arXiv:2609.28391
Quantum score matching with applications to learning thermal states
Yulong Dong, Jiaqi Leng
Abstract
Score matching has driven major advances in classical generative learning by enabling models to learn from data without evaluating intractable normalization constants, or partition functions. Yet, extending this principle to quantum learning requires rethinking its foundations, as quantum states are described by noncommuting density operators rather than scalar probabilities. The noncommutativity creates fundamental challenges not only in defining quantum scores, but also in developing a training framework with efficient circuit implementations and rigorous theoretical guarantees. In this work, we bridge this gap by establishing a general quantum score-matching framework with end-to-end theoretical guarantees. Applied to Gibbs-state learning, our approach avoids additional thermal-state preparation and achieves information-theoretically optimal sample complexity in the high-temperature regime for Hamiltonians with bounded locality and interaction degree. This positions score matching as a new route to state-of-the-art performance in learning quantum Gibbs states. Beyond these theoretical results, numerical simulations show that our method remains effective even when gradients are estimated inaccurately under limited measurement budgets. Experiments on IBM quantum hardware further demonstrate that quantum score matching is NISQ-friendly: without any error mitigation or correction, it reduces the relative Hamiltonian-parameter error from 64% to approximately 10%. Together, these results extend score matching into an experimentally realizable paradigm for quantum-state learning.
AI Summary
The paper asks whether score matching can learn quantum states directly from copies without evaluating partition functions or preparing the current model’s Gibbs state. It defines operator-valued scores through commutator derivations and logarithmic derivatives, focusing on the symmetric logarithmic derivative. A quantum Hyvärinen identity rewrites squared score discrepancy as an expectation of a model-dependent observable on the target state. A commutant criterion supplies sufficient conditions for global identifiability. For Gibbs models $H(\theta)=\sum_{j=1}^m\theta_jP_j$, randomized real-time evolution and parameter-shift circuits estimate training gradients at any finite inverse temperature $\beta$, with polynomial copy cost and maximum evolution time $O(\beta\log(\operatorname{poly}(n,\beta)/\eta))$ for gradient accuracy $\eta$. For bounded-locality, bounded-overlap Hamiltonians at sufficiently high temperature, the single-qubit frame $\{X_x,Z_x\}$ suffices. Lieb–Robinson bounds establish quasi-local gradient observables, enabling shared-copy measurements; Hessian bounds establish convergence of preconditioned projected gradient descent. The resulting coefficient-learning guarantee is $\|\widehat\theta-\theta^*\|_\infty\leq\varepsilon$ with probability $1-\delta$, using $O(\beta^{-2}\varepsilon^{-2}\log(2m/\delta))$ copies, matching minimax bounds in the stated regime. Eight-qubit simulations examine finite-shot training. A four-qubit, two-parameter IBM demonstration reaches approximately 10% relative parameter error at $\beta=0.2$ after 45 updates, using 8,192 shots per update without error mitigation. The sample-optimal measurement theorem does not bound measurement-synthesis cost. General low-temperature efficiency is not guaranteed, and copy requirements can grow exponentially with $\beta$. Applying the identities beyond Gibbs families requires separate score representations.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 57 pages, 5 figures, 3 tables, with an accompanying GitHub repository at https://github.com/dongsnaq/Quantum-Score-Matching