QPU-Mini

This repository contains the experimental code, data, and figures for the paper:

Optimize the Optimization: QPU-Efficient Quantum Graph Optimization through HPC-First Experimentation

David Vesterlund, Vesterlund Ventures / WestQuant Open Source Project

Submitted to IEEE Transactions on Quantum Engineering

The required QPU budget for quantum graph optimization is not a fixed algorithmic constant but a structured, predictable property of the problem instance and workflow policy:

QPU_requirement = f(graph, problem, n, depth, quality_target, hardware)

By separating classical parameter optimization from a single QPU evaluation, a best-practice baseline achieves:

• 510–2070× shot reduction across n=8 to n=16 qubits
• ≤0.4% quality loss relative to naive QAOA
• 4 of 6 strategies achieve 100% quality retention at p=3

MaxCut, WeightedMaxCut, Maximum Independent Set (MIS), Maximum Clique, Minimum Vertex Cover, Graph Partitioning

Erdős–Rényi (ER), Barabási–Albert (BA), Watts–Strogatz (WS), Regular, Geometric (GEO), Stochastic Block Model (SBM), Configuration (CONFIG)

• B0 (naive): 30 QPU evaluations per optimization cycle
• B1 (best practice): Classical parameter optimization + single QPU evaluation
• B2 (classical): Exact classical solver, no QPU

qpu-mini/
├── experiments/
│   ├── run_v2.py            # Experiment runner (all 4 phases)
│   ├── baselines.py         # B0/B1/B2 implementations
│   ├── strategies.py        # 6 QPU-minimization strategies
│   ├── metrics_v2.py        # Multi-metric quality and resource vector
│   ├── graphs.py            # Graph generation for 7 families
│   ├── problems.py          # Problem Hamiltonians (6 problems)
│   └── qaoa.py              # QAOA circuit construction and simulation
├── results_v2/
│   ├── raw/
│   │   ├── v2_core_factorial.json    # 11,340 results
│   │   └── v2_scaling_study.json    # 1,260 results
│   └── processed/
│       └── structural_predictor.json
├── paper/
│   ├── manuscript_v2.tex    # IEEE TQE submission
│   ├── manuscript_v2.pdf    # Compiled PDF
│   └── figures_v2/          # 5 figures at 300 DPI
└── configs/
    └── default.yaml         # Experiment configuration

# Install dependencies
pip install numpy scipy matplotlib qiskit networkx scikit-learn

# Run all experiments
cd qpu-mini
python3 -u -c "import sys; sys.path.insert(0, '.'); from experiments.run_v2 import run_v2_all; run_v2_all()"

# Recompile manuscript
cd paper
tectonic manuscript_v2.tex

The QPU budget predictor from this research is integrated into the WestQuant QCSC package:

pip install westquant-qcsc

from qcsc import QPUBudgetPredictor, ProblemProfile

profile = ProblemProfile(problem="MaxCut", n_qubits=16, p=3, graph_family="GEO")
predictor = QPUBudgetPredictor()
budget = predictor.estimate(profile)
# 1024 shots, 2070x reduction, quality=0.9961

• Small graph sizes (n ≤ 16) — exact statevector simulation only
• No noise model — ideal simulation
• Limited depth (p ≤ 3)
• Penalty-dominated problems (MIS, MaxClique, MVC) need conditional quality metric
• Structural predictor R²=1.0 reflects B1's constant cost, not genuine structural prediction

• Discussions: Join the conversation
• Contributing: See CONTRIBUTING.md

MIT

David Vesterlund — Vesterlund Ventures / WestQuant Open Source Project

• Email: david@vesterlundventures.se
• ORCID: 0009-0000-6455-1141