
338 L.P. Ammann
The weight matrix output by RSVD can be used to identify outliers in each principal
component as follows. Suppose that Z is a random variable from a standard Gaussian
distribution with mean 0 and s.d. 1. Let a denote some suitably small probability, e.g.,
a = .05, and let ua be the solution to
v(r < ~=)= ~.
Then any pixel with a weight less than or equal to ua in column k of the weight matrix
would be classified as an outlier in principal component k. For example, if r is the biweight
function with c = 4.685, a = .05, then ua = 0.6806, which would correspond ...