Box least squares (BLS) periodogram

The “box least squares (BLS) periodogram” [1] is a statistical tool used for detecting transiting exoplanets and eclipsing binaries in time series photometric data. The main interface to this implementation is the BoxLeastSquares class.

Mathematical Background

The BLS method finds transit candidates by modeling a transit as a periodic upside down top hat with four parameters: period, duration, depth, and a reference time. In this implementation, the reference time is chosen to be the mid-transit time of the first transit in the observational baseline. These parameters are shown in the following sketch:

(png, svg, pdf)

../_images/bls-1.png

Assuming that the uncertainties on the measured flux are known, independent, and Gaussian, the maximum likelihood in-transit flux can be computed as

\[y_\mathrm{in} = \frac{\sum_\mathrm{in} y_n/{\sigma_n}^2}{\sum_\mathrm{in} 1/{\sigma_n}^2}\]

where \(y_n\) are the brightness measurements, \(\sigma_n\) are the associated uncertainties, and both sums are computed over the in-transit data points. Similarly, the maximum likelihood out-of-transit flux is

\[y_\mathrm{out} = \frac{\sum_\mathrm{out} y_n/{\sigma_n}^2}{\sum_\mathrm{out} 1/{\sigma_n}^2}\]

where these sums are over the out-of-transit observations. Using these results, the log likelihood of a transit model (maximized over depth) at a given period \(P\), duration \(\tau\), and reference time \(t_0\) is

\[\log \mathcal{L}(P,\,\tau,\,t_0) = -\frac{1}{2}\,\sum_\mathrm{in}\frac{(y_n-y_\mathrm{in})^2}{{\sigma_n}^2} -\frac{1}{2}\,\sum_\mathrm{out}\frac{(y_n-y_\mathrm{out})^2}{{\sigma_n}^2} + \mathrm{constant}\]

This equation might be familiar because it is proportional to the “chi squared” \(\chi^2\) for this model and this is a direct consequence of our assumption of Gaussian uncertainties. This \(\chi^2\) is called the “signal residue” by [1], so maximizing the log likelihood over duration and reference time is equivalent to computing the box least squares spectrum from [1]. In practice, this is achieved by finding the maximum likelihood model over a grid in duration and reference time as specified by the durations and oversample parameters for the BoxLeastSquares.power() method. Behind the scenes, this implementation minimizes the number of required calculations by pre-binning the observations onto a fine grid following [1] and [2].

Basic Usage

The transit periodogram takes as input time series observations where the timestamps t and the observations y (usually brightness) are stored as NumPy arrays or Quantity. If known, error bars dy can also optionally be provided. For example, to evaluate the periodogram for a simulated data set, can be computed as follows:

>>> import numpy as np
>>> import astropy.units as u
>>> from astropy.stats import BoxLeastSquares
>>> np.random.seed(42)
>>> t = np.random.uniform(0, 20, 2000)
>>> y = np.ones_like(t) - 0.1*((t%3)<0.2) + 0.01*np.random.randn(len(t))
>>> model = BoxLeastSquares(t * u.day, y, dy=0.01)
>>> periodogram = model.autopower(0.2)

The output of the BoxLeastSquares.autopower() method is a BoxLeastSquaresResults object with several useful attributes, the most useful of which are generally the period and power attributes. This result can be plotted using matplotlib:

>>> import matplotlib.pyplot as plt                  
>>> plt.plot(periodogram.period, periodogram.power)  

(png, svg, pdf)

../_images/bls-2.png

In this figure, you can see the peak at the correct period of 3 days.

Objectives

By default, the BoxLeastSquares.power() method computes the log likelihood of the model fit and maximizes over reference time and duration. It is also possible to use the signal-to-noise ratio with which the transit depth is measured as an objective function. To do this, call BoxLeastSquares.power() or BoxLeastSquares.autopower() with objective='snr' as follows:

>>> model = BoxLeastSquares(t * u.day, y, dy=0.01)
>>> periodogram = model.autopower(0.2, objective="snr")

(png, svg, pdf)

../_images/bls-3.png

This objective will generally produce a periodogram that is qualitatively similar to the log likelihood spectrum, but it has been used to improve the reliability of transit search in the presence of correlated noise.

Period Grid

The transit periodogram is always computed on a grid of periods and the results can be sensitive to the sampling. As discussed in [1], the performance of the transit periodogram method is more sensitive to the period grid than the LombScargle periodogram. This implementation of the transit periodogram includes a conservative heuristic for estimating the required period grid that is used by the BoxLeastSquares.autoperiod() and BoxLeastSquares.autopower() methods and the details of this method are given in the API documentation for BoxLeastSquares.autoperiod(). It is also possible to provide a specific period grid as follows:

>>> model = BoxLeastSquares(t * u.day, y, dy=0.01)
>>> periods = np.linspace(2.5, 3.5, 1000) * u.day
>>> periodogram = model.power(periods, 0.2)

(png, svg, pdf)

../_images/bls-4.png

However, if the period grid is too coarse, the correct period can easily be missed.

>>> model = BoxLeastSquares(t * u.day, y, dy=0.01)
>>> periods = np.linspace(0.5, 10.5, 15) * u.day
>>> periodogram = model.power(periods, 0.2)

(png, svg, pdf)

../_images/bls-5.png

Peak Statistics

To help in the transit vetting process and to debug problems with candidate peaks, the BoxLeastSquares.compute_stats() method can be used to calculate several statistics of a candidate transit. Many of these statistics are based on the VARTOOLS package described in [2]. This will often be used as follows to compute stats for the maximum point in the periodogram:

>>> model = BoxLeastSquares(t * u.day, y, dy=0.01)
>>> periodogram = model.autopower(0.2)
>>> max_power = np.argmax(periodogram.power)
>>> stats = model.compute_stats(periodogram.period[max_power],
...                             periodogram.duration[max_power],
...                             periodogram.transit_time[max_power])

This calculates a dictionary with statistics about this candidate. Each entry in this dictionary is described in the documentation for BoxLeastSquares.compute_stats().

Literature References

[1](1, 2, 3, 4, 5) Kovacs, Zucker, & Mazeh (2002), A&A, 391, 369 (arXiv:astro-ph/0206099)
[2](1, 2) Hartman & Bakos (2016), Astronomy & Computing, 17, 1 (arXiv:1605.06811)