🔬 physics

Benchmarking Modular Optimization Strategies for Parameterized Quantum Circuits

This paper introduces a modular benchmarking framework that decouples quantum search-direction estimation from classical parameter-update rules to systematically evaluate the performance and cost-efficiency of various optimizers across diverse quantum workloads, including QAOA, QCNN, and VQE, under both finite-shot simulations and execution on a 156-qubit quantum processor.

Original authors: Carla Cotea, Stefan Balauca, Andreea Arusoaie

Published 2026-11-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Carla Cotea, Stefan Balauca, Andreea Arusoaie

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the race to build useful quantum computers, scientists are currently working with machines that are powerful but fragile. These devices, known as near-term quantum processors, can perform calculations that classical computers struggle with, but they are prone to errors and lose their delicate quantum states quickly. To get useful results from this imperfect hardware, researchers rely on a hybrid approach called variational quantum algorithms. In this method, a quantum computer acts as a specialized calculator that tests a specific setup of parameters, while a standard classical computer acts as a coach, analyzing the results and suggesting how to tweak those parameters to get a better answer. The goal is to find the best possible setting for a problem, whether that is solving a complex math puzzle, classifying an image, or simulating a molecule. However, the path to the best answer is not straightforward. The landscape of possible solutions is often rugged and filled with traps, and the measurements taken from the quantum computer are inherently noisy, meaning the coach receives a slightly different signal every time it asks a question.

A team of researchers at Alexandru Ioan Cuza University and FreeYa Mind Campus in Romania set out to understand how to navigate this difficult terrain more effectively. They focused on the two distinct parts of the coaching process: the method used to estimate which direction to move the parameters, and the rule used to actually update those parameters. While many previous studies treated different optimization strategies as complete, unchangeable packages, this team decided to separate the two. They treated the "search direction" and the "update rule" as independent modules that could be mixed and matched. By doing this, they could see exactly how the choice of one part influenced the performance of the other, rather than just observing which pre-made combination happened to work best in a single experiment.

The researchers tested their modular framework across four very different types of problems. They tackled a combinatorial optimization task known as MaxCut, which involves dividing a network of connections into two groups to maximize the number of links between them. They also tested supervised learning tasks, using a small dataset of flower measurements to classify species and a larger set of handwritten digits to distinguish between the numbers zero and one. Finally, they applied their methods to a problem in quantum chemistry, attempting to find the lowest energy state of a hydrogen molecule. For each of these challenges, they ran simulations on a noiseless computer model and selected configurations on a real, physical quantum processor with 156 qubits. They compared various ways of estimating the gradient, or the slope of the solution landscape, including methods that shift parameters slightly to measure change, methods that perturb all parameters at once, and methods that do not use gradients at all. These were then paired with different update rules, such as simple step-by-step adjustments or more complex adaptive rules that remember past movements.

The results revealed that there is no single "best" optimizer that works for every situation. The performance of a strategy depended heavily on the specific problem being solved and the resources available. For the MaxCut problem, a method that perturbs all parameters simultaneously and uses a simple update rule performed well on average, though another combination achieved the single best result in one specific test run. For the hydrogen molecule simulation, a method that smooths out the search direction and uses a simple update rule produced the closest individual result to the known correct answer, but a different combination using a standard gradient estimate and an adaptive update rule was more consistent across multiple test runs. In the image classification tasks, a strategy that adjusted the search direction based on the curvature of the landscape and used an adaptive update rule consistently achieved the highest accuracy across all test runs.

Crucially, the study highlighted that achieving a better result often came at a steep cost in terms of the number of measurements required. Some methods that reached the highest accuracy needed to run the quantum circuit thousands of times more than others to get there. For instance, in the digit classification task, one configuration reached perfect accuracy but required roughly twenty-three times more circuit executions than a different configuration that reached a slightly lower accuracy. This trade-off suggests that the most accurate method is not always the most practical one, especially when running on real hardware where time and resources are limited. The researchers also found that the choice of the update rule mattered significantly. Changing the rule used to adjust the parameters could drastically alter the final accuracy, even when the method for estimating the direction remained the same. In some cases, a simple update rule worked best with a specific estimator, while a more complex adaptive rule was required for a different estimator to perform well.

When the team moved from simulations to the physical quantum processor, the results were mixed but informative. The real hardware runs showed that the algorithms could still improve the solutions, but the progress was often less smooth and the final results were not as good as those seen in the noiseless simulations. In the hydrogen molecule experiment, the hardware run showed a temporary improvement before the results began to deteriorate, a behavior that was not seen in the simulation. This suggests that the noise and imperfections of the physical device can change the trajectory of the optimization in unpredictable ways. The researchers were careful to note that these hardware runs were case studies of specific configurations rather than a definitive ranking of all methods, as the limited number of runs on the physical device was not enough to isolate the specific effects of the hardware noise from other factors.

Ultimately, the study provides a clearer map of the optimization landscape for quantum algorithms. It demonstrates that the performance of these algorithms is not determined by a single magic bullet but by the careful pairing of how the system looks for a solution and how it decides to move. The best choice depends on the specific problem, the available measurement budget, and the need for consistency versus the hope of a peak performance. By separating the estimation of the search direction from the update rule, the researchers showed that scientists can now tune these components independently to find the most efficient path forward for their specific needs. This modular approach offers a practical way to navigate the noisy and uncertain world of near-term quantum computing, helping researchers make the most of the powerful but imperfect machines they currently have.

Technical Summary: Benchmarking Modular Optimization Strategies for Parameterized Quantum Circuits

Problem Statement

Training Parameterized Quantum Circuits (PQCs) within the Noisy Intermediate-Scale Quantum (NISQ) era requires balancing optimization progress against strict evaluation budgets and the statistical uncertainty inherent in finite-shot measurements. While Variational Quantum Algorithms (VQAs) such as the Quantum Approximate Optimization Algorithm (QAOA), Variational Quantum Eigensolver (VQE), and Quantum Machine Learning (QML) models have shown promise, their practical performance is heavily dependent on the optimization loop.

Existing comparative studies often treat optimizers as monolithic "black boxes," evaluating complete packages (e.g., a specific gradient estimator paired with a specific update rule). This approach obscures the specific contributions of individual components, making it difficult to determine whether performance variations arise from the bias or variance of the search-direction estimator, the classical update rule, or their interaction. Furthermore, the literature lacks a clear distinction between the cost of logical circuit executions and the total number of objective evaluations, particularly when dealing with finite-shot sampling and derivative-free methods.

Methodology

The authors propose a modular benchmarking framework that decouples the optimization loop into two distinct conceptual operations:

Search-Direction Estimation: The quantum component that constructs an estimated search direction ( $\hat{d}_k$ ) from circuit measurements.
Classical Parameter Update: The classical component that transforms the direction into a parameter update step ( $p_k$ ) using a learning rate and optimizer state.

Experimental Design

The study evaluates a suite of strategies across four distinct workloads:

Combinatorial Optimization: A 5-qubit MaxCut instance using QAOA ( $p=2$ ).
Supervised Learning (Low-Dim): Iris classification using a 4-qubit Hardware-Efficient Ansatz (HEA).
Supervised Learning (High-Dim): Binary MNIST classification using an 8-qubit Quantum Convolutional Neural Network (QCNN) with 63 parameters.
Quantum Chemistry: Molecular Hydrogen ( $H_2$ ) ground state estimation using VQE with a UCCSD ansatz.

Components Compared

Direction Estimators: Parameter-Shift Rule (PSR), Finite Differences (FD), Simultaneous Perturbation Stochastic Approximation (SPSA), Population-Based Gradient Estimation (PGPE), and Diagonal Observable-Curvature Preconditioning (OCP).
Update Rules: Stochastic Gradient Descent (SGD), Adam, and RMSprop.
Derivative-Free Baselines: COBYLA, Nelder–Mead, and Bayesian Optimization (retained as complete, native optimizers).

Execution Environments

Experiments were conducted under two conditions:

Finite-Shot Simulation: Using the Qiskit Aer simulator with 256 shots per circuit and no noise model, across three random seeds (42, 123, 1234).
Physical Hardware: Selected configurations were executed on the 156-qubit ibm_aachen processor.

Resource Accounting

The framework explicitly distinguishes between:

Objective Evaluations: Complete evaluations of the loss/energy function.
Logical Circuit Executions: Individual circuit runs required to compute an objective (e.g., $1+2M$ for coordinate-wise finite differences).
Shots: The total number of measurement repetitions.

Key Contributions

Modular Decomposition: The study operationalizes the separation of search-direction estimation from classical parameter updating. This allows for the explicit examination of how specific estimator–update rule pairings interact, rather than treating optimizers as opaque units.
Multi-Workload Benchmarking: Instead of deriving conclusions from a single VQA instance, the study compares these modular combinations across optimization, classification, and quantum chemistry tasks with varying parameter counts and measurement requirements.
Explicit Cost Differentiation: The framework separates objective evaluations from logical circuit costs, exposing the resource overhead introduced by parameter counts, mini-batching, and shifted evaluations.
Hardware Case Studies: Selected configurations are run on physical hardware to provide targeted case studies of simulation findings, rather than a general test of quantum advantage or a universal optimizer ranking.

Results

Simulator Performance (Across Seeds)

MaxCut (QAOA): PGPE-SGD achieved the lowest mean cost across seeds, while SPSA-SGD and PSR-Adam shared the single best peak performance. The choice of update rule significantly impacted performance; for instance, SPSA performed best with SGD, whereas PSR performed best with Adam.
H2 (VQE): FiniteDiff-RMSprop achieved the lowest mean absolute error (2.69 mHa), though no method consistently met the 1.6 mHa chemical accuracy threshold across all seeds. PGPE-SGD achieved the closest individual estimate to the Full Configuration Interaction (FCI) reference at one seed but had higher mean error.
Classification (Iris & MNIST): OCP-RMSprop achieved the highest mean test accuracy (100% on Iris, 96.67% on MNIST) and demonstrated the most consistent performance across seeds. Adaptive update rules (Adam, RMSprop) generally outperformed SGD in classification tasks, with improvements ranging from 10 to 35 percentage points.

Hardware Performance

Trajectory Behavior: Selected hardware runs showed objective improvement on both backends, but terminal values often differed from the best observed values during training. For example, the H2 hardware trajectory showed transient improvement followed by deterioration.
No Universal Ranking: The hardware runs did not establish a definitive optimizer ranking or isolate a causal effect of device noise. Differences in performance were attributed to a combination of noise, compilation, mitigation, and stochastic sampling.
Specific Outcomes:
- MaxCut: SPSA-SGD reached ~78% of the exact cut on hardware.
- Iris: COBYLA achieved 90% test accuracy on hardware (vs. 100% in simulation).
- QCNN: OCP-RMSprop achieved 95% test accuracy on hardware.
- VQE: PGPE-SGD showed a large deviation from the reference energy on hardware, highlighting the sensitivity of VQE to hardware noise.

Resource vs. Performance Trade-offs

The study found that higher measurement budgets do not always yield proportionate improvements in objective quality. For instance, OCP-RMSprop required approximately 23 times more circuit evaluations than PGPE-Adam on the QCNN task to gain only one additional correct classification on the test set. Conversely, derivative-free methods like COBYLA achieved perfect accuracy on the Iris task with significantly fewer circuit evaluations than gradient-based methods.

Significance and Claims

The paper modestly claims that its primary contribution is the operationalization of the separation between estimator and update rule, allowing researchers to trace performance gains to specific components of the optimization loop.

No Universal Optimizer: The authors explicitly state that the results do not provide a universal optimizer ranking. Performance is highly dependent on the workload, sampling budget, noise profile, and hyperparameter choices.
Context-Dependent Selection: The "best" configuration depends on the specific metric (peak vs. average performance), the task (optimization vs. classification), and the resource constraints.
Hardware Limitations: The hardware experiments serve as case studies for selected simulation findings. They do not isolate hardware noise mechanisms or prove quantum advantage. The authors emphasize that terminal objectives, best observed objectives, and classifier accuracies must be reported separately to avoid conflating different notions of performance.

In conclusion, the study provides a descriptive evidence base for how modular optimization strategies behave under finite-shot simulation and physical execution, highlighting the importance of matching estimator-update pairs to specific workloads rather than relying on a single "best" optimizer for all quantum tasks.