Writing a fake Voynich by rules: copies, habits and the link between words

[quimqu]

Zipf slope: 1.00 vs 1.00

Surprising. How do you measure the Zipf slope? Is there a standard way to do it, like averaging between something and something?

The log-log graph is a curve, and only a small portion gets close to a slope of 1 (actually -1).

This was plotted from RF1a-n:

Blue dots are (order, frequency), orange dots are (order, constant/order), in a log-log graph.

You’re right, and “1.00” was misleading. Here’s what I actually did:

I took the 1000 most frequent words and fitted a single straight line through them on the log-log plot. Each word counts as one point, so 900 of the 1000 points are words ranked 100 to 1000. That part of the curve is the steepest, and it dominates the fit. The real values were 1.005 (manuscript) and 0.993 (scribe), and I rounded both to 1.00.

As you say, the plot is a curve, not a line, so one slope doesn’t describe it well. Measured piece by piece:

• Top 10 words: 0.40 (manuscript) vs 0.45 (scribe)
  
• Ranks 10–100: 0.80 vs 0.80
  
• Ranks 100–1000: 1.15 vs 1.10
  

(Scribe values are the average of 5 runs.)

So “Zipf slope 1” was the wrong way to put it. I ammend the comment.