[sdiy] Uniformly distributed noise generator?

[cheater00 .]

Amos,

On Mon, Jun 10, 2013 at 4:53 PM, Amos <controlvoltage at gmail.com> wrote:
> what the heck kind of rocket science are you imagining, for which the
> proposals here might not be "random enough" ??

It's my experience that in measurements, as soon as there's a way for
things to repeat and correlate, very intricate relationships can
happen that will result in incorrect measurement. Managing randomness
in measurement is therefore of utmost importance, just like trackable
derivation of your laboratory standards from the international Big K
standard etc. If you have a pseudo-random source the burden is on you
to prove that it's not correlated with anything else in your
measurement in a way that it will not affect anything else. To me,
this burden is too big. I can see infinite possibilities for
interactions that I would have to meticulously exclude as
possibilities. Unfortunately, here we're dealing with a thing that
can't exactly be proven. On the other hand, analogue noise is, at
least until you get down to very minute detail, certainly going to be
independent. This is much more important than that your randomness
source has exactly perfect spectrum or distribution. Even more so
because you can't get those perfect anyways, since even a test lead
will change either the frequency response or the levels in some way.

Here's an example of what I mean of excluding possibilities.

1. Currently we don't know about any problems with the design. Daniel
made a proposal fix which was accepted.

2. Let me bring up one problem that hasn't been mentioned yet, which
indeed exists.

3. Say you have one of those randomness sources, and you've been using
it for a year. After a year you build a second one. You forget
yourself and reset the batteries in both. They now have the same seed
and state and are deeply correlated.

This is a problem which may not be found for a longer period of time,
which might destroy weeks or months of work for someone relying on the
equipment.

4. So you say, easily solved, use Non-Volatile RAM. I agree, this
problem is thus easily solved.

(list ends here)

The thing is, after 4 is done, it doesn't mean there won't be any more
problems. There is no certainty that you had thought of everything.
Or, in other words, the workflow is that you go from 4 to 1, not from
4 to some state called "5. there are no problems at all". The state 5
doesn't and cannot exist. After you're done fixing your current
problem, you go to "we don't know about any problems currently" rather
than "there are no problems".

The issue here is that one can only disprove that something
pseudorandom is uncorrelated, one cannot prove it, unless you're
inside a very simple universe such as an audio processing program.

That's one reason why for me pseudorandom is not necessarily random enough.

Cheers,
D.