⚛️ quantum physics

Operator Score Matching for Learning Quantum Hamiltonians

This paper introduces Operator Score Matching, a scalable and partition-function-free framework that learns quantum Hamiltonian parameters from low-temperature thermal states via gradient descent, effectively overcoming the computational bottlenecks of existing methods.

Original authors: Shreya Shukla, Abhijith Jayakumar, Andrey Y. Lokhov

Published 2026-09-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Shreya Shukla, Abhijith Jayakumar, Andrey Y. Lokhov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quantum world, the behavior of matter is governed by a mathematical blueprint known as a Hamiltonian. This blueprint dictates how particles interact, how energy flows, and how a system settles into its most stable state. For physicists, knowing this blueprint is essential. It allows them to predict how a material will conduct electricity, how a superconductor might function, or how a quantum computer will process information. However, there is a profound difficulty in this field: while scientists can easily predict the behavior of a system if they already know the blueprint, figuring out the blueprint itself by observing the system is incredibly hard. This is known as an inverse problem. When researchers look at a quantum system, they see the results of complex interactions, but peeling back those layers to find the specific rules that created them is like trying to reconstruct a recipe by tasting a finished cake without knowing the ingredients or the oven temperature.

The challenge is particularly acute when the system is cold. At low temperatures, quantum effects become dominant, leading to exotic phenomena like superconductivity. To learn the rules of these systems, scientists usually need to measure the system while it is in a thermal state, a condition where it has settled into a specific energy configuration. The problem is that calculating the exact probability of finding the system in any given state requires solving a massive mathematical hurdle called the partition function. This calculation is so complex that it is often impossible to perform, even for the most powerful classical computers. Consequently, many existing methods for learning quantum Hamiltonians are stuck; they are either too slow to be useful for large systems or they rely on approximations that break down when the temperature drops.

A team of researchers at Los Alamos National Laboratory has developed a new approach to bypass this bottleneck entirely. Instead of trying to calculate the impossible partition function, they adapted a technique from classical statistics called score matching. In the classical world, this method learns a system's rules by comparing how the probability of a state changes when you nudge it slightly, rather than trying to calculate the total probability itself. The researchers realized that this idea could be translated into the quantum realm, but doing so required a fundamental rethinking of what a "nudge" or a "derivative" means when dealing with quantum operators, which do not behave like ordinary numbers.

The team introduced a new concept called operator score matching. In this framework, they treat the quantum system not as a collection of probabilities, but as a set of operators—mathematical objects that act on the system's state. They defined a way to measure how the system's energy blueprint changes when it is perturbed by specific local probes. Imagine tapping a single point on a drum and listening to how the vibration spreads; in this quantum version, the researchers "tap" the system with specific mathematical operations and measure the resulting "score," or response. By comparing the response of their guessed model to the response of the actual physical system, they can adjust their model to minimize the difference. Crucially, this process does not require knowing the total energy of the system or calculating the intractable partition function. The method relies on a mathematical trick involving the cyclical nature of quantum traces, which allows the unknown parts of the equation to cancel out, leaving only the measurable differences between the model and reality.

To make this work in practice, the researchers had to solve a second problem: the quantum "score" involves complex nested interactions that can spread across the entire system. To handle this, they used a technique of truncation, essentially cutting off the calculation after a certain number of steps. They found that for many physical systems, the information needed to learn the rules is local. The influence of a single probe does not need to be tracked across the entire universe of the system; it only needs to be followed until it reaches the edge of the relevant interaction zone. By testing their method on several different models, including spin chains that mimic magnetic materials, models of electrons moving through a lattice, and theories of scalar fields, they demonstrated that this truncation works remarkably well. They showed that even with a relatively simple approximation, the method could recover the true parameters of the Hamiltonian with high precision.

The results of their numerical experiments were striking. When they applied the method to a one-dimensional chain of spins, a model for electrons in a material, and a lattice field theory, the error in their recovered parameters dropped rapidly as they increased the depth of their calculation. They found that the amount of calculation required depended on the strength of the interactions and the temperature, but not on the total size of the system. This means the method scales efficiently, making it suitable for larger, more complex systems that were previously out of reach. In one test involving a fermionic system, they used the natural language of fermions rather than converting them into a different format, proving the method is flexible enough to handle different types of particles without needing artificial translations.

The researchers also analyzed the mathematical landscape of their new loss function, which measures the error between the model and the data. They proved that near the correct answer, the function behaves in a way that guarantees a standard optimization algorithm will find the solution quickly and reliably. While the function is not perfectly smooth everywhere, it has a single, clear global minimum that corresponds to the true physical rules. In their simulations, the algorithm consistently found this minimum, suggesting that the method is robust and practical. This work establishes a new framework for learning quantum Hamiltonians that is both theoretically sound and computationally efficient. By removing the need for the partition function, it opens the door to understanding complex quantum materials and validating the performance of engineered quantum devices, offering a practical path forward for a field that has long been hindered by mathematical intractability.

Technical Summary: Operator Score Matching for Learning Quantum Hamiltonians

Problem Statement
The paper addresses the fundamental inverse problem of learning the microscopic parameters of a local quantum Hamiltonian ( $H'$ ) from experimental observations of its low-temperature thermal (Gibbs) state, $\rho \propto e^{-H'}$ . While forward simulation has advanced significantly, learning the Hamiltonian remains challenging. Existing methods face scalability bottlenecks: some require computing the partition function (which is intractable), while others, such as the recent polynomial-time algorithm by Bakshi et al. (BLMT), rely on high-degree Sum-of-Squares (SoS) relaxations. These relaxations lead to large semidefinite programs whose computational cost grows rapidly with system size, inverse temperature ( $\beta$ ), and target accuracy, limiting their practical utility in the low-temperature regime where quantum phenomena are most prominent.

Methodology: Operator Score Matching
The authors propose "Operator Score Matching," a quantum analog of the classical score matching technique used in statistics to estimate parameters without computing intractable normalization constants (partition functions).

Operator Score Definition:
In classical statistics, the score is the gradient of the log-probability, $\nabla \log p(x) = -\nabla H(x)$ , which is independent of the partition function. In the quantum setting, the authors define a generalized derivative operator based on commutators. For a direction operator $F_k$ , the "quantum score" of a model density $p_\theta = e^{-H_\theta}/Z(\theta)$ is defined as:
$D_k(-\log p_\theta) := [F_k, e^{-H_\theta}]e^{H_\theta}$
This definition eliminates the dependence on the partition function $Z(\theta)$ .
Loss Function Construction:
The authors introduce a loss function that minimizes the discrepancy between the model's operator score and the data's operator score:
$L_{\text{osm}}(\theta) = \sum_k \langle |D_k H_\theta - D_k H'|^2 \rangle_\rho$
Using a trace integration-by-parts identity (derived from the cyclicity of the trace), the term involving the unknown true Hamiltonian $H'$ is replaced by expectation values of known commutators:
$L_{\text{std}}(\theta) = \sum_k \langle |D_k H_\theta|^2 \rangle_\rho - 2 \text{Re} \sum_k \langle [F_k^\dagger, D_k H_\theta] \rangle_\rho$
This formulation allows the loss to be estimated using independent copies of the target density matrix without prior knowledge of the true parameters.
Truncation and Approximation:
The operator score involves an infinite series of nested commutators (via the Baker-Campbell-Hausdorff identity). The authors approximate this by truncating the series at a finite depth $d$ :
$D_k^{(d)} H_\theta \approx \sum_{\ell=1}^d \frac{(-1)^{\ell+1}}{\ell!} [H_\theta, F_k]_\ell$
This results in a polynomial loss function of degree $2d$ in the parameters $\theta$ , making gradient descent computationally feasible.
Direction Operators ( $F_k$ ):
To ensure the Hamiltonian is uniquely identifiable (up to an additive constant), the set of direction operators $\{F_k\}$ must generate the full matrix algebra of the system. The paper provides specific prescriptions:
- Qubits/Spins: Single-site Pauli $X$ and $Z$ operators.
- Fermions: Single-mode Majorana operators, allowing the method to work directly in Fock space without mapping to qubits.
- General Lattices: Generalized Pauli (clock-shift) operators.

Key Contributions

Partition-Function-Free Framework: The method recovers Hamiltonian parameters using simple gradient descent without requiring the computation of the partition function or solving large semidefinite programs.
Theoretical Guarantees: The authors establish a local Polyak–Łojasiewicz (PL) inequality near the true parameters. This guarantees that gradient descent converges exponentially fast to the global minimum, provided the step size is appropriate. The convergence rate is governed by the minimum eigenvalue of a Gram matrix.
Identifiability: The paper proves that if the direction operators generate the full algebra, the loss function has a unique global minimum corresponding to the true Hamiltonian (up to an energy shift).
Universality: The framework is agnostic to particle statistics, applicable to qubits, spins, bosons, and fermions (using Majorana operators) without requiring Jordan-Wigner transformations.

Results
The authors validate the framework numerically on four distinct models:

Transverse-Field Ising Model (TFIM): For a 50-qubit non-uniform chain, the method successfully recovers site-dependent couplings. The required truncation depth $d^*$ to achieve a target error depends on the inverse temperature $\beta$ and interaction strength but is independent of the chain length $L$ once $L$ exceeds the light cone of the nested commutators.
XXZ Spin Chain: Parameter recovery is demonstrated across different physical regimes (from Heisenberg to Ising limits). The required depth scales linearly with $\beta$ .
Fermi-Hubbard Model: The method recovers hopping ( $t$ ), interaction ( $U$ ), and chemical potential ( $\mu$ ) parameters directly in Fock space using Majorana directions, confirming the agnosticism to particle statistics.
Lattice $\phi^4$ Theory: The framework is extended to lattice field theories. By representing field operators in a truncated harmonic oscillator basis, the authors demonstrate successful parameter recovery for mass, coupling, and interaction terms.

In all cases, the error decreases with increasing truncation depth until it saturates due to statistical noise from finite sampling. The authors observe that for typical parameter initializations, the loss landscape often possesses a single minimum, making optimization numerically stable.

Significance and Claims
The paper claims to establish Operator Score Matching as a "practical, partition-function-free framework" for quantum Hamiltonian learning. Its primary significance lies in offering a scalable alternative to SoS-based methods for the low-temperature regime, where existing polynomial-time algorithms become computationally prohibitive due to the degree of the required relaxations. By leveraging non-commutative integration-by-parts identities, the method avoids the fundamental bottleneck of partition function computation. The authors note that while their proof of the PL inequality is local, numerical experiments suggest the favorable optimization landscape extends over large radii. They position this work as a complement to existing approaches (Bayesian inference, reduced density matrices, time evolution) by operating directly on thermal states without requiring coherent dynamics. The paper concludes by identifying open problems, such as a complete non-asymptotic error analysis and extensions to continuum quantum field theories.

More like this