This Python module implements estimation of the JSD scores for the observed data assuming some distribution. This module was developed when performing empirical analysis for the forthcomming paper by Mark Levene and Aleksejus Kononovicius (draft of the paper is available on arXiv: 1809.11052 [stat.ME]).

Feel free to reuse the code or modify it. We would like to encourage you to reference our paper, if it would be appropriate, but we surely do not require you to do it.

If you want to reference the repository itself, you can do that in the following manner:

• A. Kononovicius and M. Levene. PyJSD: Python implementation of the Jensen-Shannon divergence. http://github.com/akononovicius/pyjsd.

Here we have implemented a JSD function, which does three things:

1. It estimates distribution parameter values given the assumed (theoretical) distribution and the data using Maximum likelihood estimation.
2. It estimates Jensen-Shannon Divergence (JSD) between the empirical and the assumed distribution. Lower scores are better.
3. It estimates confidence intervals for the JSD using moving block bootstrap method.

The JSD function takes four parameters: data, empiricalDist, theorDist and bootstrap.

• data should be an array containing empirically observed values.
• empiricalDist should be a dictionary with three keys: start (minimal value reflected in the obtained empirical distribution), stop (maximum value reflected in the obtained empirical distribution) and bins (number of bins to used when estimating the empirical distribution). For example:

empiricalDist={
    "start": 0.0,
    "stop": 505.0,
    "bins": 1000,
}

• theorDist should be a dictionary with three keys: cdf (a function which returns CDF values at given points), likelihood (a function which returns likelihood of the given empirically observed values) and params (initial parameter values from which MLE algorithm will start). For example:

norm={
    "cdf": lambda params,x: scipy.stats.norm.cdf(x,loc=params[0],scale=params[1]),
    "likelihood": lambda params,data: -numpy.sum(scipy.stats.norm.logpdf(data,loc=params[0],scale=params[1])),
    "params": [1,1],
}

Note that some distributions are defined jsd.distributions submodule.

• bootstrap should be a dictionary with three keys: iterations (number of sample to obtain), blockSize (what block size to use) and percentiles (which percentiles to report). Example:

bootstrap={
    "iterations": 1000,
    "blockSize": 1,
    "percentiles": [2.5,97.5],
}