkg-config pybind11 scikit-build-core
mamba activate gwframes
python -m pip install --no-deps --no-build-isolation "git+https://github.com/moble/GWFrames"

On Apple Silicon, make sure that the environment is native: mamba info should show platform : osx-arm64. Older instructions for GWFrames recommended Intel environments (with CONDA_SUBDIR=osx-64 or a subdir: osx-64 setting), which run under Rosetta. The first build in such an environment can fail with "Accessing CMake timed out" while Rosetta translates CMake; running the same command again works.

Checking the installation

python -c 'import GWFrames; W = GWFrames.PNWaveform("TaylorT4", 0.0, [0., 0., 0.9], [0.0, 0.0, 0.9], 0.01); print("That worked!")'

Usage

The primary objects are Waveform and PNWaveform, with various methods defined for each; quaternions and functions of arrays of quaternions are in GWFrames.Quaternions. Detailed documentation of most functions may be found through Python's help function, as in help(GWFrames.Waveform). A Jupyter notebook in the Docs directory contains extensive examples, following the outline of the paper.

Changes from the SWIG-based versions

Version 2 replaces the SWIG bindings with pybind11, and the old setup.py with CMake and scikit-build-core. The Python interface is almost entirely unchanged, apart from these differences:

• The separate Quaternions module is now GWFrames.Quaternions, so import Quaternions becomes from GWFrames import Quaternions, and GWFrames.Quaternion is also available.
• Methods that modify a Waveform in place and return it (such as TransformToCorotatingFrame) return the very same Python object, including for PNWaveform.

Several bugs were fixed along the way, notably that Quaternion objects compared equal exactly when they were not equal, and that Hybridize read past the end of an array and could not handle waveforms in an inertial frame.

Development

Clone the repository with its submodules, then build and test it with uv:

git clone --recursive https://github.com/moble/GWFrames.git
cd GWFrames
uv sync --no-editable
uv run --no-sync pytest

The C++ code is in cpp/, the pybind11 bindings in cpp/bindings/, and the Python package in src/GWFrames/. The submodules Quaternions, SphericalFunctions, SpacetimeAlgebra, and PostNewtonian, along with a vendored copy of spinsfast, are in extern/. The tests include a comparison with results recorded from the last SWIG-based build; see tests/reference/generate_swig_reference.py.

Every push to main that passes the wheel builds and tests is released automatically: the version is bumped, tagged, released on GitHub, and published to PyPI. The bump is a patch release unless a commit message in the push contains #minor or #major; adding #dev makes a development release (such as 2.1.0.dev0), and a later plain push completes it. A head commit message containing [no release] builds and tests without releasing, and [skip ci] skips CI entirely.

Note on licensing: spinsfast, GSL, and FFTW are licensed under the GNU GPL, and the binary wheels link all three statically, so the wheels as a whole are covered by the GPL.

Contributions

The code is primarily written by me (Mike Boyle). Dan Hemberger helped by porting some of my older code from Triton to perform noise-weighted overlap calculations, and with numerous bug reports and helpful suggestions. Serguei Ossokine also helped substantially by cross-checking the post-Newtonian formulas and results. Dante Iozzo helped to keep things working as this code entered its dotage. Claude (Opus 5.5) resuscitated the package for Python 3.