⚛️ quantum physics

PINNverse meets Jordan--Wigner: learning the Hamiltonian of long spin chains

This paper demonstrates that applying the Jordan-Wigner transformation to inverse physics-informed neural networks (PINNverse) eliminates the exponential scaling of learning quantum Hamiltonians, enabling the accurate recovery of parameters for spin chains of up to 50 qubits using only local observables and $N^2$ network outputs.

Original author: Gubio G De Lima

Published 2026-11-11

📖 5 min read🧠 Deep dive

CC0 1.0

Original author: Gubio G De Lima

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

In the quest to build better quantum computers, scientists face a fundamental puzzle: how to know exactly what a machine is doing. A quantum computer is built from tiny units called qubits, which can exist in multiple states at once. To make these machines work, engineers must calibrate them with extreme precision, ensuring that the forces connecting the qubits behave exactly as intended. The problem is that as you add more qubits, the number of things you need to measure grows so fast that it quickly becomes impossible to check them all. Traditional methods require so many different measurement settings that for a machine with just twenty qubits, the task would take longer than the age of the universe to complete. This bottleneck has long stood in the way of scaling up quantum technology, leaving researchers with a device they cannot fully understand or control.

A team of physicists has found a way to bypass this impossible barrier for a specific and common type of quantum machine. They developed a new method that uses a smart computer program to learn the hidden settings of a long chain of qubits, even when that chain contains fifty units. The breakthrough relies on a clever mathematical trick that changes how the problem is viewed. Instead of trying to track every single possible state of the entire chain, which is an exponentially growing nightmare, the researchers realized they could describe the system's behavior using a much smaller, manageable set of numbers. By treating the qubits as if they were a stream of particles moving along a line, they reduced the complexity of the problem from something that explodes in size to something that grows slowly and steadily. This allowed them to use a neural network—a type of artificial intelligence designed to find patterns in data—to solve the equations of motion for the system and then work backward to discover the exact strengths of the connections between the qubits.

The researchers tested their approach using synthetic data, which is computer-generated information that mimics what a real experiment would produce. They simulated chains of qubits with varying strengths of connection and local frequencies, including the small, unwanted interactions that often plague real hardware. The neural network was trained to predict how the system would evolve over time based on a few simple measurements taken at different moments. Remarkably, the network not only learned to predict the future behavior of the chain but also successfully identified the underlying parameters that govern it. For chains of up to fifty qubits, the method recovered the connection strengths with an accuracy better than one-tenth of a percent. This level of precision was achieved using only five global measurement settings, regardless of how long the chain was, whereas older methods would have required millions or even billions of settings for a chain of that size.

One of the most significant aspects of this work is how it handles the messy reality of quantum hardware. Real devices are never perfect; they suffer from "crosstalk," where a qubit unintentionally influences its neighbor in a way that wasn't programmed. The researchers incorporated a technique called a time-dependent Hartree–Fock closure into their model. In simple terms, this is a way of approximating how these unwanted interactions affect the system without having to calculate every single detail, which would be too slow. The method proved robust enough to detect and quantify these parasitic couplings, identifying them with high accuracy even when they were very weak. This is crucial because knowing the exact nature of these errors is the first step toward fixing them.

The study also revealed a subtle challenge that arises as the chains get longer. While the method could pinpoint the strength of the connections between neighbors with extreme precision, the local frequencies of the qubits themselves accumulated small errors over the length of the chain. To solve this, the researchers combined their neural network with a traditional statistical fitting technique. The neural network provided a highly accurate starting point, and the traditional method refined the result, removing the accumulated errors. This hybrid approach allowed them to recover the parameters for a fifty-qubit chain with errors as low as seven-tenths of a percent, even when the data included realistic levels of noise.

This work demonstrates that it is possible to learn the Hamiltonian—the mathematical description of the energy and interactions of a system—for large quantum devices without needing an exponential amount of data. The method is not a magic wand that solves every quantum problem; it is specifically designed for chains where the qubits interact primarily with their immediate neighbors. However, for the superconducting circuits that are currently leading the race to build practical quantum computers, this limitation is not a barrier but a feature. These devices naturally form chains of exchange-coupled qubits, making the new technique directly applicable to the hardware being built today. By turning an impossible measurement problem into a solvable learning task, the researchers have provided a powerful tool for calibrating the next generation of quantum simulators, bringing us closer to machines that can truly harness the power of quantum mechanics.

Technical Summary: PINNverse meets Jordan–Wigner: learning the Hamiltonian of long spin chains

Problem Statement
Determining the Hamiltonian of a quantum device from measured expectation values is a critical task for calibrating quantum hardware and simulating quantum systems. Traditional methods, such as process tomography, require a number of measurement settings that grows exponentially with the number of qubits ( $N$ ). While Physics-Informed Neural Networks (PINNverse) have been applied to inverse problems by learning unknown coefficients of differential equations, standard formulations face a fundamental scalability barrier. In these approaches, the neural network must represent the full quantum state (density matrix $\rho$ ) or the complete set of Pauli expectation values, both of which scale as $4^N$ . For $N=20$ , this requires approximately $10^{12}$ outputs, rendering the approach intractable for large registers. Existing strategies, such as "freezing" mechanisms that learn qubits pairwise, circumvent this by requiring many independent experiments and additional control, failing to learn a whole chain simultaneously with polynomial scaling.

Methodology
This work proposes a reformulation of PINNverse for one-dimensional spin chains with nearest-neighbor exchange interactions, leveraging the Jordan–Wigner (JW) transformation to eliminate exponential scaling.

Jordan–Wigner Mapping: The spin-1/2 chain is mapped to a system of spinless fermions. For the inhomogeneous XX chain in a longitudinal field (the target model), the Hamiltonian becomes quadratic in fermionic operators. Consequently, the Heisenberg equations of motion close exactly on the $N \times N$ one-body density matrix, $\rho^{(1)}(t)$ , rather than the full $2^N \times 2^N$ density matrix.
Neural Network Architecture: Instead of outputting $4^N - 1$ Pauli expectation values, the neural network takes time $t$ as input and outputs the $N^2$ real entries of the one-body density matrix $\rho^{(1)}_{NN}(t)$ . The network is constructed to satisfy the initial condition exactly (a hard constraint) and to output a Hermitian matrix.
Loss Function: The training loss combines a data term and a physics term:
- Data Term ( $L_{data}$ ): Compares the network's predicted local observables (magnetizations $\langle \sigma^z_j \rangle$ and nearest-neighbor correlations/currents) against measured data. Crucially, these observables can be acquired using only five global measurement settings (independent of $N$ ), whereas reconstructing all Pauli values would require $3^N$ settings.
- Physics Term ( $L_{Physics}$ ): Enforces the equation of motion. For the XX chain, this is the Liouville–von Neumann equation for the one-body matrix: $d\rho^{(1)}/dt = -i[K, \rho^{(1)}]$ , where $K$ is the single-particle Hamiltonian matrix containing the unknown couplings $J_j$ and fields $h_j$ .
Handling Crosstalk (ZZ Couplings): To address residual ZZ interactions (parasitic couplings) common in superconducting qubits, the author introduces a time-dependent Hartree–Fock (TDHF) closure. This approximates the quartic interaction terms as quadratic, resulting in a nonlinear equation of motion where the error scales as $O(\Delta^2)$ for weak crosstalk $\Delta$ .
Optimization Strategy: The inverse problem is solved in two stages. First, the Adam optimizer trains the network and parameters. Second, the L-BFGS quasi-Newton method refines the solution on a dense grid. To mitigate error accumulation in long chains, the network's output serves as a high-quality initial guess for a direct least-squares fit, which removes accumulated biases in local frequencies.

Key Contributions

Polynomial Scaling Reformulation: The paper demonstrates that for nearest-neighbor exchange chains, the exponential cost of PINNverse can be reduced to polynomial scaling ( $N^2$ network outputs) without approximation for the XX model.
Minimal Measurement Protocol: It identifies a minimal set of local observables (magnetizations and spin currents) that renders all hopping amplitudes and field differences identifiable using only five global measurement settings for any chain length $N$ .
Crosstalk Detection: A TDHF extension is developed to treat ZZ crosstalk, allowing the simultaneous learning of exchange couplings, local fields, and parasitic ZZ strengths.
Scalable Demonstration: The method successfully learns Hamiltonian parameters for chains up to $N=50$ qubits, a regime far beyond the reach of previous PINNverse formulations.

Results
Using synthetic data generated from exact spin dynamics:

Forward Problem: The network accurately solves the equation of motion for the one-body density matrix, achieving maximum deviations from the reference solution on the order of $10^{-4}$ to $10^{-7}$ depending on $N$ and the presence of crosstalk.
Inverse Problem (XX Chain): For noise-free data, the network recovers exchange couplings ( $J_j$ ) with a Mean Absolute Percentage Error (MAPE) below 0.1% for chains up to $N=50$ . Local field differences ( $h_j - h_1$ ) are learned with high accuracy, though absolute field values accumulate small errors over long chains.
Noise Robustness: With measurement noise ( $\sigma=0.01$ ), the network maintains accuracy comparable to direct least-squares fits.
Hybrid Approach: A "PINN + fit" strategy, where the network estimate initializes a direct fit, eliminates the accumulated frequency errors in long chains. For $N=50$ with noise, this hybrid approach achieves MAPEs of 0.047% for couplings and 0.71% for fields.
Crosstalk Detection: The method successfully detects and quantifies ZZ couplings ( $\Delta_j$ ) in $N=8$ chains, with the TDHF model showing significantly lower misfit errors compared to a model neglecting crosstalk.

Significance and Claims
The author claims that this work provides the first formulation of PINNverse that learns a whole spin chain at once with network outputs growing polynomially with $N$ . By exploiting the exact closure of the dynamics on the one-body density matrix via the Jordan–Wigner transformation, the method overcomes the exponential barrier that has limited inverse physics-informed learning to small systems (typically $N \le 10$ ).

The paper emphasizes that this approach is not merely an academic exercise but is directly applicable to superconducting transmon qubit chains, which naturally implement the inhomogeneous XX model. The ability to calibrate parameters and detect crosstalk using only five global measurement settings offers a practical pathway for characterizing large-scale analog quantum simulators. The author modestly notes that while the method is exact for the XX model, extensions to anisotropic XY chains (requiring Majorana covariance matrices) and open systems (Lindblad dynamics) are natural next steps, while general two-body Hamiltonians remain outside this specific scope.

Technical Summary: PINNverse meets Jordan–Wigner: learning the Hamiltonian of long spin chains

More like this