⚛️ quantum physics

Gauge freedom and efficient algorithms for Lindbladian learning

This paper establishes a sharp characterization of the fundamental limitations imposed by state-preparation-and-measurement (SPAM) noise on learning local Lindbladians, identifying unlearnable gauge-dependent components while providing scalable, SPAM-robust algorithms to efficiently learn all universally gauge-invariant parameters using only trusted single-qubit operations.

Original authors: Steven T Flammia, Savar D Sinha, Yu Tong

Published 2026-09-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Steven T Flammia, Savar D Sinha, Yu Tong

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, machines do not operate in isolation. They are constantly interacting with their surroundings, exchanging energy and information in a way that blurs the line between the system itself and the environment. To understand these open systems, scientists look for a mathematical description called a Lindbladian. Think of this as a master instruction manual that dictates how a quantum system changes over time, capturing both the smooth, predictable evolution of the system and the messy, random jitters caused by its environment. Knowing this manual is crucial for building better quantum computers and simulating complex materials, as it allows engineers to diagnose errors and design systems that can withstand the noise of the real world. However, there is a significant hurdle in reading this manual: the tools used to observe the system are themselves imperfect. The process of preparing a quantum state and measuring its final condition is always tainted by noise, a problem known as state-preparation-and-measurement error. This noise can hide the true nature of the system, making it impossible to tell if a change in the data comes from the system itself or from the flaws in the measurement process.

A team of researchers has now mapped out exactly which parts of this quantum instruction manual can be learned despite these noisy measurements, and which parts remain forever hidden. They focused on systems where interactions are limited to small groups of neighboring particles, a common feature in real-world quantum devices. Their work reveals a fundamental divide in what can be known. Some specific numbers in the manual, which describe how the system loses energy or how it rotates, are robust. These values can be extracted with high precision using only simple, trusted operations on individual particles, even if the overall preparation and measurement of the system are wildly inaccurate. The researchers developed efficient algorithms to find these reliable numbers, proving that they can be learned with a number of experiments that grows only slowly as the system gets larger. This means that even for very large quantum devices, scientists can reliably characterize the most important aspects of their behavior without needing perfect equipment.

However, the study also draws a sharp line around what cannot be learned. For many other parts of the instruction manual, the noise from the measurement process creates a permanent fog. The researchers proved that for almost every physical system, there are specific coefficients that can be changed by an arbitrary amount without altering any experimental outcome. In other words, two completely different physical systems could produce identical data if the measurement noise is unknown, making it impossible to distinguish between them. This is not a limitation of current technology or a temporary gap in knowledge; it is a fundamental barrier. Even with unlimited data and perfect control over individual particles, these specific details of the system remain unlearnable. The uncertainty in these values does not shrink as the system grows; it remains a fixed, irreducible error floor.

The team achieved this clarity by identifying a specific type of transformation, which they call a gauge freedom, that can shuffle the values of the instruction manual without changing the observable results. They showed that this shuffling is possible because the noise in preparing and measuring the system can mimic the effects of changing the system's internal rules. By carefully analyzing which parts of the manual are immune to this shuffling, they separated the learnable from the unlearnable. The learnable parts include the rates at which the system decays and certain specific interactions between particles. The unlearnable parts include many of the complex, multi-particle interactions and the imaginary components of the dissipative terms. The researchers constructed specific examples to show that the uncertainty in these unlearnable parts is not just a theoretical possibility but a concrete reality that persists regardless of how much data is collected.

To find the learnable numbers, the researchers designed a protocol that relies on a technique called twirling. This involves rapidly applying random, simple rotations to the system during its evolution. This process effectively averages out the confusing parts of the noise, leaving behind a clean signal that reveals the robust coefficients. Because the protocol only requires trusted operations on single particles, it avoids the need for complex, error-prone multi-particle gates or extra helper particles. The method is efficient, requiring a number of experiments that scales logarithmically with the size of the system, meaning it remains practical even for very large devices. The researchers also identified a special case where the physical laws of the system, specifically the requirement that probabilities must remain positive, can force some of the normally hidden numbers to become visible. However, learning these special cases remains a difficult, open problem that requires further investigation.

The implications of this work are profound for the future of quantum science. It provides a clear guide for what is possible to measure and what must be accepted as unknown. By focusing on the universally learnable components, scientists can build robust characterization tools that work even with imperfect hardware. This approach offers a practical path forward for calibrating quantum processors and understanding the dynamics of open quantum systems. The work does not promise to solve every mystery of quantum noise, but it definitively separates the solvable from the unsolvable, offering a toolkit that allows researchers to extract maximum information from the noisy data they can actually collect. In doing so, it shifts the focus from trying to eliminate all noise to understanding exactly how much information can be recovered despite it.

Technical Summary: Gauge Freedom and Efficient Algorithms for Lindbladian Learning

1. Problem Statement

The paper addresses the inverse problem of learning a local Lindbladian $\mathcal{L}$ for an $N$ -qubit open quantum system from dynamical data. The Lindbladian is expanded in the Pauli basis as:
$\mathcal{L}[\rho] = -i \sum_{c \in I_h} h_c [P_c, \rho] + \sum_{(a,b) \in I_d} \alpha_{ab} D_{ab}[\rho]$
where $h_c$ are Hamiltonian coefficients and $\alpha_{ab}$ are dissipative coefficients (elements of a positive semidefinite Kossakowski matrix).

The central challenge is the presence of unknown State-Preparation-and-Measurement (SPAM) noise. Unlike idealized settings where initial states and measurements are perfect, experimental SPAM errors can render distinct Lindbladians indistinguishable. Specifically, a similarity transformation (gauge transformation) can simultaneously alter the Lindbladian, the initial state, and the measurement observable such that all experimental outcome probabilities remain identical. The paper seeks to:

Characterize which components of a local Lindbladian are fundamentally unlearnable due to this gauge freedom.
Identify which components are universally gauge-invariant and can be learned efficiently.
Develop SPAM-robust algorithms using only trusted single-qubit control, without requiring ancillary qubits or entangling gates.

2. Methodology

2.1 Gauge Transformations and Indistinguishability

The authors define a family of gauge transformations generated by local depolarizing maps $S_\theta = e^{\Gamma_\theta}$ , where $\Gamma_\theta$ acts on Pauli operators $P$ as $\Gamma_\theta[P] = -\theta \cdot s(P) P$ . Under this transformation, the Lindbladian transforms as $\mathcal{L} \to S_\theta \circ \mathcal{L} \circ S_\theta^{-1}$ . Crucially, if the SPAM noise channels are also transformed appropriately (absorbing the gauge), the experimental probabilities $p(t) = \text{Tr}[M e^{t\mathcal{L}}(\rho_0)]$ remain invariant.

The paper establishes that for any fixed support pattern (defining which qubits interact), there exists an $N$ -parameter family of such transformations that leaves all experimental data unchanged. This leads to a classification of Lindbladian components (Hamiltonian coefficients $h_c$ and real/imaginary parts of $\alpha_{ab}$ ) based on their behavior under these transformations.

2.2 Classification of Learnability

The authors distinguish between two categories of components:

Universally Gauge-Invariant: Components whose values remain unchanged under all admissible gauge transformations for any generator in the ambient space.
Generically Gauge-Dependent: Components whose values can be altered by arbitrarily small gauge transformations for almost every physical Lindbladian (preserving complete positivity and the support pattern).

The classification relies on the structure of the Pauli basis:

Diagonal Dissipative Coefficients ( $\alpha_{aa}$ ): Universally gauge-invariant.
Type I Off-Diagonal Coefficients ( $\text{Re } \alpha_{ab}$ ): Invariant if $P_a$ and $P_b$ differ on exactly one qubit and both are non-identity on that qubit.
Type II Off-Diagonal Coefficients ( $\text{Im } \alpha_{ab}$ and single-qubit $h_c$ ): Invariant if $P_a$ and $P_b$ differ on exactly one qubit, with exactly one being the identity on that qubit (or if $h_c$ corresponds to a single-qubit Pauli).

All other components are shown to be generically gauge-dependent. For these dependent components, the authors construct specific Lindbladians where the uncertainty is $\Omega(1)$ (independent of system size $N$ ), even under complete positivity constraints.

2.3 Learning Algorithms

The paper proposes efficient protocols to learn the universally gauge-invariant components using only trusted single-qubit operations. The core technique involves Pauli twirling (or partial twirling) to isolate specific dynamics and cancel SPAM factors.

Diagonal and Type I Components:
- Method: Full Pauli twirling converts the dynamics into exponential decay curves. The SPAM noise enters only as multiplicative prefactors ( $s_Q m_Q$ ) in the signal $f(t) = s_Q m_Q e^{t \ell}$ . By taking the ratio of signals at $t=0$ and $t=\tau$ , the unknown SPAM factors cancel exactly, allowing the extraction of the decay rate $\ell$ .
- Reconstruction: Local decay rates are related to the coefficients via a linear system solvable by back-substitution over the support inclusion order.
- Complexity: $O(\epsilon^{-2} \log(N/\delta))$ experiments and total evolution time.
Type II Components:
- Method: Since these cannot be reduced to simple decay rates, the authors use partial twirling (averaging over a subgroup $\{I, Z\}$ on a specific qubit) to isolate the dynamics to a two-dimensional invariant subspace in the Heisenberg picture.
- Reconstruction: The dynamics in this subspace are governed by a $2 \times 2$ matrix where the target parameter (e.g., $h_z$ or $\text{Im } \alpha_{xy}$ ) appears in the antisymmetric coupling. By measuring the off-diagonal elements of the time-evolved matrix and differentiating at $t=0$ , the parameter is extracted. The SPAM prefactor again cancels in the ratio of derivatives to the zero-time signal.
- Differentiation: Derivatives are estimated using Chebyshev polynomial interpolation on sampled time points to achieve high precision.
- Complexity: $O(\epsilon^{-2} \log^4(2/\epsilon) \log(N/\delta))$ experiments and total evolution time.

2.4 Boundary-Induced Gauge Invariance

The paper investigates whether the requirement of complete positivity (CPTP) restricts gauge freedom at the "boundary" of the physical set (e.g., when the Kossakowski matrix has zero eigenvalues). They derive sufficient conditions under which certain Hamiltonian coefficients become gauge-invariant at specific physical Lindbladians. However, they note that learning these boundary-specific invariants generally remains an open problem.

3. Key Results

Sharp Characterization of Obstruction: The paper provides a complete classification of local Lindbladian components into universally gauge-invariant and generically gauge-dependent sets. It proves that for the generically dependent set, there exist physical Lindbladians where the uncertainty is $\Omega(1)$ regardless of system size, even with infinite data.
Efficient SPAM-Robust Protocols:
- Theorem 2: Protocols exist to learn all universally gauge-invariant components (diagonal $\alpha_{aa}$ , Type I $\text{Re } \alpha_{ab}$ , Type II $\text{Im } \alpha_{ab}$ , and single-qubit $h_c$ ) with additive error $\epsilon$ and failure probability $\delta$ .
- Resource Scaling: The protocols require $O(\epsilon^{-2} \log(N/\delta))$ to $O(\epsilon^{-2} \log^4(2/\epsilon) \log(N/\delta))$ experiments and total evolution time. The single-qubit gate count scales linearly with $N$ (up to polylogarithmic factors).
- Assumptions: The algorithms require only trusted single-qubit gates, no ancillas, and no entangling control. They tolerate SPAM noise that is arbitrarily far from ideal globally, provided there is non-vanishing local visibility ( $r_p, r_m > 0$ ) for constant-sized subsets.
Boundary Conditions: Theorem 31 and 32 provide sufficient conditions for additional Hamiltonian coefficients to be gauge-invariant at the boundary of the physical set, linking this to the "Hamiltonian-not-in-Lindblad-span" condition.

4. Significance and Claims

The authors claim this work offers a practical toolkit for the scalable characterization of open quantum systems in the presence of realistic SPAM noise.

Scalability: Unlike previous methods that require full process tomography (exponential scaling) or assume ideal/known SPAM, this framework achieves logarithmic scaling in system size $N$ for the number of experiments and evolution time.
Minimal Control: The reliance solely on trusted single-qubit operations makes the protocols applicable to current hardware (e.g., trapped ions) where entangling gates are difficult to calibrate or ancillas are unavailable.
Fundamental Limits: By rigorously characterizing the "unlearnable" parts of the Lindbladian, the paper sets a fundamental error floor for learning tasks under unknown SPAM, distinguishing between limitations due to data scarcity and those due to fundamental indistinguishability.
Gauge-Aware Framework: The work moves beyond simply ignoring SPAM or assuming it is known; it explicitly models the gauge freedom, classifies what is learnable, and provides algorithms that cancel the unlearnable degrees of freedom.

The paper concludes that while boundary-induced invariants offer potential for learning additional parameters, the general problem of learning them remains open. The primary contribution is the complete solution for the universally gauge-invariant subset, which covers a significant portion of physically relevant parameters in local open systems.