building a true RNG

[David Wagner]

> An example: presume we take a simple first order statistical model. If our
> input is an 8-bit sample value from a noise source, we will build a 256
> bin histogram. When we see an input value, we look its probability up in
> the model, and discard every 1/(p(x)-1/256)'th sample with value x. When
> this happens, the sample is just eaten and nothing appears in the output;
> otherwise we copy.

I understand what you're trying to say, but this will not give a
general-purpose function that "doesn't waste entropy" regardless of the
input distribution.  This only works when the distribution on the input
stream consists of independent, memoryless samples from some distribution
on 8-bit values.

---------------------------------------------------------------------
The Cryptography Mailing List
Unsubscribe by sending "unsubscribe cryptography" to majordomo at wasabisystems.com