Set different seed for each sampling in AerStatevector

[hhorii]

Summary

Previously seed is not initialized in AerStatevector and then sampled results are always same. With this commit, a seed is initialized for each sampling and sampled results can be vary.

Details and comments

Previously, sampled counts are always same in multiple calls because AerStatevector uses the same seed to generate random values which are used in processing sample_counts(). More specifically, _seed is not initialized in state_controller if seed_simulator is not set. For example, following two sampled values count0 and count1 are the same.

circ = QuantumVolume(5, seed=1111)
state = AerStatevector(circ)

shots = 1024
counts0 = state.sample_counts(shots, qargs=range(5))
counts1 = state.sample_counts(shots, qargs=range(5))

Even if seed_simulator is set, the value is reused to initialize RngEngine and sampled counts become same.

With this PR, _seed is initialized with std::random_device() and renewed seed_simulator for each sampling, and then sampled counts are always different in the same AerStatevector instance.

[jakelishman]

Is it always going to be the case that a user will want the RNG to advance by calling sample_memory? I'm thinking probably yes, but I'm just mentioning it as a question.

[jakelishman]

This seems like a slightly odd split in responsibilities. Perhaps it would be more convenient for now, and for future development if AerState gains an advance_rng method or something like that mutates the AerState in place. It'll be easier to re-use between the different Aer quantum_info classes you're planning that way, it gives clearer hints as to what's happening, and modifying it is kind of an AerState internal feature that we don't particularly want client classes to need to know about.

[hhorii]

QuantumState advances RNG. So, AerStatevector also needs to advance it in sample_memory().
https://github.com/Qiskit/qiskit-terra/blob/main/qiskit/quantum_info/states/quantum_state.py#L269

However, instead of manually update of seed_simulator by increasing 1, it is better to succeed the state of mt19937 from the previous AerState.

[jakelishman]

What was happening before this change?

[hhorii]

Depending on environemnt. In my macbook (clang), it was always set 0.

[hhorii]

@jakelishman With 4b80871, AerStatevector uses a random number generator of mt19937. Also, AerStatevector overrides seed() to enable reproducing same returns from sample_memory() by calling seed() before each sample_memory().

[jakelishman]

One comment aside, this looks like a much cleaner split of responsibilities to me, thanks!

[jakelishman]

Do we still need to track the seed_ member after this change? I see it's still used in clear_ops(), but it's not 100% clear to me why that function should reset the RNG. Admittedly, I'm not exactly clear on the circumstances in which it's called.

[hhorii]

qiskit-aer/src/controllers/state_controller.hpp

Line 1228 in b5b0c4d

last_result_.seed = buffer_.seed;

This seed_ is to be returned as a member of a result. AerState.last_result() returns it. We may not need to construct a result as AerSimulator does for each call of methods, but it is not expensive to keep seed_, I think.

[jakelishman]

Ok, sounds good to me then!