Charles Martin

See RBM_Baseline_Ominglot.ipynb

See how the parameters affect the convergence of the free energy

The issue here is when the Power Law displays finite-size effects, causing the tail to decay slower than expected for the scale-free portion of the PDF, and yielding alpha sometimes...

There are number of models for dealing with the cutoff or finite-size-effects that arise in when trying to model real world, finite data with a simple power law distribution. Here's...

What I mean is what is the intent of the normalization to be used, since this thread is asking if there is a bug in the choice of normalization?

I take it then that there is indeed a bug in the normalization of the PDF, when xmax is set explicitly? If so, and you make a change, it would...

The current research uses the existing implementation of the normalization of the PDF, even when xmax is set explicitly: See: https://github.com/CalculatedContent/WeightWatcher/blob/b32f4876659275b8148b024f9a4fed95b1524e90/weightwatcher/weightwatcher.py#L1975 Old code versions become stale and difficult to support...

**I expected that it was finding the optimal alpha, using the Clauset MLE approach, as described in this paper** https://arxiv.org/abs/0706.1062 **and with the normalization as described in the paper** **According...

If there is a general way to cut the dropping part of the tail off (i.e to set xmax), that would be helpful

Binarized data is