gabimoog / xspec_emcee-1

Reimplementation of the MCMC algorithm emcee in XSPEC

xspec_emcee

emcee

emcee is a pure-Python implementation of Goodman & Weare’s Affine Invariant Markov chain Monte Carlo (MCMC) Ensemble sampler.

xspec

xspec is an X-Ray Spectral Fitting Package, distributed as part of the high energy astrophysics software package, HEAsoft from NASA.

xspec has its own implementation of the GW algorithm, but I find it somewhat difficult to use, so I created my own implementation, which gives more control on the chains.

INSTALL:

Choose the files relevant to your HEAsoft version. Here I will work with v6.26.

• Download all updated 5 files from the relevant folder. the orig and new refer to the original files from xspec (provided in case you want to restore the files) and the new updated files that need to be used: Chain.cxx, Chain.h, ChainManager.cxx, ChainManager.h, xsChain.cxx
• Place them in the right place inside the xspec source structure and recompile the relevant code:
  • Chain.cxx, Chain.h, ChainManager.cxx, ChainManager.h inside: heasoft-6.26/Xspec/src/XSFit/MCMC
  • then inside heasoft-6.26/Xspec/src/XSFit, run: hmake and hmake install. Ensure that HEAsoft is initialized in the standard way. See their documentaton.
  • xsChain.cxx inside: heasoft-6.26/Xspec/src/XSUser/Handler, then inside heasoft-6.26/Xspec/src/XSUser, run: hmake and hmake install
• Run the GW chain in the usual way.

Example:

Assuming the spectra and model have been setup and an initial fit is found, we do:

chain len 10000
chain burn 10000
chain walker 100
para walk 30
chain run mcmc.fits

This will run the chain, printing progress along the way:

• The chains are initialized using 0.5*sigma from the fit covariance, so a valid fit is needed.
• The progress prints:
  • percentage progress:
  • best statistic in the current run.
  • acceptance fraction. It should be around ~0.2-0.3
  • The last number is the adjustable a parameter in the GW algorithm (see the algorithm paper for details). It can be adjusted to drive the acceptance fraction towards a desired value. If the acceptance fraction is too small, a can be reduced (using chain temperature 1.5 for example) to increase the acceptance fraction.

* Initializing: Using the 0.5* Covariance **

** Done initializaing **
         5%  498.871    0.313333       2
        10%  498.868       0.307       2
        15%  498.869       0.342       2
        20%  498.866       0.342       2
        25%  498.868       0.328       2
        30%  498.865       0.308       2
        35%  498.872       0.327       2
        40%  498.866       0.324       2
        45%  498.866       0.346       2
        50%  498.87       0.331       2
        55%  498.871       0.285       2
        60%  498.868       0.252       2
        65%  498.867       0.301       2
        70%  498.867        0.33       2
        75%  498.869       0.308       2
        80%  498.869       0.317       2
        85%  498.866       0.321       2
        90%  498.873       0.312       2
        95%  498.868       0.318       2
       100%  498.867       0.299       2
  New chain tmp.fits is now loaded.