Problems

Currently, the major weakness of this approach is its sensitivity to parameter choice. The separation fitting goals (6) apply the inverse of a non-stationary PEF. If that PEF isn't stable, the separation of the multiples and primaries is not possible. To get a stable filter we can increase $\epsilon$ in our filter estimation (9). Unfortunately, increasing $\epsilon$ decreases the quality of our prediction. By changing the size of our micro-patches, we can usually get a stable filter while obtaining a good prediction. At this stage we haven't figured an algorithm that can automatically change micro-patch size to obtain the desired combination, a stable non-stationary PEF that can satisfactorily predict the data.