MATLAB: How to duplicate rand(‘twister’, 0) in current Matlab

rng

Hi everyone,
I am trying to change my old code to use the new RNG in current Matlab (R2011a).
In my old code, I usually have something like this,
>> rand('twister', 1);
>> x = rand(1, 1e5);
So I have tried the following experiment to verify that RNG will give me the same random numbers,
>> rng(1, 'twister'); x1 = rand(1,100);
>> rand('twister', 1); x2 = rand(1,100);
>> max(abs(x1 - x2))
ans =
       0
So this works for seed = 1. However when I try setting seed to 0, then they do not match,
>> rng(0, 'twister'); x1 = rand(1,100);
>> rand('twister', 0); x2 = rand(1,100);
>> max(abs(x1 - x2))
ans =
     0.817037959716122
A while back (probably year ago), I have posted some comment about random stream (back then, this is a new feature) and seed of 0 have come up in the discussion. But I cannot find the discussion any more.
So my question: how can I use RNG to duplicate the effect of rand('twister', 0)?
Also besides 0, is there any other seed having this problem?
Thanks Kevin

Best Answer

Historically, using 0 as a seed for MATLAB's random number generators meant, under the hood, "use the special seed that MATLAB uses at startup". It did not mean to literally use the value zero as a seed, because for one thing, 0 was not a valid seed for most of those generators. 'twister', starting in R14 or so, was the one exception to that, 0 meant 0. You probably used 0 in your code because all the older generators had that convenience built in. But to my knowledge, the documentation actually never showed that -- always 5489 or something else non-zero.
Beginning in R2008a (or thereabouts), the Mersenne Twister that you get via the RandStream class, and therefore via the RNG function, goes back to the old meaning for 0: "use the seed that MATLAB uses at startup", which happens to be 5489, the seed that for whatever reasons, the MT designers used as their default.
So the short answer is that you can't directly use RNG to duplicate the effect of rand('twister', 0). 0 for the MT is the only case of this. The longer answer is that you could use the old code, read the state using the old syntax, and muck around with the structure that you get back from RNG. But I recommend not getting into those low-level details.
Related Question