Running a specific quantum circuit multiple times gets different state vectors

[jiannanWang]

Environment

• Qiskit Terra version: 0.19.1
• Python version: 3.8.0
• Operating system: Ubuntu 22.04.3

What is happening?

When running a specific quantum circuit multiple times, I got different state vector results. To the best of my knowledge, the state vector should not change given a fixed quantum circuit.

How can we reproduce the issue?

The below code is a minimal reproduction example for the issue. It defines a quantum circuit with four gates, runs it 1,000 times, and counts the number of different state vector results.

from qiskit import Aer, transpile, execute
from qiskit import QuantumCircuit, ClassicalRegister, QuantumRegister
from qiskit.circuit.library.standard_gates import *
from qiskit.circuit.add_control import add_control
from qiskit.quantum_info import Statevector
import numpy as np

qr = QuantumRegister(6, name='qr')
cr = ClassicalRegister(6, name='cr')
qc = QuantumCircuit(qr, cr, name='qc')
qc.append(ECRGate(), qargs=[qr[1], qr[0]], cargs=[])
qc.append(C4XGate().control(1), qargs=[qr[5], qr[2], qr[0], qr[3], qr[4], qr[1]], cargs=[])
qc.append(add_control(UGate(4.993516280749269, 0.48372007916715914, 3.4370453288464713), num_ctrl_qubits=1, label=None, ctrl_state=None), qargs=[qr[0], qr[1]], cargs=[])
qc.append(add_control(UGate(5.61106634998484, 4.625022873964695, 0.6622664050777831), num_ctrl_qubits=1, label=None, ctrl_state=None), qargs=[qr[0], qr[1]], cargs=[])

qc = transpile(qc, basis_gates=None, optimization_level=1, coupling_map=None)
simulator = Aer.get_backend('statevector_simulator')

def add_or_update_count_list(st_count_list, st):
    for i, (key_st, count) in enumerate(st_count_list):
        # if np.allclose(key_st, st):
        if key_st.equiv(st):
            st_count_list[i] = (key_st, count + 1)
            return
    st_count_list.append((st, 1))

def count_result_statevector(result_statevector_list: list):
    """Count the number of different values of result statevectors."""
    result_statevecotr_count = []
    for st in result_statevector_list:
        add_or_update_count_list(result_statevecotr_count, st)
    for i, (key_st, count) in enumerate(result_statevecotr_count):
        np_key = np.array(key_st)
        original_len = len(np_key)
        np_key_trimmed = np.trim_zeros(np_key, trim='b')
        trimmed_len = len(np_key_trimmed)
        print(f"{i}: {count}, emitting {original_len - trimmed_len} zeros, {np_key_trimmed}, ") # trim trailing zeros
    print(f"Number of different result statevectors: {len(result_statevecotr_count)}")
    return

results_list = []
for i in range(1000):
    result = execute(qc, simulator).result().get_statevector(qc)
    results_list.append(Statevector(result))
count_result_statevector(results_list)

I ran the above code and got the following results, which show 29 types of different result state vectors.

0: 956, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -3.01633221e-01+4.07897115e-01j
  7.07106781e-01-1.29893408e-16j -2.24140623e-02+4.92072102e-01j],
1: 4, emitting 60 zeros, [5.30245156e-33+4.32978028e-17j 4.32660408e-01+2.64891729e-01j
 7.07106781e-01-1.29893408e-16j 4.92151082e-01-2.06071794e-02j],
2: 1, emitting 60 zeros, [5.30245156e-33+4.32978028e-17j 5.02607531e-01-2.47077153e-01j
 7.07106781e-01-1.29893408e-16j 4.86492157e-02-4.28919345e-01j],
3: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -5.21450038e-01-2.37083130e-01j
  7.07106781e-01-1.29893408e-16j -2.52338686e-01-3.28947769e-01j],
4: 1, emitting 60 zeros, [5.30245156e-33+4.32978028e-17j 5.17810814e-01+2.44929609e-01j
 7.07106781e-01-1.29893408e-16j 2.47343054e-01+3.32720395e-01j],
5: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -2.77601817e-01-4.86414104e-01j
  7.07106781e-01-1.29893408e-16j -4.31103095e-01-2.21059176e-02j],
6: 2, emitting 60 zeros, [5.30245156e-33+4.32978028e-17j 5.31748765e-01+1.75797328e-01j
 7.07106781e-01-1.29893408e-16j 3.35201161e-01-2.71990316e-01j],
7: 1, emitting 60 zeros, [5.30245156e-33+4.32978028e-17j 5.50452165e-01+1.03265983e-01j
 7.07106781e-01-1.29893408e-16j 2.95907541e-01-3.14288527e-01j],
8: 4, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -5.52198647e-01-1.52299728e-01j
  7.07106781e-01-1.29893408e-16j -3.00841521e-01-2.85264486e-01j],
9: 1, emitting 60 zeros, [5.30245156e-33+4.32978028e-17j 5.59937887e-01+1.14460538e-02j
 7.07106781e-01-1.29893408e-16j 2.40263216e-01-3.58625344e-01j],
10: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -4.88417950e-02-5.70730262e-01j
  7.07106781e-01-1.29893408e-16j  1.63103885e-01-3.81154260e-01j],
11: 3, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -2.86868306e-01+4.95807551e-01j
  7.07106781e-01-1.29893408e-16j -3.52015941e-01+2.19011927e-01j],
12: 2, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -4.85148835e-01-3.04547469e-01j
  7.07106781e-01-1.29893408e-16j -2.06185495e-01-3.59679008e-01j],
13: 3, emitting 60 zeros, [5.30245156e-33+4.32978028e-17j 5.31054102e-01-1.77884767e-01j
 7.07106781e-01-1.29893408e-16j 1.05381981e-01-4.18608634e-01j],
14: 1, emitting 60 zeros, [5.30245156e-33+4.32978028e-17j 4.46688893e-01-3.81766143e-01j
 7.07106781e-01-1.29893408e-16j 2.99019891e-01-2.55559679e-01j],
15: 3, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -5.72289783e-01-2.45551724e-02j
  7.07106781e-01-1.29893408e-16j -3.57163088e-01-2.10513600e-01j],
16: 1, emitting 60 zeros, [5.30245156e-33+4.32978028e-17j 1.96385111e-01+5.24494364e-01j
 7.07106781e-01-1.29893408e-16j 4.22044406e-01+9.06480570e-02j],
17: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j  1.56829525e-01+5.50929263e-01j
  7.07106781e-01-1.29893408e-16j -8.73908570e-02+4.05270632e-01j],
18: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -5.39973137e-01-1.48628600e-01j
  7.07106781e-01-1.29893408e-16j -3.20989258e-01+2.88625097e-01j],
19: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -4.86772582e-01+1.42881453e-01j
  7.07106781e-01-1.29893408e-16j -3.15090517e-01+3.78622912e-01j],
20: 2, emitting 60 zeros, [5.30245156e-33+4.32978028e-17j 5.72756047e-01+8.31045951e-03j
 7.07106781e-01-1.29893408e-16j 3.62991237e-01+2.00296803e-01j],
21: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -5.28242306e-01+2.21536946e-01j
  7.07106781e-01-1.29893408e-16j -4.12809402e-01-3.83385552e-02j],
22: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -2.73309770e-01-5.03408703e-01j
  7.07106781e-01-1.29893408e-16j -3.18319667e-03-4.14573654e-01j],
23: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j  1.56522639e-01-5.37737959e-01j
  7.07106781e-01-1.29893408e-16j -2.83892003e-01-3.25182842e-01j],
24: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -5.57921415e-01-4.88379369e-02j
  7.07106781e-01-1.29893408e-16j -2.63691064e-01+3.41768304e-01j],
25: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -4.37260206e-01-3.92530085e-01j
  7.07106781e-01-1.29893408e-16j -2.92708193e-01-2.62765215e-01j],
26: 1, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -1.54460694e-02+5.72608044e-01j
  7.07106781e-01-1.29893408e-16j -2.04803946e-01+3.60467462e-01j],
27: 2, emitting 60 zeros, [ 5.30245156e-33+4.32978028e-17j -3.38928204e-01+4.61785909e-01j
  7.07106781e-01-1.29893408e-16j -3.73684386e-01+1.79558980e-01j],
28: 1, emitting 60 zeros, [5.30245156e-33+4.32978028e-17j 5.43731324e-01-1.34230015e-01j
 7.07106781e-01-1.29893408e-16j 1.38984240e-01-4.08683168e-01j],
Number of different result statevectors: 29

What should happen?

I expected the state vector results to be consistent between multiple runs. Please let me know if my hypothesis is wrong.

Any suggestions?

No response

[jakelishman]

Your assumption is right, in the absence of noise or measurements (i.e. if the input circuit is unitary); Aer should find the same statevector each time for a fixed circuit. If you set seed_simulator (by something like simulator.set_options(seed_simulator=0)), then Aer should find the same results each time for the same input even with noise or measurements.

You're using a rather old Qiskit (Terra) version here, and I think you might be using quite an old version of Aer as well. The main bug you're seeing was in Qiskit Aer (but now fixed), where it wouldn't check that it had received the correct number of parameters to a gate. This manifests in your example because of the use of add_control, which is an internal-only method, and not intended for public use. Doing add_control(UGate(...)) doesn't correctly produce a controlled-U gate parametrised in the way that Aer expects. This meant Aer was reading effectively random memory for one of the parameters, which is what you're seeing.

The fix is:

• update your versions of Qiskit and Qiskit Aer - recent versions will detect the problem and raise an error.
• don't use add_control; it's not documented, and isn't meant to be a public function. Instead, use UGate(...).control(1), like you did with the C4XGate.

---

Just to be on the safe side in the future: you said that you're feeding it the same circuit each time, but actually your code block doesn't do that - it just happens to work out the same in this particular case. execute(qc, backend) is only a helper method that's equivalent to backend.run(transpile(qc, backend)), and transpile has (for real backends) stochastic components. You probably just meant to do simulator.run(qc) at the end, rather than execute(qc, simulator), since your qc was already transpiled.

[jiannanWang]

Thank you for your quick reply! I've updated to the latest version and the code now produces an error instead of the inconsistencies. I'll close this issue since the original bug is fixed.