Guided-SPSA: Simultaneous Perturbation Stochastic Approximation Assisted by the Parameter Shift Rule

Maniraman Periyasamy, Axel Plinge, Christopher Mutschler, Daniel D. Scherer, Wolfgang Mauerer

Abstract

Variational Quantum Circuits (VQCs) are central to the Noisy Intermediate-Scale Quantum (NISQ) era but face significant training challenges due to the high cost of exact gradients and the instability of stochastic approximations. This paper introduces Guided-SPSA, a hybrid gradient estimation technique that integrates the precision of the Parameter-Shift Rule (PSR) with the efficiency of Simultaneous Perturbation Stochastic Approximation (SPSA). By utilizing a small subset of data for exact gradient "guidance," the method achieves a 15-25% reduction in circuit evaluations while maintaining or exceeding the accuracy of standard methods. Notably, Guided-SPSA demonstrates an ability to escape local minima where traditional gradient-based methods fail.

Introduction

Training Variational Quantum Circuits (VQCs) requires calculating gradients to update circuit parameters, a process currently hampered by a trade-off between precision and scalability. The Parameter-Shift Rule (PSR) provides exact gradients but requires $2M$ circuit evaluations per data point (where $M$ is the number of parameters), making it computationally prohibitive for large models. Conversely, Simultaneous Perturbation Stochastic Approximation (SPSA) approximates gradients using a constant number of evaluations regardless of $M$, but it is often unstable and prone to suboptimal convergence. Guided-SPSA addresses this "scalability vs. stability" gap by seeking a middle ground suitable for noisy hardware.

Methodology

Guided-SPSA functions as a hybrid algorithm that employs both PSR and SPSA within each training batch. The mechanism operates through two primary components:

  • The Guide: A small portion of the training data (defined by a ratio $\tau$) is processed using the exact Parameter-Shift Rule, providing a high-fidelity directional "compass."
  • The Engine: The remaining data is processed via SPSA for efficiency.
  • Guidance Mechanism: Noisy SPSA gradients are scaled and suppressed based on the norm of the exact PSR gradients, ensuring the stochastic component remains aligned with the true gradient path.
  • Dynamic Complexity: The algorithm employs a linear increase in SPSA perturbation samples over epochs, prioritizing speed in early training and precision during convergence.

Results and Discussion

The authors evaluated Guided-SPSA across regression (Friedman and Boston Housing datasets), classification (Iris dataset), and reinforcement learning (CartPole-v0) tasks.

Efficiency and Convergence

Guided-SPSA achieved a 15% to 25% reduction in total circuit evaluations compared to PSR. On the Friedman dataset, Guided-SPSA not only matched PSR's performance but reached a lower final error at a significantly lower computational cost. It consistently outperformed standard SPSA, which frequently failed to converge.

Convergence comparison on the Friedman dataset; Guided-SPSA (black) shows superior performance.

Escaping Local Minima and Robustness

A key finding is the algorithm's ability to navigate complex cost landscapes. In scenarios where exact gradients (PSR) became trapped in local minima, the controlled stochasticity of Guided-SPSA allowed the model to "jump" to the global optimum. Furthermore, the method proved robust to poor initializations (such as all-zero states) where PSR often fails due to vanishing initial gradients.

Comparison showing Guided-SPSA escaping a local minimum where PSR gets trapped.

Conclusion

Guided-SPSA represents a viable "Goldilocks" solution for quantum machine learning, offering the stability of exact gradients with the efficiency of stochastic methods. While the method introduces new hyperparameters like the sample ratio ($\tau$), its performance under noisy conditions and with error mitigation techniques suggests it is well-suited for current NISQ processors. Future research will focus on adaptive guidance ratios and testing scalability on circuits with hundreds of qubits.

Summary updated

BibTeX
@inproceedings{periyasamy2024guided,
  author = {Maniraman Periyasamy and Axel Plinge and Christopher Mutschler and Daniel D. Scherer and Wolfgang Mauerer},
  title = {{Guided-SPSA: Simultaneous Perturbation Stochastic Approximation Assisted by the Parameter Shift Rule}},
  booktitle = {QCE},
  year = {2024},
  doi = {10.1109/qce60285.2024.00177},
}

Loading related content...

Figure