It might have started as Benford's law (actually Newcomb's - 50 years earlier) - but it quickly progressed into number theory.
I believe it does hold for all positive integers - so long as you grow a set outwards from 0, rather than inwards from infinity! I remember the proof Dr Glass of Sydney university tabled in 1982, it went along the lines of examine the first digit of all positive integers starting at 1 and counting in buckets or groups of size Constant k - hence the number of elements in your group is always = n * k as we let n approach infinity.
Start with say k = 50. You get a result you expect. Now let k = 100, then 200, then 900. The results all converge to different numbers - so with any constant bucket size no consistent convergent answer can be found.
Now do the same but set the bucket size to be any power of the base (10 ^ k in our example) - the results always converge to digit d occurs log base 10 (1+1/d).
http://www.rexswain.com/benford.html
Talks about the forensics uses.
Now when I first heard this I thought garbage - so I went to see Prof John Canon who headed the pure maths department to get a really good random number calculator. JC wrote Cayley - a number theory program for handling groups with over 10 ^50 elements and capable of - in his words "infinite precision maths".
JC advised mod a very large pseudo random number ( say 200 digits) - say electrical atmospheric noise by 5 very large primes ( totalling say 120 digits) - this gives you a number that appears to be random in a set of about 10 ^ 300 - say the cube of all atoms in the universe.
So I did - and wrote a program that every order of 10 digits printed out how well the theory match a test. By the time I got to 1,000 digits - it agreed within 3 decimal places. By 10,000 numbers 5 places - by 1 million digits it was accurate to over eight decimal places. I left it going for a month on a mainframe as the idle process and it left no room for doubt.
A few years ago I contacted Prof Leanne Rylands of the Maths department of Western Sydney - she used to be my tutor. Leanne informed the algorithm had been adopted by forensic auditors - because sliced any way an altered dataset can be determined. So take sales data - slice it by period, by geography, by product, by sales team - the trend should be followed!
It even works (slightly weaker) with the second most prominent digit.
I'll look into it more - but my (at time poor memory) records this as a well established and provable law of number theory!
Matt