This package facilitates the creation and manipulation of arbitrarily complicated (correlated) multi-dimensional Gaussian random variables. The random variables are represented by a new data type (gvar.GVar) that can be used in arithmetic expressions and pure Python functions. Such expressions/functions create new Gaussian random variables while automatically tracking statistical correlations between the new and old variables. This data type is useful for simple error propagation, but also is heavily used by the Bayesian least-squares fitting module lsqfit.py — to define priors and specify fit results, while accounting for correlations between all variables. Documentation can is in the doc/ subdirectory (see doc/html/index.html for the html version or doc/lsqfit.pdf for a pdf version).

These packages use numpy for efficient array arithmetic, and cython to compile efficient code. gvar uses automatic differentiation to track covariances through arbitrary arithmetic.

Information on how to install the components is in the INSTALLATION file.

To test the libraries try make tests. (Some tests involve random numbers and so may occasionally — less than 1 in 100 runs — fail due to rare multi-sigma fluctuations; rerun the tests if they do fail.) Some examples are give in the examples/ subdirectory.

Versioning: Version numbers for gvar are now (5.0 and later) based upon semantic versioning (http://semver.org). Incompatible changes will be signaled by incrementing the major version number, where version numbers have the form major.minor.patch. The minor number signals new features, and the patch number bug fixes.