Python implementation of Aaron Clauset's power-law distribution fitter
This is a python implementation of a power-law distribution fitter. The code here was originally hosted on agpy but was moved and re-packaged to make setup.py cleaner.
See also http://code.google.com/p/powerlaw, an alternate implementation of the same algorithm with additional bells & whistles.
I’ve attempted to make the setup.py file work nicely, but it includes some hacks, so if you run into trouble, please report it on github:
git clone email@example.com:keflavich/plfit.git cd plfit python setup.py install
If python setup.py install doesn’t work, you can try the following:
To install the cython function, run: python setup.py build_ext --inplace
To install the fortran function:
cd plfit/plfit/ f2py -c fplfit.f -m fplfit --fcompiler=gfortran
Aaron Clauset et al. address the issue of fitting power-laws to distributions on this website and in their paper Power-law distributions in empirical data. I have created a python implementation of their code because I didn’t have matlab or R and wanted to do some power-law fitting.
Power-laws are very commonly used in astronomy and are typically used to describe the initial mass function (IMF), the core mass function (CMF), and often luminosity distributions. Most distributions in astronomy tend to be apparent power-laws because the source counts are too few or too narrow to distinguish powerlaws from log-normal and other distributions. But, to this end, I’ve included the testing mechanism to test for consistency with a power law as described in the above paper.
The python internal documentation is complete. A brief description of relevant functions is included here for convenience:
plfit is implemented as a class. This means that you import plfit, and declare an instance of the plfit class:
import plfit X = rand(1000) myplfit = plfit.plfit(X)
The results of the fit are printed to the screen (if desired) and are stored as part of the object.
alpha_ and kstest_ are functions used internally to determine the ks-statistic and alpha values as a function of xmin.
The powerlaw fitter is very effective at returning the correct value of alpha but not as good at returning the correct value of xmin.
There are 3 implementations of the code internals. fplfit.f is a fortran function, cplfit.pyx is a cython function, and plfit.py is the wrapper and includes a python-only implementation that requires numpy. FORTRAN is fastest, follow closely by cython. Python is ~3x slower.
As of November 21, 2011, there is a pure python (i.e., no numpy) implementation at <https://github.com/keflavich/plfit/blob/master/plfit/plfit_py.py> - you can just put this file in your local working directory and import it, since it contains no requirements beyond pure python. It’s slower and hobbled, but it works, and perhaps will run fast with pypy.
A very simple example:
import plfit from numpy.random import rand,seed # generate a power law using the "inverse" power-law generator code X=plfit.plexp_inv(rand(1000),1,2.5) # use the numpy version to fit (usefortran=False is only needed if you installed the fortran version) myplfit=plfit.plfit(X,usefortran=False) # output should look something like this: # PYTHON plfit executed in 0.201362 seconds # xmin: 0.621393 n(>xmin): 263 alpha: 2.39465 +/- 0.0859979 Log-Likelihood: -238.959 ks: 0.0278864 p(ks): 0.986695 # generate some plots from pylab import * figure(1) myplfit.plotpdf() figure(2) myplfit.plotcdf()
If you use this code, please cite Clauset et al 2009 and consider posting a comment below.
Direction citations to the source are welcome! The python translation has been cited in the following works (and perhaps others?):
v1.0.1 - bugfix to pypi only; just adds things to MANIFEST.in v1.0 - first release