Path: utzoo!utgpu!jarvis.csri.toronto.edu!cs.utexas.edu!usc!ucsd!ucsdhub!hp-sdd!hp-pcd!hplsla!tomb
From: tomb@hplsla.HP.COM (Tom Bruhns)
Newsgroups: sci.electronics
Subject: Re: A TRUE (not pseudo) random number Chip by AT&T (T7001)
Message-ID: <5170067@hplsla.HP.COM>
Date: 8 Jan 90 23:16:25 GMT
References: <50497@bbn.COM>
Organization: HP Lake Stevens, WA
Lines: 34

aboulang@bbn.com (Albert Boulanger) writes:

>I noticed the spec sheet for this little guy in the Sams book "Video
>Scrambling & Descrambling" Rudolf Graf, & William Sheets. The specs
>for this chip say that it produces TRULY random bits based on the
>phase jitter of a free-running oscillator. It is a companion chip to
>AT&T's T7000A DES chip.
>
>Is there any analysis of the quality of the random numbers based on
>the technology used in this chip? I have been thinking of a similar
>idea but based on XORING a bank of free running oscillators (which is
>in turn based on a simple asynchronous random sequence algorithm I
>have been playing with for MIMD machines). Is there some general
>analysis and description of this generic class of methods for
>producing random sequences?

A slightly tangential comment:  usually "true random" generators have
one or more statistical characteristics that are significantly less
than ideal, and may drift with time.  PRN generators, on the other hand, 
are quite predictable in many of their lower-order statistics (which is
usually where the "true random" fail...from my meager understanding) but
not so good in some of the higher-order things.  Thus one good ploy is
to mix (e.g., XOR) a PRN and a "true random" bit stream.  I think there
are indeed well-known analysis techniques; when you discover them, you
might want to apply them to the suggested implementation.  --  As a
simple example, however, consider a "true random" generator with a bias:
although higher-order statistics are good, its probability of generating
a "1" is, say, 0.4.  If you mix this with a PRN stream with probability
of "1" equal to (2^(n-1)/(2^n-1))  (which is really close to 0.5 for a
moderate n), the probability of generating a "1" is now even closer to
0.5 than for the PRN stream.  And similarly, the "true random" helps out
the higher-order stuff of the PRN:  there is no longer a zero probability
of observing streams of 1's or of 0's longer than n, and the combination
also "never" repeats, even though the PRN sequence, in theory, does.