This repo contains code to reproduce the numerical experiments and figures in the paper "An Ordering of divergences for variational inference with a factorized Gaussian approximation" by Charles Margossian, Loucas Pillaud-Vivien and Lawrence Saul.

The files are organized as follows:

• vi_ordering.r, with support from fgvi_solve.r, contains the code to reproduce Figure 1 (ellipse and circle plots) and Figure 3 (variance, precision and entropy estimate as we violate the assumption of factorization).
• variational_collapse.M contains the code to reproduce Figure 2.
• entropy_matching.r contains the code to reproduce Figure 4 (alpha for entropy match).
• ordering.ipynb, with support from fgvi.py and factirized_bam.py, contains the code to reproduce Figure 5 (variance estimates for the 8 schools) and Figure 6 (entropies for inference gym model). This notebook contains additional code to examine various properties of the VI algorithms, including convergence.