Computing the new variances
For axis-aligned Gaussians, we just fit the
variance of the Gaussian on each dimension to
the posterior-weighted data
Its more complicated if we use a full-
covariance Gaussian that is not aligned with
the axes.