PCA

Bazman2 · September 10th, 2013, 11:14 am

Hey there,When applying PCA it is usual to throw out the eigenvectors with small eigenvalues on the premise that these contribute relatively little to the process bwing studied.However I wondered if there was a more scientific way of testing which eigenvalues to keep and which to throw out?In essence I would like to test for a "signal" in these low eigenvalue eigenvectors.I wish to prove or disprove the null-hypthesis that these are white noise processes.Anyone tried this approach?Baz

bearish · September 11th, 2013, 9:53 am

I would suggest setting up a small Monte Carlo experiment. That way you know exactly what the structure of the randomness is, and you can see for yourself how much data you need to reliably separate a signal of a given strength from the background noise.