Deteksi error dengan overhead rendah menggunakan spacetime codes

Estimasi penggunaan: 10 detik pada prosesor Heron r3 (CATATAN: Ini hanyalah estimasi. Waktu aktual bisa berbeda.)

Pengantar

Low-overhead error detection with spacetime codes [1] oleh Simon Martiel dan Ali Javadi-Abhari mengusulkan sintesis pemeriksaan spacetime berbobot rendah yang memperhatikan konektivitas untuk sirkuit yang didominasi Clifford, lalu melakukan post-selection pada pemeriksaan tersebut untuk menangkap kesalahan dengan overhead jauh lebih rendah dibanding koreksi error penuh dan jumlah shot lebih sedikit dari mitigasi error standar.

Makalah ini mengusulkan metode baru untuk deteksi error pada sirkuit kuantum (khususnya sirkuit Clifford) yang menyeimbangkan antara koreksi error penuh dan teknik mitigasi yang lebih ringan. Ide utamanya adalah menggunakan spacetime codes untuk menghasilkan "pemeriksaan" di sepanjang sirkuit yang mampu menangkap error, dengan overhead qubit dan gate yang jauh lebih rendah dibanding koreksi error toleran-kesalahan penuh. Para penulis merancang algoritme efisien untuk memilih pemeriksaan berbobot rendah (melibatkan sedikit qubit), kompatibel dengan konektivitas fisik perangkat, dan mencakup wilayah temporal dan spasial sirkuit yang luas. Mereka mendemonstrasikan pendekatan ini pada sirkuit hingga 50 qubit logis dan ~2450 gate CZ, menghasilkan peningkatan fidelitas fisik-ke-logis hingga 236 kali lipat. Perlu dicatat pula bahwa seiring bertambahnya operasi non-Clifford dalam sirkuit, jumlah pemeriksaan yang valid menyusut secara eksponensial, menandakan metode ini paling efektif untuk sirkuit yang didominasi Clifford. Secara keseluruhan, dalam jangka pendek, deteksi error melalui spacetime codes dapat menawarkan jalur praktis dan overhead rendah untuk meningkatkan keandalan pada perangkat keras kuantum.

Teknik deteksi error ini bertumpu pada konsep coherent Pauli checks dan didasarkan pada karya Single-shot error mitigation by coherent Pauli checks [2] oleh van den Berg et al.

Baru-baru ini, makalah Big cats: entanglement in 120 qubits and beyond [3] oleh Javadi-Abhari et al. melaporkan pembuatan state Greenberger-Horne-Zeilinger (GHZ) 120-qubit, state multikuantum yang paling besar dicapai sejauh ini pada platform qubit superkonduktor. Menggunakan kompiler yang memperhatikan hardware, deteksi error overhead rendah, dan teknik "temporary uncomputation" untuk mengurangi noise, para peneliti mencapai fidelitas 0,56 ± 0,03 dengan efisiensi post-selection sekitar 28%. Karya ini mendemonstrasikan entanglement sejati di semua 120 qubit, memvalidasi beberapa metode sertifikasi fidelitas, dan menandai tonggak penting bagi skalabilitas perangkat keras kuantum.

Tutorial ini dibangun di atas ide-ide tersebut, membimbingmu mengimplementasikan algoritme deteksi error pertama pada sirkuit Clifford acak skala kecil, lalu melalui tugas persiapan state GHZ, agar kamu bisa bereksperimen dengan deteksi error pada sirkuit kuantummu sendiri.

Persyaratan

Sebelum memulai tutorial ini, pastikan kamu telah menginstal:

• Qiskit SDK v2.0 atau lebih baru, dengan dukungan visualisasi
• Qiskit Runtime v0.40 atau lebih baru (pip install qiskit-ibm-runtime)
• Qiskit Aer v0.17.2 (pip install qiskit-aer)
• Qiskit Device Benchmarking (pip install "qiskit-device-benchmarking @ git+https://github.com/qiskit-community/qiskit-device-benchmarking.git")
• NumPy v2.3.2 (pip install numpy)
• Matplotlib v3.10.7 (pip install matplotlib)

Setup

# Added by doQumentation — required packages for this notebook
!pip install -q matplotlib numpy qiskit qiskit-aer qiskit-device-benchmarking qiskit-ibm-runtime

# Standard library imports
from collections import defaultdict, deque
from functools import partial

# External libraries
import matplotlib.pyplot as plt
import numpy as np

# Qiskit
from qiskit import ClassicalRegister, QuantumCircuit
from qiskit.circuit import Delay
from qiskit.circuit.library import RZGate, XGate
from qiskit.converters import circuit_to_dag, dag_to_circuit
from qiskit.quantum_info import Pauli, random_clifford
from qiskit.transpiler import AnalysisPass, PassManager
from qiskit.transpiler.passes import (
    ALAPScheduleAnalysis,
    CollectAndCollapse,
    PadDelay,
    PadDynamicalDecoupling,
    RemoveBarriers,
)
from qiskit.transpiler.passes.optimization.collect_and_collapse import (
    collect_using_filter_function,
    collapse_to_operation,
)
from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager
from qiskit.visualization import plot_histogram

# Qiskit Aer
from qiskit_aer import AerSimulator
from qiskit_aer.noise import NoiseModel, ReadoutError, depolarizing_error

# Qiskit IBM Runtime
from qiskit_ibm_runtime import QiskitRuntimeService
from qiskit_ibm_runtime import SamplerV2 as Sampler

# Qiskit Device Benchmarking
from qiskit_device_benchmarking.utilities.gate_map import plot_gate_map

Contoh pertama

Untuk mendemonstrasikan metode ini, kita mulai dengan membangun sirkuit Clifford sederhana. Tujuan kita adalah mampu mendeteksi ketika jenis error tertentu terjadi dalam sirkuit ini, sehingga kita bisa membuang hasil pengukuran yang keliru. Dalam terminologi deteksi error, ini juga dikenal sebagai payload circuit kita.

circ = random_clifford(num_qubits=2, seed=11).to_circuit()
circ.draw("mpl")

Tujuan kita adalah menyisipkan coherent Pauli check ke dalam payload circuit ini. Namun sebelum itu, kita pisahkan sirkuit ini menjadi lapisan-lapisan. Ini akan berguna nanti saat menyisipkan gate Pauli di antaranya.

# Separate circuit into layers
dag = circuit_to_dag(circ)
circ_layers = []
for layer in dag.layers():
    layer_as_circuit = dag_to_circuit(layer["graph"])
    circ_layers.append(layer_as_circuit)

# Create subplots
fig, (ax1, ax2, ax3, ax4, ax5) = plt.subplots(1, 5, figsize=(10, 4))

# Draw circuits on respective axes
circ_layers[0].draw(output="mpl", ax=ax1)
circ_layers[1].draw(output="mpl", ax=ax2)
circ_layers[2].draw(output="mpl", ax=ax3)
circ_layers[3].draw(output="mpl", ax=ax4)
circ_layers[4].draw(output="mpl", ax=ax5)

# Adjust layout to prevent overlap
plt.tight_layout()
plt.show()

Sekarang kita siap menambahkan coherent Pauli checks ke dalam payload circuit. Untuk melakukan ini, kita perlu membangun "pemeriksaan yang valid" dan menyisipkannya ke dalam sirkuit. Sebuah "pemeriksaan" dalam hal ini adalah operator yang mampu memberi sinyal apakah error terjadi dalam sirkuit dengan melakukan pengukuran pada qubit ancilla. Pemeriksaan dianggap valid ketika operator tambahan yang disisipkan ke dalam sirkuit kuantum tidak mengubah sirkuit asli secara logis.

Pemeriksaan ini mampu mendeteksi jenis error yang antikommut dengannya, dan pemeriksaan akan memicu pengukuran state $\ket{1}$ pada qubit ancilla, bukan $\ket{0}$, melalui phase kickback. Oleh karena itu, kita akan bisa membuang pengukuran di mana error telah ditandai.

Secara umum, coherent Pauli checks adalah operator Pauli terkendalikan yang disisipkan ke dalam "wire" — lokasi spacetime di antara gate. Qubit ancilla yang bertanggung jawab memberi sinyal error adalah qubit kontrol.

Di bawah ini kita membangun pemeriksaan yang valid untuk sirkuit Clifford yang dibuat di atas. Kita bisa mendemonstrasikan bahwa pemeriksaan ini tidak mengubah operasi sirkuit dengan menunjukkan bahwa ketika Pauli checks ini dipropagasi ke bagian depan sirkuit, mereka saling membatalkan. Ini mudah ditunjukkan karena operator Pauli yang melalui gate Clifford adalah operator Pauli lainnya.

Secara umum, seseorang dapat menggunakan heuristik decoding seperti yang diuraikan dalam [1] untuk mengidentifikasi pemeriksaan yang valid. Untuk keperluan contoh awal kita, kita juga dapat membangun pemeriksaan valid menggunakan kondisi perkalian gate Pauli dan Clifford secara analitik.

# Define a valid check
pauli_1 = Pauli("ZI")
pauli_2 = Pauli("XZ")

circ_1 = circ_layers[0].compose(circ_layers[1])
circ_1.draw("mpl")

pauli_1_ev = pauli_1.evolve(circ_1, frame="h")
pauli_1_ev

Pauli('-ZI')

circ_2 = circ.copy()
circ_2.draw("mpl")

pauli_2_ev = pauli_2.evolve(circ_2, frame="h")
pauli_2_ev

Pauli('-ZI')

pauli_1_ev.dot(pauli_2_ev)

Pauli('II')

Seperti yang bisa kita lihat, kita memiliki pemeriksaan yang valid, karena operator Pauli yang disisipkan hanya memiliki efek yang sama dengan operator identitas pada sirkuit. Kita sekarang bisa menyisipkan pemeriksaan-pemeriksaan ini ke dalam sirkuit dengan qubit ancilla. Qubit ancilla ini, atau qubit pemeriksaan, dimulai dalam state $\ket{+}$. Ia mencakup versi terkendalikan dari operasi Pauli yang diuraikan di atas dan akhirnya diukur dalam basis $X$. Qubit pemeriksaan ini kini mampu menangkap error dalam payload circuit tanpa mengubahnya secara logis. Ini karena jenis noise tertentu dalam payload circuit akan mengubah state qubit pemeriksaan, dan akan diukur "1" alih-alih "0" jika error semacam itu terjadi.

# New circuit with 3 qubits (2 payload + 1 ancilla for check)
circ_meas = QuantumCircuit(3)
circ_meas.h(0)
circ_meas.compose(circ_layers[0], [1, 2], inplace=True)
circ_meas.compose(circ_layers[1], [1, 2], inplace=True)
circ_meas.cz(0, 2)
circ_meas.compose(circ_layers[2], [1, 2], inplace=True)
circ_meas.compose(circ_layers[3], [1, 2], inplace=True)
circ_meas.compose(circ_layers[4], [1, 2], inplace=True)
circ_meas.cz(0, 1)
circ_meas.cx(0, 2)
circ_meas.h(0)

# Add measurement to payload qubits
c0 = ClassicalRegister(2, name="c0")
circ_meas.add_register(c0)
circ_meas.measure(1, c0[0])
circ_meas.measure(2, c0[1])

# Add measurement to check qubit
c1 = ClassicalRegister(1, name="c1")
circ_meas.add_register(c1)
circ_meas.measure(0, c1[0])

# Visualize the final circuit with the inserted checks
circ_meas.draw("mpl")

Jika qubit pemeriksaan diukur sebagai "0", kita menyimpan pengukuran tersebut. Jika diukur sebagai "1", berarti error telah terjadi dalam payload circuit, dan kita membuang pengukuran itu.

# Noiseless simulation using stabilizer method
sim_stab = AerSimulator(method="stabilizer")
res = sim_stab.run(circ_meas, shots=1000).result()
counts_noiseless = res.get_counts()
print(f"Stabilizer simulation result: {counts_noiseless}")

Stabilizer simulation result: {'0 11': 523, '0 01': 477}

# Plot the noiseless results
# Note that the first bit in the key corresponds to the check qubit
plot_histogram(counts_noiseless)

Perhatikan bahwa dengan simulator ideal, qubit pemeriksaan tidak mendeteksi error apa pun. Kita sekarang memperkenalkan model noise ke simulasi dan melihat bagaimana qubit pemeriksaan menangkap error.

# Qiskit Aer noise model
noise = NoiseModel()
p2 = 0.003  # two-qubit depolarizing per CZ
p1 = 0.001  # one-qubit depolarizing per 1q Clifford
pr = 0.01  # readout bit-flip probability

# 1q depolarizing on common 1q gates
e1 = depolarizing_error(p1, 1)
for g1 in ["id", "rz", "sx", "x", "h", "s"]:
    noise.add_all_qubit_quantum_error(e1, g1)

# 2q depolarizing on CZ
e2 = depolarizing_error(p2, 2)
noise.add_all_qubit_quantum_error(e2, "cz")

# Readout error on measure
ro = ReadoutError([[1 - pr, pr], [pr, 1 - pr]])
noise.add_all_qubit_readout_error(ro)

# Qiskit Aer simulation with noise model
aer = AerSimulator(method="automatic", seed_simulator=43210)
job = aer.run(circ_meas, shots=1000, noise_model=noise)
result = job.result()
counts_noisy = result.get_counts()
print(f"Noise model simulation result: {counts_noisy}")

Noise model simulation result: {'1 01': 5, '0 11': 478, '1 11': 6, '1 00': 2, '1 10': 1, '0 01': 500, '0 00': 5, '0 10': 3}

plot_histogram(counts_noisy)

Seperti yang bisa kita lihat, beberapa pengukuran menangkap error dengan menandai qubit pemeriksaan sebagai "1", yang terlihat pada empat kolom terakhir. Shot-shot ini dibuang.

Catatan: Qubit ancilla juga bisa memperkenalkan error baru ke dalam sirkuit. Untuk mengurangi efek ini, kita bisa menyisipkan pemeriksaan bertingkat dengan qubit ancilla tambahan ke dalam sirkuit kuantum.

Contoh nyata: Mempersiapkan state GHZ pada hardware nyata

Langkah 1: Petakan input klasik ke masalah kuantum

Sekarang kita mendemonstrasikan tugas penting untuk algoritme komputasi kuantum, yaitu mempersiapkan state GHZ. Kita akan mendemonstrasikan cara melakukan ini pada Backend nyata menggunakan deteksi error.

# Set optional seed for reproducibility
SEED = 1

if SEED:
    np.random.seed(SEED)

Algoritme deteksi error untuk persiapan state GHZ menghormati topologi hardware. Kita mulai dengan memilih hardware yang diinginkan.

# This is used to run on real hardware
service = QiskitRuntimeService()

# Choose a backend to build GHZ on
backend_name = service.least_busy(
    operational=True, simulator=False, min_num_qubits=133
)

backend = service.backend(backend_name)
coupling_map = backend.target.build_coupling_map()

State GHZ pada $n$ qubit didefinisikan sebagai $\lvert \mathrm{GHZ}_n\rangle \;=\; \frac{1}{\sqrt{2}}\Big(\lvert 0\rangle^{\otimes n} \,+\, \lvert 1\rangle^{\otimes n}\Big).$

Pendekatan yang sangat naif untuk mempersiapkan state GHZ adalah memilih qubit root dengan gate Hadamard awal, yang menempatkan qubit ke state superposisi yang setara, lalu mengentangle qubit ini dengan setiap qubit lainnya. Ini bukan pendekatan yang baik, karena memerlukan interaksi CNOT jangkauan panjang dan dalam. Dalam tutorial ini, kita akan menggunakan beberapa teknik bersama deteksi error untuk mempersiapkan state GHZ secara andal pada hardware nyata.

Langkah 2: Optimalkan masalah untuk eksekusi pada hardware kuantum

Petakan state GHZ ke hardware

Pertama, kita mencari root untuk memetakan Circuit GHZ pada hardware. Kita menghapus edge/node yang error CZ-nya, error pengukurannya, dan nilai T2-nya lebih buruk dari ambang batas di bawah ini. Ini tidak akan disertakan dalam Circuit GHZ.

def bad_cz(target, threshold=0.01):
    """Return list of edges whose CZ error is worse than threshold."""
    undirected_edges = []
    for edge in backend.target.build_coupling_map().get_edges():
        if (edge[1], edge[0]) not in undirected_edges:
            undirected_edges.append(edge)
    edges = undirected_edges
    cz_errors = {}
    for edge in edges:
        cz_errors[edge] = target["cz"][edge].error
    worst_edges = sorted(cz_errors.items(), key=lambda x: x[1], reverse=True)
    return [list(edge) for edge, error in worst_edges if error > threshold]

def bad_readout(target, threshold=0.01):
    """Return list of nodes whose measurement error is worse than threshold."""
    meas_errors = {}
    for node in range(backend.num_qubits):
        meas_errors[node] = target["measure"][(node,)].error
    worst_nodes = sorted(
        meas_errors.items(), key=lambda x: x[1], reverse=True
    )
    return [node for node, error in worst_nodes if error > threshold]

def bad_coherence(target, threshold=60):
    """Return list of nodes whose T2 value is lower than threshold."""
    t2s = {}
    for node in range(backend.num_qubits):
        t2 = target.qubit_properties[node].t2
        t2s[node] = t2 * 1e6 if t2 else 0
    worst_nodes = sorted(t2s.items(), key=lambda x: x[1])
    return [node for node, val in worst_nodes if val < threshold]

THRESH_CZ = 0.025  # exclude from BFS those edges whose CZ error is worse than this threshold
THRESH_MEAS = 0.15  # exclude from BFS those nodes whose measurement error is worse than this threshold
THRESH_T2 = 10  # exclude from BFS those nodes whose T2 value is lower than this threshold

bad_edges = bad_cz(backend.target, threshold=THRESH_CZ)
bad_nodes_readout = bad_readout(backend.target, threshold=THRESH_MEAS)
dead_qubits = bad_readout(backend.target, threshold=0.4)
bad_nodes_coherence = bad_coherence(backend.target, threshold=THRESH_T2)
bad_nodes = list(set(bad_nodes_readout) | set(bad_nodes_coherence))
print(f"{len(bad_edges)} bad edges: \n{bad_edges}")
print(f"{len(bad_nodes)} bad nodes: \n{bad_nodes}")

17 bad edges: 
[[30, 31], [112, 113], [113, 114], [113, 119], [120, 121], [130, 131], [145, 146], [146, 147], [111, 112], [55, 59], [64, 65], [131, 138], [131, 132], [119, 133], [129, 130], [47, 57], [29, 38]]
5 bad nodes: 
[1, 113, 131, 146, 120]

Menggunakan fungsi di bawah ini, kita membangun Circuit GHZ pada hardware yang dipilih mulai dari root dan menggunakan breadth-first search (BFS).

def parallel_ghz(root, num_qubits, backend, bad_edges, skip):
    """
    Build a GHZ state of size `num_qubits` on the given `backend`,
    starting from `root`, expanding in BFS order.

    At each BFS layer, every active qubit adds at most one new neighbor
    (so that two-qubit operations can run in parallel with no qubit conflicts).

    It grows the entanglement tree outward layer-by-layer.
    """

    # -------------------------------------------------------------
    # (1) Filter usable connections from the backend coupling map
    # -------------------------------------------------------------
    # The coupling map lists all directed hardware connections as (control, target).
    # We remove edges that are:
    #   - listed in `bad_edges` (or their reversed form)
    #   - involve a qubit in the `skip` list
    cmap = backend.configuration().coupling_map
    edges = [
        e
        for e in cmap
        if e not in bad_edges
        and [e[1], e[0]] not in bad_edges
        and e[0] not in skip
        and e[1] not in skip
    ]

    # -------------------------------------------------------------
    # (2) Build an undirected adjacency list for traversal
    # -------------------------------------------------------------
    # Even though coupling_map edges are directed, BFS expansion just needs
    # connectivity information (so we treat edges as undirected for search).
    adj = defaultdict(list)
    for u, v in edges:
        adj[u].append(v)
        adj[v].append(u)

    # -------------------------------------------------------------
    # (3) Initialize the quantum circuit and BFS state
    # -------------------------------------------------------------
    n = backend.configuration().num_qubits
    qc = QuantumCircuit(
        n
    )  # create a circuit with same number of qubits as hardware
    visited = [
        root
    ]  # record the order qubits are added to the GHZ chain/tree
    queue = deque([root])  # BFS queue (start from root)
    explored = defaultdict(
        set
    )  # to track which neighbors each node has already explored
    layers = []  # list of per-layer (control, target) gate tuples
    qc.h(root)  # GHZ states start with a Hadamard on the root qubit

    # -------------------------------------------------------------
    # (4) BFS expansion: build the GHZ tree one layer at a time
    # -------------------------------------------------------------
    # Loop until we’ve added the desired number of qubits to the GHZ
    while queue and len(visited) < num_qubits:
        layer = []  # collect new (control, target) pairs for this layer
        current = list(
            queue
        )  # snapshot current frontier (so queue mutations don’t affect iteration)
        busy = (
            set()
        )  # track qubits already used in this layer (to avoid conflicts)

        for node in current:
            queue.popleft()

            # find one unvisited neighbor of this node not already explored
            unvisited_neighbors = [
                nb
                for nb in adj[node]
                if nb not in visited and nb not in explored[node]
            ]

            if unvisited_neighbors:
                nb = unvisited_neighbors[
                    0
                ]  # pick the first available neighbor
                visited.append(nb)  # mark it as part of the GHZ structure
                queue.append(
                    node
                )  # re-enqueue current node (can keep growing)
                queue.append(nb)  # enqueue the newly added qubit
                explored[node].add(nb)  # mark that edge as explored
                layer.append(
                    (node, nb)
                )  # schedule a CNOT between node and neighbor
                busy.update([node, nb])  # reserve both qubits for this layer

                # stop early if we've reached the desired number of qubits
                if len(visited) == num_qubits:
                    break
            # else: node has no unused unvisited neighbors left → skip

        if layer:
            # add all pairs (node, nb) scheduled this round to layers
            layers.append(layer)
        else:
            # nothing new discovered this pass → done
            break

    # -------------------------------------------------------------
    # (5) Emit all layers into the quantum circuit
    # -------------------------------------------------------------
    # For each layer:
    #   - apply a CX gate for every (control, target) pair
    #   - insert a barrier so transpiler keeps layer structure
    for layer in layers:
        for q1, q2 in layer:
            qc.cx(q1, q2)
        qc.barrier()

    # -------------------------------------------------------------
    # (6) Return outputs
    # -------------------------------------------------------------
    # qc: the built quantum circuit
    # visited: order of qubits added
    # layers: list of parallelizable two-qubit operations per step
    return qc, visited, layers

Kita sekarang berulang kali mencari root terbaik, tempat Circuit GHZ akan berasal.

ROOT = None  # root for BFS search
GHZ_SIZE = 100  # number of (data) qubits in the GHZ state
SKIP = []  # nodes to intentionally skip so that we have a better chance for finding checks

# Search for the best root (yielding the shallowest GHZ)
if ROOT is None:
    best_root = -1
    base_depth = 100
    for root in range(backend.num_qubits):
        qc, ghz_qubits, _ = parallel_ghz(
            root, GHZ_SIZE, backend, bad_edges, SKIP
        )
        if len(ghz_qubits) != GHZ_SIZE:
            continue
        depth = qc.depth(lambda x: x.operation.num_qubits == 2)
        if depth < base_depth:
            best_root = root
            base_depth = depth
    ROOT = best_root

Kita sekarang membangun Circuit GHZ mulai dari node tertentu — yaitu root terbaik — mencari kedalaman terpendek menggunakan breadth-first search.

# Build a GHZ starting at the best root
qc, ghz_qubits, _ = parallel_ghz(
    ROOT, GHZ_SIZE, backend, bad_edges, SKIP + bad_nodes
)
base_depth = qc.depth(lambda x: x.operation.num_qubits == 2)
base_count = qc.size(lambda x: x.operation.num_qubits == 2)
print(f"base depth: {base_depth}, base count: {base_count}")
print(f"ROOT: {ROOT}")
if len(ghz_qubits) != GHZ_SIZE:
    raise Exception("No GHZ found. Relax error thresholds.")

base depth: 17, base count: 99
ROOT: 50

Kita perlu satu pertimbangan terakhir sebelum menyisipkan pemeriksaan yang valid. Ini berkaitan dengan konsep "coverage", yang merupakan ukuran berapa banyak wire dalam sirkuit kuantum yang bisa dicakup oleh sebuah pemeriksaan. Dengan coverage yang lebih tinggi, kita bisa mendeteksi error pada bagian sirkuit yang lebih luas. Dengan ukuran ini, kita bisa memilih di antara pemeriksaan valid yang memiliki coverage sirkuit tertinggi. Dengan kata lain, kita akan menggunakan fungsi weighted_coverage untuk menilai berbagai pemeriksaan untuk Circuit GHZ.

def weighted_coverage(layers, parities, w_idle=0.2, w_gate=0.8):
    """
    Compute weighted fraction (idle + gate) of wires that are
    covered by at least one parity to all active wires.
    """
    wires = active_wires(layers)  # defined below
    covered_by_any = {n_layer: set() for n_layer in range(len(layers))}
    for parity in parities:
        trace = z_trace_backward(layers, parity)  # defined below
        for n_layer, qs in trace.items():
            covered_by_any[n_layer] |= qs
    covered_weight = 0
    total_weight = 0
    for n_layer in range(len(layers)):
        idle = wires[n_layer]["idle"]
        gate = wires[n_layer]["gate"]
        total_weight += w_idle * len(idle) + w_gate * len(gate)
        covered_idle = covered_by_any[n_layer] & idle
        covered_gate = covered_by_any[n_layer] & gate
        covered_weight += w_idle * len(covered_idle) + w_gate * len(
            covered_gate
        )
    return covered_weight / total_weight if total_weight > 0 else 0

def active_wires(layers):
    """
    Returns per-layer dict with two sets:
    - 'idle': activated wires that are idle in this layer
    - 'gate': activated wires that are control/target of a CNOT at this layer
    """
    first_activation = {}
    for n_layer, layer in enumerate(layers):
        for c, t in layer:
            first_activation.setdefault(c, n_layer)
            first_activation.setdefault(t, n_layer)
    result = {}
    for n_layer in range(len(layers)):
        active = {
            q
            for q, n_layer0 in first_activation.items()
            if n_layer >= n_layer0
        }
        gate = {q for c, t in layers[n_layer] for q in (c, t)}
        idle = active - gate
        result[n_layer] = {"idle": idle, "gate": gate}
    return result

def z_trace_backward(layers, initial_Zs):
    """
    Backward propagate Zs with parity cancellation.
    Returns {layer: set of qubits with odd parity Z at that layer}.
    """
    wires = active_wires(layers)
    support = set(initial_Zs)
    trace = {}
    for n_layer in range(len(layers) - 1, -1, -1):
        active = wires[n_layer]["idle"] | wires[n_layer]["gate"]
        trace[n_layer] = support & active
        # propagate backwards
        new_support = set()
        for q in support:
            hit = False
            for c, t in layers[n_layer]:
                if q == t:  # Z on target: copy to control
                    new_support ^= {t, c}  # toggle both
                    hit = True
                    break
                elif q == c:  # Z on control: passes through
                    new_support ^= {c}
                    hit = True
                    break
            if not hit:  # unaffected
                new_support ^= {q}
        support = new_support
    return trace

Kita sekarang bisa menyisipkan pemeriksaan ke Circuit GHZ. Menemukan pemeriksaan yang valid sangat mudah untuk state GHZ, karena setiap operator Pauli $Z$ dua-qubit $Z_i Z_j$ yang bekerja pada dua qubit mana pun $i,j$ dari Circuit GHZ adalah support dan oleh karena itu merupakan pemeriksaan yang valid.

Perlu dicatat juga bahwa pemeriksaan dalam kasus ini adalah operator $Z$ terkendalikan yang bersebelahan dengan gate Hadamard dari kiri dan kanan pada qubit ancilla. Ini setara dengan gate CNOT yang diterapkan pada qubit ancilla. Kode di bawah ini menyisipkan pemeriksaan ke dalam sirkuit.

# --- Tunables controlling the search space / scoring ---
MAX_SKIPS = 10  # at most how many qubits to skip (in addition to the bad ones and the ones forced to skip above)
SHUFFLES = 200  # how many times to try removing nodes for checks
MAX_DEPTH_INCREASE = 10  # how far from the base GHZ depth to go to include checks (increase this for more checks at expense of depth)

W_IDLE = 0.2  # weight of errors to consider during idle timesteps
W_GATE = 0.8  # weight of errors to consider during gates

# Remove random nodes from the GHZ and build from the root again to increase checks
degree_two_nodes = [
    i
    for i in ghz_qubits
    if all(n in ghz_qubits for n in coupling_map.neighbors(i))
    and len(coupling_map.neighbors(i)) >= 2
]

# --- Best-so-far tracking for the randomized search ---
num_checks = 0
best_covered_fraction = -1
best_qc = qc
best_checks = []
best_parities = []
best_layers = []

# Outer loop: vary how many GHZ nodes we try skipping (0..MAX_SKIPS-1)
for num_skips in range(MAX_SKIPS):
    # Inner loop: try SHUFFLES random choices of 'num_skips' nodes to skip
    for _ in range(SHUFFLES):
        # Construct the skip set:
        #   - pre-existing forced SKIP
        #   - plus a random sample of 'degree_two_nodes' of size 'num_skips'
        skip = SKIP + list(np.random.choice(degree_two_nodes, num_skips))

        # Rebuild the GHZ using the current skip set and bad_nodes
        qc, ghz_qubits, layers = parallel_ghz(
            ROOT, GHZ_SIZE, backend, bad_edges, skip + bad_nodes
        )

        # Measure circuit cost as 2-qubit-gate depth only
        depth = qc.depth(lambda x: x.operation.num_qubits == 2)

        # If we failed to reach the target GHZ size, discard this attempt
        if len(ghz_qubits) != GHZ_SIZE:
            continue

        # --- Build "checks" around the GHZ we just constructed ---
        # A check qubit is a non-GHZ, non-dead qubit that has ≥2 neighbors inside the GHZ
        # and all those incident edges are usable (i.e., not in bad_edges).
        checks = []
        parities = []
        for i in range(backend.num_qubits):
            neighbors = [
                n for n in coupling_map.neighbors(i) if n in ghz_qubits
            ]

            if (
                i not in ghz_qubits
                and i not in dead_qubits
                and len(neighbors) >= 2
                and not any(
                    [
                        [neighbor, i] in bad_edges
                        or [i, neighbor] in bad_edges
                        for neighbor in neighbors
                    ]
                )
            ):
                # Record this qubit as a check qubit
                checks.append(i)
                parities.append((neighbors[0], neighbors[1]))
                # Physically couple the check qubit 'i' to the two GHZ neighbors via CNOTs
                # (This is the actual "check" attachment in the circuit.)
                qc.cx(neighbors[0], i)
                qc.cx(neighbors[1], i)

        # Score this design using the weighted coverage metric over the GHZ build layers
        covered_fraction = weighted_coverage(
            layers=layers, parities=parities, w_idle=W_IDLE, w_gate=W_GATE
        )

        # Keep it only if:
        #   - coverage improves over the best so far, AND
        #   - the 2q depth budget isn't blown by more than MAX_DEPTH_INCREASE
        if (
            covered_fraction > best_covered_fraction
            and depth <= base_depth + MAX_DEPTH_INCREASE
        ):
            best_covered_fraction = covered_fraction
            best_qc = qc
            best_ghz_qubits = ghz_qubits
            best_checks = checks
            best_parities = parities
            best_layers = layers

Kita sekarang bisa mencetak qubit yang digunakan dalam Circuit GHZ dan qubit pemeriksaan.

# --- After search, report the best design found ---
qc = best_qc
checks = best_checks
parities = best_parities
layers = best_layers
ghz_qubits = best_ghz_qubits
if len(ghz_qubits) != GHZ_SIZE:
    raise Exception("No GHZ found. Relax error thresholds.")

print(f"GHZ qubits: {ghz_qubits} {len(ghz_qubits)}")
print(f"Check qubits: {checks} {len(checks)}")

covered_fraction = weighted_coverage(
    layers=layers, parities=parities, w_idle=W_IDLE, w_gate=W_GATE
)
print(
    "Covered fraction (no idle): ",
    weighted_coverage(
        layers=layers, parities=parities, w_idle=0.0, w_gate=1.0
    ),
)

GHZ qubits: [50, 49, 51, 38, 52, 48, 58, 53, 47, 71, 39, 46, 70, 54, 33, 45, 72, 69, 55, 32, 37, 73, 68, 34, 31, 44, 25, 74, 78, 67, 18, 24, 79, 75, 89, 57, 11, 23, 93, 59, 88, 66, 10, 22, 92, 90, 87, 65, 12, 9, 21, 94, 91, 86, 77, 13, 8, 20, 95, 98, 97, 14, 7, 36, 99, 111, 107, 15, 6, 41, 115, 110, 106, 19, 17, 5, 40, 114, 109, 108, 105, 27, 4, 42, 118, 104, 28, 3, 129, 117, 103, 29, 2, 128, 125, 96, 30, 127, 124, 102] 100
Check qubits: [16, 26, 35, 43, 85, 126] 6
Covered fraction (no idle):  0.4595959595959596

Kita juga bisa mencetak beberapa statistik error.

def circuit_errors(target, circ, error_type="cz"):
    """
    Pull per-resource error numbers from a Qiskit Target
    for ONLY the qubits/edges actually used by `circ`.

    Args:
        target: qiskit.transpiler.Target (e.g., backend.target)
        circ:   qiskit.QuantumCircuit
        error_type: one of {"cz", "meas", "t1", "t2"}:
            - "cz"   -> 2q CZ gate error on the circuit's used edges
            - "meas" -> measurement error on the circuit's used qubits
            - "t1"   -> T1 (converted to microseconds) on used qubits
            - "t2"   -> T2 (converted to microseconds) on used qubits

    Returns:
        list[float] of the requested quantity for the active edges/qubits.
    """

    # Get all 2-qubit edges that appear in the circuit (as undirected pairs).
    active_edges = active_gates(circ)  # e.g., {(0,1), (2,3), ...}

    # Intersect those with the device coupling map (so we only query valid edges).
    # Note: target.build_coupling_map().get_edges() yields directed pairs.
    edges = [
        edge
        for edge in target.build_coupling_map().get_edges()
        if tuple(sorted(edge)) in active_edges
    ]

    # Deduplicate direction: keep only one orientation of each edge.
    undirected_edges = []
    for edge in edges:
        if (edge[1], edge[0]) not in undirected_edges:
            undirected_edges.append(edge)
    edges = undirected_edges  # (not used later—see note below)

    # Accumulators for different error/physics quantities
    cz_errors, meas_errors, t1_errors, t2_errors = [], [], [], []

    # For every active (undirected) edge in the circuit, fetch its CZ error.
    # NOTE: Uses active_gates(circ) again (undirected tuples). This assumes
    # `target['cz']` accepts undirected indexing; many Targets store both directions.
    for edge in active_gates(circ):
        cz_errors.append(target["cz"][edge].error)

    # For every active qubit, fetch measure error and T1/T2 (converted to µs).
    for qubit in active_qubits(circ):
        meas_errors.append(target["measure"][(qubit,)].error)
        t1_errors.append(
            target.qubit_properties[qubit].t1 * 1e6
        )  # seconds -> microseconds
        t2_errors.append(
            target.qubit_properties[qubit].t2 * 1e6
        )  # seconds -> microseconds

    # Select which set to return.
    if error_type == "cz":
        return cz_errors
    elif error_type == "meas":
        return meas_errors
    elif error_type == "t1":
        return t1_errors
    else:
        return t2_errors

def active_qubits(circ):
    """
    Return a list of qubit indices that participate in at least one
    non-delay, non-barrier instruction in `circ`.
    """
    active_qubits = set()
    for inst in circ.data:
        # Skip scheduling artifacts that don’t act on state
        if (
            inst.operation.name != "delay"
            and inst.operation.name != "barrier"
        ):
            for qubit in inst.qubits:
                q = circ.find_bit(
                    qubit
                ).index  # map Qubit object -> integer index
                active_qubits.add(q)
    return list(active_qubits)

def active_gates(circ):
    """
    Return a set of undirected 2-qubit edges (i, j) that appear in `circ`.
    """
    used_2q_gates = set()
    for inst in circ:
        if inst.operation.num_qubits == 2:
            qs = inst.qubits
            # map Qubit objects -> indices, then sort to make the edge undirected
            qs = sorted([circ.find_bit(q).index for q in qs])
            used_2q_gates.add(tuple(sorted(qs)))
    return used_2q_gates

# ---- Print summary statistics ----
cz_errors = circuit_errors(backend.target, qc, error_type="cz")
meas_errors = circuit_errors(backend.target, qc, error_type="meas")
t1_errors = circuit_errors(backend.target, qc, error_type="t1")
t2_errors = circuit_errors(backend.target, qc, error_type="t2")

np.set_printoptions(linewidth=np.inf)
print(
    f"cz errors: \n mean: {np.round(np.mean(cz_errors), 3)}, max: {np.round(np.max(cz_errors), 3)}"
)
print(
    f"meas errors: \n mean: {np.round(np.mean(meas_errors), 3)}, max: {np.round(np.max(meas_errors), 3)}"
)
print(
    f"t1 errors: \n mean: {np.round(np.mean(t1_errors), 1)}, min: {np.round(np.min(t1_errors), 1)}"
)
print(
    f"t2 errors: \n mean: {np.round(np.mean(t2_errors), 1)}, min: {np.round(np.min(t2_errors), 1)}"
)

cz errors: 
 mean: 0.002, max: 0.012
meas errors: 
 mean: 0.014, max: 0.121
t1 errors: 
 mean: 267.9, min: 23.6
t2 errors: 
 mean: 155.9, min: 13.9

Seperti sebelumnya, kita bisa mensimulasikan sirkuit terlebih dahulu tanpa noise untuk memastikan kebenaran Circuit persiapan state GHZ.

# --- Simulate to ensure correctness ---

qc_meas = qc.copy()

# Add measurements to the GHZ qubits
c1 = ClassicalRegister(len(ghz_qubits), "c1")
qc_meas.add_register(c1)
for q, c in zip(ghz_qubits, c1):
    qc_meas.measure(q, c)

# Add measurements to the check qubits
if len(checks) > 0:
    c2 = ClassicalRegister(len(checks), "c2")
    qc_meas.add_register(c2)
    for q, c in zip(checks, c2):
        qc_meas.measure(q, c)

# Simulate the circuit with stabilizer method
sim_stab = AerSimulator(method="stabilizer")
res = sim_stab.run(qc_meas, shots=1000).result()
counts = res.get_counts()
print("Stabilizer simulation result:")
print(counts)

# Rename keys to "0 0" and "0 1" for easier plotting
# First len(checks) bits are check bits, rest are GHZ bits
keys = list(counts.keys())
for key in keys:
    check_bits = key[: len(checks)]
    ghz_bits = key[(len(checks) + 1) :]
    if set(check_bits) == {"0"} and set(ghz_bits) == {"0"}:
        counts["0 0"] = counts.pop(key)
    elif set(check_bits) == {"0"} and set(ghz_bits) == {"1"}:
        counts["0 1"] = counts.pop(key)
    else:
        continue

plot_histogram(counts)

Stabilizer simulation result:
{'000000 1111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111': 525, '000000 0000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000': 475}

Seperti yang diharapkan, qubit pemeriksaan diukur semuanya sebagai nol, dan kita berhasil mempersiapkan state GHZ.

Langkah 3: Eksekusi menggunakan Qiskit primitives

Sekarang kita siap menjalankan Circuit di hardware nyata dan mendemonstrasikan bagaimana protokol deteksi error dapat menangkap kesalahan dalam persiapan state GHZ.

BAD_QUBITS = []  # specify any additional bad qubits to avoid (this is specific to the chosen backend)
SHOTS = 10000  # number of shots

Kita mendefinisikan fungsi bantu untuk menambahkan pengukuran ke GHZ Circuit.

def add_measurements(qc, ghz_qubits, checks):
    # --- Measure each set of qubits into different classical registers to facilitate post-processing ---

    # Add measurements to the GHZ qubits
    c1 = ClassicalRegister(len(ghz_qubits), "c1")
    qc.add_register(c1)
    for q, c in zip(ghz_qubits, c1):
        qc.measure(q, c)

    # Add measurements to the check qubits
    c2 = ClassicalRegister(len(checks), "c2")
    qc.add_register(c2)
    for q, c in zip(checks, c2):
        qc.measure(q, c)

    return qc

Sebelum eksekusi, kita menggambar tata letak GHZ Qubit dan check Qubit pada hardware yang dipilih.

# Plot the layout of GHZ and check qubits on the device
plot_gate_map(
    backend,
    label_qubits=True,
    line_width=20,
    line_color=[
        "black"
        if edge[0] in ghz_qubits + checks and edge[1] in ghz_qubits + checks
        else "lightgrey"
        for edge in backend.coupling_map.graph.edge_list()
    ],
    qubit_color=[
        "blue"
        if i in ghz_qubits
        else "salmon"
        if i in checks
        else "lightgrey"
        for i in range(0, backend.num_qubits)
    ],
)
plt.show()

qc.draw("mpl", idle_wires=False, fold=-1)

Sekarang kita menambahkan pengukuran.

qc = add_measurements(qc, ghz_qubits, checks)

Pipeline penjadwalan di bawah ini mengunci waktu, menghapus barrier, menyederhanakan delay, dan menyuntikkan dynamical decoupling, semuanya sambil mempertahankan waktu operasi aslinya.

# The scheduling consists of first inserting delays while barriers are still there
# Then removing the barriers and consolidating the delays, so that the operations do not move in time
# Lastly we replace delays with dynamical decoupling
collect_function = partial(
    collect_using_filter_function,
    filter_function=(lambda node: node.op.name == "delay"),
    split_blocks=True,
    min_block_size=2,
    split_layers=False,
    collect_from_back=False,
    max_block_width=None,
)

collapse_function = partial(
    collapse_to_operation,
    collapse_function=(
        lambda circ: Delay(sum(inst.operation.duration for inst in circ))
    ),
)

class Unschedule(AnalysisPass):
    """Removes a property from the passmanager property set so that the circuit looks unscheduled, so we can schedule it again."""

    def run(self, dag):
        del self.property_set["node_start_time"]

def build_passmanager(backend, dd_qubits=None):
    pm = generate_preset_pass_manager(
        target=backend.target,
        layout_method="trivial",
        optimization_level=2,
        routing_method="none",
    )

    pm.scheduling = PassManager(
        [
            ALAPScheduleAnalysis(target=backend.target),
            PadDelay(target=backend.target),
            RemoveBarriers(),
            Unschedule(),
            CollectAndCollapse(
                collect_function=collect_function,
                collapse_function=collapse_function,
            ),
            ALAPScheduleAnalysis(target=backend.target),
            PadDynamicalDecoupling(
                dd_sequence=[XGate(), RZGate(-np.pi), XGate(), RZGate(np.pi)],
                spacing=[1 / 4, 1 / 2, 0, 0, 1 / 4],
                target=backend.target,
                qubits=dd_qubits,
            ),
        ]
    )

    return pm

Kita sekarang bisa menggunakan custom pass manager untuk men-transpile Circuit untuk Backend yang dipilih.

# Transpile the circuits for the backend
pm = build_passmanager(backend, ghz_qubits)

# Instruction set architecture (ISA) level circuit after scheduling and DD insertion
isa_circuit = pm.run(qc)

# Draw after scheduling and DD insertion
# timeline_drawer(isa_circuit, show_idle=False, time_range=(0, 1000), target=backend.target)

isa_circuit.draw("mpl", fold=-1, idle_wires=False)

Kemudian kita mengirimkan job menggunakan Qiskit Runtime Sampler primitive.

# Select the sampler options
sampler = Sampler(mode=backend)
sampler.options.default_shots = SHOTS
sampler.options.dynamical_decoupling.enable = False
sampler.options.execution.rep_delay = 0.00025

# Submit the job
print("Submitting sampler job")
ghz_job = sampler.run([isa_circuit])

print(ghz_job.job_id())

d493f17nmdfs73abf9qg

Langkah 4: Post-process dan kembalikan hasil dalam format klasik yang diinginkan

Kita sekarang bisa mengambil dan menganalisis hasil dari Sampler job.

# Retrieve the job results
job_result = ghz_job.result()

# Get the counts from GHZ and check qubit measurements
ghz_counts = job_result[0].data.c1.get_counts()
checks_counts = job_result[0].data.c2.get_counts()

# Post-process to get unflagged GHZ counts (i.e., check bits are all '0')
joined_counts = job_result[0].join_data().get_counts()
unflagged_counts = {}
for key, count in joined_counts.items():
    check_bits = key[: len(checks)]
    ghz_bits = key[len(checks) :]
    if set(check_bits) == {"0"}:
        unflagged_counts[ghz_bits] = count

# Get top 20 outcomes by frequency from the unflagged counts
top_counts = dict(
    sorted(unflagged_counts.items(), key=lambda x: x[1], reverse=True)[:20]
)

# Rename keys for better visualization
top_counts_renamed = {}
i = 0
for key, count in top_counts.items():
    if set(key) == {"0"}:
        top_counts_renamed["all 0s"] = count
    elif set(key) == {"1"}:
        top_counts_renamed["all 1s"] = count
    else:
        top_counts_renamed[f"other_{i}"] = count
        i += 1

plot_histogram(top_counts_renamed, figsize=(12, 7))

Pada histogram di atas, kita memplot 20 pengukuran bitstring dari GHZ Qubit yang tidak ditandai oleh check Qubit. Seperti yang diharapkan, bitstring semua-0 dan semua-1 memiliki jumlah terbanyak. Perlu dicatat bahwa beberapa bitstring yang keliru dengan bobot error rendah tidak tertangkap oleh deteksi error. Jumlah tertinggi masih ditemukan pada bitstring yang diharapkan.

Diskusi

Dalam tutorial ini, kita menunjukkan cara mengimplementasikan teknik deteksi error dengan overhead rendah menggunakan spacetime codes, dan mendemonstrasikan penerapannya di dunia nyata untuk mempersiapkan state GHZ pada hardware. Lihat [3] untuk mengeksplorasi lebih lanjut detail teknis persiapan state GHZ. Selain deteksi error, para penulis memanfaatkan readout error mitigation dengan M3 dan TREX serta melakukan teknik uncomputation sementara untuk mempersiapkan state GHZ dengan fidelitas tinggi.

Referensi

• [1] Martiel, S., & Javadi-Abhari, A. (2025). Low-overhead error detection with spacetime codes. arXiv preprint arXiv:2504.15725.
• [2] van den Berg, E., Bravyi, S., Gambetta, J. M., Jurcevic, P., Maslov, D., & Temme, K. (2023). Single-shot error mitigation by coherent Pauli checks. Physical Review Research, 5(3), 033193.
• [3] Javadi-Abhari, A., Martiel, S., Seif, A., Takita, M., & Wei, K. X. (2025). Big cats: entanglement in 120 qubits and beyond. arXiv preprint arXiv:2510.09520.