Frequency principle for quantum machine learning via Fourier analysis

Abstract: Quantum machine learning is one of the most exciting potential applications of quantum technology. While under intensive studies, the training process of quantum machine learning is relatively ambiguous and its quantum advantages are not very completely explained. Here we investigate the training process of quantum neural networks from the perspective of Fourier analysis. We empirically propose a frequency principle for parameterized quantum circuits that preferentially train frequencies within the primary frequency range of the objective function faster than other frequencies. We elaborate on the frequency principle in a curve fitting problem by initializing the parameterized quantum circuits as low, medium, and high-frequency functions and then observing the convergence behavior of each frequency during training. We further explain the convergence behavior by investigating the evolution of residues with quantum neural tangent kernels. Moreover, the frequency principle is verified with the discrete logarithmic problem for which the quantum advantage is provable. Our work suggests a new avenue for understanding quantum advantage from the training process.