Compare differential entropy of multivariate Gaussian with different dimensions

494 Views Asked by At

Since joint entropy is commonly used to quantify the average information of random variables, in order to compare the amount of information, can we directly compare the joint entropy of multivariate normal distributions with different dimensions? (e.g., one MVN with 2x2 covariance matrix while another MVN with 4x4 covariance matrix)

1

There are 1 best solutions below

0
On

No, because there is no operational interpretation of differential entropy in the first place (and thus no common ground on which one could reasonably make a comparison). Cf. the Cross Validated question on differential entropy.

If you have two pairs of multivariate Gaussians, one for each dimension, then you may interpret differences of KL divergence as how much faster Neyman–Pearson hypothesis testing distinguishes one pair relative to the other, à la Stein's lemma.