Hi @XL258W,
The discrepancy between what’s described in that abstract, and the current documentation / software capabilities, arises from the fact that I get to play with my own toys before other people do 
You can read about the relevant fixel correspondence method here. One of the many jobs on my todo list is to finalise that method, integrate it into MRtrix3, and update the FBA documentation accordingly, to be included in the future 3.1.0 update. Oh, and ideally publish it at some point too. 
Fibre density threshold
With the current suggested pipeline, there is already a kind of fibre density threshold. But that threshold is applied at the stage of performing FOD segmentation in order to define the set of fixels in template space, controlling which FOD lobes do or do not yield fixels in the template.
(My language here diverges from what’s currently in the documentation. There, the data resulting from segmenting FODs in template space is called a “fixel mask”. This to me is incorrect: a “fixel mask” is a fixel data file for which the value stored per fixel is a 1 or a 0 indicating inclusion in / exclusion from the mask. The output from fod2fixel is simply “a set of fixels” / “a fixel dataset”, or perhaps less preferably “a fixel image”, not a “mask”.)
With the aforementioned advanced fixel correspondence method, applying a relatively stringent fibre density threshold during FOD segmentation is detrimental: it discards information that could otherwise be important for the fixel correspondence step. As such, when utilising this method, I use a more lenient fibre density threshold during FOD segmentation, allowing smaller fixels to contribute to establishing fixel correspondence, then when I subsequently generate a mask defining which fixels will and will not be involved in statistical inference, I apply a more stringent fibre density threshold to exclude smaller fixels from subsequent analysis.
Consistency of subject-template fixel correspondence
The aforementioned method accounts for the prospect of the “appropriate” mapping of fixel data between subject space and template space being more complex than a one-to-one mapping. Multiple subject fixels can push data to a single template fixel, and multiple template fixels can pull data from a single template fixel. This information gets explicitly stored in a data format that I have devised to encode such information on the file system. So for instance, in the process of defining a fixel mask for statistical inference, one may choose to exclude from that mask those fixels for which more than a certain fraction of subjects have no fixel that correspond to the template, or those fixels for which the number of subject fixels that map to the template fixel varies too much across participants. There’s quite a few tricks that this method will enable; the hard part for me will be coming up with the “recommendations” that everyone will then go on to adopt as hard rules… 
Number of fixels
200k is entirely reasonable; I’ve seen ~ 100k - 600k. More important than the raw number is to make sure that all of the WM pathways that could conceivably have an association with your effect of interest are sensibly contained within the fixel template & can be reconstructed using streamlines tractography (since statistical inference depends on such).
Cheers
Rob