arXiv:2610.26833
Supervised Learning with Quantum Hamiltonian Embeddings
Riccardo Molteni, Minjun Jeon, Paolo Braccia, Abhinav Deshpande, Vedran Dunjko, Liang Jiang, Su-un Lee, Swarnadeep Majumder, Javier Robledo Moreno, Parthasarathy Suryanarayanan, Kunal Sharma
Abstract
Machine learning is widely regarded as a promising application in which quantum computing could offer a computational advantage. However, general quantum machine learning models with provable guarantees on performance, trainability, and efficient inference remain limited. In this work, we introduce quantum Hamiltonian embeddings, a supervised learning framework in which a classical training dataset is encoded into a single effective, parameterized Hamiltonian. Training optimizes the parameters of this data-dependent Hamiltonian, whose approximate ground state defines the classification rule. Once the ground state has been approximated using a quantum algorithm, inference can be performed either classically using a classical representation of the learned state, or with a quantum representation of the state. We establish efficient methods for gradient estimation and prove favorable generalization guarantees. Furthermore, under standard cryptographic assumptions, we show that Hamiltonian embeddings can realize hypothesis functions beyond the reach of efficient classical algorithms. We then demonstrate the framework at scale on IBM Heron processors using a standard benchmark dataset. Together, these results establish Hamiltonian embeddings as a new framework for quantum machine learning, combining theoretical guarantees with a path for experimental investigations at scale on quantum hardware.
AI Summary
The paper asks whether supervised quantum learning can combine generalization guarantees, accessible gradients, and classical inference while retaining quantum computational advantages. It introduces a data-dependent Hamiltonian $H_{\mathcal T}(\alpha)=I-N^{-1}\sum_\ell U^\dagger(x_\ell,\alpha)O(y_\ell)U(x_\ell,\alpha)$. Training alternates approximate ground-state preparation with updates to the encoding parameters; the learned state defines label scores through dressed-observable expectation values. Matrix Hoeffding concentration and parameter-space covering arguments give classification risk $R\leq2\widehat C+\varepsilon$ with sample complexity $N=\widetilde O((n+M+\log(1/\delta))/\varepsilon^2)$, assuming bounded, Lipschitz Hamiltonian terms. Hellmann–Feynman and parameter-shift identities enable frozen-state gradient estimation. A conditional barren-plateau result bounds coordinate-wise gradient variance for a subset of final-layer parameters once energy is sufficiently below $1/2$; it does not establish optimization convergence. Shallow local encodings permit classical-shadow-based inference. Constructed tasks exhibit sample-efficiency gains over particular classical surrogates, and a discrete-cube-root construction gives a quantum–classical learning separation under a cryptographic assumption. Binary Fashion-MNIST experiments use 448 training examples and models up to 50 qubits. Classical ground-state approximations reach roughly 92.4% mean accuracy at intermediate sizes, versus 94.6% for the best classical baseline. Five 50-qubit hardware runs use sample-based Krylov diagonalization, with initializations and splits selected through classical validation runs. Efficient ground-state preparation is not guaranteed generally. Some shallow-model relaxations admit efficient classical inner optimization and comparable generalization bounds. Advantages on generic datasets, suitable Hamiltonian families, and benefits beyond cryptographically constructed tasks remain open.
Error found: The abstract describes hypothesis functions beyond efficient classical algorithms, whereas the cryptographic construction establishes hardness of learning a classifier that is efficiently classically evaluable once its learned key is supplied. The statement that the barren-plateau theorem guarantees a useful training direction overinterprets a frozen-state, coordinate-resampling variance bound that does not control gradients at the current optimization parameters.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 59 pages, 7 figures