⚛️ quantum physics

The stationarity test: a framework for learning quantum many-body systems from their thermal states

This paper introduces a rigorous, time-efficient "stationarity test" algorithm that learns the interaction graph and coefficients of lattice Hamiltonians from thermal and metastable states by measuring the rate-of-change of local observables under detailed-balanced quantum Markov chains, while simultaneously refining the area law for thermal metastable states.

Original authors: Thiago Bergamaschi

Published 2026-10-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Thiago Bergamaschi

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, matter is not made of static, silent particles but of complex, interacting systems where every piece influences its neighbors. To understand how these systems behave, scientists rely on a mathematical blueprint called a Hamiltonian. This blueprint lists every interaction between the particles, dictating how energy flows and how the system settles down. When a quantum system is left alone to cool, it eventually reaches a state of thermal equilibrium, known as a Gibbs state. In this state, the system has settled into a predictable pattern of behavior, much like a cup of coffee cooling to room temperature. The central challenge for physicists is to work backward: if they can only observe the final, cooled-down state of a system, can they figure out the exact blueprint that created it? This is the problem of learning the Hamiltonian. It is crucial for calibrating quantum devices and understanding the fundamental laws of condensed matter, yet it has been notoriously difficult because quantum systems do not follow the simple rules of classical physics, and the data required to solve the puzzle is often incomplete or noisy.

A researcher has now introduced a new, streamlined method to solve this puzzle, one that works even when the system has not fully settled into its final equilibrium. Their approach, which they call the "stationarity test," operates on a simple but powerful idea: if you have the correct blueprint, the system should appear perfectly still under the rules of that blueprint. Imagine trying to guess the rules of a complex machine by watching its gears. If you guess the rules correctly, the gears should not be accelerating or decelerating; they should be in a state of balance. The researcher's method involves taking a guess at the Hamiltonian, simulating how the system would change under that guess, and then checking if the actual observed state is indeed stationary, or unchanging, under those rules. If the state is changing, the guess was wrong. If it is stationary, the guess is likely correct. This process allows them to iteratively refine their guess, peeling away incorrect possibilities until the true underlying structure and strength of the interactions are revealed.

The paper presents a rigorous framework that uses this test to achieve two major goals that were previously out of reach. First, the researcher developed an algorithm that can learn not just the strength of the interactions, but also the very structure of the system—specifically, which particles are connected to which. In many quantum materials, the connections form a specific lattice, like a grid, but the exact pattern might be unknown. The new method can deduce this connectivity map from thermal samples, a feat that had remained an open question for non-commuting quantum systems. Second, and perhaps more surprisingly, the method works even when the system is not in a perfect equilibrium. In the real world, preparing a perfect thermal state is often computationally impossible, and nature frequently leaves systems in "metastable" states. These are local energy minima, or temporary resting spots, where the system gets stuck before it can reach the true equilibrium. The researcher proved that their stationarity test is robust enough to learn the Hamiltonian from these imperfect, metastable states, provided the system is close enough to being stationary.

The significance of this work lies in its efficiency and its ability to handle the messy reality of quantum systems. The algorithms are designed to be fast and to require a near-optimal number of samples, meaning they do not need an impossible amount of data to succeed. By connecting the learning problem to the physics of how quantum states evolve over time, the author bypassed the need for complex, abstract mathematical constructions that plagued previous attempts. They showed that by simply measuring how fast local properties change under a guessed set of rules, one can extract the precise coefficients of the Hamiltonian. This approach is not limited to high temperatures; it functions across the entire temperature spectrum, from the heat of a warm system to the near-absolute zero of a cold one.

Furthermore, the paper addresses a subtle but critical issue regarding the nature of the data used. Previous methods often assumed that the input data came from a perfect, equilibrium state. The new work demonstrates that this assumption is unnecessary. Even if the data comes from a system that is only approximately stationary—a state that is merely "hanging" in a local energy valley—the learning algorithm can still recover the correct Hamiltonian parameters. The researcher established that the error in the final answer is directly tied to how far the system is from being perfectly stationary, providing a clear, quantitative bound on the accuracy of the result. This finding suggests that the "local minima" of free energy, which are often seen as obstacles in physics, can actually serve as sufficient data sources for learning the underlying laws of the system.

The technical heart of the discovery relies on a deep connection between the stationarity test and a concept known as the quantum Fisher information. In simple terms, this information measures how sensitive a system is to changes in its parameters. The author proved that if the stationarity test indicates the system is not changing, then the Fisher information is small, which mathematically guarantees that the guessed Hamiltonian is close to the true one. They utilized properties of the system's evolution, specifically how information spreads through the lattice, to ensure that these measurements could be performed locally and in parallel. This locality is what makes the method scalable, allowing it to handle large systems without the computational cost exploding.

In addition to the learning algorithms, the paper offers a refined understanding of the structure of these metastable states. It provides a new proof for an "area law," a principle stating that the amount of information shared between two parts of a system depends on the size of the boundary between them, rather than the volume of the parts themselves. The researcher showed that this law holds even for metastable states, with the error depending only on the size of the boundary and not the total size of the system. This result is significant because it confirms that the structural simplicity of equilibrium states extends to these non-equilibrium, metastable configurations, reinforcing the idea that the fundamental geometry of quantum interactions is robust.

The work stands as a bridge between learning theory and many-body physics, offering a practical tool for the future of quantum device calibration and material science. By demonstrating that one can learn the rules of a quantum system from its thermal or near-thermal states without needing to know the system's structure in advance, the researcher has opened a new path for understanding complex quantum matter. The method is rigorous, mathematically proven, and efficient, providing a concrete answer to the question of whether nature's thermal samples are sufficient to decode the Hamiltonian. It suggests that even when a system is not perfectly at rest, the faint signals of its underlying rules are still there, waiting to be detected by the right kind of test.

Technical Summary: The Stationarity Test for Learning Quantum Many-Body Systems

Problem Statement
The paper addresses the fundamental problem of learning the Hamiltonian interactions ( $H$ ) governing a quantum many-body system, given samples of its thermal (Gibbs) state $\rho_H = e^{-\beta H}/\text{tr}(e^{-\beta H})$ . This task sits at the intersection of quantum learning theory and many-body physics. While classical analogs (learning Ising or graphical models) are well-studied, the quantum setting presents unique challenges due to non-commuting interactions and the lack of exact conditional independence properties found in classical systems.

A critical practical limitation in this field is the assumption that nature or an experimental black-box provides perfect equilibrium Gibbs states. Preparing Gibbs states at low temperatures is generally computationally intractable (NP-hard). Consequently, the paper investigates whether learning is possible from thermal metastable states—states that are approximately stationary under the dissipative dynamics of the system but may be far from the true Gibbs state in statistical distance.

The paper targets three specific objectives:

Parameter Learning: Recovering the coefficients of a lattice Hamiltonian given its interaction graph.
Structure Learning: Recovering both the interaction graph (connectivity) and the coefficients of a lattice Hamiltonian.
Metastable Learning: Performing the above tasks using samples from thermal metastable states rather than exact Gibbs states.

Methodology: The Stationarity Test
The core contribution is a framework built around a primitive the author coins the stationarity test. This test leverages recent advances in quantum Gibbs sampling, specifically the [CKG23] Lindbladian dynamics, which satisfy detailed balance with respect to the Gibbs state.

Definition: Given a candidate "guess" Hamiltonian $G$ and a sampled state $\sigma$ , the stationarity test measures the rate of change of a local observable $O$ under the adjoint Lindbladian dynamics $L_G^\dagger$ associated with $G$ :
$\text{Test}_i[O, G; \sigma] := \text{tr}[\sigma L_{G,i}^\dagger[O]]$
Completeness: If $G$ is the true Hamiltonian $H$ and $\sigma$ is the Gibbs state $\rho_H$ , the test yields zero (due to detailed balance).
Soundness: The paper establishes that if the test yields approximately zero for all local observables, then $G$ must be close to $H$ . This soundness is proven by relating the stationarity test to the Quantum Fisher Information (QFI). Specifically, the test measures the QFI of the sample state relative to the guess state.
Approximate Convexity: A key technical innovation is proving an "approximate convexity" property for the Fisher information. This allows the author to bound the distance between the true Hamiltonian and the guess Hamiltonian based on the magnitude of the stationarity test, even without knowing the true Hamiltonian.

Key Contributions and Results

1. Structure Learning of Lattice Hamiltonians
The paper presents the first algorithm for learning the underlying interaction graph (structure) of lattice Hamiltonians at all temperatures, given copies of their Gibbs states.

Algorithm: The method iteratively refines a guess Hamiltonian $G^{(r)}$ . In each step, it enumerates local perturbations (neighborhoods) and selects the perturbation that minimizes the stationarity test.
Complexity: The algorithm is time-efficient and nearly sample-optimal. The sample complexity is $N_{\text{samp}} \approx \frac{e^{\text{poly}(\beta)}}{\beta^2} \cdot O\left(\frac{\text{polylog}(\eta^{-1})}{\eta^2} \log n\right)$ , matching known lower bounds for parameter learning up to logarithmic factors.
Mechanism: To achieve sample efficiency in structure learning (where the number of possible neighborhoods is large), the author combines short-time Lindbladian evolution with Pauli shadow tomography. They use a "baseline" estimate (truncated to a logarithmic radius) and a "correction" term computed via shadow tomography on constant-sized regions, allowing parallel measurements.

2. Parameter Learning from Metastable States
The paper addresses the realistic scenario where samples are drawn from thermal metastable states (local minima of free energy) rather than exact Gibbs states.

Definition: A state $\sigma$ is $\epsilon$ -locally metastable if $\|L_{H,i}[\sigma]\|_1 \leq \epsilon$ .
Result: The author proves that the same learning algorithms used for Gibbs states succeed for metastable states, provided the target accuracy $\eta$ exceeds a threshold $\eta_{\text{thr}} \propto \epsilon$ .
Significance: This implies that Hamiltonian parameters can be learned from states that are efficiently preparable (via polynomial-time Lindbladian evolution) even at low temperatures, resolving the dilemma of whether "nature's" efficient thermal samples are sufficient for learning.

3. Refined Area Law for Metastable States
As a corollary to their locality analysis, the paper refines the area law for mutual information in thermal metastable states (extending [BCV25]).

Result: They prove that the mutual information between a region $A$ and its complement satisfies an approximate thermal area law: $I(A:A^c)_\sigma \leq 2\beta \|\partial_A H\| + e^{\mu|A|} \epsilon^\lambda$ .
Improvement: Unlike previous results, this bound removes the dependence on the total system size $n$ , making the statement meaningful in the thermodynamic limit ( $n \to \infty$ ).

Technical Foundations
The proofs rely on several advanced technical tools:

Lieb-Robinson Bounds: Used to establish the quasi-locality of the Lindbladian dynamics and the stationarity test, ensuring that measurements can be truncated to finite radii with exponentially decaying error.
Operator Fourier Transforms (OFTs): Used to decompose operators and analyze the Dirichlet form of the Lindbladian.
Approximate Detailed Balance: The author utilizes the approximate detailed-balance condition for metastable states to relate the stationarity test directly to the gradient of the free energy (Fisher information), bypassing the need for the state to be a true Gibbs state.
Convexity of Fisher Information: New arguments establishing that the Fisher information behaves convexly with respect to Hamiltonian perturbations, even in the presence of metastability errors.

Significance and Claims
The paper claims to provide a rigorous, time-efficient, and nearly sample-optimal framework for learning quantum Hamiltonians.

It unifies the learning of parameters and structure under a single "stationarity test" primitive.
It extends the scope of Hamiltonian learning to metastable states, offering a practical solution for systems where preparing exact Gibbs states is infeasible.
It provides the first structure learning algorithm for non-commuting lattice Hamiltonians at all temperatures.
The work draws connections between quantum learning theory, quantum Gibbs sampling, and quantum optimal transport, suggesting new interdisciplinary directions.

The author explicitly notes that while their algorithms are efficient, the sample complexity dependence on inverse temperature $\beta$ (exponential in $\beta$ ) is not necessarily optimal, and the runtime for structure learning is currently $n^{\text{poly}(\beta)}$ , leaving room for future improvement to quadratic time.

Technical Summary: The Stationarity Test for Learning Quantum Many-Body Systems

More like this