/CBVshrink

Kepler systematics correction using co-trending basis vectors (CBV), Variational Bayes and shrinkage priors

Primary LanguagePythonGNU General Public License v2.0GPL-2.0

CBVshrink

Kepler systematics correction using public co-trending basis vectors (CBV), Variational Bayes and shrinkage priors.

Description

This module uses the co-trending basis vectors (CBVs) derived by the Kepler PDC-MAP pipeline and published on the MAST archive to detrend individual Kepler light curves. Like the PDC-MAP pipeline, each light curve is modelled as a linear combination of CBVs. However, here this model is implemented in a Variational Bayes (VB) framework, where the priors over the weights associated with each CBV are optimized to mximse the marginal likelihood of hte model. Because we use zero-mean Gaussian priors for the weights, with hyper-priors on the widths of those priors centred on zero, any CBV not strongly supported by the data is not used in the final model. This approach, known as automatic relevance determination (ARD) or shrinkage, reduces the risk of overfitting. As the CBVs are derived from the Kepler light curves and contain some noise, it also reduces the amount of noise injected into the light curves by the correction, including on planetary transit (few hour) timescales. Finally, it also preserves intrinsic stellar variability more successfully than the standard PDC-MAP pipeline.

For more info, see Aigrain, Parvianen, Roberts & Evans (in prep.). An older version of our Kepler systematics correction, using the same VB-ARD framework, but our own basis vector discovery, was described in Roberts, McQuillan, Reece & Aigrain (2013).

Dependencies

numpy, pylab, pyfits, argparse

Installation

Simply clone the repo and place it in your Python path.

Usage

See the notebook example1.ipynb in the examples directory.

TBD

  • implement recipe for selecting number of CBVs used automatically
  • demonstrate use of "exclude_func" keyword to mask specific regions of the light curve during the CBV fitting process.