Last active
March 24, 2019 01:04
-
-
Save fehiepsi/5ef8e09e61604f10607380467eb82006 to your computer and use it in GitHub Desktop.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Re: normalization: In the current design,
Gaussian
s are essentially numerically stable representations of strictly concave quadratic polynomials, i.e. they are intentionally non-normalized. Thus thelog(2 pi)
andlog det
terms appear only when reducing via.logaddexp()
. This is a bit tricky since we often reduce only a subset of dims, and only that subset get thelog(2 pi)
terms. I don't yet know how to compute thelog det
terms of the partial reduction.