1106.1989 (Ingo P. Waldmann)
Ingo P. Waldmann
The characterisation of ever smaller and fainter extrasolar planets requires
an intricate understanding of one's data and the analysis techniques used.
Correcting the raw data at the 10^-4 level of accuracy in flux is one of the
central challenges. This can be difficult for instruments that do not feature a
calibration plan for such high precision measurements. Here, it is not always
obvious how to de-correlate the data using auxiliary information of the
instrument and it becomes paramount to know how well one can disentangle
instrument systematics from one's data, given nothing but the data itself. We
propose a non-parametric machine learning algorithm, based on the concept of
independent component analysis, to de-convolve the systematic noise and all
non-Gaussian signals from the desired astrophysical signal. Such a `blind'
signal de-mixing is commonly known as the `Cocktail Party problem' in
signal-processing. Given multiple simultaneous observations of the same
exoplanetary eclipse, as in the case of spectrophotometry, we show that we can
often disentangle systematic noise from the original light curve signal without
the use of any complementary information of the instrument. In this paper, we
explore these signal extraction techniques using simulated data and two data
sets observed with the Hubble-NICMOS instrument. Another important application
is the de-correlation of the exoplanetary signal from time-correlated stellar
variability. Using data obtained by the Kepler mission we show that the desired
signal can be de-convolved from the stellar noise using a single time series
spanning several eclipse events. Such non-parametric techniques can provide
important confirmations of the existent parametric corrections reported in the
literature, and their associated results. Additionally they can substantially
improve the precision exoplanetary light curve analysis in the future.
View original:
http://arxiv.org/abs/1106.1989
No comments:
Post a Comment