⚛️ quantum physics

Learning Local Fermionic Lindbladians under Parity Superselection

This paper presents a protocol for learning time-independent, parity-covariant, local fermionic Lindbladians using only even preparations and measurements by employing internal markers to convert odd probes into even observables, achieving efficient sample complexity for both known and unknown interaction graphs while respecting parity superselection rules.

Original authors: Tim Möbus, Daniel Stilck França, Cambyse Rouzé

Published 2026-10-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Tim Möbus, Daniel Stilck França, Cambyse Rouzé

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 matter, particles known as fermions—such as electrons—obey a strict rule of privacy. They cannot be observed directly in certain states without breaking the fundamental symmetry of the system. This rule, called parity superselection, acts like a gatekeeper that forbids the direct measurement of specific odd-numbered properties, forcing scientists to rely only on even-numbered combinations. For decades, this limitation made it incredibly difficult to map out how these particles interact and lose energy over time. Understanding these interactions is crucial for simulating complex materials, like the Fermi-Hubbard model used to study high-temperature superconductors, and for building future quantum computers that can handle particle loss and noise. Without a way to see the full picture of how these systems evolve, researchers were often left guessing at the underlying laws that govern their behavior.

A team of physicists has now developed a method to bypass this gatekeeper and learn the complete rules of motion for these fermionic systems, even when they are losing energy to their environment. They achieved this by treating the system not as a static object to be photographed, but as a dynamic process to be probed with gentle, short bursts of time. Instead of trying to measure the forbidden odd properties directly, which is impossible, the researchers used a clever trick involving "markers." They introduced specific, harmless reference points within the system that could be measured safely. By observing how these markers interacted with the rest of the system over very short intervals, they could mathematically reconstruct the hidden, forbidden information. It is similar to how a blind person might use a cane to feel the shape of a room; the cane does not see the walls, but the way it bounces back reveals the room's structure.

The researchers focused on systems where particles interact only with their immediate neighbors, a common scenario in physical materials. They showed that by preparing the system in a specific, measurable state and letting it evolve for a tiny fraction of a second, they could extract the exact mathematical coefficients that describe the system's behavior. These coefficients tell us the strength of the forces between particles and how quickly they lose energy. The method works without needing any external reference frames or complex control sequences during the evolution, relying instead on a series of independent, short experiments. The team proved that by repeating these experiments a number of times that grows only with the square of the desired precision, they could recover the full description of the system's dynamics with high confidence.

A key innovation in their approach was the use of internal markers that are part of the system itself, rather than external tools. In many previous attempts, scientists had to rely on external reference modes, which are difficult to implement in real-world experiments. This new method uses three specific points within the system to calibrate the measurements. By measuring how pairs of these markers influence each other, the researchers could calculate and subtract the noise introduced by the markers themselves. This calibration allowed them to isolate the true signal of the system's evolution. Once the raw data was collected and cleaned, they used a mathematical process to invert the information, turning the measured even properties back into the original, hidden odd properties that define the system's unique behavior.

The study demonstrates that this learning process is efficient and scalable. For systems arranged on a grid, such as atoms in a crystal lattice, the number of experiments required to learn the system's rules grows slowly as the system gets larger. The researchers provided a rigorous guarantee that their method works with a probability of success that can be made arbitrarily high by simply collecting more data. They also showed that the method remains effective even when the interactions between particles do not stop abruptly at a certain distance but instead fade away exponentially. In these cases, the method can still reconstruct the system's behavior by focusing on the most significant interactions and ignoring the faint, distant ones.

This work is significant because it provides a practical blueprint for characterizing open quantum systems, which are systems that interact with their environment. Previous methods often required assumptions about the system's structure or relied on idealized conditions that do not exist in the lab. By proving that the full generator of the dynamics can be learned using only even preparations and measurements, the researchers have removed a major barrier to understanding dissipative fermionic systems. Their results apply to a wide range of physical models, including those describing particle loss, injection, and dephasing, which are common in ultracold atom experiments and solid-state devices. The ability to learn these dynamics without needing to know the locations of the interactions beforehand means the method can be applied to unknown materials or complex engineered systems.

The researchers also addressed the challenge of systems where the interaction graph is not known in advance. In such cases, they showed that by adding a single idle auxiliary mode for every system mode, they could still learn the local coefficients of the interactions. While this requires slightly more resources, it ensures that the method is robust enough to handle completely unknown structures. The team verified that their approach matches the theoretical lower bounds for how much data is needed, meaning it is as efficient as physically possible for short-time experiments. This efficiency is crucial for experimentalists who have limited time and resources to probe fragile quantum states.

Ultimately, this paper establishes a general, scalable scheme for learning the laws of motion for interacting fermions. It combines local even preparations, uninterrupted short-time evolution, and standard measurement techniques to achieve a level of precision that was previously thought difficult to reach. The findings offer a new path for experimentalists to characterize the complex, noisy dynamics of quantum matter, paving the way for more accurate simulations of materials and better control over quantum technologies. By turning the problem of hidden parity into a solvable calibration task, the researchers have turned a fundamental limitation of quantum mechanics into a manageable engineering challenge.

Technical Summary: Learning Local Fermionic Lindbladians under Parity Superselection

1. Problem Statement

The paper addresses the challenge of learning time-independent, parity-covariant, $k$ -mode-local Lindbladians (generators of Markovian open quantum dynamics) for fermionic systems. The learning task is constrained by parity superselection, a fundamental physical principle stating that odd Majorana observables cannot be measured directly; only even operators (those commuting with the total fermion parity) are physically accessible.

Key difficulties include:

Measurement Constraints: Physical states and measurements must be even. Odd probes, which are necessary to fully characterize fermionic dynamics, are inaccessible without an external parity reference.
Locality vs. Qubit Mapping: While fermionic operators may be local in terms of modes, they can map to long non-local qubit strings under standard Jordan–Wigner transformations, complicating the definition of "locality" for learning algorithms.
Scalability: Standard quantum process tomography scales exponentially with the number of modes. The goal is to achieve sample complexity that scales polynomially (specifically, near-linearly or polylogarithmically) with system size $m$ for local generators.
Unknown Structure: The learning protocol must operate without prior knowledge of the specific locations of non-zero coefficients or the underlying interaction graph, relying only on bounds on interaction strength and locality. Note: The geometric learning scenario explicitly assumes a known bounded-degree mode graph to define allowed supports, while the structure learning scenario handles unknown graphs using ancillary modes.

2. Methodology

The proposed framework combines parity dressing, internal marker calibration, and signed Fierz inversion to recover canonical generator coefficients from short-time evolution data.

2.1 Parity Dressing and Internal Markers

To circumvent the inability to measure odd observables, the authors employ a "parity dressing" technique:

Dressing: An odd Majorana monomial $\Gamma_A$ is converted into an even Hermitian observable by multiplying it with a disjoint Majorana marker $\eta$ (i.e., $i\Gamma_A\eta$ ).
Internal Markers: Unlike protocols requiring external ancillary modes, this method uses internal markers (specific Majorana modes within the system itself) that are spatially separated from the target region.
Calibration: The dynamics of these internal markers are influenced by the system's Lindbladian. By measuring pair-response slopes between three separated markers ( $\eta_1, \eta_2, \eta_3$ ), the protocol calibrates the dynamical contributions of the markers. This allows the extraction of the "bare" derivative of the target observable, correcting for the marker's own evolution.

2.2 Signed Fierz Inversion

Once the dressed transfer derivatives (initial time derivatives of even transfer curves) are estimated, the authors recover the canonical coefficients using a signed Fierz inversion:

Completeness: The inversion relies on the completeness of the Majorana basis.
Restricted Inversion: For a target coefficient supported on a region $R$ , the algorithm uses only derivatives with labels supported on $R$ .
Descending Recursion: The relationship between the restricted dressed coefficients and the global canonical coefficients is triangular. The algorithm recovers coefficients by descending from larger support sizes ( $k$ ) to smaller ones ($0$), subtracting contributions from larger supports (diagonal extensions) at each step. This recursion is stable, with error propagation controlled by the locality parameter $k$ .

2.3 Experimental Protocols

The learning protocols utilize even short-time experiments:

Preparation: Even states are prepared using local fermionic Gaussian controls, randomized occupation-state preparations, and pairing rotations.
Evolution: The system evolves under the unknown Lindbladian for a short, preassigned time $t \le (4k\bar{\alpha})^{-1}$ without intermediate control.
Measurement: Even observables are measured.
Pair-Block Shadows: To estimate many input-output overlaps from a single dataset, the authors use a "pair-block process shadow" technique. Modes are paired into blocks, and random even settings are applied to each block. This allows the estimation of multiple transfer curves simultaneously, significantly reducing sample complexity.

3. Key Results

The paper provides rigorous learning guarantees for two distinct scenarios:

3.1 Geometric Learning (Known Graph, No External Modes)

Assuming a known bounded-degree mode graph and finite interaction range $r$ :

Sample Complexity: To achieve entrywise error $\varepsilon$ for all canonical coefficients, the protocol requires $\tilde{\mathcal{O}}(\bar{\alpha}^2 \varepsilon^{-2} \log(m/\delta))$ samples.
Diamond-Norm Recovery: To return a valid physical generator with diamond-norm error $\varepsilon$ , the cost increases by a factor of $m^2$ , i.e., $\tilde{\mathcal{O}}(\bar{\alpha}^2 m^2 \varepsilon^{-2} \log(m/\delta))$ .
Support Recovery: Thresholding allows for exact support recovery if a gap exists between zero and non-zero coefficients.
Lower Bounds: The paper establishes that the $m^2$ scaling for diamond-norm recovery matches known lower bounds for short-time experiments on fresh systems, up to logarithmic and geometric factors.
Note: While the graph is known to fix allowed supports, the protocol does not require prior knowledge of the specific locations of non-zero coefficients or a local decomposition.

3.2 Structure Learning (Unknown Graph, Ancillary Modes)

When the interaction graph is unknown, the protocol allows for one idle ancillary mode per system mode and potentially nonlocal pair operations:

Approximate Degree: The method assumes an "approximate coefficient-degree" bound $d$ (at most $d$ coefficients involve any single mode, after discarding a small tail of coefficients).
Sample Complexity: The protocol achieves local error $\varepsilon$ using $\tilde{\mathcal{O}}_k(d^2 m^{\lfloor k/2 \rfloor} \varepsilon^{-2} \log(m/\delta))$ samples.
Output: The estimated coefficients have a local error bounded by $\varepsilon$ and involve at most $\mathcal{O}_k(d)$ non-zero entries per mode. A subsequent convex fit can return a valid generator with comparable error.

3.3 Extensions

Exponential Tails: The framework extends to models with exponentially decaying interactions, incurring only polylogarithmic overhead in sample complexity.
Prediction: For geometric models, fixed-support even observables can be predicted with logarithmic system-size sample complexity at fixed time and accuracy.

4. Significance and Claims

The authors claim to establish the first general, scalable scheme for learning interacting local fermionic Lindbladians that satisfies the following criteria simultaneously:

Parity Compatibility: It respects parity superselection using only even preparations and measurements, without requiring external parity references.
Locality Preservation: It maintains fermionic locality (mode support) rather than relying on qubit-string locality.
Uninterrupted Evolution: It operates without interleaved control during the evolution, relying on short-time data.
Standard Quantum Limit Scaling: It achieves standard-quantum-limit sample scaling ( $\varepsilon^{-2}$ ) for coefficient learning.

The work bridges the gap between theoretical learning bounds and physical constraints in fermionic systems (such as the Fermi-Hubbard model), demonstrating that full dynamical characterization is possible using only local even operations and short-time evolution. The results are particularly significant for platforms like ultracold atoms and superconducting circuits where parity constraints are intrinsic.

The paper explicitly notes that prior work had not established such a combination of local even access, uninterrupted short-time evolution, and optimal sample scaling for interacting fermionic systems. The methodology adapts techniques from continuum free-fermion potential learning and extends them to interacting Lindbladians via internal marker calibration and signed inversion.