Why and how the text could be Bavarian

[Jorge_Stolfi]

I myself use a cleaned-up EVA file, in which some labels are also missing, provided they are single letters. All special characters in EVA were also removed for the calculation.

I suggest that you do all statistics separately for each of the major sections (Herbal-A, Herbal-B,Bio, and Starred Parags).  Statistics are properties of a text, not of a language; and each of those four sections seems to be a text of a very different nature.   

Statistics are heavily influenced by the nature of the text, as you can see by comparing the Bavarian recipe book with the other three.  If you run statistics on King Arthur and The Joy of Cooking together, you may find that the numbers best match the History of Cannibalism among the Knights Templar than either of the two actual books.

The other VMS sections are likely to different from these four, too, but are too small to yield meaningful data.  You may come back to them later, once you have got the main conclusions from the meaty sections.

You should also consider only running text in paragraphs -- excluding labels, titles, etc.  Labels are likely to have their own statistics, very different from those of the main text.

And you may want to treat the letters p and f as '?'.  They are likely to be fancified versions of other glyphs, like our capital letters, or the initials and finals of Arabic script.  A text in mixed case will have higher entropy and vocabulary than the same text mapped all to lowercase, even though the meaningful information is practically the same. 

Vocab — Number of different word types
Hw — Word entropy (bit)

These statistics are affected by the size of the sample text.  If word frequencies follow the Zipf distribution, the size of the vocabulary is expected to grow proportionally to the square root of the number N of tokens, or something like that.   The word entropy should slightly increase too with N.  Unless some correction is applied when estimating word probabilities from their observed frequencies.

H1 norm: VMS (0.865) lies in the middle of the MHD texts (0.871–0.901) → no anomaly signal at character level
H(X₂|X₁): VMS (2.503) significantly lower than MHD (3.22–3.50) → high predictability of character strings, explainable by fixed prefix patterns of the absorption cipher

As I argued several times, character and bigram statistics are hardly meaningful. They are affected not only by the nature of the text, but even more so by the spelling system.  Merely spelling German with "x" instead of  "sch", "v" instead of "ch", etc will increase its character and bigram entropies.

Hw: VMS (10.267) higher than all MHD texts → the cipher generates many word types with relatively few tokens through prefix combinations

This number is more significant, but one must normalize it for the sample size before comparing it to other texts.  And it too can be inflated by trivial factors -- like the presence of spelling and transcription errors (affecting maybe 5% or more of the VMS tokens), inclusion of p and f as distinct letters, the inclusion of labels, the mixing of texts of different nature, etc.

But I see that, in spite of all those entropy-boosting factors, the VMS Hw is only slightly higher than that of other Bavarian texts.  I would call that a match.

The combination of low H(X₂|X₁) and high Hw corresponds exactly to the entropy profile that absorption encryption could generate – the highly predictable character strings result from fixed prefixes, but it increased the diversity at the word level through combinatorial prefix attachment. This profile is not to be expected when generating meaningless text, during self-copying processes, or in unencrypted natural language.

Ahem!  You mean in unencrypted European natural language...

All the best, --stolfi