MCMC 2D surface brightness modeling for quasar / host galaxy decompositions
Usage is provided in docstrings, and an example model is provided in the
examples directory. My Ph.D. thesis contains a more detailed discussion of
the problem (simultaneous modeling of quasar point sources and host galaxies)
and why I have made specific design decisions. My dissertation is available
open-access at the following URL:
[Markov Chain Monte Carlo Modeling of High-Redshift Quasar Host Galaxies in Hubble Space Telescope Imaging] (http://repository.asu.edu/attachments/126037/content/Mechtley_asu_0010E_13571.pdf)
A journal article is forthcoming, but if you use psfMC in the meantime, please cite:
[Mechtley, M. 2014, PhD Thesis, Arizona State Univ.] (https://ui.adsabs.harvard.edu/#abs/2014PhDT.........1M)
The current master branch is a beta release (until I finalize formal unit tests). I've used it myself extensively, but please be aware that there may still be bugs lurking around.
Beta release 2.0b1 is a backwards-incompatible release that switches to
emcee for sampling (and is compatible with Python3). The pymc requirement
has dropped entirely, as I now also store trace databases as a FITS table and
handle the priors myself in a more lightweight manner. Note that because priors
in psfMC.distributions now wrap scipy.stats distributions directly, some
parameter names have changed (e.g., Normal now expects loc= and scale=
instead of mu= and tau=).
The rationale behind such a large change is that the Affine-Invariant ensemble
sampler used by emcee is much more efficient for this particular problem
(high-dimensional parameter space, reasonably convex single-mode posterior).
Traditional Metropolis-Hastings chains required 50,000 samples or more of
burn-in each to approach convergence, severely limiting the number of chains
that could be used, and thus our confidence in the final result. The sampler now
reliably approaches convergence with a similar number of total burn-in samples
(e.g., iterations=250 * nwalkers=200 = 50000), but the ensemble sampler means
this only has to be run once, rather than once for each chain. The large number
of walkers replaces the primary purpose of multiple independent chains in the
previous implementation: to improve our confidence in convergence.
As a side effect, the sampling is now overall faster (takes less time to run the
same number of samples), since emcee is a lightweight sampler only and doesn't
attempt clever caching / lazy evaluation like pymc (which the previous pymc
implementation was not taking advtange of anyway).
Beta release 1.0b5 adds a new analysis submodule. FITS image production has been
moved there, and various statistical and plotting methods have been added for
analyzing posterior chains. This includes MCMC covariance corner plots via Dan
Foreman-Mackey's corner.py module. Also
includes two new command line scripts: corner_plot and plot_chain to
generate these plots from a terminal.
The software depends on the numpy, scipy, emcee, astropy, and six (for
Python 2 backward compatibility) modules. Additionally the matplotlib and
corner modules are required for posterior analysis plots.
The pyregion module is optional but strongly recommended for ease of masking
out foreground or background objects unrelated to the quasar being modeled.
The numexpr module is also optional, and is used to parallelize certain
calculations (generating Sersic profiles), providing a slight increase in
overall speed.
I only test/support Anaconda, use other Python distributions at your own risk.
Note: everything else is included with the default Anaconda distribution.
conda install --channel astropy emcee pyregion corner
pip install git+https://github.com/mmechtley/psfMC.git