Abstract
Variational Quantum Algorithms (VQAs) are central to the Noisy Intermediate-Scale Quantum (NISQ) era, yet their performance is often evaluated using metrics that ignore real-world hardware constraints. This paper introduces Model Gradient Descent (MGD), a surrogate-model-based optimizer designed to minimize wall-clock time in high-latency, noisy environments. We demonstrate that accounting for network latency and circuit batching significantly shifts the landscape of "optimal" classical optimizers for quantum hardware.
Introduction
The hybrid quantum-classical loop is the foundation of VQAs, where a classical optimizer iteratively updates parameters for a quantum circuit. However, standard benchmarks often focus on iteration counts or noiseless simulations, neglecting the practical overheads of cloud-based quantum computing: sampling noise, gate errors, and significant network latency. A theoretically efficient optimizer may fail in practice if it requires excessive sequential communication with the quantum processor. This work bridges the gap between theory and practice by evaluating optimizers through a realistic wall-clock time model.
Methods: Model Gradient Descent (MGD)
To address the challenges of noisy landscapes, we propose Model Gradient Descent (MGD). Unlike traditional gradient descent, MGD utilizes a surrogate model to navigate the parameter space.
- Surrogate Modeling: MGD maintains a history of sampled points and fits a local quadratic model using least-squares within a "moving trusted region."
- Data Reuse: By leveraging historical data rather than discarding it after each step, MGD builds a more robust local map of the energy landscape, reducing the need for fresh, noisy evaluations.
We compare MGD against deterministic (Nelder-Mead, BOBYQA) and stochastic (SPSA, SGD) optimizers, specifically tuned for a wall-clock time model that accounts for sampling rates, circuit switching overhead, and network latency.
Key Insights: The Importance of Tuning and Batching
A critical finding of this study is that "out-of-the-box" optimizer settings are frequently inadequate for noisy quantum hardware. Hyperparameter tuning is essential; when properly tuned, even algorithms previously considered weak can become competitive and stable.
Furthermore, the ability to batch multiple circuit evaluations into a single request to the quantum cloud is a transformative factor. Batching mitigates the impact of network latency, favoring gradient-based methods like MGD and SGD that can evaluate multiple points in parallel.
Key Results and Discussion
Our benchmarks on Max-Cut, the Sherrington-Kirkpatrick model, and the 2D Hubbard model reveal that MGD consistently outperforms traditional methods in terms of wall-clock efficiency and reliability.
In higher-depth circuits, MGD demonstrated the highest probability of reaching the global minimum. While deterministic methods like BOBYQA require a massive number of measurement shots to suppress noise (e.g., 125,000 shots vs. 1,000 for SGD), MGD maintains stability with far fewer resources. Experimental validation on Google’s Sycamore processor confirms MGD's practical utility in real-world NISQ applications.
Conclusion
Evaluating optimizers based on wall-clock time rather than abstract iterations reveals that surrogate-model-based approaches like MGD are superior for current quantum hardware. Future work will extend these methods to address global search challenges like "Barren Plateaus" and refine noise models to account for systematic hardware errors.
Summary updated