Learning to Maximize Quantum Neural Network Expressivity via Effective Rank
Juan Yao · 2025
Read Learning to Maximize Quantum Neural Network Expressivity via Effective Rank by Juan Yao (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-15375.txt.
Introduction
A quantum neural network can only solve what it can express, yet designers have lacked a clean gauge of that reach. Juan Yao proposes one in this 2025 paper: the effective rank, a count of how many of a circuit's knobs truly act independently. Redundant knobs inflate the parts list without adding power; the rank strips them away.
Analysis shows the rank can reach its ceiling, set by the number of qubits, when the circuit layout, the input data spread, and the measurement scheme are each fully expressive. Yao then hands the gauge to a reinforcement learning agent with a self-attention brain, which searches circuit layouts on its own, chasing high rank. The work joins a design rule to an automated designer, aiming at quantum circuits that waste no parameters.
Catalog record
| CREATOR | Juan Yao |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 26 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-15375
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Learning to Maximize Quantum Neural Network Expressivity via Effective Rank Juan Yao Abstract Quantum neural networks (QNNs) are widely employed as Ansätze 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 κ \kappa , as a novel quantitative measure of expressivity. Specifically, κ \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 κ \kappa can saturate its theoretical upper bound, d n = 4 n − 1 d_{n}=4^{n}-1 , for an n 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 κ \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 κ \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. I Introduction Quantum neural networks (QNNs) have emerged as powerful tools in quantum machine learning Biamonte et al. 2017 ; Cerezo et al. 2022 ; Wang and Liu 2024 ; Du et al. 2025 ; Zaman et al. 2025 and variational quantum algorithms Benedetti et al. 2019 ; Cerezo et al. 2021 ; De Palma 2023 ; Qi et al. 2024 , playing a pivotal role in diverse applications, including quantum chemistry Lourenço et al. 2023 , quantum sensing Degen et al. 2017 ; Ohshima 2022 and financial modeling Orús et al. 2019 . As parameterized quantum circuits (PQC), QNNs serve as ansätze for optimizing cost functions over high-dimensional Hilbert spaces. Their ability to efficiently represent complex functions directly impacts their performance in solving variational problems. Enhancing the performance and scalability of QNNs remains a key focus in ongoing research. Various types of parameterized quantum circuits (PQCs) have been proposed to enhance the expressivity and effectiveness of quantum neural networks (QNNs) in diverse applications. These include convolutional QNNs Cong et al. 2019 ; Henderson et al. 2020 ; Herrmann et al. 2022 , recurrent QNNs Bausch 2020 ; Li et al. 2023 , quantum generative adversarial networks (QGANs) Huang et al. 2021 , and randomness-enhanced QNNs Wu et al. 2024 . Each of these architectures leverages unique design principles to improve representation power and tackle complex quantum learning tasks. Various metrics have been explored to assess and compare the expressivity of different quantum neural network (QNN) architectures. In Ref. Sim et al. 2019 , a statistical measure, Expr, was introduced to quantify expressivity. While Expr provides insights into a QNN’s ability to explore the Hilbert space, its computation requires extensive sampling over the entire parameter space, similar to measures such as effective dimension Abbas et al. 2021 and effective capacity Wilkinson and Hartmann 2022 , making it computationally expensive. Additionally, Expr is defined as the Kullback-Leibler divergence between the estimated fidelity distribution and that of the Haar-distributed ensemble, restricting its applicability to unitary circuits and making it unsuitable for non-unitary architectures such as dissipative QNNs Beer et al. 2020 . Other studies have assessed expressivity based on Fourier analysis Schuld et al. 2021 , entangling power Hubregtsen et al. 2021 ; Ballarin et al. 2023 or learning performance Wu et al. 2021 ; Wu et al. 2024 . Despite these advances, a precise and generalizable framework for evaluating QNN expressivity remains an open problem, and the lack of standardized metrics continues to obscure a clear understanding of QNN performance. To address this challenge, we introduce the effective rank ( \kappa ) as a novel quantitative measure of expressivity, derived from the Fisher information matrix Liu et al. 2019 ; Meyer 2021 ; Karakida et al. 2020 . Ref. Haug et al. 2021 proposes that the rank of the quantum Fisher information matrix can be used to characterize the capacity of a quantum circuit. However, it focuses on a fixed pure input state and neglects the role of the measurement protocol. To comprehensively account for both the input data distribution and the measurement protocol, we model the quantum neural network as a statistical system and employ the classical Fisher information metric to evaluate its overall expressivity. Generally, increasing the number of variational parameters can enhance the expressivity of a QNN Sim et al. 2019 . However, not all parameters contribute equally, and some may be redundant or ineffective due to circuit structure or parameter correlations. The effective rank \kappa captures the number of independent parameters, reflecting the true degrees of freedom that contribute to expressivity, and provides a precise quantitative measure of QNN expressivity. We systematically analyze the roles of circuit architecture, input data distribution, and measurement protocol, and quantify their impact on expressivity through the effective rank \kappa . By optimizing these factors individually for maximal expressivity, \kappa can reach its theoretical upper bound, d_{n}=4^{n}-1 , which corresponds to the number of generators in the Lie algebra of an n -qubit system. We show that \kappa serves as a direct metric for assessing the degree of controllability in parameterized quantum circuits. Beyond its theoretical characterization, we utilize \kappa as a practical guide for designing highly expressive QNN architectures using reinforcement learning. Constrained by a particular quantum platform, there exists a vast space of configurations for circuit design. Within reinforcement learning framework, we train a self-attention transformer agent to efficiently generate highly expressive configurations. By bridging theoretical insights with practical optimization, our work establishes \kappa as a robust expressivity metric and demonstrates how reinforcement learning can facilitate the automated design of expressive quantum circuits. II fisher information and effective rank As shown in Fig. 1 (a), a quantum neural network (QNN) is typically constructed using a parameterized quantum circuit (PQC), which encodes and processes the input quantum states. The output of the QNN is obtained through statistical measurements of the evolved quantum state, determined by the chosen measurement operators. Treating the quantum neural network as a statistical model, the elements of the Fisher information matrix are defined as F_{ij}({\bm{\theta}})\equiv\mathbb{E}_{(x,y)\sim P}\left[\frac{\partial}{\partial\theta_{i}}\log P(y|{\bm{\theta}},x)\frac{\partial}{\partial\theta_{j}}\log P(y|{\bm{\theta}},x)\right], (1) which quantifies the sensitivity of the output distribution with respect to the model parameters. Here, x represents the input quantum state, while y denotes the measurement outcome, which occurrs with probability P(y|{\bm{\theta}},x) . The quantum neural network is parameterized by {\bm{\theta}}=\{\theta_{j}|j=1,2,3,...,p\} which governs the evolution of the input state. As shown in Eq. ( 1 ), the Fisher information matrix is a p\times p square matrix whose eigenvalues are non-negative real numbers. The spectral decomposition of F({\bm{\theta}}) is strongly correlated with the feasibility of training Karakida et al.