Noise model doesn't work with identity operations

[jon-wurtz]

Running the following snippet of code (to explore the interplay between doing local gates, vs. global gates and inverting a subset). However, even with to_native_gateset=False the transformer still cuts out the identity qubits. And, I have to write the identity as a PhXZ gate otherwise there is an error.

sqrt_Y = cirq.PhasedXZGate(axis_phase_exponent=0.5,x_exponent=0.5,z_exponent=0)
sqrt_Y_dag = cirq.PhasedXZGate(axis_phase_exponent=0.5,x_exponent=-0.5,z_exponent=0)
identity = cirq.PhasedXZGate(axis_phase_exponent=0.0,x_exponent=0.0,z_exponent=0)

qubits = cirq.LineQubit.range(8)
fidelity_local = []
fidelity_global = []
for k in range(1,8):
    circuit_global = cirq.Circuit.from_moments(
        cirq.Moment([identity.on(q) for q in qubits]),
        cirq.Moment([sqrt_Y.on(q) for q in qubits]),
        cirq.Moment([sqrt_Y_dag.on(q) for q in qubits[k::]]))
    circuit_local = cirq.Circuit.from_moments(
        cirq.Moment([identity.on(q) for q in qubits]),
        cirq.Moment([sqrt_Y.on(q) for q in qubits[0:k]]))
    noisy_circuit_local = transform_circuit(circuit_local, model=model,to_native_gateset=False)
    noisy_circuit_global = transform_circuit(circuit_global, model=model,to_native_gateset=False)
    
    # Simulate noisy circuits
    clean = simulator_dm.simulate(circuit_local).final_density_matrix
    noisy = simulator_dm.simulate(noisy_circuit_local).final_density_matrix
    fidelity_local.append(np.trace(clean @ noisy).real)
    
    clean = simulator_dm.simulate(circuit_global).final_density_matrix
    noisy = simulator_dm.simulate(noisy_circuit_global).final_density_matrix
    fidelity_global.append(np.trace(clean @ noisy).real)
plt.plot(fidelity_local,label='Local')
plt.plot(fidelity_global,label='Global')
plt.legend()
plt.show()

[cduck]

I suppose that means identity gates have no noise.

[jon-wurtz]

The point here is that it should not merge gates -- here, I do a sqrtY rotation globally and then a sqrtY_dagger on a subset to remove those gates. They merge to the identity but in a nontrivial way. The noise model should not do any merging of gates, instead only annotating noise channels.

[cduck]

That is an issue. Looking at the code, I don't see anything that should be merging gates when to_native_gateset and parallelize_circuit are False.

[cduck]

@jon-wurtz Can you write up a minimum example where you do a tiny circuit with just sqrtY, sqrtY_dagger, and Identity and print the output circuit?

[david-pl]

@cduck @jon-wurtz it's this line here:

bloqade-circuit/src/bloqade/cirq_utils/noise/model.py

Line 414 in ed2c0ae

compressed_circuit = cirq.merge_single_qubit_moments_to_phxz(

It's a bit hidden, but yeah this optimization in the noise pass merges the gates. This leads to unexpected results here, I guess no one's really tested circuits that weren't "optimized" in that sense already.
It's also a bit inconsistent: we only do this for the GeminiOneZoneNoiseModel, but not the two-zone one.

Pinging @tcochran-quera here, since I think this is from his original implementation. I'm not sure what the purpose of it is, though.

At the very least, we can add a toggle to switch off the merging.

[tcochran-quera]

I think we're safe removing the merge_single_qubit_moments_to_phxz in this line. It doesn't cause any tests to fail in cirq_utils.

We'll have to be clear that if users want to take advantage of parallelism by hand (like in this global vs local trick) they'll have to set to_native_gateset = False AND parallelize_circuit = False AND program in PhasedXZ

I also am trying to remember why it was important that we parallelize after the interleaving for the one zone model (next lines in model.py). I'll explore as I do some simulations in the next couple of days.