Optimizer-Dependent Generalization Bound for Quantum Neural Networks

Chenghong Zhu, Hongshun Yao, Yingjian Liu, Xin Wang · 2025

Read Optimizer-Dependent Generalization Bound for Quantum Neural Networks by Chenghong Zhu, Hongshun Yao, Yingjian Liu, Xin Wang (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-2501-16228.txt.

Introduction

Chenghong Zhu, Hongshun Yao, Yingjian Liu, and Xin Wang present a 2025 paper on how well quantum neural networks generalize to new data. These networks are the quantum kin of deep learning models, and steady performance on fresh data is key to their use. Their generalization traits, though, had not been fully mapped.

The team studies the question through the stability of the learning process, which avoids scanning every possible model. A link between these networks and a quantum information tool lets them read behavior in a fresh way. Their error bound rests on the count of tunable values, data loading passes, data size, and the settings of the classical trainer. Computer tests backed the bound.

Catalog record

CREATORChenghong Zhu, Hongshun Yao, Yingjian Liu, Xin Wang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE78 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-16228

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Optimizer-Dependent Generalization Bound for Quantum Neural Networks Thanks: Important notice : A preliminary version of our paper was submitted to ICLR 2025 and was publicly available on OpenReview since October 4, 2024. It has come to our notice that an recent paper, arXiv:2501.12737 , which investigates the same problem, was submitted to arXiv on January 22, 2025. For reasons unknown to us, arXiv:2501.12737 contains substantial and inexplicable overlaps with our earlier version that has been accessible on OpenReview since October 4, 2024. Chenghong Zhu, Hongshun Yao, Yingjian Liu, Xin Wang Abstract Quantum neural networks (QNNs) play a pivotal role in addressing complex tasks within quantum machine learning, analogous to classical neural networks in deep learning. Ensuring consistent performance across diverse datasets is crucial for understanding and optimizing QNNs in both classical and quantum machine learning tasks, but remains a challenge as QNN’s generalization properties have not been fully explored. In this paper, we investigate the generalization properties of QNNs through the lens of learning algorithm stability, circumventing the need to explore the entire hypothesis space and providing insights into how classical optimizers influence QNN performance. By establishing a connection between QNNs and quantum combs, we examine the general behaviors of QNN models from a quantum information theory perspective. Leveraging the uniform stability of the stochastic gradient descent algorithm, we propose a generalization error bound determined by the number of trainable parameters, data uploading times, dataset dimension, and classical optimizer hyperparameters. Numerical experiments validate this comprehensive understanding of QNNs and align with our theoretical conclusions. As the first exploration into understanding the generalization capability of QNNs from a unified perspective of design and training, our work offers practical insights for applying QNNs in quantum machine learning. 1 Introduction Quantum computing leverages the laws of quantum mechanics to solve complex problems more efficiently than classical computers, offering notable quantum speedups in areas such as cryptography [ 1 ] , and quantum simulations [ 2 , 3 ] . Recent advancements in quantum hardware have demonstrated quantum advantages in specific tasks [ 4 , 5 , 6 ] , catalyzing the exploration of quantum computing’s potential in artificial intelligence. This interdisciplinary connection has given rise to quantum machine learning [ 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 ] . One of the leading frameworks in quantum machine learning is the quantum neural network (QNN), which represents the quantum analog of classical artificial neural networks and typically refers to parameterized quantum circuits that are trainable based on quantum measurement results. A well-known architecture within QNNs is the data re-uploading QNN [ 22 , 23 ] , which integrates multiple training and data encoding layers within a single quantum circuit. This approach significantly enhances the expressivity of the models, allowing them to approximate functions more effectively [ 22 , 24 , 25 , 26 , 27 , 28 ] . Such characteristics make the data re-uploading QNN a suitable quantum machine model for supervised learning tasks. Despite these advancements in quantum machine learning, critical challenges remain, particularly in understanding and predicting the performance of models in practical settings. A fundamental criterion for evaluating the performance of any learning algorithm is the generalization gap [ 29 ] . This gap essentially measures the difference between a model’s accuracy on training data and its expected accuracy on unseen data, providing essential theoretical guidance on determining the amount of training data required and informing the design of the model’s architecture to ensure robust generalization capabilities. QNNs typically update trainable parameters using classical optimizers, where classical optimizers play a crucial role in the training process by adjusting the parameters based on quantum measurement results. However, the integration of various classical optimizers with quantum processes introduces additional complexity, particularly in ensuring that these updates contribute positively to minimizing the generalization gap. Hence, providing theoretical guidance on the design and training of QNNs with regard to the strategic use of specific optimizers is of fundamental importance. Among these optimizers, stochastic gradient descent (SGD) algorithm is the most commonly used. Understanding the effects of SGD is vital as it not only enhances our comprehension of this optimizer but also informs the uses of its advanced variants in QNN applications. To deepen our understanding of QNNs, our work establishes theoretical guarantees on how classical optimizers impact their generalization performance. To summarize, our contributions include: Fig 1: Overview of our work. (a) We investigate the relationship between QNNs and quantum combs, highlighting how this connection informs our understanding of QNN dynamics. (b) We demonstrate that QNNs utilizing the SGD algorithm and trained for T iterations exhibit \beta_{m} -uniform stability. This stability metric quantifies the maximum change in the QNN’s output due to alterations in a single training example. (c) Leveraging the uniform stability, we derive a generalization error bound that assesses the QNN’s performance on unseen data, thereby better understanding the performance of QNNs in practical applications. • We introduce a more general perspective in the study of QNNs by conceptualizing them as a special form of trainable quantum combs (cf. Sec. 3.1 ). This approach allows us to leverage the rich theoretical framework of quantum combs to analyze the properties and dynamics of QNNs. • We then investigate the stability of QNNs based on the quantum comb perspective to show that QNNs are uniformly stable (cf. Fig. 1 ). We further establish an upper bound on generalization error for data reuploading model (cf. Sec. 3.3 ). This bound not only informs the control of QNN expressivity power through the number of layers and trainable parameters during the designing of QNNs, but also offers new insights into training QNNs with input data dimensions and classical optimizer-dependent parameters. Importantly, the bound introduces a new training guideline suggesting that the learning rate and the number of trainable gates should be designed to be inversely proportional to enable stable training. • To substantiate our theoretical claims, we have conducted extensive numerical experiments focused on assessing changes in the expressivity and stability of QNNs. These experiments underscores the importance of stable learning for the practical training of these networks and hence draw a guideline for future QNN developments. 1.1 Related Work Currently, the theory of generalization in QNNs mainly focuses on the complexity measures. [ 30 ] derives an upper bound for generalization error with the dependence of the number of trainable quantum gates and the operator norm of the observable by leveraging covering number [ 31 ] to quantify the expressivity of VQAs. Later, [ 32 ] uses quantum channels to derive more general results and [ 33 ] extends it to multi-class classification tasks. [ 34 ] uses the effective dimension [ 35 , 36 ] as another complexity measure for the parameter space of QNN. [ 37 ] uses Vapnik-Chervonenkis dimension [ 38 ] that investigates the balance of empirical and generalization performance on the dimension of inputs and the Frobenius norm of the observable. Besides, [ 39 ] , [ 40 , 41 ] and [ 42 ] derives an upper bound based on Rademacher complexity [ 43 ] , where the

Read the complete text