Universal Approximation Theorem and Error Bounds for Quantum Neural Networks and Quantum Reservoirs

Area of Science:

• Quantum Computing
• Machine Learning
• Artificial Intelligence

Background:

• Universal approximation theorems underpin classical neural networks' ability to approximate functions.
• Recent advancements demonstrate parameterized quantum circuits can achieve similar function approximation capabilities.
• Extending these concepts to quantum settings is crucial for developing quantum machine learning.

Purpose of the Study:

• To provide precise error bounds for function approximation by quantum neural networks.
• To extend these theoretical guarantees to randomized quantum circuits, analogous to classical reservoir networks.
• To establish the resource requirements (qubits and weights) for achieving a target approximation error.

Main Methods:

• Analysis of parameterized quantum circuits for function approximation.
• Development of error bounds for specific function classes, including those with integrable Fourier transforms.
• Investigation of randomized quantum circuits, drawing parallels with classical reservoir computing models.

Main Results:

• Precise error bounds are established for quantum neural networks approximating functions.
• The study demonstrates that quantum neural networks can effectively mimic classical reservoir networks through randomization.
• A key finding shows that O(ε−2) weights and O(⌈log2(ε−1)⌉) qubits are sufficient for achieving an approximation error of ε for functions with integrable Fourier transforms.

Conclusions:

• Quantum neural networks offer a viable approach for function approximation with theoretical guarantees.
• Randomized quantum circuits present a promising direction for quantum machine learning, inspired by classical reservoir computing.
• The established resource scaling provides practical insights into the implementation of quantum neural networks for approximation tasks.