Abstract
Quantum neural networks (QNNs) have emerged as a leading strategy toestablish applications in machine learning, chemistry, and optimization. Whilethe applications of QNN have been widely investigated, its theoreticalfoundation remains less understood. In this paper, we formulate a theoreticalframework for the expressive ability of data re-uploading quantum neuralnetworks that consist of interleaved encoding circuit blocks and trainablecircuit blocks. First, we prove that single-qubit quantum neural networks canapproximate any univariate function by mapping the model to a partial Fourierseries. Beyond previous works' understanding of existence, we in particularestablish the exact correlations between the parameters of the trainable gatesand the working Fourier coefficients, by exploring connections to quantumsignal processing. Second, we discuss the limitations of single-qubit nativeQNNs on approximating multivariate functions by analyzing the frequencyspectrum and the flexibility of Fourier coefficients. We further demonstratethe expressivity and limitations of single-qubit native QNNs via numericalexperiments. As applications, we introduce natural extensions to multi-qubitquantum neural networks, which exhibit the capability of classifying real-worldmulti-dimensional data. We believe these results would improve ourunderstanding of QNNs and provide a helpful guideline for designing powerfulQNNs for machine learning tasks.