Join Nostr
2024-10-20 01:13:55 UTC
in reply to

Stephen Brooks 🦆 on Nostr: npub1c9m22…z9xjd Alright, principal components is a volume-preserving rotation into ...

Alright, principal components is a volume-preserving rotation into a set of un-correlated Gaussians distributions (in your case). Say they have standard deviation sigma_n (n=1...).

"Entropy" information would be sum_n(-log_2(sigma_n)). This might diverge and need renormalising.

Effective degrees of freedom should be 2^entropy. That's
product_n(1/sigma_n).