Hey Sameer.

Very interesting paper. Some day I hope I can play with math to that level. I have some questions about the intuition of how the algorithm works.

There are several differences between gaussian noise and noise dteremined by another distribution, such as exponential or poisson. One of the major differences is that gaussian noise is symmetric about the mean. Does this algorithm assume that the best fit for the data is not necissarily symmetrically distant from is principal component projection?

Does this explain why the exponential distribution is less susceptible to the outliers in the first example?

The paper definately raises more questions and possibilities than it tries to answer.

Any thoughts on how you would use non-gaussian PCA?