arXiv:2610.10254
Benchmarking Modular Optimization Strategies for Parameterized Quantum Circuits
Carla Cotea, Stefan Balauca, Andreea Arusoaie
Abstract
Training parameterized quantum circuits requires balancing optimization progress with limited evaluation budgets and the statistical uncertainty of near-term quantum hardware. We present a modular benchmarking framework that separates quantum search-direction estimation from classical parameter-update rules, allowing their interactions and sensitivities to be examined independently. We evaluate a suite of gradient-based, stochastic, and derivative-free optimizers across combinatorial optimization using the Quantum Approximate Optimization Algorithm (QAOA), supervised quantum machine learning using Iris classification and a quantum convolutional neural network (QCNN) for binary MNIST classification, and quantum chemistry using the variational quantum eigensolver (VQE) for molecular hydrogen. By separating objective evaluations from sample-circuit costs under finite-shot simulation and physical execution on a 156-qubit processor, we compare optimizer behavior across workloads with different parameter counts and measurement requirements. Four selected simulator and hardware case studies report terminal and best observed objectives separately, alongside post-training classifier accuracies. The comparisons include selected MaxCut and hydrogen trajectories and illustrate workload-dependent objective changes and evaluation costs, while the simulator benchmark compares both peak and mean performance across three seeds. The hardware runs do not establish an optimizer ranking or isolate a causal effect of device noise.
AI Summary
The paper asks how quantum search-direction estimators interact with classical update rules when training parameterized circuits under finite-shot evaluation costs. Its benchmark pairs parameter-shift, finite differences, SPSA, PGPE, and diagonal observable-curvature preconditioning (OCP) with SGD, Adam, and RMSprop; COBYLA, Nelder–Mead, and Bayesian optimization retain their native dynamics. OCP is a curvature-rescaling heuristic, not a quantum Fisher metric. Simulations use 256 shots per setting and three seeds on five-qubit MaxCut QAOA, four-qubit Iris classification, an eight-qubit binary-MNIST QCNN with 63 parameters, and four-qubit $H_2$ VQE. At the 100-step modular checkpoint or earlier termination, PGPE-SGD has the lowest mean best MaxCut cost, $-4.345052$, against an exact target of $-5$. FiniteDiff-RMSprop has the lowest mean absolute VQE error, $2.6949$ mHa. OCP-RMSprop has the highest mean test accuracies: $100\%$ for Iris and $96.67\%$ for QCNN. Rankings based on individual peaks differ from mean rankings. Resource accounting separates objective queries from logical circuits; QCNN OCP uses 404,800 training circuits over 100 updates, versus 17,600 for PGPE. Four shorter, single-trajectory hardware case studies use 4–8 qubits on a 156-qubit processor. Each shows some monitored-objective improvement, but they do not establish hardware optimizer rankings or isolate noise effects. Stated limitations include fixed hyperparameters, unequal measurement budgets, only three simulation seeds, small classifier test sets, changing mini-batches, missing VQE measurement-group counts, and selection bias from noisy minima. Reported VQE agreement is not independently certified chemical accuracy.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.