arXiv:2609.25246
Operator Score Matching for Learning Quantum Hamiltonians
Shreya Shukla, Abhijith Jayakumar, Andrey Y. Lokhov
Abstract
Learning quantum Hamiltonians from low-temperature thermal state measurements is a fundamental problem in quantum physics. Scalability of existing methods is limited by the complexity of semidefinite optimization problems or partition function computation. Here, we develop a quantum analog of classical score matching method that exploits generalized notions of derivatives and integration by parts in operator algebras. We introduce a new \emph{Operator Score Matching} loss function that recovers Hamiltonian parameters by a simple gradient descent without the need to compute intractable normalization constants. Numerical experiments on the Gibbs states of the transverse-field Ising model, XXZ spin chain, Fermi-Hubbard model, and lattice $φ^4$ theory demonstrate efficient parameter recovery using finite-depth truncations of nested commutator expansions. Our results establish Operator Score Matching as a practical, partition-function-free framework for quantum Hamiltonian learning.
AI Summary
The paper addresses learning a Hamiltonian $H_\theta=\sum_i\theta_i h_i$, with known operator terms and unknown coefficients, from measurements of its Gibbs state. It introduces operator score matching to avoid partition-function evaluation and replace semidefinite optimization with gradient-based minimization. The score along a probe operator $F_k$ is $\mathscr D_kH=[F_k,e^{-H}]e^H$. A squared score discrepancy, weighted by the target state, is rewritten using trace integration by parts into expectations involving only model operators. An algebraic commutant condition establishes that zero loss identifies the Hamiltonian up to an additive constant; single-site Pauli or clock–shift probes, and single-mode Majorana probes for fermions, provide sufficient choices. A positive-definite Jacobian Gram matrix establishes a local Polyak–Łojasiewicz inequality and linear gradient-descent convergence near the true parameters, with constants determined by conditioning and local smoothness. Implementation truncates nested-commutator expansions at depth $d$, producing a degree-$2d$ polynomial loss. Tests include a nonuniform 50-qubit transverse-field Ising chain using METTS-generated expectations, a four-site XXZ chain, a two-site Fermi–Hubbard model, and five-site lattice $\phi^4$ theory with local dimension eight. Reported behavior includes approximately $N^{-1/2}$ statistical errors, empirically size-independent required depth after finite-size saturation, and Hubbard parameter errors around $10^{-6}$. The exact loss can have spurious local minima, demonstrated numerically on two qubits. No global optimization or end-to-end finite-sample recovery guarantee is established. Controlling truncation bias, measurement noise, and commutator proliferation remains unresolved; infinite-dimensional applications require regularization.
Error found: The stated gradient-descent iteration bound omits the gradient-Lipschitz constant: with step size proportional to 1/L, the displayed contraction yields an O((L/mu) log(1/epsilon)) bound, up to the initial-loss factor.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.