Re: Randomness from a Quantum source?

David Eather <eather@xxxxxxxxxx> writes:

Unruh wrote:
Wolfgang Draxinger <wdraxinger@xxxxxxxxxxxxxxxx> writes:

David Eather wrote:

Please have a look at this thing:

It makes a claim of 4 or 16 Mbits per second of random data based a
quantum process. I don't dispute the method but is that kind of data
rate feasible / compatible with the concept of measuring one photon at a

Of course. All you need is a very weak laser that produces that amount of
photons per second with a total random polarisation. At a wavelength of
640nm this amounts to a power of 0.7pW.

Getting a laser to produce "totally random polarizations is hard.

This laser is directed to a 50:50 beamsplitter, and at each branch a
avalance photodiode (those can detect single photons). One diode yields a
0-bit the other a 1-bit.

The question is whether or not the photodiode can recover on the 10s of
nanoseconds timescale. Also are you sure that the beamsplitter is really
50-50( biasing the 1s and 0s)? How are you sure? Are you sure that the photodiodes do not become
more/less sensitive after detecting a photon for a while (biasing the
consecutive 1 or 0), or that the dark noise changes after a detection?
Or that the diode produces 1/f noise? All physical systems have biases of some sort
or another. Now they are often predictable or correctable. But if you
are not careful, they bite you.

We have such a thingy at my university, and there's even a service for
students, that they can bring a mobile storage (HDD, USB flashdrive, etc.)
which is filled up with a perfect random bitstream.



If I may ask, how big a doubt would you have that this thing -

might not be producing true random data at the data rate suggested?

No idea. That is an experimental question which you need someone versed in the
art to really know. I would worry about biases as I mentioned, and see if they
have thought of them and checked for them. I suppose that they could distill
randomness out of even a biases source if they have an estimate of those biases.
(eg take 256 bits and run them through say MD5 or SHA to distill 128 bits which
should be unbiased, and uncorrelated.)

I think you need to talk to them, to get at least an idea of what they have
actually done to substantiate their claims.