Discussion of "A possible generating algorithm of the Voynich manuscript"

[MarcoP]

I don't understand why self-citation should cause the initial glyph of one word to depend on the final glyph of the previous word.

The previous word is the last written word. To copy an element of the previous word is therefore an obvious choice. The only problem is that this choice is indeed obvious. The shape of <q> is quite similar to the shape of <y>. See for instance the difference between <y> and <q> in <yoky> and <qoty> on page f79r.

It is possible to demonstrate that indeed an element of the previous word was copied. Consecutive tokens start more often than expected with the same initial glyphs or end with the same final glyphs. See table 1. "Initial-Initial dependency" and table 3 "Final-Final dependency" in the Analysis of word dependency by Vladimir Sazonov: "As it is clear from the tables 1 and 3, the words very often reproduce the same form (the same initial or the same final) in the next word. For example, -aiin or qo- or -dy can repeat many times with good density" (V. Sazonov 2003).

Thank you, Torsten.
I will first comment your second argument. I interpret it as a suggestion that initial-initial repetition and/or final-final repetition can cause a bias in the distribution of word-break combinations. While in principle this is true (i.e. a sufficient amount of repetition can cause considerable shifts in word-break combinations) it is not what happens in the VMS.
In the case of -y.q- only a subset of initial-initial repetitions can cause this word break combination: q*y.q, i.e. a word starting with q- and ending with -y, followed by a second word starting with q-.
If one considers the whole manuscript, about 2000 occurrences of -y.q- would be expected, while about 3500 actually occur. The difference between the two (1500) is the number of unexpected occurrences. This graph shows how different X*y.q- sequences contribute to the unexpected combinations (I only included the most frequent "prefixes").

   

As you can see, q*y accounts for about 400 of the unexpected occurrences: they are many, but q- repetition does not nearly explain the observations. The other 1100 unexpected occurrences are due to other prefixes. The graph also shows that what "triggers" the presence of q- is -y alone: the preference is quite uniform independent of the prefix of the previous word.

Turning to your first argument, if I understand correctly what you write, it appears to be more in agreement with the observations: the word after -y is modified or selected in such a way that it starts with q-.
I understand your statement that "The shape of <q> is quite similar to the shape of <y>" as your explanation for why q- appears to be systematically chosen to be written immediately after -y. 
Graphical similarity is somehow subjective. One can also say that <y> looks like <o>, but the word-break combination -y.o- is slightly dis-preferred (i.e. occurs less often than expected).

Similarly, -o.y- is dis-preferred. Moreover, all the immediate repetitions of the same character across words are dis-preferred (in various degrees): if the idea were having similar characters across word-breaks, -y.y- could have been an option.

If one examines word-break preferences in their entirety, it is possible to point out a different explanation that applies to a large part of the observations. In word-break combinations, both -o and -y prefer to be followed by the gallows characters and <d> <s> <r> <l>. These same preferences can be observed for combinations "inside" words. Words containing the sequence "yo" are rarer than expected, while words containing "ok", "od", "yk", "yd" etc are more frequent than expected. With the exception of <e> and <i>, characters tend not to occur consecutively inside words, as it happens in word-break combinations.

In general, the preferred combinations do not seem to encourage the consecutive occurrence of graphically similar characters. E.g. o/y are different from the gallows and from r/s. The general principle could possibly be the reverse: similar characters do not typically follow each other.

What can be said is that the preferences in word-break combinations largely reflect the preferences for consecutive characters inside words.

But of course this is not a hard rule, there are exceptions and various problems. Still this seems to work much better than assuming that word-breaks combinations favour the presence of similar characters across word-breaks.

Also, please correct me if I am wrong, but I think the generated text here doesn't feature any clear dependency between the last glyph of a word and the first glyph of the next word?

"Keep in mind that the VMS was created by a human writer who had complete freedom to vary some details of the generating algorithm on the spur of a moment. ..." (Timm & Schinner 2019, p. 16). With other words we didn't argue that the text was created by a computer program and we didn't argue that our program is able to simulate the complexity of a human mind.

Beside line breaks numerous similar rules for describing some kind of local repetition exists. Even if I would try to add each rule it would still be possible to point to an another similar rule. At the same time each additional rule would increase the complexity of the program and it would be argued that the program is too sophisticated. Therefore our goal was to keep the algorithm as simple as possible.

Thank you for clarifying!
Since in the paper you state that your algorithm "is able to reproduce all the intriguing statistical key properties of the original VMS", I was unsure if you considered the unexpected word-break combinations a key property of Voynichese or not. You made things clear now. I also understand your reasons for not modelling a larger subset of the known properties of Voynichese.