This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| celerite2 | |
|---|---|
| Name | celerite2 |
| Programming language | Python, C++ |
| Operating system | Cross-platform |
| License | MIT License |
celerite2
celerite2 is a software library for fast Gaussian process modeling tailored to one-dimensional time series. It extends techniques developed in Gaussian process literature and builds on computational advances popularized in projects such as SciPy and NumPy. The library is used in studies published in venues like the Astrophysical Journal and presented at conferences including NeurIPS and AAS meetings.
celerite2 implements a class of scalable Gaussian process kernels optimized for irregularly sampled temporal data common in Kepler and TESS time series analyses. The package targets astronomers working with datasets from facilities such as the Hubble Space Telescope, James Webb Space Telescope, and ground-based observatories like Subaru Telescope and Very Large Telescope. It follows software design practices promoted by communities around GitHub, Conda, and PyPI. Key contributors and users include researchers affiliated with institutions such as Harvard University, University of California, Berkeley, and Princeton University.
The mathematical core uses linear algebra techniques drawn from classical works by Carl Friedrich Gauss and modern expositions like those found in texts by Andrew Ng and Christopher Bishop. The method represents covariance functions as sums of exponentials, enabling semi-analytic forms of the covariance matrix inverse and determinant computed using algorithms inspired by Cholesky decomposition and the Sherman–Morrison–Woodbury formula. The formulation connects to stochastic processes described by authors such as Norbert Wiener and Andrey Kolmogorov and leverages state-space representations related to works from Rudolf Kalman and Peter Swerling. Kernel parameter estimation commonly employs likelihood optimization techniques developed in the tradition of Fisher information analysis and maximum-likelihood estimators used in COSMOS survey analyses.
The implementation is written in Python with performance-critical components in C++ and interfaces compatible with OpenMP and modern compiler toolchains. Computational strategies use banded matrix algebra, low-rank updates, and fast recurrence relations reminiscent of algorithms in Numerical Recipes and libraries like LAPACK and BLAS. The codebase integrates automatic differentiation patterns similar to those from JAX and TensorFlow for gradient computations and supports optimization routines popularized by scipy.optimize and LMFIT. Testing and continuous integration practices follow patterns established by projects on Travis CI and GitHub Actions.
celerite2 is applied in time-domain astronomy for tasks such as exoplanet transit modeling in datasets from Kepler Space Telescope and Transiting Exoplanet Survey Satellite, stellar variability studies in surveys like Gaia and Pan-STARRS, and radial velocity filtering in programs at facilities such as Keck Observatory and European Southern Observatory. It has seen use in cosmological time series analyses engaging teams from Sloan Digital Sky Survey and Dark Energy Survey. Outside astronomy, practitioners in finance at institutions like Goldman Sachs and J.P. Morgan have adapted related Gaussian process tools for irregular financial time series, while signal processing groups at MIT and Stanford University apply state-space approximations to telemetry from missions such as Voyager.
Benchmarks compare celerite2 to dense Gaussian process implementations exemplified by frameworks used in studies at Google and Microsoft Research, showing substantial reductions in computational cost for long one-dimensional datasets arising in missions like Kepler and TESS. Performance reports often reference scaling analyses analogous to those in papers from NeurIPS and speed comparisons against implementations in GPflow and Scikit-learn. Empirical tests performed on hardware from vendors such as Intel and NVIDIA highlight throughput gains when compiled with optimizations from GCC and Clang.
celerite2 is contrasted with full-rank Gaussian process approaches in the tradition of Rasmussen and Williams and with sparse approximations developed by groups at University of Cambridge and University College London. It differs from inducing-point methods used in GPyTorch and from multi-dimensional kernels employed in analyses by teams at Imperial College London and ETH Zurich by providing exact computations for its kernel class rather than approximate factorizations. The method shares conceptual space with state-space autoregressive models popularized in Box–Jenkins time series analysis and with Kalman-filtering implementations used by researchers at JPL.
celerite2 integrates with scientific ecosystems centered on Astropy, Pandas, Matplotlib, Jupyter Notebook, and workflow tools from Snakemake. Packaging and distribution fit into channels like Conda-Forge and PyPI with source repositories hosted on GitHub under permissive licenses similar to those used by scikit-image and scikit-learn. Interoperability with probabilistic programming frameworks such as Stan and PyMC has been demonstrated in community notebooks presented at workshops organized by Institute for Computational and Experimental Research in Mathematics and departmental seminars at Caltech.
Category:Scientific software