ndarray-rand with Complex and default PRNG

[termoshtt]

I am considering a plan to port ndarray_linalg::generate::random feature into ndarray-rand crate. It is not depends on non-Ruat code (=LAPACK), and I think it will be useful for non-linalg users.

I'd like to improve the current API in the following two points. These are partially done in ndarray-linalg implementation.

Complex support

• num-complex has rand feature which impl Distribution<Complex> for Standard
• It will be rand 0.5 while num-complex 0.2, i.e. incompatible to ndarray-rand doesn't work with rand 0.7 #658, Upgrade to rand 0.7.0 #659 until 0.3 will be released

Default PRNG

• Most ndarray users will seek ndarray_rand::random((n, n)) which does not need to initialize PRNG manually.
• They will not need CSPRNG, but rand::thread_rng returns CSPRNG. For Monte-Calro or randomized algorithm usage, Mersenne twister will be good choice. I've re-implement SFMT in pure-Rust with std::arch.

[LukeMathWalker]

For complex support, how should we proceed (once they update to rand 0.7)? Should we let the end-user pull in num-complex to get complex distributions or are you suggesting we should pull in num-complex in ndarray-rand directly?

On default PRNG, I do agree that a CSPRNG is probably overkill as the default RNG for random - if we were to use SFMT I assume we would need a fallback for non x86/x86_64 architectures, am I correct?

[bluss]

The thread_rng doesn't seem like a high quality CSPRNG but the point is taken - that one is probably overkill. Ndarray-rand has never defaulted to it - as is documented. ("The default RNG is a fast automatically seeded rng (currently rand::rngs::SmallRng seeded from rand::thread_rng).")

And picking a new default with a good motivation sounds like a great idea.