When dimension d>10, the variability of the estimator becomes very large.

[pbeste18]

• It seems that the error is multiplicative wrt the true MI. So as number of variables increase, the estimator breaks down. Mind sharing how you stabilized it?

• also, the loss function its the DV bound, not what's in the paper? Am I missing something?

I've always loved your work going back to your idea of learning portfolios directly from data using GPs (you were well ahead of the curve on this one). I'd love to use this project, but it seems it's been abandoned, I've resorted to doing my own implementation. Any hints would be much appreciated. :)