Repetition codes

Usage estimate: less than 10 seconds on a Heron processor (NOTE: This is an estimate only. Your runtime might vary.)

Learning outcomes

After completing this tutorial, you can expect to understand the following information:

How to implement a bit-flip error correction code using dynamic circuits
How to measure stabilizers to detect quantum errors without destroying encoded information
How to evaluate the performance of quantum error correction by comparing corrected and uncorrected results

Prerequisites

It is recommended that you familiarize yourself with these topics:

Background

To enable real-time quantum error correction (QEC), you need to be able to dynamically control quantum program flow during execution so that quantum gates can be conditioned on measurement results. This tutorial runs the bit-flip code, which is a very simple form of QEC. It demonstrates a dynamic quantum circuit that can protect an encoded qubit from a single bit-flip error, and then evaluates the bit-flip code performance.

You can exploit additional ancilla qubits and entanglement to measure stabilizers that do not transform encoded quantum information, while still informing you of some classes of errors that might have occurred. A quantum stabilizer code encodes $k$ logical qubits into $n$ physical qubits. Stabilizer codes critically focus on correcting a discrete error set with support from the Pauli group $\Pi^n$ .

In this tutorial, we demonstrate the bit-flip code using a simple quantum memory experiment. We will prepare the encoded qubit in the logical state $|\bar{1}\rangle \equiv |111\rangle$ and then implement multiple cycles of idle time (to accrue errors) followed by error detection and correction. We then quantify the probability of a logical error as a function of the number of cycles (time), where the logical error probability is the probability of finding the qubits in a state that doesn't recover the $|\bar{1}\rangle$ after final measurement (concretely, states corresponding to an error are $|000\rangle$ , $|001\rangle$ , $|010\rangle$ , and $|100\rangle$ ).

We will compare the error rate versus time to the error rates of individual, unencoded qubits, and also to use of the repetition code where we only detect and correct errors after final measurement, but not dynamically during the circuit.

Note: the repetition code only allows for correction of bit-flip errors and therefore is not a complete error correction code. However, because of its simplicity, it is a good starting point to illustrate how to implement error correction on a quantum computer. The memory experiment below only tests for a single type of error (a $|\bar{1}\rangle$ decaying to $|\bar{0}\rangle$ ) and technically only demonstrates the protection of classical information.

Requirements

Before starting this tutorial, ensure that you have the following installed:

Qiskit SDK v2.0 or later, with visualization support
Qiskit Runtime v0.40 or later (pip install qiskit-ibm-runtime)

Setup

import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
from scipy.optimize import curve_fit

sns.set()
plt.rc("xtick", labelsize=20)
plt.rc("ytick", labelsize=20)
plt.rc("lines", linewidth=3)
plt.rc("font", size=20)
plt.rc("legend", fontsize="large")
plt.rc("axes", labelsize=20)
plt.rcParams["figure.figsize"] = 15, 6
plt.rcParams["legend.title_fontsize"] = 25

# Qiskit imports
from qiskit import (
    QuantumCircuit,
    QuantumRegister,
    ClassicalRegister,
)

# qiskit-ibm-runtime
from qiskit_ibm_runtime import QiskitRuntimeService, SamplerV2 as Sampler
from qiskit_ibm_runtime.circuit import MidCircuitMeasure

service = QiskitRuntimeService()

Small-scale simulator example

We will forgo this step since the goal of this experiment is to measure the lifetime of a logical qubit under real hardware noise, such as amplitude damping during the idle delays; a noiseless simulator would show no errors to detect and correct.

Large-scale hardware example

Step 1: Map classical inputs to a quantum problem

Choose a backend

To detect errors during the circuit, we need to choose a backend that has access to the MidCircuitMeasure method (see the documentation).

# You can see all backends that support mid-circuit measurements by running the following code.
service.backends(filters=lambda b: "measure_2" in b.supported_instructions)

Output:

[<IBMBackend('ibm_pittsburgh')>,
 <IBMBackend('ibm_boston')>,
 <IBMBackend('ibm_kingston')>]

# Choose the least busy backend that supports mid-circuit measurements (`measure_2`).

backend = service.least_busy(
    filters=lambda b: "measure_2" in b.supported_instructions,
    operational=True,
    simulator=False,
    dynamic_circuits=True,
)
# backend = service.backend(backend_name) # alternatively, you could choose a specific backend
print(backend.name)

Output:

ibm_boston

Build a sequence of bit-flip stabilizer circuits implementing multiple rounds of error detection and correction

The bit-flip code is among the simplest examples of a stabilizer code. It protects the state against a single bit-flip (X) error on any of the encoding qubits. Consider the action of bit-flip error $X$ , which maps $|0\rangle \rightarrow |1\rangle$ and $|1\rangle \rightarrow |0\rangle$ on any of our qubits, then we have $\epsilon = \{E_0, E_1, E_2 \} = \{IIX, IXI, XII\}$ . The code requires five qubits: three are used to encode the protected state (the "data qubits"), and the remaining two are used as stabilizer measurement ancillas.

Below you will build circuits that (1) prepare the data qubits in the logical $|\bar{1} \rangle$ state, then (2) run multiple cycles of a $5 \mu s$ delay followed by error correction (including reset of the syndrome qubits), and (3) read out the state of the data qubits.

We will also test the lifetime of the $|1\rangle$ state without using error correction by including three reference qubits that we prepare in the $|1\rangle$ state, leave idle, and then read out.

def build_qc(
    qreg_data,
    qreg_syndrome,
    creg_data,
    creg_syndrome,
    qreg_ref=None,
    creg_ref=None,
):
    """Build a typical error correction circuit"""
    if qreg_ref:
        return QuantumCircuit(
            qreg_data,
            qreg_syndrome,
            creg_data,
            creg_syndrome,
            qreg_ref,
            creg_ref,
        )
    else:
        return QuantumCircuit(
            qreg_data, qreg_syndrome, creg_data, creg_syndrome
        )


def encode_bit_flip(circuit, qreg_data, qreg_ref=None) -> QuantumCircuit:
    """Encode bit-flip. This is done by simply adding a cx"""

    for q in qreg_data:
        circuit.x(q)

    if qreg_ref:
        for q in qreg_ref:
            circuit.x(q)

    circuit.barrier()

    return circuit


def measure_syndrome_bit(
    circuit, qreg_data, qreg_syndrome, creg_syndrome, qreg_ref=None
):
    """
    Measure the syndrome by measuring the parity.
    We reset our ancilla qubits after measuring the stabilizer
    so we can reuse them for repeated stabilizer measurements.
    Because we have already observed the state of the qubit,
    we can write the conditional reset protocol directly to
    avoid another round of qubit measurement if we used
    the `reset` instruction.
    """
    circuit.cx(qreg_data[0], qreg_syndrome[0])
    circuit.cx(qreg_data[1], qreg_syndrome[0])
    circuit.cx(qreg_data[0], qreg_syndrome[1])
    circuit.cx(qreg_data[2], qreg_syndrome[1])
    circuit.barrier()

    for q_measure, c_measure in zip(qreg_syndrome, creg_syndrome):
        circuit.append(MidCircuitMeasure(), [q_measure], [c_measure])

    with circuit.if_test((creg_syndrome[0], 1)):
        circuit.x(qreg_syndrome[0])
    with circuit.if_test((creg_syndrome[1], 1)):
        circuit.x(qreg_syndrome[1])

    circuit.barrier()
    return circuit


def apply_correction_bit(circuit, qreg_data, creg_syndrome):
    """We can detect where an error occurred and correct our state"""
    with circuit.if_test((creg_syndrome, 3)):
        circuit.x(qreg_data[0])
    with circuit.if_test((creg_syndrome, 1)):
        circuit.x(qreg_data[1])
    with circuit.if_test((creg_syndrome, 2)):
        circuit.x(qreg_data[2])
    circuit.barrier()
    return circuit


def apply_final_readout(
    circuit, qreg_data, creg_data, qreg_ref=None, creg_ref=None
):
    """Read out the final measurements"""

    circuit.barrier()
    if qreg_ref:
        circuit.measure(qreg_ref, creg_ref)
    circuit.measure(qreg_data, creg_data)

    return circuit

def build_error_correction_sequence(
    num_cycles, cycles_per_circuit, nq_ref=3, delay=None
) -> QuantumCircuit:
    circuits = []
    reps = []

    qreg_data = QuantumRegister(3, name="qdata")
    qreg_syndrome = QuantumRegister(2, name="qsyndrome")
    creg_data = ClassicalRegister(3, name="cdata")
    creg_syndrome = ClassicalRegister(2, name="csyndrome")
    qreg_ref = QuantumRegister(nq_ref, name="qreference")
    creg_ref = ClassicalRegister(nq_ref, name="creference")

    circuit = build_qc(
        qreg_data,
        qreg_syndrome,
        creg_data,
        creg_syndrome,
        qreg_ref=qreg_ref,
        creg_ref=creg_ref,
    )
    circuit = encode_bit_flip(circuit, qreg_data, qreg_ref=qreg_ref)

    circuit_n = circuit.copy()
    circuit_n = apply_final_readout(
        circuit_n, qreg_data, creg_data, qreg_ref=qreg_ref, creg_ref=creg_ref
    )
    circuits.append(circuit_n)
    reps.append(0)

    for i in range(1, num_cycles + 1):
        if delay:
            circuit.delay(delay, unit="us")
            circuit.barrier()
        circuit = measure_syndrome_bit(
            circuit,
            qreg_data,
            qreg_syndrome,
            creg_syndrome,
            qreg_ref=qreg_ref,
        )
        circuit = apply_correction_bit(circuit, qreg_data, creg_syndrome)
        circuit_n = circuit.copy()
        if i % cycles_per_circuit == 0:
            circuit_n = apply_final_readout(
                circuit_n,
                qreg_data,
                creg_data,
                qreg_ref=qreg_ref,
                creg_ref=creg_ref,
            )
            circuits.append(circuit_n)
            reps.append(i)

    return circuits, np.array(reps)


num_cycles = 40
cycles_per_circuit = 4
nq_ref = 3
circuits, rep_array = build_error_correction_sequence(
    num_cycles, cycles_per_circuit, nq_ref=3, delay=5
)
circuits[1].draw(output="mpl", cregbundle=False, fold=50)

Output:

Step 2: Optimize the problem for quantum hardware execution

To reduce the total job execution time, Qiskit primitives only accept circuits and observables that conform to the instructions and connectivity supported by the target system (referred to as instruction set architecture (ISA) circuits and observables). Learn more about transpilation.

Generate ISA circuits

We will start by finding an initial layout (as in, a selection of physical qubits to map our circuits to) by transpiling the longest of our circuits using the preset pass manager with optimization level 3.

from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager

pm = generate_preset_pass_manager(backend=backend, optimization_level=3)
isa_circuit_ref = pm.run(circuits[-1])

init_layout = isa_circuit_ref.layout.initial_index_layout(
    filter_ancillas=True
)
print(init_layout)

Output:

[56, 44, 62, 43, 63, 22, 48, 67]

For the reference qubits that we will compare our logical quantum memory to, we will choose them to be the best available qubits in terms of amplitude damping coherence time ( $T_1$ ).

# get all qubits ordered by T1
t1_data = []
for i in range(backend.num_qubits):
    try:
        t1_us = backend.properties().t1(i) * 1e6
    except Exception:
        t1_us = 0.0
    t1_data.append((i, t1_us))

t1_data_sorted = sorted(t1_data, key=lambda x: x[1], reverse=True)

# exclude the qubits we have already mapped the error correcting code to
t1_data_sorted = [
    t1_data for t1_data in t1_data_sorted if t1_data[0] not in init_layout[:5]
]

# use the best qubits in terms of T1 for the reference qubits
init_layout = init_layout[:5] + [t1_data[0] for t1_data in t1_data_sorted[:3]]
print(init_layout)

Output:

[56, 44, 62, 43, 63, 143, 131, 31]

# These are the resulting T1 times
properties = backend.properties()
print("Amplitude damping decoherence times for code data qubits:")
for q in init_layout[:3]:
    t1 = properties.t1(q)
    print(f"qubit {q}: T1 = {t1 * 1e6:.0f} mus")

print("\nAmplitude damping decoherence times for reference qubits:")
for q in init_layout[-3:]:
    t1 = properties.t1(q)
    print(f"qubit {q}: T1 = {t1 * 1e6:.0f} mus")

Output:

Amplitude damping decoherence times code data qubits:
qubit 56: T1 = 322 mus
qubit 44: T1 = 263 mus
qubit 62: T1 = 290 mus

Amplitude damping decoherence times reference qubits:
qubit 143: T1 = 442 mus
qubit 131: T1 = 410 mus
qubit 31: T1 = 401 mus

# now we transpile all circuits to this initial layout; this way each circuit is run on the same qubits and we can make a fair comparison
pm = generate_preset_pass_manager(
    backend=backend,
    optimization_level=3,
    initial_layout=init_layout,
)
isa_circuits = pm.run(circuits)

isa_circuits[1].draw("mpl", cregbundle=False, fold=50)

Output:

Step 3: Execute using Qiskit primitives

sampler = Sampler(mode=backend)

sampler.options.environment.job_tags = ["TUT-REPCODE"]
sampler.options.max_execution_time = 600  # this workload is expected to be under 10s, but it is generally a good habit to set a max execution time (here 600s = 10m)

job = sampler.run(isa_circuits, shots=1000)
print(job.job_id())

job.status()

Output:

'DONE'

Step 4: Post-process and return result in desired classical format

We will now compare the error rates vs. time between the logical memory using the 3-qubit repetition code on the one hand, and the individual unencoded reference qubits on the other hand.

results = job.result()

def correct_counts(counts_dict):
    """
    Corrects the measured logical qubit encoded in the repetition code using majority vote
    """

    result = {"000": 0, "111": 0}
    for bitstring, count in counts_dict.items():
        key = "111" if bitstring.count("1") > 1 else "000"
        result[key] += count
    return result


accuracy = []  # logical qubit
accuracies_ref = np.zeros(
    (len(results), nq_ref)
)  # accuracies on individual reference qubits

for n, pub_result in enumerate(results):
    # logical accuracy (one minus error probability) for active error correction with repetition code
    counts = pub_result.data.cdata.get_counts()
    shots = sum(counts.values())
    counts_corrected = correct_counts(counts)
    accuracy.append(counts_corrected.get("111", 0) / shots)

    # accuracy for individual physical reference qubits without any error correction
    for i in range(nq_ref):
        counts = pub_result.data.creference.slice_bits(i).get_counts()
        accuracies_ref[n, i] = counts.get("1", 0) / shots

accuracy = np.array(accuracy)

def error_proba_t1(N, t1):
    """
    Exponential fitting function for amplitude damping vs. number of cycles
    """

    t_cycle = 7.3e-6  # approximate time per cycle = 5 mus delay + 2.3 mus for error correction
    return 1 - np.exp(-t_cycle * N / t1)


fig, ax = plt.subplots(1, 1, figsize=(15, 5))
ax.plot(
    rep_array, (1.0 - accuracy) * 100.0, "ko-", linewidth=3, label="rep code"
)

for i in range(nq_ref):
    accuracy_1q = accuracies_ref[:, i]
    if i == 0:
        ax.plot(
            rep_array,
            (1.0 - accuracy_1q) * 100.0,
            "go-",
            linewidth=1,
            label="1q reference",
        )
    else:
        ax.plot(rep_array, (1.0 - accuracy_1q) * 100.0, "go-", linewidth=1)

params_bf, pcov = curve_fit(
    error_proba_t1, rep_array, 1.0 - accuracy, bounds=([0, 5e-3])
)
t1_bf = params_bf[0]
print(f"Best-fit effective T1 = {t1_bf * 1e6:.0f} us")
error_prob_bf = error_proba_t1(
    rep_array, t1_bf
)  # np.array([1 - np.exp(-t_cycle * nt/t1) for nt in rep_array])
ax.plot(rep_array, error_prob_bf * 100.0, "k--", linewidth=1)

ax.set_xlabel("error correction rounds")
ax.set_ylabel("error [%]")
ax.set_ylim(bottom=0)
ax.legend(fontsize=15);

Output:

Best-fit effective T1 = 1397 us

We clearly see an improved lifetime of the $|\bar{1}\rangle$ state, even compared to the best (unencoded) physical qubits. However, keep in mind that this experiment just corrects one type of error, namely bit-flips. Can you improve the lifetime of the logical qubit? You might want to explore optimizing the delay time, scaling up the size of the repetition code beyond $n=3$ , etc.

Next steps

If you found this work interesting, you might be interested in the following material:

Recommendations

Foundations of quantum error correction course - learn more about quantum error correction
Low-overhead error detection with spacetime codes tutorial - learn how to use Pauli checks to detect errors and post-select samples

このページは役に立ちましたか？

バグや誤字の報告、またはコンテンツの要求はGitHubで行ってください。