Spectral methods: crucial for machine learning, natural for quantum computers?

Abstract

This article presents an argument for why quantum computers could unlock new methods for machine learning. We argue that spectral methods, in particular those that learn, regularise, or otherwise manipulate the Fourier spectrum of a machine learning model, are often natural for quantum computers. For example, if a generative machine learning model is represented by a quantum state, the Quantum Fourier Transform allows us to manipulate the Fourier spectrum of the state using the entire toolbox of quantum routines, an operation that is usually prohibitive for classical models. At the same time, spectral methods are surprisingly fundamental to machine learning: A spectral bias has recently been hypothesised to be the core principle behind the success of deep learning; support vector machines have been known for decades to regularise in Fourier space, and convolutional neural nets build filters in the Fourier space of images. Could, then, quantum computing open fundamentally different, much more direct and resource-efficient ways to design the spectral properties of a model? We discuss this potential in detail here, hoping to stimulate a direction in quantum machine learning research that puts the question of ``why quantum?'' first.

Can quantum computers support useful machine-learning regularisation by manipulating model spectra more directly than classical methods? This perspective, tutorial and review connects Fourier-based simplicity biases in learning with quantum Fourier transforms (QFTs), rather than establishing an end-to-end learning advantage. It develops group Fourier analysis for continuous, binary and permutation-valued data, explaining how low-order spectra encode smoothness, moments, interaction effects or permutation marginals. It relates these ideas to stationary kernels, maximum mean discrepancy, convolutional networks and neural-network spectral bias, then reviews quantum approaches involving Born machines, data-encoding circuits, permutation models and resource-theoretic decompositions. A worked construction prepares a uniform superposition of training bitstrings, applies Hadamards, filters amplitudes by Hamming weight, transforms back and samples. The key distinction is that QFTs manipulate amplitudes, not probabilities: the probability spectrum is an autocorrelation, $\hat p(k)=2^{-n/2}\sum_s\hat\psi(s)\hat\psi^*(s+k)$ on $\mathbb Z_2^n$. The example filter $(1-2\theta)^{|k|}$ applied to probabilities is also classically implementable by sampling a training point and independently flipping bits with probability $\theta$, illustrating a dequantisation route. The proposed advantages remain conditional. Nonunitary filtering incurs postselection costs; useful unitary filters and scalable training remain open problems. Efficient QFTs neither expose all coefficients nor establish useful learning speedups, and amplitude manipulation need not preserve desired probability-spectrum properties. The discussion also identifies classical kernel competitors, measurement-access restrictions and the absence of large-scale benchmarking as obstacles to demonstrating practical benefit.