Adaptive directional gradients for parameterised quantum circuits

2026-06-08 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

A new framework of forward gradient estimators is proposed for training parameterised quantum circuits (PQCs), directly addressing the high measurement cost of traditional gradient estimation methods like the parameter-shift rule. This framework, based on the forward mode of automatic differentiation, provides an unbiased gradient estimator by averaging a freely tunable number of random directional derivatives. It recovers SPSA, random coordinate descent, and the parameter-shift rule as limiting cases, without requiring ancilla qubits or controlled-gate overhead. The work introduces QUIVER (Quantum Iterative V-adaptive Estimator Rule), an adaptive optimizer derived from a closed-form minimum measurement-cost allocation. Numerically, forward gradients train 60-qubit, 1770-parameter orthogonal quantum neural networks on ECG5000 and MNIST datasets orders of magnitude more efficiently, and QUIVER outperforms iCANS and gCANS on quantum approximate optimisation algorithm (QAOA) and variational quantum eigensolver (VQE) problems.

Key takeaway

For Machine Learning Engineers optimizing parameterised quantum circuits on quantum hardware, adopting the QUIVER adaptive optimizer or forward gradient estimators can drastically reduce the measurement cost. This enables more efficient training of larger models, such as 60-qubit orthogonal quantum neural networks. You should explore integrating these methods to overcome the parameter-shift rule's scaling limitations and accelerate quantum algorithm development.

Key insights

Forward gradient estimators significantly reduce measurement cost for training PQCs by averaging random directional derivatives.

Principles

• Measurement cost bottlenecks PQC training.
• Forward gradients offer unbiased estimation.
• Adaptive optimizers can minimize measurement cost.

Method

QUIVER derives an update rule from a closed-form minimum measurement-cost allocation for adaptive PQC optimization, based on forward gradient estimators.

In practice

• Train orthogonal quantum neural networks efficiently.
• Optimize QAOA and VQE problems with fewer measurements.

Topics

• Parameterised Quantum Circuits
• Quantum Gradient Estimation
• Forward Gradient Estimators
• QUIVER Optimizer
• Quantum Machine Learning
• Measurement Cost Reduction

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Scalable On-Hardware Training of Quantum Neural Networks and Application to Clinical Data Imputation — A novel QNN training framework reduces hardware gradient estimation costs from O(n^2) to O(log n), enabling scalable quantum computing applications.
• Quantum Feature Selection with Higher-Order Binary Optimization on Trapped-Ion Hardware — A quantum HUBO framework on trapped-ion hardware enhances feature selection by capturing higher-order data dependencies.
• The Quantum Leap in AI: Building Hybrid Classical-Quantum Neural Networks — Hybrid classical-quantum neural networks combine classical data processing with quantum classification for advanced AI.
• AlphaCNOT: Learning CNOT Minimization with Model-Based Planning — AlphaCNOT combines RL with Monte Carlo Tree Search for efficient CNOT gate minimization in quantum circuits.
• Long Range Frequency Tuning for QML — Trainable-frequency QML models have limited frequency reachability, which ternary grid initialization overcomes for robust convergence.
• Qilimanjaro Pushes Analog Quantum as AI Compute Demands Surge — Analog quantum computing offers a path to reduced errors and accelerated AI by avoiding extensive gate operations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.