Multivariate Gaussianization for Data Processing
Multivariate Gaussianization for Data Processing
Multivariate Gaussianization for Data Processing
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Intro Iterative <strong>Gaussianization</strong> Experiments ConclusionsOn the suitable rotationWhich is the suitable rotation?Closed-<strong>for</strong>m Theoretical Convergence Comput.Rotation convergence√rate costICA√×√Max ∆J O(2md(d + 1)n)PCA√√∆J = 2nd order O(d 2 (d + 1)n)RND∆J ≥ 0 O(d 3 )n samples of dimension d, FastICA running m iterationsICA guarantees the theoretical convergence of the <strong>Gaussianization</strong> processsince it seeks <strong>for</strong> the maximally non-Gaussian marginal PDFsPCA leads to non-optimal convergence rate because it does reduceredundancy to a certain extent (it removes correlation), but it does notmaximize the marginal non-Gaussianity J m(x)PCA is closed-<strong>for</strong>m and is faster than ICAPCA require more iterations but Gaussianizes fasterUsing RND trans<strong>for</strong>ms guarantees the theoretical convergence, butconverges slowly