Mga pamamaraan ng kompilasyon para sa mga Hamiltonian simulation circuit
Tinatayang paggamit: wala pang isang minuto sa isang IBM Heron processor (TANDAAN: Ito ay tinantya lamang. Ang iyong runtime ay maaaring mag-iba.)
Mga layunin sa pagkatuto
Pagkatapos ng tutorial na ito, maiintindihan mo ang:
- Kung paano gamitin ang Qiskit transpiler na may SABRE para sa layout at routing optimization
- Kung paano gamitin ang AI-powered transpiler para sa advanced circuit optimization
- Kung paano gamitin ang Rustiq plugin para sa pag-synthesize ng mga operasyon ng
PauliEvolutionGatesa mga Hamiltonian simulation circuit - Kung paano mag-benchmark at ikumpara ang mga pamamaraan ng kompilasyon gamit ang two-qubit depth, kabuuang gate count, at runtime
Mga kinakailangan
Inirerekomenda naming pamilyar ka sa mga sumusunod na paksa bago simulan ang tutorial na ito:
Background
Ang quantum circuit compilation ay nagbabago ng isang high-level quantum algorithm patungo sa isang physical circuit na sumusunod sa mga limitasyon ng target hardware. Ang epektibong kompilasyon ay maaaring makabuluhang bawasan ang circuit depth at gate count, na parehong direktang nakakaapekto sa kalidad ng mga resulta sa mga near-term quantum device.
Ang tutorial na ito ay nag-be-benchmark ng tatlong pamamaraan ng kompilasyon sa mga Hamiltonian simulation circuit na binuo gamit ang PauliEvolutionGate. Ang mga circuit na ito ay nagmomodelo ng pairwise na pakikipag-ugnayan ng qubit (tulad ng mga terminong , , at ) at karaniwan sa quantum chemistry, condensed matter physics, at materials science.
Ang mga benchmark circuit ay nagmula sa koleksyon ng Hamlib, na na-access sa pamamagitan ng repository ng Benchpress. Nagbibigay ang Hamlib ng standardisadong hanay ng mga kinatawang Hamiltonian, na nagbibigay-daan sa pagkukumpara ng mga estratehiya ng kompilasyon sa mga realistikong simulation workload.
Pangkalahatang-ideya ng mga pamamaraan ng kompilasyon
Qiskit transpiler na may SABRE
Ang Qiskit transpiler ay gumagamit ng SABRE (SWAP-based BidiREctional heuristic search) algorithm upang i-optimize ang circuit layout at routing. Nakatuon ang SABRE sa pagpapaliit ng mga SWAP gate at ang kanilang epekto sa circuit depth habang sumusunod sa mga hardware connectivity constraint. Ito ay isang pangkalahatang pamamaraan na nagbibigay ng magandang balanse sa pagitan ng performance at oras ng kompilasyon. Para sa higit pang detalye, tingnan ang [1]. Ang mga kalamangan at paggalugad ng parameter ng SABRE ay saklaw nang malalim sa hiwalay na tutorial.
AI-powered transpiler
Ang AI-powered transpiler ay gumagamit ng machine learning upang hulaan ang pinakamahusay na mga estratehiya ng transpilation sa pamamagitan ng pagsusuri ng mga pattern sa istruktura ng circuit at mga hardware constraint. Maaari rin itong mag-apply ng AIPauliNetworkSynthesis pass, na naglalayong sa mga Pauli network circuit gamit ang isang reinforcement learning-based synthesis approach. Para sa higit pang impormasyon, tingnan ang [2] at [3].
Rustiq plugin
Ang Rustiq plugin ay nagbibigay ng mga advanced synthesis technique na partikular para sa mga operasyon ng PauliEvolutionGate, na kumakatawan sa mga Pauli rotation na karaniwang ginagamit sa Trotterized dynamics. Dinisenyo ito upang makagawa ng mga low-depth circuit decomposition para sa mga Hamiltonian simulation workload. Para sa higit pang detalye, tingnan ang [4].
Mga pangunahing sukatan
Ikukumpara natin ang tatlong pamamaraan sa mga sumusunod na sukatan:
- Two-qubit depth: Ang depth ng circuit na bibilangin lamang ang mga two-qubit gate. Ito ay kadalasang bottleneck para sa fidelity sa tunay na hardware.
- Laki ng circuit (kabuuang gate count): Ang kabuuang bilang ng mga gate sa transpiled circuit.
- Runtime: Ang wall-clock na oras para sa transpilation.
Mga kinakailangan
Bago simulan ang tutorial na ito, siguraduhing naka-install mo ang mga sumusunod:
- Qiskit SDK v2.0 o mas bago, na may visualization support
- Qiskit Runtime v0.22 o mas bago (
pip install qiskit-ibm-runtime) - Qiskit Aer (
pip install qiskit-aer) - Qiskit IBM Transpiler (
pip install qiskit-ibm-transpiler) - Qiskit AI Transpiler local mode (
pip install qiskit_ibm_ai_local_transpiler) - Networkx (
pip install networkx)
Setup
# Added by doQumentation — required packages for this notebook
!pip install -q matplotlib numpy qiskit qiskit-aer qiskit-ibm-runtime qiskit-ibm-transpiler requests scipy
from qiskit.circuit import QuantumCircuit
from qiskit_ibm_runtime import QiskitRuntimeService, SamplerV2
from qiskit.circuit.library import PauliEvolutionGate
from qiskit_ibm_transpiler import generate_ai_pass_manager
from qiskit.quantum_info import SparsePauliOp
from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager
from qiskit.transpiler.passes.synthesis.high_level_synthesis import HLSConfig
from qiskit_aer import AerSimulator
from qiskit_aer.noise import NoiseModel, depolarizing_error
from collections import Counter
from statistics import mean, stdev
from scipy.sparse import SparseEfficiencyWarning
import time
import warnings
import matplotlib.pyplot as plt
import matplotlib.ticker as ticker
import numpy as np
import json
import requests
import logging
# Suppress noisy loggers and warnings
logging.getLogger(
"qiskit_ibm_transpiler.wrappers.ai_local_synthesis"
).setLevel(logging.ERROR)
warnings.filterwarnings("ignore", category=FutureWarning)
warnings.filterwarnings("ignore", category=SparseEfficiencyWarning)
seed = 42 # Seed for reproducibility
Kumonekta sa isang backend
Pumili ng backend na gagamitin para sa parehong maliit at malaking halimbawa. Tinutukoy ng backend ang coupling map at mga basis gate na tina-target ng transpiler.
# QiskitRuntimeService.save_account(channel="ibm_quantum_platform",
# token="<YOUR-API-KEY>", overwrite=True, set_as_default=True)
service = QiskitRuntimeService(channel="ibm_quantum_platform")
backend = service.least_busy(operational=True, simulator=False)
print(f"Using backend: {backend.name}")
Using backend: ibm_pittsburgh
Tukuyin ang mga pass manager
I-set up ang tatlong pamamaraan ng kompilasyon.
# SABRE pass manager (Qiskit default at optimization level 3)
pm_sabre = generate_preset_pass_manager(
optimization_level=3, backend=backend, seed_transpiler=seed
)
# AI transpiler pass manager (local mode)
pm_ai = generate_ai_pass_manager(
backend=backend, optimization_level=3, ai_optimization_level=3
)
Fetching 127 files: 0%| | 0/127 [00:00<?, ?it/s]
# Rustiq pass manager for PauliEvolutionGate synthesis
hls_config = HLSConfig(
PauliEvolution=[
(
"rustiq",
{
"nshuffles": 400,
"upto_phase": True,
"fix_clifford": True,
"preserve_order": False,
"metric": "depth",
},
)
]
)
pm_rustiq = generate_preset_pass_manager(
optimization_level=3,
backend=backend,
hls_config=hls_config,
seed_transpiler=seed,
)
Tukuyin ang mga helper function
Ang sumusunod na function ay nag-ta-transpile ng isang listahan ng mga circuit gamit ang isang ibinigay na pass manager, at nagtatala ng mga pangunahing sukatan (two-qubit depth, laki ng circuit, at runtime) para sa bawat circuit.
def capture_transpilation_metrics(
results, pass_manager, circuits, method_name
):
"""
Transpile circuits and append one metrics record per circuit to
``results``.
Args:
results (list): List of dicts to append the metrics records to.
pass_manager: Pass manager used for transpilation.
circuits (list): List of quantum circuits to transpile.
method_name (str): Name of the transpilation method.
Returns:
list: List of transpiled circuits.
"""
transpiled_circuits = []
for i, qc in enumerate(circuits):
start_time = time.time()
transpiled_qc = pass_manager.run(qc)
end_time = time.time()
# Decompose swaps for consistency across methods
transpiled_qc = transpiled_qc.decompose(gates_to_decompose=["swap"])
transpilation_time = end_time - start_time
two_qubit_depth = transpiled_qc.depth(
lambda x: x.operation.num_qubits == 2
)
circuit_size = transpiled_qc.size()
results.append(
{
"method": method_name,
"qc_name": qc.name,
"qc_index": i,
"num_qubits": qc.num_qubits,
"two_qubit_depth": two_qubit_depth,
"size": circuit_size,
"runtime": transpilation_time,
}
)
transpiled_circuits.append(transpiled_qc)
print(
f"[{method_name}] Circuit {i} ({qc.name}): "
f"2Q depth={two_qubit_depth}, size={circuit_size}, "
f"time={transpilation_time:.2f}s"
)
return transpiled_circuits
def _method_order(results):
"""Return the distinct method names in their first-seen order."""
order = []
for r in results:
if r["method"] not in order:
order.append(r["method"])
return order
def print_summary_table(results):
"""
Print the mean and standard deviation of each metric per compilation
method, followed by the mean percent improvement relative to SABRE.
"""
metrics = [
("two_qubit_depth", "2Q Depth"),
("size", "Gate Count"),
("runtime", "Runtime (s)"),
]
methods = _method_order(results)
by_method = {m: [r for r in results if r["method"] == m] for m in methods}
sabre_by_index = {r["qc_index"]: r for r in by_method.get("SABRE", [])}
col_w = 22
name_w = max(len(m) for m in methods)
header = f"{'Method':<{name_w}}" + "".join(
f" {label:>{col_w}}" for _, label in metrics
)
print("Mean +/- std per compilation method")
print(header)
print("-" * len(header))
for method in methods:
cells = []
for key, _ in metrics:
values = [r[key] for r in by_method[method]]
std = stdev(values) if len(values) > 1 else 0.0
cells.append(f"{mean(values):,.1f} +/- {std:,.1f}")
print(
f"{method:<{name_w}}" + "".join(f" {c:>{col_w}}" for c in cells)
)
others = [m for m in methods if m != "SABRE"]
if others and sabre_by_index:
print()
print("Mean % improvement vs SABRE (positive = better than SABRE)")
print(header)
print("-" * len(header))
for method in others:
cells = []
for key, _ in metrics:
pct = [
(sabre_by_index[r["qc_index"]][key] - r[key])
/ sabre_by_index[r["qc_index"]][key]
* 100
for r in by_method[method]
if sabre_by_index.get(r["qc_index"])
and sabre_by_index[r["qc_index"]][key]
]
if pct:
std = stdev(pct) if len(pct) > 1 else 0.0
cells.append(f"{mean(pct):+.1f}% +/- {std:.1f}%")
else:
cells.append("n/a")
print(
f"{method:<{name_w}}"
+ "".join(f" {c:>{col_w}}" for c in cells)
)
def print_per_circuit_comparison(results, num_rows=5):
"""
Print a per-metric comparison of the compilation methods for the
first ``num_rows`` circuits (sorted by qubit count). The best
(lowest) value for each metric is marked with an asterisk.
"""
metrics = [
("two_qubit_depth", "2Q Depth"),
("size", "Gate Count"),
("runtime", "Runtime (s)"),
]
methods = _method_order(results)
by_index = {}
for r in results:
by_index.setdefault(r["qc_index"], {})[r["method"]] = r
ordered = sorted(
by_index.items(),
key=lambda kv: (next(iter(kv[1].values()))["num_qubits"], kv[0]),
)[:num_rows]
for key, label in metrics:
print(f"{label} (first {num_rows} circuits by qubit count); * = best")
header = f"{'Idx':>3} {'Circuit':<16} {'Q':>3}" + "".join(
f"{m:>9}" for m in methods
)
print(header)
print("-" * len(header))
for idx, method_map in ordered:
any_record = next(iter(method_map.values()))
present = {
m: method_map[m][key] for m in methods if m in method_map
}
best = min(present.values())
line = (
f"{idx:>3} {any_record['qc_name'][:16]:<16} "
f"{any_record['num_qubits']:>3}"
)
for m in methods:
value = method_map[m][key]
text = f"{value:.2f}" if key == "runtime" else f"{int(value)}"
if value == best:
text += "*"
line += f"{text:>9}"
print(line)
print()
Mag-load ng mga Hamiltonian circuit mula sa Hamlib
Naglo-load tayo ng isang kinatawang hanay ng mga Hamiltonian mula sa Benchpress repository at nagtatayo ng mga PauliEvolutionGate circuit. Ang mga circuit na lumalagpas sa bilang ng qubit ng backend ay tinatanggal, kasama ang mga circuit na ang decomposed na laki ay lumagpas sa 1,500 gate (upang mapanatiling makatwirang oras ng transpilation).
# Obtain the Hamiltonian JSON from the benchpress repository
url = "https://raw.githubusercontent.com/Qiskit/benchpress/e7b29ef7be4cc0d70237b8fdc03edbd698908eff/benchpress/hamiltonian/hamlib/100_representative.json"
response = requests.get(url)
response.raise_for_status()
ham_records = json.loads(response.text)
# Remove circuits that are too large for the backend
ham_records = [
h for h in ham_records if h["ham_qubits"] <= backend.num_qubits
]
# Build PauliEvolutionGate circuits
qc_ham_list = []
for h in ham_records:
terms = h["ham_hamlib_hamiltonian_terms"]
coeff = h["ham_hamlib_hamiltonian_coefficients"]
num_qubits = h["ham_qubits"]
name = h["ham_problem"]
evo_gate = PauliEvolutionGate(SparsePauliOp(terms, coeff))
qc = QuantumCircuit(num_qubits)
qc.name = name
qc.append(evo_gate, range(num_qubits))
qc_ham_list.append(qc)
# Remove circuits whose decomposed size exceeds 1500 gates so that transpilation completes in a reasonable time frame
qc_ham_list = [qc for qc in qc_ham_list if qc.decompose().size() <= 1500]
print(f"Total Hamiltonian circuits loaded: {len(qc_ham_list)}")
print(
f"Qubit range: {min(qc.num_qubits for qc in qc_ham_list)} to {max(qc.num_qubits for qc in qc_ham_list)}"
)
Total Hamiltonian circuits loaded: 42
Qubit range: 2 to 112
Hatiin ang mga circuit sa maliit na antas (wala pang 20 qubit) at malalaking antas (20 o higit pa na qubit) na mga grupo.
qc_small = [qc for qc in qc_ham_list if qc.num_qubits < 20]
qc_large = [qc for qc in qc_ham_list if qc.num_qubits >= 20]
print(f"Small-scale circuits (<20 qubits): {len(qc_small)}")
print(f"Large-scale circuits (>=20 qubits): {len(qc_large)}")
Small-scale circuits (<20 qubits): 20
Large-scale circuits (>=20 qubits): 22
I-preview ang isa sa mga maliit na Hamiltonian circuit bago ang transpilation.
# We decompose the circuit here, otherwise it would just be a PauliEvolutionGate box,
# which isn't very informative to look at!
qc_small[0].decompose().draw("mpl", fold=-1)
Maliit na antas na halimbawa
Sa seksyong ito, ibi-benchmark natin ang tatlong pamamaraan ng kompilasyon sa mga Hamiltonian circuit na may wala pang 20 qubit. Ang mga circuit na ito ay mabilis na nag-ta-transpile at nagbibigay ng malinaw na pagtingin kung paano pinapamahalaan ng bawat pamamaraan ang mga circuit ng katamtamang kumplikasyon.
Hakbang 1: I-map ang mga classical input sa isang quantum problem
Ang bawat Hamiltonian ay naka-encode bilang isang PauliEvolutionGate circuit. Ang mga circuit ay itinayo na sa seksyon ng setup mula sa Hamlib benchmark data.
Hakbang 2: I-optimize ang problema para sa quantum hardware execution
Ina-transpile natin ang lahat ng maliit na circuit gamit ang bawat isa sa tatlong pass manager, pagkatapos ay kinokolekta ang mga sukatan.
results_small = []
tqc_sabre_small = capture_transpilation_metrics(
results_small, pm_sabre, qc_small, "SABRE"
)
tqc_ai_small = capture_transpilation_metrics(
results_small, pm_ai, qc_small, "AI"
)
tqc_rustiq_small = capture_transpilation_metrics(
results_small, pm_rustiq, qc_small, "Rustiq"
)
[SABRE] Circuit 0 (all-vib-bh): 2Q depth=3, size=30, time=2.09s
[SABRE] Circuit 1 (all-vib-c2h): 2Q depth=18, size=111, time=0.01s
[SABRE] Circuit 2 (all-vib-o3): 2Q depth=6, size=58, time=0.00s
[SABRE] Circuit 3 (all-vib-c2h): 2Q depth=2, size=37, time=0.01s
[SABRE] Circuit 4 (graph-gnp_k-2): 2Q depth=24, size=126, time=0.01s
[SABRE] Circuit 5 (LiH): 2Q depth=66, size=285, time=0.01s
[SABRE] Circuit 6 (all-vib-fccf): 2Q depth=66, size=339, time=0.01s
[SABRE] Circuit 7 (all-vib-ch2): 2Q depth=88, size=413, time=0.01s
[SABRE] Circuit 8 (all-vib-f2): 2Q depth=180, size=1000, time=0.02s
[SABRE] Circuit 9 (all-vib-bhf2): 2Q depth=18, size=223, time=0.03s
[SABRE] Circuit 10 (graph-gnp_k-4): 2Q depth=122, size=675, time=0.02s
[SABRE] Circuit 11 (Be2): 2Q depth=343, size=1628, time=0.03s
[SABRE] Circuit 12 (all-vib-fccf): 2Q depth=14, size=134, time=0.00s
[SABRE] Circuit 13 (uf20-ham): 2Q depth=50, size=341, time=0.01s
[SABRE] Circuit 14 (TSP_Ncity-4): 2Q depth=118, size=615, time=0.01s
[SABRE] Circuit 15 (graph-complete_bipart): 2Q depth=232, size=1420, time=0.03s
[SABRE] Circuit 16 (all-vib-cyclo_propene): 2Q depth=18, size=354, time=0.93s
[SABRE] Circuit 17 (all-vib-hno): 2Q depth=6, size=174, time=0.14s
[SABRE] Circuit 18 (all-vib-fccf): 2Q depth=30, size=286, time=0.01s
[SABRE] Circuit 19 (tfim): 2Q depth=31, size=232, time=0.03s
[AI] Circuit 0 (all-vib-bh): 2Q depth=3, size=30, time=0.01s
Fetching 4 files: 0%| | 0/4 [00:00<?, ?it/s]
[AI] Circuit 1 (all-vib-c2h): 2Q depth=18, size=101, time=0.18s
[AI] Circuit 2 (all-vib-o3): 2Q depth=6, size=58, time=0.01s
[AI] Circuit 3 (all-vib-c2h): 2Q depth=2, size=37, time=0.01s
[AI] Circuit 4 (graph-gnp_k-2): 2Q depth=24, size=133, time=0.07s
[AI] Circuit 5 (LiH): 2Q depth=62, size=267, time=8.00s
[AI] Circuit 6 (all-vib-fccf): 2Q depth=65, size=300, time=0.18s
[AI] Circuit 7 (all-vib-ch2): 2Q depth=79, size=353, time=0.16s
[AI] Circuit 8 (all-vib-f2): 2Q depth=176, size=998, time=0.43s
[AI] Circuit 9 (all-vib-bhf2): 2Q depth=18, size=194, time=0.11s
[AI] Circuit 10 (graph-gnp_k-4): 2Q depth=114, size=668, time=0.18s
[AI] Circuit 11 (Be2): 2Q depth=292, size=1382, time=0.88s
[AI] Circuit 12 (all-vib-fccf): 2Q depth=14, size=134, time=0.01s
[AI] Circuit 13 (uf20-ham): 2Q depth=40, size=330, time=0.16s
[AI] Circuit 14 (TSP_Ncity-4): 2Q depth=96, size=600, time=0.29s
[AI] Circuit 15 (graph-complete_bipart): 2Q depth=231, size=1531, time=0.46s
[AI] Circuit 16 (all-vib-cyclo_propene): 2Q depth=18, size=309, time=0.25s
[AI] Circuit 17 (all-vib-hno): 2Q depth=10, size=198, time=0.15s
[AI] Circuit 18 (all-vib-fccf): 2Q depth=34, size=402, time=0.02s
[AI] Circuit 19 (tfim): 2Q depth=44, size=311, time=0.15s
[Rustiq] Circuit 0 (all-vib-bh): 2Q depth=3, size=30, time=0.01s
[Rustiq] Circuit 1 (all-vib-c2h): 2Q depth=13, size=69, time=0.00s
[Rustiq] Circuit 2 (all-vib-o3): 2Q depth=13, size=82, time=0.01s
[Rustiq] Circuit 3 (all-vib-c2h): 2Q depth=2, size=40, time=0.01s
[Rustiq] Circuit 4 (graph-gnp_k-2): 2Q depth=31, size=132, time=0.01s
[Rustiq] Circuit 5 (LiH): 2Q depth=59, size=285, time=0.01s
[Rustiq] Circuit 6 (all-vib-fccf): 2Q depth=34, size=193, time=0.00s
[Rustiq] Circuit 7 (all-vib-ch2): 2Q depth=49, size=302, time=0.01s
[Rustiq] Circuit 8 (all-vib-f2): 2Q depth=141, size=807, time=0.02s
[Rustiq] Circuit 9 (all-vib-bhf2): 2Q depth=13, size=146, time=0.02s
[Rustiq] Circuit 10 (graph-gnp_k-4): 2Q depth=129, size=683, time=0.02s
[Rustiq] Circuit 11 (Be2): 2Q depth=220, size=1101, time=0.02s
[Rustiq] Circuit 12 (all-vib-fccf): 2Q depth=53, size=333, time=0.01s
[Rustiq] Circuit 13 (uf20-ham): 2Q depth=63, size=425, time=0.01s
[Rustiq] Circuit 14 (TSP_Ncity-4): 2Q depth=123, size=767, time=0.02s
[Rustiq] Circuit 15 (graph-complete_bipart): 2Q depth=309, size=2107, time=0.05s
[Rustiq] Circuit 16 (all-vib-cyclo_propene): 2Q depth=16, size=283, time=0.32s
[Rustiq] Circuit 17 (all-vib-hno): 2Q depth=19, size=291, time=0.32s
[Rustiq] Circuit 18 (all-vib-fccf): 2Q depth=44, size=546, time=0.02s
[Rustiq] Circuit 19 (tfim): 2Q depth=24, size=416, time=0.01s
Ang talahanayan sa ibaba ay nagbubuod ng average at standard deviation ng bawat sukatan sa lahat ng maliit na circuit, kasama ang porsyentong pagpapabuti kumpara sa SABRE. Dahil nagkakaiba-iba ang laki ng circuit, ang standard deviation ay nagbibigay ng mahalagang konteksto para sa interpretasyon ng mga average.
print_summary_table(results_small)
Mean +/- std per compilation method
Method 2Q Depth Gate Count Runtime (s)
------------------------------------------------------------------------------
SABRE 71.8 +/- 89.6 424.1 +/- 446.0 0.2 +/- 0.5
AI 67.3 +/- 80.2 416.8 +/- 426.7 0.6 +/- 1.8
Rustiq 67.9 +/- 80.0 451.9 +/- 484.7 0.0 +/- 0.1
Mean % improvement vs SABRE (positive = better than SABRE)
Method 2Q Depth Gate Count Runtime (s)
------------------------------------------------------------------------------
AI -2.1% +/- 19.8% -0.6% +/- 14.7% -5635.1% +/- 20725.2%
Rustiq -25.3% +/- 85.4% -16.3% +/- 50.4% -7.0% +/- 60.6%
Ang per-circuit na talahanayan ay nagpapakita kung paano nagkukumpara ang bawat pamamaraan sa mga indibidwal na circuit. Ang pinakamahusay na halaga para sa bawat sukatan ay minarkahan ng asterisk. Pansinin na para sa mga pinakasimpleng circuit, ang lahat ng tatlong pamamaraan ay kadalasang nagco-converge sa parehong resulta.
print_per_circuit_comparison(results_small, num_rows=8)
2Q Depth (first 8 circuits by qubit count); * = best
Idx Circuit Q SABRE AI Rustiq
----------------------------------------------------
0 all-vib-bh 2 3* 3* 3*
1 all-vib-c2h 3 18 18 13*
2 all-vib-o3 4 6* 6* 13
3 all-vib-c2h 4 2* 2* 2*
4 graph-gnp_k-2 4 24* 24* 31
5 LiH 4 66 62 59*
6 all-vib-fccf 4 66 65 34*
7 all-vib-ch2 4 88 79 49*
Gate Count (first 8 circuits by qubit count); * = best
Idx Circuit Q SABRE AI Rustiq
----------------------------------------------------
0 all-vib-bh 2 30* 30* 30*
1 all-vib-c2h 3 111 101 69*
2 all-vib-o3 4 58* 58* 82
3 all-vib-c2h 4 37* 37* 40
4 graph-gnp_k-2 4 126* 133 132
5 LiH 4 285 267* 285
6 all-vib-fccf 4 339 300 193*
7 all-vib-ch2 4 413 353 302*
Runtime (s) (first 8 circuits by qubit count); * = best
Idx Circuit Q SABRE AI Rustiq
----------------------------------------------------
0 all-vib-bh 2 2.09 0.01 0.01*
1 all-vib-c2h 3 0.01 0.18 0.00*
2 all-vib-o3 4 0.00* 0.01 0.01
3 all-vib-c2h 4 0.01 0.01 0.01*
4 graph-gnp_k-2 4 0.01* 0.07 0.01
5 LiH 4 0.01* 8.00 0.01
6 all-vib-fccf 4 0.01 0.18 0.00*
7 all-vib-ch2 4 0.01 0.16 0.01*
I-visualize ang mga resulta
Ang mga plot sa ibaba ay nagkukumpara ng tatlong pamamaraan sa bawat sukatan sa bawat circuit na batayan. Ang mga circuit ay nakaayos ayon sa bilang ng qubit at may label na index sa x-axis, dahil ang maraming circuit ay maaaring magkaroon ng parehong bilang ng qubit.
def plot_transpilation_comparison(results, title_prefix):
"""
Create a three-panel figure comparing compilation methods on
two-qubit depth, circuit size, and runtime.
Circuits are sorted by qubit count and plotted by circuit index.
"""
methods = _method_order(results)
palette = {"SABRE": "#1f77b4", "AI": "#ff7f0e", "Rustiq": "#2ca02c"}
markers = {"SABRE": "o", "AI": "^", "Rustiq": "s"}
# Order circuits by qubit count (then index) and map to plot positions
ref = sorted(
[r for r in results if r["method"] == methods[0]],
key=lambda r: (r["num_qubits"], r["qc_index"]),
)
pos_map = {r["qc_index"]: pos for pos, r in enumerate(ref)}
tick_positions = [pos_map[r["qc_index"]] for r in ref]
tick_labels = [
f"{pos_map[r['qc_index']]} ({r['num_qubits']}q)" for r in ref
]
metrics = [
("two_qubit_depth", "Two-Qubit Depth"),
("size", "Total Gate Count (Circuit Size)"),
("runtime", "Transpilation Runtime (s)"),
]
fig, axes = plt.subplots(1, 3, figsize=(20, 5.5))
fig.suptitle(title_prefix, fontsize=15, fontweight="bold", y=1.02)
for ax, (metric, ylabel) in zip(axes, metrics):
for method in methods:
subset = sorted(
[r for r in results if r["method"] == method],
key=lambda r: pos_map[r["qc_index"]],
)
ax.plot(
[pos_map[r["qc_index"]] for r in subset],
[r[metric] for r in subset],
marker=markers.get(method, "o"),
label=method,
color=palette.get(method, None),
linewidth=1.5,
markersize=6,
alpha=0.85,
)
ax.set_xlabel("Circuit Index (num qubits)", fontsize=11)
ax.set_ylabel(ylabel, fontsize=11)
ax.legend(frameon=True, fontsize=9)
ax.grid(True, linestyle="--", alpha=0.4)
step = max(1, len(tick_positions) // 15)
ax.set_xticks(tick_positions[::step])
ax.set_xticklabels(
[tick_labels[i] for i in range(0, len(tick_labels), step)],
fontsize=7,
rotation=45,
ha="right",
)
plt.tight_layout()
plt.show()
def plot_pct_improvement_vs_sabre(results, title_prefix):
"""
Plot the per-circuit percent improvement of each non-SABRE method
relative to SABRE, for each metric. A positive value means the
method improved on SABRE; negative means SABRE was better.
"""
metrics = [
("two_qubit_depth", "2Q Depth"),
("size", "Gate Count"),
("runtime", "Runtime"),
]
palette = {"AI": "#ff7f0e", "Rustiq": "#2ca02c"}
markers = {"AI": "^", "Rustiq": "s"}
methods = _method_order(results)
sabre = sorted(
[r for r in results if r["method"] == "SABRE"],
key=lambda r: (r["num_qubits"], r["qc_index"]),
)
other_methods = [m for m in methods if m != "SABRE"]
tick_positions = list(range(len(sabre)))
tick_labels = [
f"{i} ({sabre[i]['num_qubits']}q)" for i in range(len(sabre))
]
fig, axes = plt.subplots(1, 3, figsize=(20, 5.5))
fig.suptitle(
f"{title_prefix}: % Improvement over SABRE",
fontsize=15,
fontweight="bold",
y=1.02,
)
for ax, (metric, label) in zip(axes, metrics):
ax.axhline(0, color="#1f77b4", linewidth=2, label="SABRE (baseline)")
for method in other_methods:
data = sorted(
[r for r in results if r["method"] == method],
key=lambda r: (r["num_qubits"], r["qc_index"]),
)
pct = [
(sabre[i][metric] - data[i][metric]) / sabre[i][metric] * 100
for i in range(len(sabre))
]
ax.plot(
tick_positions,
pct,
marker=markers.get(method, "o"),
label=method,
color=palette.get(method, None),
linewidth=1.5,
markersize=6,
alpha=0.85,
)
ax.set_xlabel("Circuit Index (num qubits)", fontsize=11)
ax.set_ylabel(f"% Improvement ({label})", fontsize=11)
ax.legend(frameon=True, fontsize=9)
ax.grid(True, linestyle="--", alpha=0.4)
step = max(1, len(tick_positions) // 15)
ax.set_xticks(tick_positions[::step])
ax.set_xticklabels(
[tick_labels[i] for i in range(0, len(tick_labels), step)],
fontsize=7,
rotation=45,
ha="right",
)
ylims = ax.get_ylim()
ax.axhspan(0, max(ylims[1], 1), alpha=0.04, color="green")
ax.axhspan(min(ylims[0], -1), 0, alpha=0.04, color="red")
plt.tight_layout()
plt.show()
plot_transpilation_comparison(
results_small,
"Small-Scale Hamiltonian Circuits: Compilation Comparison",
)
plot_pct_improvement_vs_sabre(
results_small,
"Small-Scale Hamiltonian Circuits",
)
Sa sukat na ito, ang lahat ng tatlong pass manager ay gumaganap nang maayos, at ang kanilang mga average na resulta ay malapit sa isa't isa. Ito ay dahil ang mga maliit na circuit ay nag-iiwan ng limitadong espasyo para sa karagdagang optimization, kaya ang mga pamamaraan ay may tendensiyang magco-converge sa mga katulad na solusyon.
Sa halimbawang ito, ang Rustiq ay gumagawa ng pinaka-variable na resulta, na may pinakamalaking outlier sa parehong two-qubit depth at gate count. Bagama't ang variability na ito ay nangangahulugang paminsan-minsan ay nahahuli ito, nangangahulugan din ito na paminsan-minsan ay nakakakita ang Rustiq ng mas mahusay na solusyon kaysa sa dalawang iba pang pamamaraan. Ang AI transpiler ay mas matatag sa mga resulta nito kumpara sa SABRE at Rustiq, na malapit na sumusunod sa karamihan ng mga circuit nang wala masyadong mga outlier.
Para sa runtime, ang SABRE at Rustiq ay parehong mabilis, habang ang AI-powered transpiler ay kapansin-pansing mas mabagal sa ilang partikular na circuit.
Pinakamahusay na pamamaraan ayon sa sukatan
Ang chart sa ibaba ay nagpapakita kung gaano kadalas nakamit ng bawat pamamaraan ang pinakamahusay (pinakamababa) na halaga para sa bawat sukatan. Posible ang mga tie: para sa mga mas simpleng circuit, maraming pamamaraan ang maaaring maabot ang parehong optimal na two-qubit depth o gate count. Kapag nagkaroon ng tie, lahat ng mga pamamaraang nakatali ay nakakakuha ng kredit, kaya ang mga porsyento para sa isang ibinigay na sukatan ay maaaring magsuma ng higit sa 100%.
def plot_best_method_bars(results, metrics_list=None):
"""
Plot a grouped bar chart showing the percentage of circuits
where each method achieved the best (lowest) value for each metric.
Ties are counted for all tied methods, so percentages per metric
can sum to more than 100%.
"""
if metrics_list is None:
metrics_list = ["two_qubit_depth", "size", "runtime"]
labels = {
"two_qubit_depth": "2Q Depth",
"size": "Gate Count",
"runtime": "Runtime",
}
methods = _method_order(results)
palette = {"SABRE": "#1f77b4", "AI": "#ff7f0e", "Rustiq": "#2ca02c"}
by_index = {}
for r in results:
by_index.setdefault(r["qc_index"], []).append(r)
n_circuits = len(by_index)
win_data = {m: [] for m in methods}
tie_counts = []
metric_labels = []
for metric in metrics_list:
metric_labels.append(
labels.get(metric, metric.replace("_", " ").title())
)
counts = Counter()
ties = 0
for group in by_index.values():
min_val = min(r[metric] for r in group)
best = [r["method"] for r in group if r[metric] == min_val]
if len(best) > 1:
ties += 1
counts.update(best)
tie_counts.append(ties)
for m in methods:
win_data[m].append(counts.get(m, 0) / n_circuits * 100)
x = np.arange(len(metric_labels))
width = 0.22
fig, ax = plt.subplots(figsize=(8, 5))
for i, method in enumerate(methods):
bars = ax.bar(
x + i * width,
win_data[method],
width,
label=method,
color=palette.get(method, None),
edgecolor="black",
linewidth=0.5,
)
for bar in bars:
height = bar.get_height()
if height > 0:
ax.text(
bar.get_x() + bar.get_width() / 2,
height + 1.5,
f"{height:.0f}%",
ha="center",
va="bottom",
fontsize=9,
)
# Annotate tie counts below each metric label
for j, ties in enumerate(tie_counts):
if ties > 0:
ax.text(
x[j] + width,
-8,
f"({ties} tie{'s' if ties != 1 else ''})",
ha="center",
va="top",
fontsize=8,
color="gray",
)
ax.set_xticks(x + width)
ax.set_xticklabels(metric_labels, fontsize=11)
ax.set_ylabel("Circuits with best value (%)", fontsize=11)
ax.set_title(
"Best-Performing Method by Metric (ties counted for all tied methods)",
fontsize=12,
fontweight="bold",
)
ax.legend(frameon=True, fontsize=10)
ax.set_ylim(-12, 120)
ax.yaxis.set_major_formatter(ticker.PercentFormatter())
ax.grid(axis="y", linestyle="--", alpha=0.4)
plt.tight_layout()
plt.show()
plot_best_method_bars(results_small)
Sa halimbawang ito, ang tatlong pamamaraan ay gumaganap nang magkakatulad sa mga maliit na circuit. Sa two-qubit depth at gate count, ang bahagi ng mga circuit kung saan pinakamahusay ang bawat pamamaraan ay malapit (humigit-kumulang 35–55%), at maraming circuit ang nagtatapos sa mga tie dahil ang mga pinakasimpleng circuit ay kadalasang may iisang optimal na solusyon na natutuklasan ng maraming pamamaraan. Ang pinakamalinaw na pagkakaiba ay runtime: ang SABRE at Rustiq ay bawat isa ay pinakamabilis sa humigit-kumulang kalahati ng mga circuit, habang ang AI-powered transpiler ay bihirang ang pinakamabilis. Kung isasaalang-alang ang lahat ng tatlong sukatan nang sama-sama, ang Rustiq ay may kaunting kalamangan sa kabuuan ito ay ang pinaka-madalas na nagwawagi sa two-qubit depth at nananatiling mapagkumpitensya sa gate count at runtime.
Hakbang 3: Isagawa gamit ang mga Qiskit primitive
Upang masuri kung paano nakakaapekto ang kalidad ng transpilation sa execution sa ilalim ng ingay, gumagamit tayo ng mirror circuit na teknik. Para sa bawat transpiled circuit , idinadagdag natin ang kabaligtaran nito upang ang pinagsama-samang circuit ay teoretically ang identity. Simula sa estado ng , ang isang perpekto (walang ingay) na execution ay magbabalik ng all-zeros bitstring na may probability 1.
Sa katotohanan, nagtitipon ang mga gate error sa buong circuit, kaya ang probability ng pagbawi ng ay bumababa. Ang isang pamamaraan ng kompilasyon na gumagawa ng mas mababaw na circuit na may mas kaunting gate ay mag-iipon ng mas kaunting ingay.
Ang mirror circuit approach ay simpleng kapansin-pansin at nag-scale sa anumang laki ng circuit, dahil ang inaasahang output ay laging at walang classical simulation ng ideal na estado ang kinakailangan. Gayunpaman, tandaan ang mga sumusunod na caveat: ang mirror circuit ay isang proxy para sa aktwal na circuit (hindi ang circuit mismo), nino-double nito ang gate count (na nagpapalaki ng epekto ng ingay), at maaari nitong maunas ang ilang partikular na error kapag ang ingay ay nagkakansela nang simetriko sa kabila ng mirror boundary.
Pinipili natin ang circuit index 6 mula sa maliit na set at pinapatakbo ang mga mirror circuit sa isang Aer simulator na may simpleng depolarizing noise model.
# Select circuit index 6 from the small-scale transpiled circuits
test_idx = 6
test_circuit = qc_small[test_idx]
print(f"Test circuit: {test_circuit.name}, {test_circuit.num_qubits} qubits")
# Get the transpiled versions
tqc_methods_small = {
"SABRE": tqc_sabre_small[test_idx],
"AI": tqc_ai_small[test_idx],
"Rustiq": tqc_rustiq_small[test_idx],
}
# Show transpilation metrics for this circuit
print(f"\nTranspilation metrics for circuit index {test_idx}:")
for method, tqc in tqc_methods_small.items():
depth_2q = tqc.depth(lambda x: x.operation.num_qubits == 2)
size = tqc.size()
print(f" {method:8s} 2Q depth={depth_2q:5d} size={size:6d}")
Test circuit: all-vib-fccf, 4 qubits
Transpilation metrics for circuit index 6:
SABRE 2Q depth= 66 size= 339
AI 2Q depth= 65 size= 300
Rustiq 2Q depth= 34 size= 193
Itayo ang mga mirror circuit (idagdag ang ), i-remap sa magkakasunod na qubit index upang ang simulator ay humawak lamang ng mga aktibong qubit, at patakbuhin sa isang maingay na Aer simulator.
def remap_to_contiguous(tqc):
"""Remap a transpiled circuit to use contiguous qubit indices.
Transpiled circuits target specific physical qubits (e.g., qubit 45, 67)
on a large backend. This remaps them to 0, 1, 2, ... so Aer only
simulates the active qubits.
"""
active = sorted(
{tqc.find_bit(q).index for inst in tqc.data for q in inst.qubits}
)
qubit_map = {old: new for new, old in enumerate(active)}
new_qc = QuantumCircuit(len(active))
for inst in tqc.data:
old_indices = [tqc.find_bit(q).index for q in inst.qubits]
new_qc.append(inst.operation, [qubit_map[i] for i in old_indices])
return new_qc
def build_mirror_circuit(tqc):
"""Build a mirror circuit: U followed by U-dagger, with measurements.
The combined circuit U-dagger @ U should be the identity, so measuring
all zeros indicates a noise-free execution.
"""
tqc_compact = remap_to_contiguous(tqc)
mirror = tqc_compact.compose(tqc_compact.inverse())
mirror.measure_all()
return mirror
# Build a simple depolarizing noise model
noise_model = NoiseModel()
noise_model.add_all_qubit_quantum_error(
depolarizing_error(0.001, 1),
["sx", "x", "rz"], # ~0.1% per 1Q gate
)
noise_model.add_all_qubit_quantum_error(
depolarizing_error(0.01, 2),
["cx", "ecr"], # ~1% per 2Q gate
)
aer_sim = AerSimulator(noise_model=noise_model)
shots = 10000
fidelities = {}
for method, tqc in tqc_methods_small.items():
mirror = build_mirror_circuit(tqc)
sampler = SamplerV2(mode=aer_sim)
job = sampler.run([mirror], shots=shots)
result = job.result()
counts = result[0].data.meas.get_counts()
# Fidelity = fraction of all-zeros (error-free) outcomes
n_qubits = mirror.num_qubits - mirror.num_clbits # active qubits
all_zeros = "0" * mirror.num_qubits
fidelity = counts.get(all_zeros, 0) / shots
fidelities[method] = fidelity
print(
f"{method:8s} P(|00...0>) = {fidelity:.4f} ({counts.get(all_zeros, 0)}/{shots})"
)
SABRE P(|00...0>) = 0.7796 (7796/10000)
AI P(|00...0>) = 0.8073 (8073/10000)
Rustiq P(|00...0>) = 0.8923 (8923/10000)
def plot_mirror_results(tqc_methods, fidelities, circuit_name):
"""
Plot a three-panel comparison: fidelity, 2Q depth,
and gate count for each compilation method.
"""
methods = list(tqc_methods.keys())
palette = {"SABRE": "#1f77b4", "AI": "#ff7f0e", "Rustiq": "#2ca02c"}
colors = [palette.get(m, "gray") for m in methods]
fidelity_vals = [fidelities[m] for m in methods]
depth_vals = [
tqc_methods[m].depth(lambda x: x.operation.num_qubits == 2)
for m in methods
]
size_vals = [tqc_methods[m].size() for m in methods]
fig, axes = plt.subplots(1, 3, figsize=(16, 5))
fig.suptitle(
f"Mirror Circuit Results: {circuit_name}",
fontsize=14,
fontweight="bold",
y=1.02,
)
def _annotate_bars(ax, bars, values, fmt="{}"):
ymax = ax.get_ylim()[1]
for bar, val in zip(bars, values):
label = fmt.format(val)
y = val + ymax * 0.03
ax.text(
bar.get_x() + bar.get_width() / 2,
y,
label,
ha="center",
va="bottom",
fontsize=10,
fontweight="bold",
)
# Panel 1: Survival Probability
bars = axes[0].bar(
methods, fidelity_vals, color=colors, edgecolor="black", linewidth=0.5
)
axes[0].set_ylabel("Fidelity P(|00...0>)", fontsize=11)
axes[0].set_title("Fidelity (higher is better)", fontsize=12)
axes[0].set_ylim(
0, max(fidelity_vals) * 1.18 if max(fidelity_vals) > 0 else 1.0
)
axes[0].grid(axis="y", linestyle="--", alpha=0.4)
_annotate_bars(axes[0], bars, fidelity_vals, fmt="{:.4f}")
# Panel 2: Two-Qubit Depth
bars = axes[1].bar(
methods, depth_vals, color=colors, edgecolor="black", linewidth=0.5
)
axes[1].set_ylabel("Two-Qubit Depth", fontsize=11)
axes[1].set_title("2Q Depth (lower is better)", fontsize=12)
axes[1].set_ylim(0, max(depth_vals) * 1.18)
axes[1].grid(axis="y", linestyle="--", alpha=0.4)
_annotate_bars(axes[1], bars, depth_vals)
# Panel 3: Gate Count
bars = axes[2].bar(
methods, size_vals, color=colors, edgecolor="black", linewidth=0.5
)
axes[2].set_ylabel("Total Gate Count", fontsize=11)
axes[2].set_title("Gate Count (lower is better)", fontsize=12)
axes[2].set_ylim(0, max(size_vals) * 1.18)
axes[2].grid(axis="y", linestyle="--", alpha=0.4)
_annotate_bars(axes[2], bars, size_vals)
plt.tight_layout()
plt.show()
plot_mirror_results(tqc_methods_small, fidelities, test_circuit.name)
Mga obserbasyon
Ang pamamaraan na may pinakamababang two-qubit depth at pinakamaunting gate ay nakakamit ng pinakamataas na fidelity, na naaayon sa inaasahan na ang mas maikling circuit ay nag-iipon ng mas kaunting ingay. Kahit ang katamtamang pagkakaiba sa depth at gate count ay nagsasalin sa nasusukat na pagkakaiba sa fidelity sa ilalim ng depolarizing noise model.
Tandaan na ang mga resultang ito ay para sa isang solong circuit. Ang relatibong ranggo ng mga pamamaraan ay maaaring magbago mula sa circuit patungong circuit depende sa istruktura ng Hamiltonian.
Malalaking antas na halimbawa sa hardware
Sa seksyong ito, ibi-benchmark natin ang parehong tatlong pamamaraan ng kompilasyon sa mga Hamiltonian circuit na may 20 o higit pang qubit. Ang mga circuit na ito ay mas kinatawan ng mga praktikal na Hamiltonian simulation workload at sinusubukan kung paano nag-scale ang bawat pamamaraan sa mga tuntunin ng kalidad ng circuit at oras ng kompilasyon.
Mga Hakbang 1-4 na pinagsama
Ang workflow ay sumusunod sa parehong istruktura tulad ng maliit na halimbawa. Ina-transpile natin ang lahat ng malalaking circuit sa bawat pamamaraan, kinokolekta ang mga sukatan, at nagsusumite ng mirror circuit sa tunay na quantum hardware.
results_large = []
tqc_sabre_large = capture_transpilation_metrics(
results_large, pm_sabre, qc_large, "SABRE"
)
tqc_ai_large = capture_transpilation_metrics(
results_large, pm_ai, qc_large, "AI"
)
tqc_rustiq_large = capture_transpilation_metrics(
results_large, pm_rustiq, qc_large, "Rustiq"
)
[SABRE] Circuit 0 (all-vib-hc3h2cn): 2Q depth=2, size=258, time=0.16s
[SABRE] Circuit 1 (ham-graph-gnp_k-5): 2Q depth=345, size=4036, time=0.08s
[SABRE] Circuit 2 (TSP_Ncity-5): 2Q depth=187, size=2045, time=0.04s
[SABRE] Circuit 3 (tfim): 2Q depth=100, size=489, time=0.21s
[SABRE] Circuit 4 (all-vib-h2co): 2Q depth=30, size=570, time=0.18s
[SABRE] Circuit 5 (uuf100-ham): 2Q depth=414, size=4779, time=0.09s
[SABRE] Circuit 6 (uuf100-ham): 2Q depth=523, size=5667, time=0.11s
[SABRE] Circuit 7 (graph-gnp_k-4): 2Q depth=3028, size=24885, time=0.39s
[SABRE] Circuit 8 (uf100-ham): 2Q depth=700, size=8271, time=0.15s
[SABRE] Circuit 9 (uf100-ham): 2Q depth=698, size=8957, time=0.15s
[SABRE] Circuit 10 (TSP_Ncity-7): 2Q depth=432, size=6353, time=0.12s
[SABRE] Circuit 11 (all-vib-cyclo_propene): 2Q depth=30, size=1144, time=0.20s
[SABRE] Circuit 12 (TSP_Ncity-8): 2Q depth=704, size=10287, time=0.18s
[SABRE] Circuit 13 (uf100-ham): 2Q depth=2454, size=30195, time=0.46s
[SABRE] Circuit 14 (tfim): 2Q depth=245, size=3670, time=0.08s
[SABRE] Circuit 15 (flat100-ham): 2Q depth=154, size=3836, time=0.12s
[SABRE] Circuit 16 (graph-regular_reg-4): 2Q depth=863, size=14063, time=0.22s
[SABRE] Circuit 17 (tfim): 2Q depth=581, size=8810, time=0.15s
[SABRE] Circuit 18 (FH_D-1): 2Q depth=1704, size=9528, time=0.35s
[SABRE] Circuit 19 (TSP_Ncity-10): 2Q depth=1091, size=22041, time=0.38s
[SABRE] Circuit 20 (TSP_Ncity-10): 2Q depth=1091, size=22005, time=0.38s
[SABRE] Circuit 21 (ham-unary-color02-queen13_13_k-4): 2Q depth=224, size=8321, time=0.17s
[AI] Circuit 0 (all-vib-hc3h2cn): 2Q depth=2, size=258, time=0.17s
[AI] Circuit 1 (ham-graph-gnp_k-5): 2Q depth=323, size=4418, time=3.13s
[AI] Circuit 2 (TSP_Ncity-5): 2Q depth=161, size=2229, time=1.47s
[AI] Circuit 3 (tfim): 2Q depth=20, size=402, time=0.34s
[AI] Circuit 4 (all-vib-h2co): 2Q depth=38, size=661, time=0.19s
[AI] Circuit 5 (uuf100-ham): 2Q depth=391, size=5130, time=3.27s
[AI] Circuit 6 (uuf100-ham): 2Q depth=463, size=6095, time=4.23s
[AI] Circuit 7 (graph-gnp_k-4): 2Q depth=3207, size=25641, time=15.15s
[AI] Circuit 8 (uf100-ham): 2Q depth=637, size=8267, time=5.87s
[AI] Circuit 9 (uf100-ham): 2Q depth=632, size=9330, time=7.29s
[AI] Circuit 10 (TSP_Ncity-7): 2Q depth=452, size=7418, time=6.02s
[AI] Circuit 11 (all-vib-cyclo_propene): 2Q depth=38, size=1323, time=0.27s
[AI] Circuit 12 (TSP_Ncity-8): 2Q depth=609, size=11131, time=10.07s
[AI] Circuit 13 (uf100-ham): 2Q depth=2251, size=31128, time=38.77s
[AI] Circuit 14 (tfim): 2Q depth=165, size=3460, time=1.64s
[AI] Circuit 15 (flat100-ham): 2Q depth=91, size=3497, time=2.49s
[AI] Circuit 16 (graph-regular_reg-4): 2Q depth=664, size=15256, time=12.35s
[AI] Circuit 17 (tfim): 2Q depth=583, size=9157, time=6.28s
[AI] Circuit 18 (FH_D-1): 2Q depth=1193, size=7754, time=4.54s
[AI] Circuit 19 (TSP_Ncity-10): 2Q depth=1134, size=22577, time=25.64s
[AI] Circuit 20 (TSP_Ncity-10): 2Q depth=1172, size=23851, time=28.97s
[AI] Circuit 21 (ham-unary-color02-queen13_13_k-4): 2Q depth=219, size=8600, time=8.85s
[Rustiq] Circuit 0 (all-vib-hc3h2cn): 2Q depth=2, size=257, time=0.16s
[Rustiq] Circuit 1 (ham-graph-gnp_k-5): 2Q depth=640, size=5831, time=0.13s
[Rustiq] Circuit 2 (TSP_Ncity-5): 2Q depth=408, size=3985, time=0.08s
[Rustiq] Circuit 3 (tfim): 2Q depth=31, size=688, time=0.07s
[Rustiq] Circuit 4 (all-vib-h2co): 2Q depth=65, size=1058, time=2.91s
[Rustiq] Circuit 5 (uuf100-ham): 2Q depth=633, size=6757, time=0.14s
[Rustiq] Circuit 6 (uuf100-ham): 2Q depth=795, size=8495, time=0.17s
[Rustiq] Circuit 7 (graph-gnp_k-4): 2Q depth=13768, size=139793, time=2.92s
[Rustiq] Circuit 8 (uf100-ham): 2Q depth=1099, size=11878, time=0.25s
[Rustiq] Circuit 9 (uf100-ham): 2Q depth=911, size=11111, time=0.22s
[Rustiq] Circuit 10 (TSP_Ncity-7): 2Q depth=1183, size=13197, time=0.27s
[Rustiq] Circuit 11 (all-vib-cyclo_propene): 2Q depth=67, size=2491, time=13.56s
[Rustiq] Circuit 12 (TSP_Ncity-8): 2Q depth=1615, size=21358, time=0.48s
[Rustiq] Circuit 13 (uf100-ham): 2Q depth=2920, size=40465, time=0.91s
[Rustiq] Circuit 14 (tfim): 2Q depth=489, size=6552, time=0.15s
[Rustiq] Circuit 15 (flat100-ham): 2Q depth=378, size=5906, time=0.14s
[Rustiq] Circuit 16 (graph-regular_reg-4): 2Q depth=12163, size=168679, time=2.94s
[Rustiq] Circuit 17 (tfim): 2Q depth=1208, size=17042, time=0.36s
[Rustiq] Circuit 18 (FH_D-1): 2Q depth=1061, size=24000, time=0.47s
[Rustiq] Circuit 19 (TSP_Ncity-10): 2Q depth=2565, size=41340, time=1.38s
[Rustiq] Circuit 20 (TSP_Ncity-10): 2Q depth=2565, size=41275, time=1.38s
[Rustiq] Circuit 21 (ham-unary-color02-queen13_13_k-4): 2Q depth=808, size=17548, time=0.42s
print_summary_table(results_large)
Mean +/- std per compilation method
Method 2Q Depth Gate Count Runtime (s)
------------------------------------------------------------------------------
SABRE 709.1 +/- 783.8 9,100.5 +/- 8,493.1 0.2 +/- 0.1
AI 656.6 +/- 777.5 9,435.6 +/- 8,853.0 8.5 +/- 10.2
Rustiq 2,062.5 +/- 3,631.1 26,804.8 +/- 43,403.1 1.3 +/- 2.9
Mean % improvement vs SABRE (positive = better than SABRE)
Method 2Q Depth Gate Count Runtime (s)
------------------------------------------------------------------------------
AI +9.6% +/- 22.8% -3.4% +/- 9.4% -3620.0% +/- 2405.5%
Rustiq -154.5% +/- 273.9% -137.1% +/- 233.2% -527.0% +/- 1405.5%
print_per_circuit_comparison(results_large, num_rows=8)
2Q Depth (first 8 circuits by qubit count); * = best
Idx Circuit Q SABRE AI Rustiq
----------------------------------------------------
0 all-vib-hc3h2cn 24 2* 2* 2*
1 ham-graph-gnp_k- 24 345 323* 640
2 TSP_Ncity-5 25 187 161* 408
3 tfim 26 100 20* 31
4 all-vib-h2co 32 30* 38 65
5 uuf100-ham 40 414 391* 633
6 uuf100-ham 40 523 463* 795
7 graph-gnp_k-4 40 3028* 3207 13768
Gate Count (first 8 circuits by qubit count); * = best
Idx Circuit Q SABRE AI Rustiq
----------------------------------------------------
0 all-vib-hc3h2cn 24 258 258 257*
1 ham-graph-gnp_k- 24 4036* 4418 5831
2 TSP_Ncity-5 25 2045* 2229 3985
3 tfim 26 489 402* 688
4 all-vib-h2co 32 570* 661 1058
5 uuf100-ham 40 4779* 5130 6757
6 uuf100-ham 40 5667* 6095 8495
7 graph-gnp_k-4 40 24885* 25641 139793
Runtime (s) (first 8 circuits by qubit count); * = best
Idx Circuit Q SABRE AI Rustiq
----------------------------------------------------
0 all-vib-hc3h2cn 24 0.16 0.17 0.16*
1 ham-graph-gnp_k- 24 0.08* 3.13 0.13
2 TSP_Ncity-5 25 0.04* 1.47 0.08
3 tfim 26 0.21 0.34 0.07*
4 all-vib-h2co 32 0.18* 0.19 2.91
5 uuf100-ham 40 0.09* 3.27 0.14
6 uuf100-ham 40 0.11* 4.23 0.17
7 graph-gnp_k-4 40 0.39* 15.15 2.92
plot_transpilation_comparison(
results_large,
"Large-Scale Hamiltonian Circuits: Compilation Comparison",
)
plot_pct_improvement_vs_sabre(
results_large,
"Large-Scale Hamiltonian Circuits",
)
plot_best_method_bars(results_large)
# Select circuit index 3 from the large-scale transpiled circuits
test_idx_large = 3
test_circuit_large = qc_large[test_idx_large]
print(
f"Test circuit: {test_circuit_large.name}, {test_circuit_large.num_qubits} qubits"
)
tqc_methods_large = {
"SABRE": tqc_sabre_large[test_idx_large],
"AI": tqc_ai_large[test_idx_large],
"Rustiq": tqc_rustiq_large[test_idx_large],
}
print(f"\nTranspilation metrics for circuit index {test_idx_large}:")
for method, tqc in tqc_methods_large.items():
depth_2q = tqc.depth(lambda x: x.operation.num_qubits == 2)
size = tqc.size()
print(f" {method:8s} 2Q depth={depth_2q:5d} size={size:6d}")
Test circuit: tfim, 26 qubits
Transpilation metrics for circuit index 3:
SABRE 2Q depth= 100 size= 489
AI 2Q depth= 20 size= 402
Rustiq 2Q depth= 31 size= 688
pm_mirror = generate_preset_pass_manager(
optimization_level=0, backend=backend
)
for method, tqc in tqc_methods_large.items():
# print the count ops for each circuit
mirror = tqc.copy()
mirror.compose(tqc.inverse(), inplace=True)
mirror.measure_all()
mirror = pm_mirror.run(mirror)
print(f"\n{method} transpiled circuit:")
print(tqc.count_ops())
print(f"{method} mirror circuit count ops:")
print(mirror.count_ops())
SABRE transpiled circuit:
OrderedDict({'sx': 211, 'rz': 163, 'cz': 104, 'x': 11})
SABRE mirror circuit count ops:
OrderedDict({'rz': 1170, 'sx': 422, 'cz': 208, 'measure': 156, 'x': 22, 'barrier': 1})
AI transpiled circuit:
OrderedDict({'sx': 165, 'rz': 162, 'cz': 68, 'x': 7})
AI mirror circuit count ops:
OrderedDict({'rz': 984, 'sx': 330, 'measure': 156, 'cz': 136, 'x': 14, 'barrier': 1})
Rustiq transpiled circuit:
OrderedDict({'sx': 316, 'rz': 225, 'cz': 140, 'x': 7})
Rustiq mirror circuit count ops:
OrderedDict({'rz': 1714, 'sx': 632, 'cz': 280, 'measure': 156, 'x': 14, 'barrier': 1})
# Build mirror circuits and submit to real hardware
# The inverse may introduce gates (e.g., sxdg) not in the backend's
# basis gate set, so we re-transpile the mirror circuit.
pm_mirror = generate_preset_pass_manager(
optimization_level=0, backend=backend
)
shots_hw = 10000
hw_jobs = {}
for method, tqc in tqc_methods_large.items():
mirror = tqc.copy()
mirror.compose(tqc.inverse(), inplace=True)
mirror.measure_all()
# Re-transpile at opt level 0 to decompose into basis gates
# without changing the layout or routing
mirror = pm_mirror.run(mirror)
sampler = SamplerV2(mode=backend)
sampler.options.environment.job_tags = ["TUT_CMHSC"]
job = sampler.run([mirror], shots=shots_hw)
hw_jobs[method] = job
print(f"{method}: submitted job {job.job_id()}")
SABRE: submitted job d8gvgq66983c73dqe5og
AI: submitted job d8gvgqe6983c73dqe5pg
Rustiq: submitted job d8gvgqm6983c73dqe5q0
# Retrieve results and compute fidelities
fidelities_large = {}
for method, job in hw_jobs.items():
result = job.result()
counts = result[0].data.meas.get_counts()
n_qubits = backend.num_qubits
all_zeros = "0" * n_qubits
fidelity = counts.get(all_zeros, 0) / shots_hw
fidelities_large[method] = fidelity
print(
f"{method:8s} P(|00...0>) = {fidelity:.4f} ({counts.get(all_zeros, 0)}/{shots_hw})"
)
SABRE P(|00...0>) = 0.0005 (5/10000)
AI P(|00...0>) = 0.3267 (3267/10000)
Rustiq P(|00...0>) = 0.1845 (1845/10000)
plot_mirror_results(
tqc_methods_large, fidelities_large, test_circuit_large.name
)
Pagsusuri ng mga resulta ng kompilasyon
Ang mga benchmark sa itaas ay nagkukumpara ng SABRE, ng AI-powered transpiler, at ng Rustiq sa mga Hamiltonian simulation circuit mula sa koleksyon ng Hamlib sa parehong maliit at malaking antas.
Two-qubit depth at gate count
Sa malaking antas, ang SABRE at ang AI-powered transpiler ang dalawang pinakamalakas na gumaganap, at ang bawat isa ay nangunguna sa ibang sukatan. Tulad ng ipinapakita ng chart na pinakamahusay na pamamaraan ayon sa sukatan, ang SABRE ay gumagawa ng pinakamababang gate count sa malaking bahagi ng mga circuit at ang pinakamabilis na pamamaraan sa halos lahat ng mga ito, na naaayon sa isang heuristic na dinisenyo upang mabawasan ang mga SWAP gate na ipinasok, at sa mga kamakailang pagpapabuti sa layout at routing nito. Ang AI-powered transpiler ay gumagawa ng pinakamababang two-qubit depth sa karamihan ng mga circuit, na naaayon sa bahagi ng reinforcement learning objective nito na naglalayong sa circuit depth. Ang summary table ay sumasalamin sa parehong split: ang SABRE ay may mas mababang mean gate count, habang ang AI transpiler ay may mas mababang mean two-qubit depth. Ang parehong pamamaraan ay pare-pareho at maaasahan sa buong hanay ng mga circuit.
Ang Rustiq, na partikular na itinayo para sa PauliEvolutionGate synthesis, ay gumagawa ng iisang pinakamahusay na resulta sa maliit na bahagi ng mga malalaking circuit. Ang average na sukatan nito ay lubos na naskewed ng ilang makabuluhang outlier, na nakikita bilang malalaking spike sa compilation comparison plot, kung saan ang Rustiq ay gumagawa ng mas mataas na depth at gate count kaysa sa ibang mga pamamaraan. Kung wala ang mga outlier na ito, ang average na performance nito ay magiging mas malapit sa SABRE at sa AI-powered transpiler.
Ang pangunahing obserbasyon ay walang iisang pamamaraan ang nangingibabaw sa bawat circuit. Ang bawat pamamaraan ay lumalagpas sa iba sa mga partikular na kaso, na ginagawang sulit ang subukan ang lahat ng available na kasangkapan at piliin ang pinakamahusay na resulta para sa bawat circuit.
Runtime
Ang SABRE ay pare-parehong pinakamabilis na pamamaraan. Ang Rustiq ay karaniwang tumatakbo sa katulad na bilis, ngunit maaari itong gumawa ng mga outlier kung saan ang kompilasyon ay tumatagal nang mas matagal. Ito ay lalo na makikita sa mga malalaking resulta, kung saan ang ilang circuit ay nagdudulot ng pagtaas ng runtime ng Rustiq. Ang mga outlier na ito ay lubos na nakakaapekto sa average na runtime, kaya ang median ay maaaring maging mas kinatawang buod para sa Rustiq. Ang AI-powered transpiler ang pinakamabagal sa tatlo, na may runtime na kapansin-pansing lumalaki sa mas malalaking at mas kumplikadong circuit.
Mga resulta ng mirror circuit
Ang mga eksperimento ng mirror circuit ay nagpapatunay ng inaasahang trend: ang mga pamamaraan na gumagawa ng mas mababang two-qubit depth at mas kaunting gate ay nakakamit ng mas mataas na fidelity sa ilalim ng ingay. Ito ay nangunguna sa parehong maingay na simulator (maliit na antas) at tunay na hardware (malalaking antas).
Tandaan na ang bawat mirror-circuit plot ay sumasalamin sa isang solong circuit, hindi sa aggregate. Ang halimbawa ng hardware sa itaas ay gumagamit ng isang 26-qubit na tfim circuit, na nangyaring isang kaso kung saan ang SABRE ay gumagawa ng mas mataas na two-qubit depth kaysa sa AI-powered transpiler at Rustiq, kaya ang fidelity nito ay naaayon na mas mababa. Hindi ito kinatawan ng mas malawak na mga resulta: sa buong hanay ng mga malalaking circuit, ang two-qubit depth ng SABRE ay karaniwang malapit sa sa AI-powered transpiler, at ang dalawang pamamaraan ay bawat isa ay nangunguna sa iba't ibang sukatan (ang AI-powered transpiler sa two-qubit depth, ang SABRE sa gate count at runtime). Ang isang solong mirror na resulta ay sumusubok ng isang doble na bersyon ng isang circuit sa halip ng buong workload, kaya hindi ito dapat basahin bilang isang hatol sa kabuuang kalidad ng pamamaraan.
Mga rekomendasyon
Walang iisang pinakamahusay na estratehiya ng transpilation para sa lahat ng circuit. Ang pinakamahusay na pagpipilian ay nakasalalay sa istruktura ng circuit, ang layunin ng optimization, at ang available na badyet ng oras ng kompilasyon:
- SABRE ang inirerekomendang default. Ito ay mabilis at maaasahan, at gumagawa ng mahusay na mga resulta sa malawak na hanay ng mga circuit. Para sa karagdagang pag-tune, maaaring dagdagan ng mga user ang layout at routing trial (tingnan ang SABRE optimization tutorial).
- Ang AI-powered transpiler ay sulit na subukan kapag ang oras ng kompilasyon ay hindi isang limitasyon, lalo na kapag ang pagbabawas ng two-qubit depth ang priyoridad: gumagawa ito ng pinakamababang two-qubit depth sa karamihan ng mga malalaking circuit sa benchmark na ito.
- Ang Rustiq ay partikular na itinayo para sa mga
PauliEvolutionGatecircuit at makakakita ng mga napakababang-depth, mababang-gate-count na solusyon, lalo na sa mas maliliit na circuit. Sa mas malalaking circuit, paminsan-minsan ay maaari itong gumawa ng mas malalaking resulta, kaya pinakamabuti itong gamitin bilang isa sa ilang pamamaraan upang subukan sa halip na bilang default.
Sa praktika, ang pinakamahusay na diskarte ay patakbuhin ang lahat ng available na pamamaraan at piliin ang pinakamahusay na resulta para sa bawat circuit. Ang overhead ng kompilasyon ng pagsubok ng maraming pamamaraan ay maliit kumpara sa potensyal na pagpapabuti ng kalidad ng execution sa tunay na hardware.
Mga susunod na hakbang
Kung natuklasan mong kapaki-pakinabang ang tutorial na ito, maaari kang maging interesado sa mga sumusunod:
Mga sanggunian
[1] "LightSABRE: A Lightweight and Enhanced SABRE Algorithm". H. Zou, M. Treinish, K. Hartman, A. Ivrii, J. Lishman et al. https://arxiv.org/abs/2409.08368
[2] "Practical and efficient quantum circuit synthesis and transpiling with Reinforcement Learning". D. Kremer, V. Villar, H. Paik, I. Duran, I. Faro, J. Cruz-Benito et al. https://arxiv.org/abs/2405.13196
[3] "Pauli Network Circuit Synthesis with Reinforcement Learning". A. Dubal, D. Kremer, S. Martiel, V. Villar, D. Wang, J. Cruz-Benito et al. https://arxiv.org/abs/2503.14448
[4] "Faster and shorter synthesis of Hamiltonian simulation circuits". T. Goubault de Brugiere, S. Martiel et al. https://arxiv.org/abs/2404.03280