Counts

Counts are the measurement histogram: how many shots produced each bitstring.

Reading counts

from qiskit import QuantumCircuit

from menos.qiskit import MENOSBackend

qc = QuantumCircuit(3, 3)
qc.x(0)
qc.x(2)
qc.measure([0, 1, 2], [0, 1, 2])

counts = MENOSBackend().run(qc, shots=50, seed_simulator=1).result().get_counts()
print(counts)
assert counts == {"101": 50}

The key is written with classical bit 2 leftmost and classical bit 0 rightmost, which is Qiskit's convention. Here qubits 0 and 2 are set, so the key is 101.

Seeds and reproducibility

The same circuit, shots and seed_simulator give identical counts on every run. Different seeds give different samples from the same distribution. See the counts example and Determinism.

Partial measurement

Keys are as wide as the circuit's classical register, and a classical bit that is never written stays 0 (see Circuits and Gates). If the register has one classical bit, the keys are one character wide:

from qiskit import QuantumCircuit

from menos.qiskit import MENOSBackend

qc = QuantumCircuit(3, 1)
qc.x(2)
qc.measure(2, 0)
counts = MENOSBackend().run(qc, shots=100, seed_simulator=1).result().get_counts()
assert counts == {"1": 100}

Other qubits still take part in the simulation if they influence the measured one. If they cannot, light-cone reduction removes them before planning.

Native counts

The native API returns counts as a dictionary keyed by hexadecimal strings, because classical registers may be arbitrarily wide:

from menos import Circuit, Opcode, RunOptions, run

circuit = Circuit(3, 3)
circuit.append(Opcode.X, [0])
circuit.append(Opcode.X, [2])
for q in range(3):
    circuit.append_measure(q, q)

options = RunOptions()
options.shots = 10
options.seed = 1
counts = run(circuit, options).counts
print(counts)
assert counts == {"0x5": 10}   # 0b101

MENOSBackend converts these to Qiskit's binary keys for you.

How counts are sampled

Counts are drawn from the exact outcome distribution. The dense engine draws from the final statevector; the stabilizer engine draws from the affine space of reachable outcomes; the MPS engine draws one shot at a time by walking its tensor chain. On a GPU, sampling still happens on the host from the copied-back state, which is why counts are identical between CPU and CUDA for a given seed.

Cost

Sampling is part of a run's cost. For the MPS engine it can be the larger part: in one measured case sampling took 69% of the run's wall time. See Monitoring.