Density matrix simulator EVs

[95-martin-orion]

Fixes #3964. CC @zaqqwerty

This is a near-exact copy of the behavior and tests for sparse_simulator. If we see the need in the future, this code could probably be folded into a SimulatesEVsAndFullState interface as the default simulate_expectation_values_sweep behavior, but that doesn't seem particularly urgent at the moment.

Relevant changes from the sparse_simulator code:

• obs.expectation_from_density_matrix(result.final_density_matrix, qmap) instead of state-vector equivalent
• added test_simulate_noisy_expectation_values test
• that's it

[95-martin-orion]

Assigning to @balopat (the next triage expert) for review, since I can't review my own PRs 😞

[balopat]

I would like to reduce the duplication introduced here.

1.) production code: I would experiment with pulling out the repeated lines and have a new "protected" method expectation_value_from_result(obs: PauliSum, result:TSimulationResult) -> float which then can be called from the abstract method

2.) test code: I would move factor out the expectation value calculation logic to a separate test.py file and parameterized it with the different simulator types.

[95-martin-orion]

1.) production code: I would experiment with pulling out the repeated lines and have a new "protected" method expectation_value_from_result(obs: PauliSum, result:TSimulationResult) -> float which then can be called from the abstract method

Packing the duplicated code in simulate_expectation_values_sweep into a Result-targeted method would force us to repeat the construction of pslist for every result. Probably not a huge overhead relative to simulation costs, but if we can avoid it I think we should.

I stand by my top-of-PR comment: the correct way to de-dupe this would be something closer to #3990. (Probably the cleanest solution would involve converting *_sweep methods to return Iterable, but that's a separate discussion.)

[balopat]

I stand by my top-of-PR comment: the correct way to de-dupe this would be something closer to #3990.

Ah, cool, I missed that, and oh, nice, yes that is exactly the direction I was thinking - having an expectation_value_from_state (I don't like expectation_from_state) implemented - though you're also right that maybe not all Simulators should be mandated to implement EV calculation, and isolating that responsibility to a separate interface like SimulatesEVandFullState or similar could be useful to make EV calculation implementation opt-in.

Also, note that #3990 adds yet another test with exactly the same cases at a first blink...maybe we should think about extracting a cirq.testing.assert_consistently_calculates_expectation_value(simulator) function that could be reused in tests for all the different implementations!

I'm okay to get this merged but I would love to have an issue created for deduping both impl and tests for this and referenced in a comment in all these places.

(Probably the cleanest solution would involve converting *_sweep methods to return Iterable, but that's a separate discussion.)

This sounds interesting but I just don't see why yet, can you elaborate on this idea?

[balopat]

LGTM - given we add an issue and a comment for deduping tests & design

[balopat]

ping @95-martin-orion we could merge this, just add in a comment so we don't forget to dedupe

[95-martin-orion]

Opened #4209 to track deduplication.