In both cases the convergence rate is exponential (considered as a
function of the leg number), the remaining distance halving at each
successive leg.
There is an explanation behind the apparent pattern - it is due to 2^10
= 1024, which above, but not far off 10^3.
Note that your n'th entry is just (1 - 2^-n). So the 2^10 entry causes
the sequence to break (just) above (1 - 10^-3). Then the next two
entries follow the observed pattern, which is down to the closeness of
the 2^10 to 10^3.
Then the 2^20 entry = (2^10)^2 is again just above (but close to) 10^6
and so on. But notice that the relative closeness with 2^20 is less
than with 2^10. It's still close enough to maintain the pattern for the
next few entries, but 2^30 will have an even larger relative
discrepancy, and continuing the sequence the discrepancies grow and the
pattern breaks down. It should be easy to program this and find where
the pattern breaks, but I haven't done this. (I think the "why" is more
interesting than the "when, exactly".)
In the long run, on average we expect it would take log_2 (10) steps to
get another 9 digit, and log_2 (10) = ln(10)/ln(2) = 3.321928.... The
average for your observed pattern would be exactly 10/3 = 3.3333333...
which is too high. (i.e. the 4's in your pattern are slightly too
frequent for the long run)
Regards,
Mike.