SPAM errors introduce large simulation overhead

[darcangelomauro]

Doing simulations with SPAM errors introduces a significant time overhead dependent on the samples_per_run argument, which was intended to give extra statistics (almost) for free.

The reason is that detection errors (false positives and false negatives) cycle on the sampled bitstrings and randomly flip the bits. In the following simple example execution time increases tenfold when increasing samples_per_run from 10 to 1000:

reg = Register.square(side=2)
pulse = Pulse.ConstantDetuning(BlackmanWaveform(duration=1000, area=np.pi),0,0)
seq = Sequence(reg, MockDevice)
seq.declare_channel('ch', 'rydberg_global')
seq.add(pulse, 'ch')
sim = Simulation(seq)
noise_config = SimConfig(noise=('SPAM', 'doppler', 'amplitude'),
                             eta=0.005, epsilon=0.03, epsilon_prime=0.08,
                             temperature=30, laser_waist=148,
                             runs=10, samples_per_run=10)
sim.set_config(noise_config)
sim.show_config()
res = sim.run()

A first issue is that reducing the epsilon and epsilon_prime parameters to zero does not get rid of the overhead.
Secondly, the time overhead could probably be improved following a suggestion of @Louis-PaulHenry.

As @sebgrijalva pointed out, reducing the evaluation times helps dramatically, so it could be used as a workaround.

[Louis-PaulHenry]

@darcangelomauro When you're mentioning a suggestion from me, do you mean computing the "reweighting" matrix?

[darcangelomauro]

Yes @Louis-PaulHenry, that would eliminate the need to flip bit by bit in each string right?

[Louis-PaulHenry]

yes, I think so (we can discuss it further in private).

[dakk]

I'm trying to work on this issue, however I don't promising any result.