It turns out that for the same reason the NumPy people also have not
used the ziggurat so far. Whereas Octave does use it according to its
documentation. Don't know whether they take into account Doornik's
caveat, but it doesn't sound like it and this seems to explain the speed
advantage over Ox (just like in the above new gretl numbers).
What about introducing the (standard, non-Doornik) ziggurat as an
option, configurable through 'set'?
There exists studies showing how bad
things can happen when using a
deficient RNG for a given purpose. Accuracy and reliability should
always come first. While many programs have more than one RNG, I am
not sure whether a more limited algorithm should be offered as an
option. I know an expert (L'Ecuyer) saying that there is no reason to
use a RNG if a better one is available.
“Remember not only to say the right thing in the right place, but far
more difficult still, to leave unsaid the wrong thing at the tempting
moment.” - Benjamin Franklin (1706-1790)