I have encountered a problem, basically related to the http://arxiv.org/abs/1206.6679 .
If I want to minimize the normalized KL-divergence KL(Q||P) with Q a multivariate Gaussian distribution. but in the paper, a method is to minimize the unnormalized KL-divergence, can I ask if they hold the same stationary point for the precision matrix?
If yes, could someone give me a proof? (I tried to prove this, but it seems there is one term difference if making use the normalized version compared to the unnormalized version.)