Optimasi multi-objektif aproksimasi kuantum

Perkiraan penggunaan: 10 menit di prosesor Heron r2 (CATATAN: Ini hanya perkiraan. Waktu jalanmu mungkin berbeda.)

Hasil pembelajaran

Tutorial ini menyelesaikan masalah optimasi portofolio dengan batasan kardinalitas: dengan syarat memegang tepat $K$ aset, seimbangkan tujuan risiko, imbal hasil, dan diversifikasi untuk menemukan himpunan portofolio optimal.

Setelah menyelesaikan tutorial ini, kamu diharapkan memahami:

• Cara menyatakan masalah pemilihan portofolio dengan tiga tujuan yang saling bersaing — risiko rendah, imbal hasil tinggi, dan diversifikasi yang baik — sebagai masalah optimasi kuantum.

• Bagaimana satu circuit QAOA, yang disapu pada sekumpulan bobot objektif, menelusuri front Pareto dari portofolio trade-off optimal.

• Bagaimana mixer XY menjaga pencarian tetap di dalam subruang "pilih tepat K aset", sehingga tidak perlu suku penalti dan kelayakan ditegakkan dengan post-selection pada bitstring hasil pengukuran.

• Cara melatih sudut-sudut circuit dengan simulator matrix-product-state pada skala yang terlalu besar untuk dioptimalkan secara eksak, mengikuti Kotil et al. (arXiv:2503.22797).

Prasyarat

Disarankan kamu sudah familier dengan:

• Workflow pola Qiskit (map, optimize, execute, post-process).

• Dasar-dasar QAOA.

Latar belakang

Seorang manajer portofolio jarang mengoptimalkan satu angka saja. Mereka ingin imbal hasil tinggi, risiko (varians dari imbal hasil itu) rendah, dan kepemilikan tersebar di berbagai sektor supaya portofolio tidak terlalu terekspos pada satu bagian pasar. Tujuan-tujuan ini saling tarik-menarik: aset dengan imbal hasil tertinggi sering kali paling volatil, dan menumpuk di satu sektor yang sedang panas merugikan diversifikasi.

Tidak ada satu portofolio "terbaik". Yang ada adalah front Pareto: himpunan portofolio yang tidak bisa diperbaiki satu tujuannya tanpa mengorbankan tujuan lain. Target kita adalah memetakan front itu supaya pengambil keputusan bisa memilih trade-off yang disukainya.

Kita merumuskan masalahnya sebagai memilih tepat $K$ aset dari $N$ (tiap aset ada di dalam atau di luar — satu qubit per aset). Tiga Hamiltonian menyandikan tiga tujuan. Kita menggabungkannya dengan bobot $c$ yang berada di sebuah simpleks (jumlahnya satu), dan sampler QAOA mengembalikan portofolio bagus untuk tiap pilihan bobot. Menyapu bobot berarti menyapu kepentingan relatif risiko versus imbal hasil versus diversifikasi, dan gabungan semua portofolio hasil sampling menelusuri front Pareto.

Pertama kita telusuri seluruh workflow pada contoh kecil delapan aset yang bisa kita verifikasi dengan brute force, lalu menjalankan metode yang sama persis pada instance 40 aset yang ukurannya disesuaikan untuk hardware kuantum.

Tutorial ini mengajarkan workflow-nya (pemetaan, pelatihan sudut, sampling terbatas, dan pasca-proses Pareto), bukan mendemonstrasikan keunggulan kuantum. Pada skala 40 aset yang dipakai di sini, sampling acak seragam berkinerja hampir sama baiknya dengan sampler QAOA, dan kita menunjukkan perbandingan itu secara eksplisit.

Persyaratan

Sebelum memulai tutorial ini, pastikan kamu sudah menginstal yang berikut:

• Qiskit SDK v2.0 atau lebih baru, dengan dukungan visualisasi

• Qiskit Runtime v0.22 atau lebih baru (pip install qiskit-ibm-runtime)

• Qiskit Aer (pip install qiskit-aer)

• Addon Qiskit Optimization mapper (pip install qiskit-addon-opt-mapper) dan pipeline pelatihan QAOA, di-pin ke tag v0.1.0: pip install "git+https://github.com/qiskit-community/qaoa_training_pipeline.git@v0.1.0"

• moocore untuk perhitungan front Pareto dan hypervolume (pip install moocore)

Setup

Impor library yang dipakai sepanjang tutorial dan tetapkan random seed supaya hasilnya reproduksibel.

# Added by doQumentation — installs the packages this notebook needs if they are missing
import importlib.util

_needed = {"matplotlib": "matplotlib", "moocore": "moocore", "numpy": "numpy", "qaoa_training_pipeline": "qaoa-training-pipeline", "qiskit": "qiskit", "qiskit_addon_opt_mapper": "qiskit-addon-opt-mapper", "qiskit_aer": "qiskit-aer", "qiskit_ibm_runtime": "qiskit-ibm-runtime", "scipy": "scipy"}
_missing = [pip for module, pip in _needed.items()
            if importlib.util.find_spec(module) is None]
# One at a time, so a package that fails to install does not block the others
for _pip in _missing:
    %pip install -q {_pip}
if not _missing:
    print("\u2713 All required packages are installed")

import numpy as np
import matplotlib.pyplot as plt
from math import comb
from moocore import hypervolume, filter_dominated, is_nondominated

from qiskit import QuantumCircuit
from qiskit.circuit import ParameterVector
from qiskit.circuit.library import qaoa_ansatz
from qiskit.quantum_info import SparsePauliOp
from qiskit.transpiler import generate_preset_pass_manager
from qiskit_aer.primitives import SamplerV2 as AerSampler
from qiskit_addon_opt_mapper.problems import OptimizationProblem
from qaoa_training_pipeline.training import ScipyTrainer
from qaoa_training_pipeline.evaluation import (
    StatevectorEvaluator,
    MPSAerEvaluator,
)

np.random.seed(42)
sampler = AerSampler(seed=42)  # local simulator for the small-scale example
print("Setup complete.")

Setup complete.

Contoh simulator skala kecil

Kita mulai dengan delapan aset dari enam sektor dan memilih tepat $K=4$ di antaranya. Dengan hanya delapan aset, hanya ada $\binom{8}{4}=70$ portofolio valid, sehingga nanti kita bisa memeriksa hasil kuantum terhadap pencarian menyeluruh.

Langkah 1: Petakan input klasik ke masalah kuantum

Setiap aset adalah satu qubit; bitstring seperti 10110010 adalah sebuah portofolio (angka 1 adalah aset yang kita pegang). Kita butuh tiga bahan: data pasar, tiga Hamiltonian objektif, dan circuit yang hanya pernah mengusulkan portofolio dengan tepat $K$ aset.

# --- Small-scale universe: 8 assets across 6 sectors ---
tickers = ["AAPL", "XOM", "JPM", "JNJ", "KO", "AMT", "AMZN", "SLB"]
sectors = [
    "Tech",
    "Energy",
    "Finance",
    "Health",
    "Staples",
    "REIT",
    "Tech",
    "Energy",
]
n_assets = len(tickers)
K = 4  # choose exactly K assets
n_obj = 3  # risk, return, diversification

# Annualized expected returns
mu = np.array([0.28, 0.12, 0.22, 0.05, 0.08, 0.10, 0.32, 0.15])

# Annualized covariance matrix (the "risk" model)
sigma = np.array(
    [
        [0.070, 0.010, 0.020, 0.008, 0.005, 0.012, 0.045, 0.011],
        [0.010, 0.065, 0.015, 0.006, 0.004, 0.008, 0.009, 0.050],
        [0.020, 0.015, 0.055, 0.010, 0.007, 0.015, 0.018, 0.014],
        [0.008, 0.006, 0.010, 0.030, 0.012, 0.009, 0.007, 0.005],
        [0.005, 0.004, 0.007, 0.012, 0.025, 0.006, 0.004, 0.003],
        [0.012, 0.008, 0.015, 0.009, 0.006, 0.045, 0.011, 0.007],
        [0.045, 0.009, 0.018, 0.007, 0.004, 0.011, 0.085, 0.010],
        [0.011, 0.050, 0.014, 0.005, 0.003, 0.007, 0.010, 0.072],
    ]
)

# Diversification score: number of cross-sector pairs in the portfolio.
# D[i,j] = 0.5 when assets i and j are in different sectors, so x^T D x counts
# the cross-sector pairs. More cross-sector pairs = better diversified.
D = np.array(
    [
        [1.0 if sectors[i] != sectors[j] else 0.0 for j in range(n_assets)]
        for i in range(n_assets)
    ]
)
np.fill_diagonal(D, 0.0)
D = D / 2

print(f"{n_assets} assets, choose K={K}, {n_obj} objectives")
for t, s, m in zip(tickers, sectors, mu):
    print(f"  {t:5s} ({s:8s})  expected return {m:5.0%}")

8 assets, choose K=4, 3 objectives
  AAPL  (Tech    )  expected return   28%
  XOM   (Energy  )  expected return   12%
  JPM   (Finance )  expected return   22%
  JNJ   (Health  )  expected return    5%
  KO    (Staples )  expected return    8%
  AMT   (REIT    )  expected return   10%
  AMZN  (Tech    )  expected return   32%
  SLB   (Energy  )  expected return   15%

# Each objective becomes a Hamiltonian whose lowest-energy bitstrings are the
# best portfolios for that objective. The opt-mapper turns a plain
# min/max problem over binary variables into the equivalent Ising operator.
def build_risk_hamiltonian(sigma, n):
    """Minimize portfolio variance x^T sigma x  (quadratic -> ZZ terms)."""
    prob = OptimizationProblem("risk")
    prob.binary_var_list(n)
    prob.minimize(quadratic=sigma)
    op, _ = prob.to_ising()
    return op.simplify()

def build_return_hamiltonian(mu, n):
    """Maximize expected return mu . x  (linear -> Z terms)."""
    prob = OptimizationProblem("return")
    prob.binary_var_list(n)
    prob.maximize(linear=mu)
    op, _ = prob.to_ising()
    return op.simplify()

def build_diversity_hamiltonian(D, n):
    """Maximize cross-sector pairs x^T D x  (quadratic -> ZZ terms)."""
    prob = OptimizationProblem("diversity")
    prob.binary_var_list(n)
    prob.maximize(quadratic=D)
    op, _ = prob.to_ising()
    return op.simplify()

H_risk = build_risk_hamiltonian(sigma, n_assets)
H_return = build_return_hamiltonian(mu, n_assets)
H_diversity = build_diversity_hamiltonian(D, n_assets)
cost_ops = [H_risk, H_return, H_diversity]

for name, op in zip(["risk", "return", "diversity"], cost_ops):
    print(f"H_{name:10s}: {op.size} Pauli terms")

H_risk      : 36 Pauli terms
H_return    : 8 Pauli terms
H_diversity : 34 Pauli terms

Menegakkan "tepat K aset" tanpa penalti. Trik yang umum adalah menambahkan suku penalti yang menghukum portofolio dengan ukuran salah, tetapi itu mengopel setiap qubit dengan setiap qubit lain, yang menghasilkan circuit jauh lebih dalam setelah transpilasi. Sebagai gantinya kita memakai mixer XY, yang hanya memindahkan state QAOA di antara bitstring dengan Hamming weight yang sama. Kalau kita mulai dari state yang sudah punya $K$ aset terpilih, setiap portofolio yang dijelajahi circuit juga punya tepat $K$ aset. Batasan ini tertanam di struktur circuit, bukan ditegakkan lewat suku penalti.

Kita menyiapkan state awal dengan murah: putar setiap qubit sehingga "menyala" dengan probabilitas $K/N$. Dibatasi pada hasil dengan $K$ aset, ini mereproduksi state berbobot sama yang ideal (state Dicke), jadi kita cukup menyimpan bitstring hasil pengukuran yang punya tepat $K$ angka satu — langkah yang disebut post-selection.

def xy_mixer(n):
    """Line XY mixer: couples neighboring qubits with XX+YY. Conserves the number
    of selected assets (Hamming weight), so cardinality is preserved automatically.
    Using a line (not a full ring) keeps the circuit shallow and hardware-friendly."""
    terms = [
        (pauli, [i, i + 1], 1) for i in range(n - 1) for pauli in ("XX", "YY")
    ]
    return SparsePauliOp.from_sparse_list(terms, n)

def product_init(n, k):
    """Cheap initial state: each qubit rotated so P(selected) = k/n. Zero two-qubit
    gates. Post-selecting its weight-k outcomes reproduces the ideal Dicke state."""
    qc = QuantumCircuit(n)
    theta = 2 * np.arcsin(np.sqrt(k / n))
    for q in range(n):
        qc.ry(theta, q)
    return qc

# Combine the three objectives with weights c (bound later, at sampling time).
p_layers = 1
c = ParameterVector("c", n_obj)
# Negate the objective so the sampling phase separator matches the sign the angles
# were trained under (the trainer maximizes the negated sum); binding below keeps +gamma.
combined_cost_op = sum(
    -c[k] * H_k for k, H_k in enumerate(cost_ops)
).simplify()

ansatz = qaoa_ansatz(
    combined_cost_op,
    reps=p_layers,
    initial_state=product_init(n_assets, K),
    mixer_operator=xy_mixer(n_assets),
)
ansatz.measure_all()

betas = [p for p in ansatz.parameters if p.name.startswith("β")]
gammas = [p for p in ansatz.parameters if p.name.startswith("γ")]
print(f"Qubits: {ansatz.num_qubits} | QAOA layers: {p_layers}")
print(
    f"Tunable angles: {len(betas)} beta + {len(gammas)} gamma, plus {n_obj} objective weights"
)

Qubits: 8 | QAOA layers: 1
Tunable angles: 1 beta + 1 gamma, plus 3 objective weights

Langkah 2: Optimalkan masalah untuk eksekusi di hardware kuantum

Sebelum dijalankan, circuit abstrak ditranspile menjadi gate native hardware. Pada skala kecil ini kita hanya memeriksa biayanya: seberapa dalam circuit-nya dan berapa banyak gate dua-qubit yang dipakainya? (Gate dua-qubit adalah sumber utama noise di device nyata.)

# Bind dummy angle values so we can transpile and measure the circuit's size.
dummy = {p: 0.1 for p in ansatz.parameters}
test_pm = generate_preset_pass_manager(optimization_level=1)
test_qc = test_pm.run(ansatz.assign_parameters(dummy))

print(f"Circuit depth          : {test_qc.depth()}")
print(
    f"Two-qubit gate depth   : {test_qc.depth(lambda x: len(x.qubits) > 1)}"
)
print(f"Two-qubit gate count   : {test_qc.num_nonlocal_gates()}")

Circuit depth          : 25
Two-qubit gate depth   : 23
Two-qubit gate count   : 42

Langkah 3: Jalankan dengan primitif Qiskit

Dua tahap. Pertama kita melatih sudut QAOA sekali, dengan bobot objektif yang sama, memakai simulator statevector eksak untuk menemukan nilai $\beta,\gamma$ yang bagus. Lalu kita menyapu banyak vektor bobot di seluruh simpleks dan men-sampling circuit di tiap titik, mengumpulkan portofolio kandidat. Karena hanya bobot objektif yang berubah di antara sapuan (bukan sudut hasil pelatihan), semua vektor bobot dikirim dalam satu job batch.

# Train the angles with equal objective weights.
# The trainer maximizes energy, so we negate the (to-be-minimized) objective sum.
training_op = sum(-1.0 / n_obj * H_k for H_k in cost_ops).simplify()

# Linear-ramp initialization (Sack and Serbyn, arXiv:2101.05742)
dt = 0.75
grid = np.arange(1, p_layers + 1) - 0.5
init_params = np.concatenate((1 - grid * dt / p_layers, grid * dt / p_layers))

trainer = ScipyTrainer(
    StatevectorEvaluator(), minimize_args={"options": {"maxiter": 300}}
)
print("Training QAOA angles (exact statevector)...")
result_train = trainer.train(
    cost_op=training_op,
    mixer=xy_mixer(n_assets),
    initial_state=product_init(n_assets, K),
    params0=init_params,
)
opt = result_train["optimized_params"]
opt_betas, opt_gammas = opt[:p_layers], opt[p_layers:]
print(f"Trained beta : {opt_betas}")
print(f"Trained gamma: {opt_gammas}")

Training QAOA angles (exact statevector)...
Trained beta : [3.329186967386619]
Trained gamma: [3.4449804324291033]

def random_uniform_simplex(n_samples, n_obj=3):
    """n_samples weight vectors spread uniformly over the (n_obj-1)-simplex."""
    s = np.zeros((n_samples, n_obj + 1))
    s[:, 1:-1] = np.random.rand(n_samples, n_obj - 1)
    s[:, -1] = 1
    s = np.sort(s, axis=1)
    return np.diff(s, axis=1)

# Bind the trained angles, leaving the objective weights c free for the sweep.
param_map = {betas[i]: opt_betas[i] for i in range(p_layers)}
param_map.update({gammas[i]: opt_gammas[i] for i in range(p_layers)})
ansatz_bound = ansatz.assign_parameters(param_map)

n_samples, shots = 200, 500
c_vecs = random_uniform_simplex(n_samples, n_obj)

print(f"Sampling {n_samples} weight vectors x {shots} shots...")
result = sampler.run([(ansatz_bound, c_vecs)], shots=shots).result()

# Collect every distinct bitstring seen across all weight vectors.
all_bitstrings = set()
for s in range(n_samples):
    for bs in result[0].data.meas.get_counts(s):
        # get_counts is little-endian; reverse so bit i = asset i
        all_bitstrings.add(bs.replace(" ", "")[::-1])
print(f"Distinct portfolios sampled: {len(all_bitstrings)}")

Sampling 200 weight vectors x 500 shots...
Distinct portfolios sampled: 256

Langkah 4: Pasca-proses dan kembalikan hasil dalam format klasik yang diinginkan

Kita hanya menyimpan portofolio yang layak (tepat $K$ aset dari langkah post-selection), menilai masing-masing pada ketiga tujuan, dan mengekstrak front Pareto: portofolio yang tidak dikalahkan pada semua tujuan sekaligus. Hypervolume adalah satu angka yang merangkum seberapa besar ruang objektif yang didominasi front, dan makin besar makin baik.

Dengan 100.000 shot atas hanya 256 bitstring, run ini melihat ke-70 portofolio valid semuanya, jadi pada ukuran ini ia pada dasarnya adalah pemeriksaan brute force bahwa pipeline terpasang dengan benar, bukan bukti bahwa QAOA menemukan front tersebut.

def evaluate_portfolio(bitstring, sigma, mu, D):
    """Score one portfolio on all three objectives (all framed as 'bigger is better')."""
    x = np.array([int(b) for b in bitstring])
    # negative risk, return, diversification (cross-sector pairs)
    return np.array([-(x @ sigma @ x), x @ mu, x @ D @ x])

# Post-select feasible portfolios, then score them.
feasible = [bs for bs in all_bitstrings if bs.count("1") == K]
fis = np.array([evaluate_portfolio(bs, sigma, mu, D) for bs in feasible])

pareto_front = filter_dominated(fis, maximise=True)
ref_point = fis.min(axis=0)
qmoo_hv = hypervolume(fis, ref=ref_point, maximise=True)

print(
    f"Feasible portfolios found : {len(feasible)} of {comb(n_assets, K)} possible"
)
print(f"Pareto-front portfolios   : {len(pareto_front)}")
print(f"Hypervolume               : {qmoo_hv:.4f}")

Feasible portfolios found : 70 of 70 possible
Pareto-front portfolios   : 26
Hypervolume               : 0.2487

fig = plt.figure(figsize=(8, 6))
ax = fig.add_subplot(111, projection="3d")
ax.scatter(
    fis[:, 0],
    fis[:, 1],
    fis[:, 2],
    c="lightgray",
    s=12,
    label="All feasible portfolios",
)
ax.scatter(
    pareto_front[:, 0],
    pareto_front[:, 1],
    pareto_front[:, 2],
    c="steelblue",
    s=45,
    label="Pareto front",
)
ax.set_xlabel("Negative risk")
ax.set_ylabel("Return")
ax.set_zlabel("Diversification")
ax.set_title("Risk / return / diversification Pareto front (8 assets)")
ax.legend()
plt.tight_layout()
plt.show()

Contoh hardware skala besar

Sekarang workflow yang sama pada 40 aset (8 sektor × 5), memilih $K=6$. Empat puluh qubit terlalu besar untuk disimulasikan secara eksak (statevector dengan $2^{40}$ amplitudo), jadi sudutnya tidak bisa dioptimalkan seperti pada delapan aset, dan circuit yang padat akan terlalu dalam untuk hardware saat ini. Sebanyak $\binom{40}{6} \approx 3.8$M portofolio valid masih cukup sedikit untuk dienumerasi secara klasik, yang kita pakai di akhir sebagai tolok ukur eksak. Beberapa hal berubah, dan tidak ada hal lain dari metodenya yang berubah:

1. Latih sudut dengan simulator matrix-product-state (MPS), bukan statevector eksak. Mengikuti referensi (Kotil et al.), kita menetapkan bobot objektif sama, mengoptimalkan satu β, γ di simulator MPS, dan memakainya ulang untuk setiap vektor pembobotan dalam sapuan. (Kita melatih pada ukuran yang kita jalankan — tanpa transfer sudut dari kecil ke besar.)

2. Jarangkan model risiko agar muat di hardware. Kovarians penuh mengopel semua 780 pasangan aset. Kita hanya menyimpan kopling terkuat yang paling murah di-routing dengan pemangkasan QAP yang sadar-kepentingan, dan mengopel setiap sektor dalam cincin ringan untuk suku diversitas. Ini menjaga tujuan tetap bermakna sambil menahan circuit di ukuran yang ramah hardware.

3. Jaga circuit tetap dangkal dan nilai dengan jujur. Routing bersifat stokastik, jadi kita mentranspile dengan beberapa seed dan menyimpan yang paling dangkal (dan tidak ada waktu kuantum yang terpakai). Portofolio selalu dinilai terhadap tujuan penuh yang sebenarnya. Penjarangan hanya membentuk circuit, bukan cara portofolio dinilai.

Kenapa pemangkasan QAP?

Menyimpan kopling terbesar tiap aset berdasarkan magnitudo saja bisa menghasilkan circuit yang jarang tetapi tetap sulit di-routing. Pemangkasan QAP justru menyimpan kopling yang besar dan dekat secara fisik di chip, sehingga anggaran gate yang sama menghasilkan circuit yang lebih dangkal dan lebih ramah hardware.

Langkah 1: Petakan input (dijarangkan untuk hardware)

import csv
import urllib.request

# Download the committed market-data snapshot from the repo.
# --- 40-asset universe: 8 GICS sectors x 5 tickers (real market data) ---
# Load the committed market-data snapshot (real annualized returns and covariance).
# Values are stored at the precision used to train the shipped QAOA angles
# (mu: 3 dp, sigma: 4 dp), so the pre-trained parameters in instances/ stay exactly valid.

url = "https://raw.githubusercontent.com/Qiskit/documentation/main/datasets/tutorials/qmoo/market_data.csv"
urllib.request.urlretrieve(url, "market_data.csv")

with open("market_data.csv", newline="") as _f:
    _rows = list(csv.reader(_f))

# Covariance column order
_tickers_csv = _rows[0][2:]
# Asset tickers
tickers_40 = [r[0] for r in _rows[1:]]
# Annualized expected returns
mu_40 = np.array([float(r[1]) for r in _rows[1:]])
# Covariance (risk model)
sigma_40 = np.array([[float(v) for v in r[2:]] for r in _rows[1:]])

sectors_40 = [
    "Tech",
    "Tech",
    "Tech",
    "Tech",
    "Tech",
    "Energy",
    "Energy",
    "Energy",
    "Energy",
    "Energy",
    "Finance",
    "Finance",
    "Finance",
    "Finance",
    "Finance",
    "Health",
    "Health",
    "Health",
    "Health",
    "Health",
    "Staples",
    "Staples",
    "Staples",
    "Staples",
    "Staples",
    "Industrials",
    "Industrials",
    "Industrials",
    "Industrials",
    "Industrials",
    "Utilities",
    "Utilities",
    "Utilities",
    "Utilities",
    "Utilities",
    "REIT",
    "REIT",
    "REIT",
    "REIT",
    "REIT",
]
n_assets_40 = len(tickers_40)
K_40 = 6  # choose exactly K assets

sector_names_40 = list(dict.fromkeys(sectors_40))
sect_idx_40 = np.array([sector_names_40.index(s) for s in sectors_40])
# True cross-sector diversification matrix (used for scoring)
D_40 = np.array(
    [
        [
            0.5 if sectors_40[i] != sectors_40[j] else 0.0
            for j in range(n_assets_40)
        ]
        for i in range(n_assets_40)
    ]
)
np.fill_diagonal(D_40, 0.0)
print(
    f"{n_assets_40} assets, {len(sector_names_40)} sectors, choose K={K_40}"
)

40 assets, 8 sectors, choose K=6

# Sparsify the covariance so the risk circuit fits on hardware. A small diagonal
# shift (added after truncation) keeps the risk model positive semidefinite; at fixed
# K it adds the same constant to every portfolio, so it never changes the ranking.
# Importance-aware QAP truncation (Kotil et al. style): place the qubits on a line
# and use a Quadratic Assignment Problem to choose the layout that keeps the
# strongest covariance couplings within routing distance k of the swap network,
# then drop the rest. Unlike a fixed top-k cap, it keeps couplings that are both
# large AND cheap to route.
from scipy.optimize import quadratic_assignment as qap
from qiskit.transpiler.passes.routing.commuting_2q_gate_routing import (
    SwapStrategy,
)

# Truncation level: larger k keeps more couplings (deeper circuit)
k_truncate = 2
_dist = np.array(
    SwapStrategy.from_line(list(range(n_assets_40))).distance_matrix
)

def qap_truncate(Q, k):
    w = np.abs(Q.copy())
    np.fill_diagonal(w, 0.0)
    mask = (_dist <= k).astype(float)
    # Seed the QAP solver explicitly (by default it draws from NumPy's global
    # RNG, which SciPy is deprecating) so the truncation is reproducible.
    perm = qap(-w, mask, options={"rng": np.random.default_rng(42)}).col_ind
    keep = mask[np.ix_(perm, perm)]
    Qt = Q * keep
    np.fill_diagonal(Qt, np.diag(Q))
    return Qt

sigma_sparse = qap_truncate(sigma_40, k_truncate)
print(
    f"QAP truncation (k={k_truncate}): risk edges kept = "
    f"{(np.count_nonzero(sigma_sparse) - n_assets_40) // 2}"
)
ridge = max(0.0, -np.linalg.eigvalsh(sigma_sparse)[0]) + 1e-6
sigma_sparse = sigma_sparse + ridge * np.eye(n_assets_40)

# Diversity: couple each sector's assets in a ring (sparse stand-in for the
# same-sector pair count). Scoring still uses the true cross-sector matrix D_40.
def build_same_sector_hamiltonian(D_same, n):
    prob = OptimizationProblem("diversity_sparse")
    prob.binary_var_list(n)
    prob.minimize(quadratic=D_same)
    op, _ = prob.to_ising()
    return op.simplify()

D_ring = np.zeros((n_assets_40, n_assets_40))
for s in set(sect_idx_40):
    members = np.where(sect_idx_40 == s)[0]
    for k in range(len(members)):
        i, j = members[k], members[(k + 1) % len(members)]
        D_ring[i, j] = D_ring[j, i] = 0.5

H_risk_40 = build_risk_hamiltonian(sigma_sparse, n_assets_40)
H_return_40 = build_return_hamiltonian(mu_40, n_assets_40)
H_diversity_40 = build_same_sector_hamiltonian(D_ring, n_assets_40)
cost_ops_40 = [H_risk_40, H_return_40, H_diversity_40]
n_zz = sum(
    1 for p in sum(cost_ops_40).simplify().paulis if str(p).count("Z") == 2
)
print(
    f"Cost-layer interactions: {n_zz} (dense would be {n_assets_40*(n_assets_40-1)//2})"
)

QAP truncation (k=2): risk edges kept = 78
Cost-layer interactions: 102 (dense would be 780)

Langkah 2-3: latih sudut, lalu bangun dan kirim job hardware

Empat puluh qubit terlalu besar untuk mengoptimalkan sudut secara eksak, jadi kita melatih satu $\beta, \gamma$ di simulator matrix-product-state pada bobot objektif yang sama dan memakainya ulang di seluruh sapuan. Sel di bawah memuat nilai hasil pelatihan dari sebuah file. Sudut lapisan biaya hasil pelatihan kecil ($\gamma \approx 0.32$), jadi circuit menerapkan bias lembut, bukan proyeksi tajam.

import json
from qiskit_ibm_runtime import QiskitRuntimeService, SamplerV2

# Pre-trained angles loaded from a file (training is slow; QDC pattern).
# Set load_params_file = False to retrain in-notebook.
load_params_file = True
params_url = "https://raw.githubusercontent.com/Qiskit/documentation/main/datasets/tutorials/qmoo/qaoa_params.json"
params_path = "qaoa_params.json"
if load_params_file:
    urllib.request.urlretrieve(params_url, params_path)
    qaoa_params = json.load(open(params_path))
    p_layers_hw = qaoa_params["p_layers"]
    opt_betas_40, opt_gammas_40 = qaoa_params["betas"], qaoa_params["gammas"]
else:
    # Same workflow as the small-scale example: qaoa_training_pipeline's MPSAerEvaluator
    # evaluates the QAOA energy on Aer's MPS simulator and supports the XY mixer.
    # As in the small-scale example the trainer maximizes energy, so we negate the
    # (to-be-minimized) objective sum; the 1/n_obj scaling matches the equal-weight
    # point of the sweep, so the trained gamma transfers directly to the weighted circuits.
    p_layers_hw = 1
    training_op_40 = sum(-1.0 / n_obj * H_k for H_k in cost_ops_40).simplify()

    dt = 0.75
    grid = np.arange(1, p_layers_hw + 1) - 0.5
    init_params_40 = np.concatenate(
        (1 - grid * dt / p_layers_hw, grid * dt / p_layers_hw)
    )

    trainer_40 = ScipyTrainer(
        MPSAerEvaluator({"matrix_product_state_max_bond_dimension": 24}),
        minimize_args={"options": {"maxiter": 80}},
    )
    print("Training QAOA angles (MPS simulator)...")
    result_train_40 = trainer_40.train(
        cost_op=training_op_40,
        mixer=xy_mixer(n_assets_40),
        initial_state=product_init(n_assets_40, K_40),
        params0=init_params_40,
    )
    opt_40 = result_train_40["optimized_params"]
    opt_betas_40, opt_gammas_40 = (
        list(opt_40[:p_layers_hw]),
        list(opt_40[p_layers_hw:]),
    )
    import os

    os.makedirs(os.path.dirname(params_path) or ".", exist_ok=True)
    json.dump(
        {
            "p_layers": p_layers_hw,
            "betas": opt_betas_40,
            "gammas": opt_gammas_40,
        },
        open(params_path, "w"),
        indent=2,
    )
    print(
        f"Trained angles saved to {params_path} (set load_params_file=True to reuse)."
    )

c40 = ParameterVector("c", n_obj)
# Negated to match the trained angles' sign, as in the small-scale cell; binding keeps +gamma.
combined_cost_op_40 = sum(
    -c40[k] * H_k for k, H_k in enumerate(cost_ops_40)
).simplify()
qc_40 = qaoa_ansatz(
    combined_cost_op_40,
    reps=p_layers_hw,
    initial_state=product_init(n_assets_40, K_40),
    mixer_operator=xy_mixer(n_assets_40),
)
qc_40.measure_all()
b40 = [p for p in qc_40.parameters if p.name.startswith("β")]
g40 = [p for p in qc_40.parameters if p.name.startswith("γ")]
pmap = {b40[i]: opt_betas_40[i] for i in range(p_layers_hw)}
pmap.update({g40[i]: opt_gammas_40[i] for i in range(p_layers_hw)})
ansatz_qc_40 = qc_40.assign_parameters(pmap)

service = QiskitRuntimeService()

# only use Heron devices
backend = service.least_busy(min_num_qubits=156)

# SABRE routing is stochastic: different seeds give different depths. Transpilation
# is classical (it costs no QPU time), so we transpile many seeds and keep only the
# shallowest circuit -- a free reduction in two-qubit depth before anything is sent
# to hardware. Only this single best circuit is ever executed.
n_seeds = 24
best = None
depths = []
for seed in range(n_seeds):
    pm = generate_preset_pass_manager(
        optimization_level=3, backend=backend, seed_transpiler=seed
    )
    qc = pm.run(ansatz_qc_40)
    d2 = qc.depth(lambda x: len(x.qubits) > 1)
    depths.append(d2)
    if best is None or d2 < best[0]:
        best = (d2, seed, qc)
isa_qc = best[2]
sd = sorted(depths)
print(
    f"Backend: {backend.name} | {n_seeds} seeds | two-qubit depth "
    f"best/median/worst = {sd[0]}/{sd[len(sd)//2]}/{sd[-1]} (best seed {best[1]})"
)
print(
    f"Selected circuit -> two-qubit gates: {isa_qc.num_nonlocal_gates()}, "
    f"two-qubit depth: {isa_qc.depth(lambda x: len(x.qubits) > 1)}"
)

Backend: ibm_kingston | 24 seeds | two-qubit depth best/median/worst = 220/261/300 (best seed 22)
Selected circuit -> two-qubit gates: 787, two-qubit depth: 220

# Submit one batched job (job mode; a single batch needs no Session).
# Extra shots: noise lowers the post-selection yield
n_samples_40, shots_40 = 24, 1500
c_vecs_40 = random_uniform_simplex(n_samples_40, n_obj)
sampler_hw = SamplerV2(mode=backend)
# Tag hardware jobs for tracking
sampler_hw.options.environment.job_tags = ["TUT_QAMOO"]
# Seconds; guard against runaway jobs
sampler_hw.options.max_execution_time = 600

# The QAOA angles are already bound; only the objective weights c remain free.
# Assign each weight vector to get one concrete circuit per point on the simplex.
bound_circuits_40 = [
    isa_qc.assign_parameters({c40[k]: cv[k] for k in range(n_obj)})
    for cv in c_vecs_40
]
job_hw = sampler_hw.run([(qc,) for qc in bound_circuits_40], shots=shots_40)
print(
    f"Submitted to {backend.name}: job id {job_hw.job_id()} ({len(bound_circuits_40)} circuits)"
)

Submitted to ibm_kingston: job id darcaalvr3kc73einokg (24 circuits)

Langkah 4: Pasca-proses menjadi front Pareto dan baca portofolio optimalnya

result_hw = job_hw.result()

# Post-select feasible portfolios (exactly K assets), score on the TRUE objectives.
feasible_40 = set()
for s in range(n_samples_40):
    for bs in result_hw[s].data.meas.get_counts():
        # get_counts is little-endian; reverse so bit i = asset i
        bs = bs.replace(" ", "")[::-1]
        if bs.count("1") == K_40:
            feasible_40.add(bs)

def score_40(P):
    """Score portfolios given as an (n, K) array of asset indices.

    Returns an (n, 3) array of [negative risk, return, diversification], all framed
    as 'bigger is better'. This is the single definition of the true objectives,
    used for the hardware samples, the exact enumeration, and the random baseline.
    Looping over the K x K index pairs keeps memory O(n) even for millions of rows.
    """
    risk = np.zeros(len(P))
    div = np.zeros(len(P))
    for a in range(P.shape[1]):
        for b in range(P.shape[1]):
            risk += sigma_40[P[:, a], P[:, b]]
            div += D_40[P[:, a], P[:, b]]
    return np.column_stack([-risk, mu_40[P].sum(1), div])

def evaluate_40(bs):
    """Score one bitstring (bit i = asset i) with score_40."""
    idx = np.flatnonzero([b == "1" for b in bs])
    return score_40(idx[None, :])[0]

fis_40 = np.array([evaluate_40(bs) for bs in feasible_40])
pareto_40 = filter_dominated(fis_40, maximise=True)
print(f"Feasible portfolios collected : {len(feasible_40)}")
print(f"Pareto-front portfolios       : {len(pareto_40)}")

Feasible portfolios collected : 2359
Pareto-front portfolios       : 20

# --- Honest benchmark: QAOA and random vs the EXACT Pareto front ---
# 40 choose 6 = 3,838,380 feasible portfolios -- few enough to enumerate exactly, score
# every one on the TRUE objectives, and get the exact Pareto front. That front is an
# absolute ceiling, and its feasible nadir is a FIXED hypervolume reference point, so the
# numbers are comparable across runs instead of depending on what happened to be sampled.
import itertools

# Stream the combinations straight into an (n, K) index array, without first
# building millions of Python tuples.
combos = np.fromiter(
    itertools.chain.from_iterable(
        itertools.combinations(range(n_assets_40), K_40)
    ),
    dtype=np.int16,
).reshape(-1, K_40)
fis_exact = score_40(combos)
front_exact = filter_dominated(fis_exact, maximise=True)

# Fixed reference = worst value of each objective over ALL feasible portfolios (the nadir).
ref_fixed = fis_exact.min(axis=0)
# The hypervolume of a point set equals the hypervolume of its front.
hv_ceiling = hypervolume(front_exact, ref=ref_fixed, maximise=True)
hv_qaoa = hypervolume(fis_40, ref=ref_fixed, maximise=True)

def random_feasible_hv(n_draw, seed):
    """Hypervolume of n_draw uniformly-random feasible portfolios, same fixed reference."""
    rng = np.random.default_rng(seed)
    picks = set()
    while len(picks) < n_draw:
        picks.add(tuple(sorted(rng.choice(n_assets_40, K_40, replace=False))))
    P = np.array(list(picks))
    return hypervolume(score_40(P), ref=ref_fixed, maximise=True)

hv_rand = np.array(
    [random_feasible_hv(len(feasible_40), seed) for seed in range(20)]
)

print(f"Exact Pareto front            : {len(front_exact)} portfolios")
print(f"Hypervolume ceiling (optimum) : {hv_ceiling:.3f}")
print(
    f"QAOA (hardware)               : {100 * hv_qaoa / hv_ceiling:5.1f}% of optimum"
)
print(
    f"Random ({len(feasible_40)} draws, 20 seeds) : "
    f"{100 * hv_rand.mean() / hv_ceiling:5.1f}% +/- {100 * hv_rand.std() / hv_ceiling:.1f}% of optimum"
)

Exact Pareto front            : 61 portfolios
Hypervolume ceiling (optimum) : 64.211
QAOA (hardware)               :  85.5% of optimum
Random (2359 draws, 20 seeds) :  84.4% +/- 1.3% of optimum

Pada ukuran masalah ini, QAOA berkinerja hampir sama baiknya dengan sampling acak seragam. Keduanya memulihkan sebagian besar hypervolume optimum eksak, dan tidak ada yang jelas unggul. Hasil ini wajar untuk satu lapisan QAOA dangkal dengan operator biaya yang dipangkas besar-besaran di hardware yang noisy; nilai contoh ini ada pada workflow multi-objektif ujung-ke-ujung (pemetaan, pelatihan sudut, sampling terbatas, dan pasca-proses Pareto), bukan pada percepatan kuantum. Mempersempit selisih dengan optimum akan membutuhkan circuit yang lebih dalam (lebih banyak lapisan QAOA), pemangkasan yang lebih lembut, atau hardware dengan noise lebih rendah.

# The best trade-offs found by the sampler: no other sampled portfolio beats these
# on every objective. They approximate the exact front computed above; a
# decision-maker picks the trade-off they prefer.
bs_list = list(feasible_40)
# Boolean mask over fis_40; keep_weakly=True also keeps portfolios whose
# objective values tie with a front point (what the strict filter would drop).
mask = is_nondominated(fis_40, maximise=True, keep_weakly=True)
front_bs = [b for b, m in zip(bs_list, mask) if m]
front_f = fis_40[mask]
order = np.argsort(-front_f[:, 1])  # show a span sorted by return
print(
    f"{mask.sum()} non-dominated sampled portfolios. A representative span:\n"
)
print(
    f"{'tickers held':40s} {'risk':>7s} {'return':>7s} {'cross-sector':>12s}"
)
for idx in order[:: max(1, len(order) // 12)]:
    held = [tickers_40[i] for i, b in enumerate(front_bs[idx]) if b == "1"]
    print(
        f"{', '.join(held):40s} {-front_f[idx,0]:7.3f} {front_f[idx,1]:7.2f} {int(front_f[idx,2]):12d}"
    )

fig = plt.figure(figsize=(8, 6))
ax = fig.add_subplot(111, projection="3d")
ax.scatter(
    fis_40[:, 0],
    fis_40[:, 1],
    fis_40[:, 2],
    c="lightgray",
    s=8,
    label="Sampled portfolios",
)
ax.scatter(
    pareto_40[:, 0],
    pareto_40[:, 1],
    pareto_40[:, 2],
    c="tomato",
    marker="D",
    s=40,
    label="Pareto front",
)
ax.set_xlabel("Negative risk")
ax.set_ylabel("Return")
ax.set_zlabel("Diversification")
ax.set_title("40-asset Pareto front (quantum hardware)")
ax.legend()
plt.tight_layout()
plt.show()

20 non-dominated sampled portfolios. A representative span:

tickers held                                risk  return cross-sector
AAPL, NVDA, XOM, GS, BLK, CAT              1.715    2.22           13
NVDA, GS, MS, PFE, CAT, PLD                1.772    2.20           14
NVDA, MS, PFE, WMT, CAT, DUK               1.040    2.14           15
NVDA, CVX, GS, JNJ, KO, WMT                0.737    2.01           14
NVDA, COP, GS, JNJ, WMT, EQIX              1.014    1.98           15
NVDA, ABT, WMT, CAT, AEP, EQIX             0.913    1.92           15
NVDA, CVX, WMT, RTX, D, EQIX               0.835    1.82           15
NVDA, XOM, GS, JNJ, DUK, EQIX              0.763    1.80           15
AAPL, NVDA, JNJ, KO, RTX, SO               0.621    1.71           14
NVDA, CVX, BLK, JNJ, RTX, AEP              0.719    1.71           15
NVDA, JNJ, KO, COST, RTX, AEP              0.545    1.69           14
NVDA, CVX, ABT, WMT, HON, AEP              0.702    1.43           15
AAPL, GS, JNJ, PEP, AEP, EQIX              0.645    1.34           15
MSFT, XOM, BLK, JNJ, CAT, SO               0.617    1.34           15
MSFT, KO, WMT, RTX, SO, SPG                0.526    1.29           14
MSFT, XOM, BLK, JNJ, WMT, DUK              0.483    1.25           15
MSFT, CVX, JNJ, RTX, DUK, D                0.482    1.02           14
MSFT, XOM, JNJ, KO, HON, EQIX              0.479    0.93           15
MSFT, JNJ, PG, KO, RTX, DUK                0.423    0.86           14
MSFT, XOM, JNJ, PEP, DUK, AMT              0.476    0.57           15

Langkah berikutnya

Rekomendasi

Kalau tutorial ini menarik buatmu, coba hal berikut:

• Ganti data pasar yang diunduh (market_data.csv) dengan estimasi imbal hasil dan kovarians milikmu sendiri dari riwayat harga yang nyata.

• Tambah jumlah lapisan QAOA, atau latih pada 12–16 aset dan transfer sudut-sudutnya, untuk mendekatkan front hardware ke optimal.

• Baca Kotil et al., Quantum Approximate Multi-Objective Optimization (Nature Computational Science, 2025), studi max-cut yang diadaptasi tutorial ini untuk portofolio.

Referensi

1. Kotil et al., "Quantum Approximate Multi-Objective Optimization," Nature Computational Science (2025). arXiv:2503.22797

2. S. H. Sack and M. Serbyn, "Quantum annealing initialization of the quantum approximate optimization algorithm," Quantum 5, 491 (2021). arXiv:2101.05742