[Article] "The Strange Quest to Crack the Voynich Code"

[MarcoP]

Did they assume that their variances (sigmas) in the other languages were the same as for the two languages they had measured many texts of - English and Portuguese? Furthermore, they chose the variance that was smallest (of the English cmp. to Portuguese texts)? Am I understanding correctly?

Then, as I understand, the compatibility was calculated from integration of the upper/lower tails from the measured value, of the interpolated distribution resulting from the use of these sigmas? If these integrals were < 0.05 it is not compatible since it is too far off a reasonable probability. 

I don't know about you, but to me this seems a bit bold... assuming the same sigma from the lower of only two measurements (languages). Not that I think that this would affect much the overall conclusions of the study though.

Hi Jonas,
sadly I don't understand enough of all this to contribute a meaningful opinion.

Another thing that strikes me is that one of the measures discussed by Amancio et al. is the Clustering Coefficient "C", something that was recently disccussed You are not allowed to view links. Register or Login to view.. If I understand correctly, they claim that C is not significant for telling an original text from it shuffled version (while C*, computed only over the 50 most frequent words, is).

Distinguishing Books from Shuffled Sequences
[...]
[measures that] do not fully satisfy  ζ1 [...] cannot be used to distinguishing a manuscript [sic] from its shuffled version.
[...]
When we performed the informativeness analysis over the most frequent words, we found that ζ1 is satisfied for the clustering coefficient and for the shortest paths (note that C* and L* are informative while C and L are not).

The paper discussed in the other thread (You are not allowed to view links. Register or Login to view. Cardenas et al. 2015) uses an almost identical approach in their analysis of the VMS: they compare its Clustering Coefficient with a randomly scrambled ordering of the text (which they call a "cyphered" text). Interestingly, Amancio et al. is not referenced: it appears that Cardenas et al. were not awere of this earlier work. Differently from what Amancio et al. found, the graph they provide (Fig.5 top) suggests that C is quite effective in separating a text from its shuffled versions.
Another puzzling thing: Amancio et. al measured a lower C (and C*) for the original text than for its scrambled versions: their normalized values are <1.  Cardenas' plot shows that C is higher for the original than for scrambled versions. For what is worth, when I tried replicating Cardenas' experiments I found something similar to what reported by Amancio.

My preliminary attempts (using the networkx Python library) are not being very successful: in particular, scrambled texts appear to result in higher mean clustering C than original texts, while in the paper the opposite happens.

Of course, the easiest explanation is that I am misunderstanding something basic in one of the two papers (or both).