(11-09-2025, 07:42 AM)quimqu wrote: You are not allowed to view links. Register or Login to view.The H₂ in my post is a bit high, but that’s easy to fix because the system is fully parameterizable.
In practice I just tighten per-position alphabets to reduce next-symbol options; this directly lowers H₂. Then I run a targeted H₂ pass that collapses only the positions contributing most to conditional entropy. And finally I do a light bigram touch-up so correlations/JS and the Zipf curve stay on track.
With those knobs, pushing H₂ down toward 2.3 is straightforward while keeping token lengths and overall grapheme/bigram shape aligned with Voynich. Here some plots:
I think it is trivial to push entropy down when encoding any particular short string using this method. For example, if I need to encode "voynich manuscript will resist decoding attempts" and I prepare the table which maps each Nth character of each word to the same letter, the ciphertext will read something close to: vvvvvvv mmmmmmmmmm wwww rrrrrr dddddddd aaaaaaaa. There will be a clash somewhere, when Nth letter of word A will be the same as the Nth letter of word B, but as long as the text is short and the table is much larger than the string, it's possible to move a lot of uncertainty from the text into the table.
So, if you have a table of N cells, I think to be safe you will need a text of size ~N * N to estimate true entropy of the encoding. So 500 250000 symbols of natural Latin text (and not just a hand picked list of words). Edit: this is a cautious estimate, maybe a few thousand characters will be more than enough, but in any case there is no lack of large texts in Latin.
