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

[MarcoP]

This method says that all first lines of paragraphs are not copied from the same page but from previous first lines of paragraphs. That means that these lines should not be searched for matching previous words, but should be searched as sources for matching words from later text on the same page. 

On average, there are probably more than 13% of words on first lines, keeping in mind that the average paragraph length is strongly affected by the short paragraphs in quire 20.

This would be a more complicated test. For the modified Macchiavelly text this is a meaningless point.

However, we can still learn more. In the modified Macchiavelli, the page breaks are arbitrary. The text is not page-oriented and in any case the page breaks were introduced arbitrarily.
So, one can predict that the minimum edit distance will vary over each page as follows:
Since the appearance of similar words is purely a matter of chance, near the top of each page the average will be higher, since there are simply fewer words to compare with. As one goes down the page, the average will decrease as there are more and more words to compare with.
This would be the typical behaviour of a text that was not created using auto-copying.

For a text that was created using auto-copying, such a trend should not exist, and the minimum edit distance should be much more equal over each page.
In fact, if this could be visualised (e.g. using colour-coded words), one should see a more or less constant  behaviour over the text, except the first lines of paragraphs which should be higher as the words were also taken from earlier pages.

Hi Rene,
again I have tried something along the lines you suggested. I am not sure I got everything right, but here are my results.
I have simplified the method you proposed, since I did not want to go into the details of detecting paragraph boundaries (though that is certainly feasible with some more effort).

As a first experiment, I processed each page, collecting the average distance of words from the closest words in previous lines. I considered the first 8 lines of each page. The plot starts from line 2 (since we need a previous line to compare with). The graph on the right is normalized by dividing by the line:2 measure. I removed pliny_mod1 and Q13 and added Q20, which seems closer to Timm's generated text, for this particular experiment (more pages, text and paragraphs than Q13).

   

It seems that, even for auto-copying, the growing set of matchable words leads to progressively smaller and smaller values. All curves have a clear descending trend. The exception is what happens with Timm's data at line:5. I believe this is due to paragraph ends / new paragraphs, since most of the paragraphs in this data-set appear to be 4 or 5 lines long.


I tried an even simpler variant: matching each line against the first line of the page. Here there is a clearer difference: Timm's data appear to steadily increase for the first lines. Line 2 is by far closer to Line 1 than all other lines. I believe this is due to line 2 being derived from line 1 only, while line 3 is derived from line 2 and line 3, and so on. Each new line introduces more random variants, so words progressively diverge from those in line 1. Again, things change when new paragraphs begin.

   

All the other data-sets oscillate around 1 with deviations of about 10%. Interestingly, mod2 also seems to progressively increase, though at a much lower rate than Timm's generated text. I guess this is due to the "counter" assuming values that are progressively higher than those in the first line.
The yellow line for the whole VMS appears to be almost perfectly flat. 
But Q20 might have some "paragraph-effect", possibly due to the features of the first lines of paragraphs (e.g. 'p' and 'f', Grove words) that make them less comparable with the rest. Similarly to Timm's data, there is a rising trend on the first few lines, but clearly the deviation is much smaller (11% vs 48%).