Incorrect Simulation Results

[glanzz]

Environment

• Qiskit version: 1.2.0
• Python version: 3.11
• Operating system: Apple M2 Air

What is happening?

The following circuit is expected to provide a 110 as dominant state. But it is resulting in state 100 for all shots randomly for optimisation level 2 and 3. The issue could possibly be in the transpilation of the circuit.

How can we reproduce the issue?

from qiskit_aer import StatevectorSimulator
from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager
from qiskit import QuantumCircuit
from qiskit.circuit.library import CU3Gate

cu3 = CU3Gate(2.7955,5.1982,3.3315,)
cu32 = CU3Gate(-2.7955,-3.3315,-5.1982)

qc = QuantumCircuit(3)
qc.x(1)
qc.x(2)
qc.id(1)
qc.swap(0,2)
qc.csx(0,1)
qc.u(5.6631,3.9488,1.066, 2)
qc.tdg(0)
qc.p(2.8007, 1)
qc.r(0.71942,4.7156, 2)


qc.ry(1.8777, 0)
qc.rxx(4.836, 1,2)
qc.rz(3.5166, 0)
qc.r(0.6993,2.7793, 1)
qc.u(0.3461,0.42597,3.202, 2)
qc.z(1)
qc.append(cu3, [2,0])
#qc.cu3( )
qc.ry(4.7962, 0)
qc.rx(2.4125, 1)
qc.r(2.4677,3.5935, 2)
qc.r(-2.4677,3.5935, 2)
qc.rx(-2.4125, 1)
qc.ry(-4.7962, 0)
qc.append(cu32, [2,0])
qc.z(1)
qc.u(-0.3461,-3.202,-0.42597, 2)
qc.r(-0.6993,2.7793, 1)
qc.rz(-3.5166, 0)
qc.rxx(-4.836, 1,2)
qc.ry(-1.8777, 0)
qc.r(-0.71942,4.7156, 2)
qc.p(-2.8007, 1)

qc.t(0)
qc.u(-5.6631,-1.066,-3.9488, 2)
qc.csx(0,1).inverse()
qc.swap(0,2)
qc.id(1)

qc.measure_all()


backend = StatevectorSimulator()
for i in range(4):
  pass_manager=generate_preset_pass_manager(backend=backend, optimization_level=i)
  transpiled_circuit = pass_manager.run(qc)
  #print(transpiled_circuit)
  job = backend.run(transpiled_circuit, shots=11000)
  results = job.result().get_counts()
  print(f"Optimisation Level: {i}, Results: {results}")

What should happen?

Running the script for 1st time provides correct results

python3 issue.py
Optimisation Level: 0, Results: {'110': 11000}
Optimisation Level: 1, Results: {'110': 11000}
Optimisation Level: 2, Results: {'110': 11000}
Optimisation Level: 3, Results: {'110': 11000}

But running the scirpt for the 3rd time gave me the following result:

python3 issue.py
Optimisation Level: 0, Results: {'110': 11000}
Optimisation Level: 1, Results: {'110': 11000}
Optimisation Level: 2, Results: {'100': 11000}
Optimisation Level: 3, Results: {'100': 11000}

But running the same result on Qiskit 1.1.0 is working properly.

Any suggestions?

No response

[glanzz]

Seems like the issue is specific to statevector simulation methods. Running the same script for the backend = AerSimulator(method='matrix_product_state') does work properly.

[Cryoris]

Hmm I can't reproduce this using either of qiskit==1.2.0 or main and qiskit-aer==0.15.0 on Python 3.11 and Mac M1 (not M2 though). What version of Qiskit Aer were you using?

[glanzz]

I am running the same version of qiskit-aer and qiskit !

[Cryoris]

Hmm interesting. If you change the code to run on Qiskit's Statevector class, does the problem remain? This means changing:

from qiskit.quantum_info import Statevector

# your code.... 
qc.csx(0,1).inverse()
qc.swap(0,2)
qc.id(1)

# no measurements!
# qc.measure_all()

for i in range(4):
  pass_manager=generate_preset_pass_manager(backend=backend, optimization_level=i)
  transpiled_circuit = pass_manager.run(qc)
  results = Statevector(transpiled_circuit).probabilities_dict(decimals=6)
  print(f"Optimisation Level: {i}, Results: {results}")

[glanzz]

These are the results:

Optimisation Level: 0, Results: {np.str_('110'): np.float64(1.0)}
Optimisation Level: 1, Results: {np.str_('110'): np.float64(1.0)}
Optimisation Level: 2, Results: {np.str_('100'): np.float64(1.0)}
Optimisation Level: 3, Results: {np.str_('100'): np.float64(1.0)}

[Cryoris]

The problem was reproducable on Qiskit 1.2.0, but disappears on 1.2.1 (I also tried on an M2 mac). This was likely a bug with the CU3Gate which we fixed in 1.2.1. That version was released yesterday, so updating Qiskit should fix the issue 🙂

[glanzz]

@Cryoris Can i know why this issue was specific to optimisation level 2 or 3 ?

[Cryoris]

The issue with the CU3Gate/CUGate (#13121) would only occur when we ran ConsolidateBlocks, which we did on optimization levels 2 and 3 to collect and re-synthesize blocks of 2-qubit gates.

[glanzz]

@Cryoris But, why did it giving correct output in some runs and incorrect output in some results ?

[jakelishman]

The transpiler has stochastic components to it, so unless you set the seed_transpiler argument to generate_preset_pass_manager or transpile, the results of the output circuit might be a bit different each time, though (bugs like this one aside) it should always be an equivalent circuit.

I'll close the issue now since it appears to have been fixed in 1.2.1, but feel free to re-open if there's more to discuss.