maxentropy: Maximum entropy and minimum divergence models in Python

Status

The code is currently in pre-release while being refactored to use the scikit-learn estimator interface.

Purpose

This package helps you to construct a probability distribution (Bayesian prior) from prior information that you encode as generalized moment constraints.

You can use it to either:

1. find the flattest distribution that meets your constraints, using the maximum entropy principle (discrete distributions only)

2. or find the "closest" model to a given prior model (in a KL divergence sense) that also satisfies your additional constraints.

Background

The maximum entropy principle has been shown [Cox 1982, Jaynes 2003] to be the unique consistent approach to constructing a discrete probability distribution from prior information that is available as "testable information".

If the constraints have the form of linear moment constraints, then the principle gives rise to a unique probability distribution of exponential form. Most well-known probability distributions are special cases of maximum entropy distributions. This includes uniform, geometric, exponential, Pareto, normal, von Mises, Cauchy, and others: see here.

Examples: constructing a prior subject to known constraints

See the notebooks folder.

Quickstart guide

This is a good place to start: Loaded die example (scikit-learn estimator API)

History

This package previously lived in SciPy (http://scipy.org) as scipy.maxentropy from versions v0.5 to v0.10. It was under-maintained and removed from SciPy v0.11.

Copyright

(c) Ed Schofield, 2003-2019