Learning to Maximize Quantum Neural Network Expressivity via Effective Rank

TL;DR

This paper introduces the effective rank as a new measure of quantum neural network expressivity, and uses reinforcement learning to optimize circuit designs based on this metric.

Contribution

It defines the effective rank to quantify QNN expressivity and employs a transformer-based reinforcement learning approach for automated circuit design.

Findings

Effective rank $ppa$ can reach the theoretical maximum $d_n=4^n-1$ for optimal circuit factors.

Reinforcement learning with self-attention effectively explores and optimizes quantum circuit architectures.

The framework enhances the design of highly expressive quantum neural networks.

Abstract

Quantum neural networks (QNNs) are widely employed as ans\"atze for solving variational problems, where their expressivity directly impacts performance. Yet, accurately characterizing QNN expressivity remains an open challenge, impeding the optimal design of quantum circuits. In this work, we introduce the effective rank, denoted as $\kappa$, as a novel quantitative measure of expressivity. Specifically, $\kappa$ captures the number of effectively independent parameters among all the variational parameters in a parameterized quantum circuit, thus reflecting the true degrees of freedom contributing to expressivity. Through a systematic analysis considering circuit architecture, input data distributions, and measurement protocols, we demonstrate that $\kappa$ can saturate its theoretical upper bound, $d_n=4^n-1$, for an $n$-qubit system when each of the three factors is optimally expressive. This result provides a rigorous framework for assessing QNN expressivity and quantifying their functional capacity. Building on these theoretical insights, and motivated by the vast and highly structured nature of the circuit design space, we employ $\kappa$ as a guiding metric for the automated design of highly expressive quantum circuit configurations. To this end, we develop a reinforcement learning framework featuring a self-attention transformer agent that autonomously explores and optimizes circuit architectures. By integrating theoretical characterization with practical optimization, our work establishes $\kappa$ as a robust tool for quantifying QNN expressivity and demonstrates the effectiveness of reinforcement learning in designing high-performance quantum circuits. This study paves the way for building more expressive QNN architectures, ultimately enhancing the capabilities of quantum machine learning.