Ways to check if book dog ears are random? I this article the royal society claims that dog ears left by Newton in his books are non-random and meant to point to specific parts of the text as a bookmark of sorts. I was wondering if there was some kind of test that one could try to see how probable it is to make ears like these randomly. I have pictures of the ears in one book, and was thinking that perhaps it would be possible to quantify the X/Y position based on the line (y) and character (x) and then taking the std of the lot? Is there some smarter way to do this? 
 A: If the issue is to confirm that Newton consistently turned
the corner down to point to a particular word or equation, I
don't see how to formulate either a satisfactory statement
of the null hypothesis or of a null distribution to go with it.
If the issue is whether Newton's 23 dog-eared pages in 
a particular 400-page book have numbers that are randomly distributed,
then the null hypothesis is that the page numbers are consistent with
a sample from $\mathsf{Unif}(0, 400).$ 
Specifically suppose page numbers of his dog-eared pages in
this book are as follows:
sort(x)
 [1]  31  47  53  54  77  83  84 103 112 113 115 122
[13] 135 165 177 190 197 232 241 252 257 319 329

length(x)
[1] 23

summary(x)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
   31.0    83.5   122.0   151.7   214.5   329.0 

Then a Kolmogorov-Smirnov test of goodness fit shows that the
page numbers are not random. Presumably, from the summary above
this is because dog-eared pages seem to be more prevalent towards the beginning of the book. 
ks.test(x, dunif, 0, 400)

        One-sample Kolmogorov-Smirnov test

data:  x
D = 0.9975, p-value = 4.441e-16
alternative hypothesis: two-sided

Notes:  (1) The vector x was generated as 400 times a sample of size
23 from $\mathsf{Beta}(2,4)$ and rounded up to the next higher integer. 
(2) The test is approximate because data are discrete and 
the distribution $\mathsf{Unif}(0, 400)$ of comparison is continuous. Also, the K-S test does not envision ties, which will occur in slightly more than half of the samples. But errors are slight. 
prod(400:(400-22)/400)
[1] 0.5248176

(3) In this particular example, out of 23 dog-eared pages, there are
17 with numbers below 200. This is not
consistent with the distribution $\mathsf{Binom}(23, .5).$
sum(dbinom(17:23, 23, .5))
[1] 0.01734483

