⚛️ quantum physics

Revealing Physical Redundancy in the Two-dimensional Fermi-Hubbard Model via Transferable Observable Reconstruction

This study demonstrates that distinct observables in the two-dimensional Fermi-Hubbard model—specifically total density, double occupancy, and spin-spin correlation—contain substantial mutually transferable physical information, as evidenced by a neural-network framework capable of accurately reconstructing one observable's phase diagram from another.

Original authors: Ao-Ning Wang, Min-Quan He, Z. D. Wang

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ao-Ning Wang, Min-Quan He, Z. D. Wang

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 microscopic world of quantum materials, scientists often look for patterns in how electrons behave when they are crowded together. Imagine a grid of tiny traps, like a chessboard, where electrons hop from square to square. When these particles interact strongly, they can form strange new states of matter, such as superconductors or magnetic insulators. To understand these complex systems, researchers measure specific quantities, like how many electrons are in a given area, how often two electrons share the same spot, or how their spins align with their neighbors. For decades, it was assumed that these different measurements offered distinct, non-overlapping windows into the system's behavior. Each was thought to reveal a unique piece of the puzzle, and to get the full picture, one had to measure them all separately.

A team of researchers has now challenged this assumption by asking a simple but profound question: if you know one of these measurements perfectly, can you predict the others? They focused on a famous theoretical model used to describe these crowded electron systems. Using powerful computer simulations and a type of artificial intelligence designed to find hidden patterns, they tested whether the information contained in one observable could be used to reconstruct the others. Their findings suggest that these seemingly independent measurements are actually deeply intertwined, carrying a surprising amount of shared information. This discovery implies that the complex behavior of these quantum systems might be governed by a simpler, underlying structure that connects all these different views, potentially offering a new way to study materials without needing to measure every single detail.

The researchers began by setting up a digital simulation of a small, two-dimensional grid where electrons could move and interact. In this virtual environment, they tracked three key properties: the total number of electrons, the frequency with which two electrons occupied the same spot, and the alignment of their magnetic spins. They generated thousands of scenarios, varying the temperature and the strength of the interactions between the electrons, to create a vast library of data. For each scenario, they recorded what the system looked like according to each of the three measurements, creating detailed maps that showed how these properties changed across different conditions.

To test if these maps were truly independent, the team trained a neural network, a computer program capable of learning complex relationships, to perform a specific task. They fed the network the map of one property, such as the total number of electrons, and asked it to predict the map of a different property, like the spin alignment. The network was not given any direct formulas or rules; it had to learn the connection purely from the data. If the properties were truly distinct and unrelated, the network would fail to make an accurate prediction. However, if the properties shared a deep physical connection, the network should be able to reconstruct the missing map with high precision.

The results were striking. The neural network successfully reconstructed the phase diagrams of one observable from another with an accuracy that was nearly as good as if it had been given the target observable itself as the input. In other words, knowing the total density of electrons allowed the computer to predict the spin patterns almost as well as if it had measured the spins directly. This held true even when the researchers tried to trick the system by scrambling the data labels or adding random noise. When the labels were shuffled so that the input and output no longer corresponded to the same physical state, the network's performance collapsed immediately, proving that it was learning real physical connections rather than just memorizing numbers. Similarly, when small amounts of noise were added to the input, the predictions remained robust, suggesting the relationship was based on broad, smooth physical trends rather than fragile, specific details.

The study also explored how this connection held up under different conditions. The researchers found that the ability to predict one property from another remained strong across a wide range of temperatures. However, the connection did weaken in specific regions where the system underwent dramatic changes, known as phase transitions. In these critical zones, where the material shifts from one state to another, the information becomes more complex and the redundancy between the measurements decreases. This makes sense, as these are the moments when the system is most sensitive and the distinct behaviors of the electrons are most pronounced.

To ensure these findings were not just a quirk of the tiny grid they used, the team tested their models on larger, more complex lattices. Even when the system size increased, the ability to transfer information between different measurements persisted. This suggests that the redundancy is not a local accident but a fundamental feature of how these quantum systems organize themselves. The researchers propose that these different observables are not separate entities but rather different projections of a single, higher-dimensional structure of physical information.

This work offers a new perspective on how we might study quantum materials in the future. In many experiments, measuring certain properties is extremely difficult or requires expensive, delicate equipment, while others are relatively easy to observe. If these properties are indeed so deeply connected, scientists might be able to infer the hard-to-measure quantities by simply measuring the easy ones. This could streamline the process of exploring new materials and understanding the complex dance of electrons that gives rise to the exotic behaviors of the quantum world. The study does not claim to have solved the mysteries of these systems, but it reveals that the information they hold is more interconnected than previously thought, suggesting that nature may be more economical in its description than we imagined.

Technical Summary: Revealing Physical Redundancy in the Two-Dimensional Fermi-Hubbard Model via Transferable Observable Reconstruction

Problem Statement
The Fermi-Hubbard model serves as a fundamental framework for studying strongly correlated quantum matter, particularly Mott physics. Traditionally, distinct physical observables—specifically total density ( $N$ ), double occupancy ( $D$ ), and spin-spin correlation ( $S$ )—are used to probe different aspects of the system (charge, interaction, and spin). While these observables are theoretically expected to be correlated, deriving one observable directly from another is generally considered analytically intractable. The central question addressed in this work is whether these seemingly distinct observables contain mutually transferable physical information beyond their apparent distinctions, and if so, to what extent.

Methodology
The authors employ a numerical framework utilizing a neural network to perform transferability tests among the three target observables.

System and Hamiltonian: The study focuses on a 2D Fermi-Hubbard model. The primary dataset is generated from a $2 \times 2$ single-band lattice (four sites), which serves as the basic building block of a 2D system. The Hamiltonian includes hopping ( $t$ ), local Coulomb interaction ( $U$ ), and chemical potential ( $\mu$ ). To facilitate simulation, the fermion-operator Hamiltonian is converted into a spin-model Hamiltonian using the Jordan-Wigner transformation.
Observables: The three observables are defined as:
1. Total Density ( $N$ ): A local measurement of the total number of electrons.
2. Double Occupancy ( $D$ ): The number of sites fully occupied by two opposite-spin electrons (local interaction).
3. Spin-Spin Correlation ( $S$ ): A staggered longitudinal spin-spin correlation involving neighboring sites (long-range interaction).
Dataset Generation: Data points are expectation values calculated from the system's Gibbs state ( $\rho(T) = e^{-\hat{H}/T} / \text{Tr}(e^{-\hat{H}/T})$ ) across a range of temperatures ( $T$ ), interaction strengths ( $U$ ), and chemical potentials ( $\mu$ ). Each data sample consists of phase diagrams where $U$ and $\mu$ are axes, and the observable value is the color map.
Neural Network Architecture: A Multi-Layer Perceptron (MLP) is used to learn a mapping function. The architecture includes:
- A trainable encoder ( $E_\phi$ ) that processes the input observable's phase diagram.
- A local concatenation step to capture smooth spatial trends by combining a vector with its eight nearest neighbors on the phase diagram grid.
- A trainable decoding function ( $f_\theta$ ) that reconstructs the target observable's phase diagram. To ensure spatial continuity, the reconstruction at any target coordinate uses a weighted average of the four nearest input vectors based on relative area.
Training Protocol: The network is trained to reconstruct observable $B$ given observable $A$ as input. Crucially, the training is supervised using the input observable $A$ 's own phase diagram alongside the target $B$ to prevent the network from "getting lost," a process verified to contain no information leakage.
Validation: The study employs shuffle tests (destroying physical labels between $A$ and $B$ ) and noise tests (adding Gaussian noise to inputs) to ensure that successful mappings rely on physical redundancy rather than numerical fluctuations or overfitting.

Key Contributions and Results

High Transferability: The primary finding is that the phase diagram of one observable can be reconstructed from another with accuracy approaching that of self-reconstruction benchmarks. For instance, reconstructing $S$ from $N$ yields Peak Signal-to-Noise Ratio (PSNR) levels comparable to reconstructing $S$ from $S$ .
Bidirectional Information Flow: The transferability holds regardless of the information hierarchy (e.g., from local $N$ to long-range $S$ , or vice versa). This suggests that the observables reside on a common higher-order many-body manifold where information is shared bidirectionally.
Robustness: The mappings remain robust under noisy inputs (up to $\sigma=0.3$ ) and persist across finite-temperature regimes.
Temperature and Phase Dependence:
- Performance degrades as the system approaches the ground state (low temperature), likely due to sharper phase transitions making discrimination harder. However, the gap between self-mapping and transferred mapping remains small across the temperature range.
- Redundancy is strongest in "trivial" phase regimes with smooth physical information.
- In critical regimes (e.g., Mott crossover, antiferromagnetic onset), the transferability weakens, and error increases significantly, indicating that distinct physical information rises rapidly in these regions, decoupling the observables.
Finite-Size Generalization: To address finite-size effects, the authors tested the learned models on larger lattices ( $2 \times 3$ , $2 \times 4$ , and $3 \times 3$ ) using the Finite-Temperature Lanczos Method (FTLM) to generate sample data. The results indicate that the physical redundancy and transferability persist across plaquette boundaries and larger system architectures, suggesting the effect is not strictly local to the $2 \times 2$ plaquette.

Significance and Claims
The paper claims to provide numerical evidence for "observable-level redundancy" in the two-dimensional Fermi-Hubbard system. The authors argue that separate observables carry a substantial fraction of one another's physical information, challenging the notion that they are merely correlated but distinct.

From a theoretical perspective, the results suggest that these observables may be viewed as different lower-order projections of a shared higher-order many-body manifold. From an experimental perspective, the authors propose that if such redundancy is identified and exploited, it could allow researchers to bypass the measurement of difficult-to-evaluate observables (like double occupancy) by inferring them from easier-to-measure ones (like total density) in quantum simulation experiments, provided the system is not in a highly critical regime.

The authors remain modest regarding the scope, acknowledging that the study is limited to finite-size systems due to computational complexity and that further theoretical work is needed to fully clarify the mechanism and limiting factors of this redundancy in the thermodynamic limit.

Technical Summary: Revealing Physical Redundancy in the Two-Dimensional Fermi-Hubbard Model via Transferable Observable Reconstruction

More like this