Subspace learning in non-Gaussian log-concave noise
M. Desai, Rami Mangoubi · 2005
We consider subspace learning from measurements corrupted by log-concave random noise. The class includes, but is not limited to, generalized Gaussian (GG) noise with shape parameter greater than or equal to unity, log-concave spherically invariant random processes (SIRPs), and their generalizations to norm invariant random processes (NIRPs). The noise properties need not be constant in the independent time and/or space variable. Necessary conditions are derived and they are computationally simpler when factorability properties are applicable, as is the case with GG's, SIRP's, and NIRPs when the signal space is one-dimensional.