Random Numbers

[KINNEY WILLIAM H <kinney@spot.Colorado.EDU>]

There's was some traffic on sci.crypt today about generating random 
numbers by reading noise off a sound port, which ties in to discussion
here of using a Zener diode device. The question is, if you have a 
noise source that is likely to create, say, long strings of zeros or
to have some other statistical bias, how do you fix it up to create
a good distribution?  

Certainly, if your only problem is that you have an input stream where
the ones are randomly distributed but _rare_, in the sense that the
stream is mostly zeros, you can just count ones for a period of time
and create an output stream like

output[i] = 1 if the parity of N input bits is odd
            0 if the parity of N input bits is even

Then the ouput stream will be very high-entropy.

Something similar, but more complicated, would probably apply to reading
thermal noise as well, since you know the input has a Boltzmann 
distribution or whatever, and can transform it to a distribution of
your choice. The problem seems to boil down to having random input
with a distribution f() and transforming it to random output with
another distribuion g(). Or if you want to make it worse, having
some not-really-random input f() and transforming it to random output
g().

But this is probably naive -- what are the pitfalls here? What is the
best way to do it for cryptographic purposes? 


                              -- Will