⚛️ quantum physics

Demonstrating and Benchmarking Classical Shadows for Lindblad Tomography

This paper demonstrates that shadow Lindblad tomography, utilizing randomized measurements and locality assumptions, can efficiently characterize decoherence and spurious couplings in a five-qubit superconducting processor with exponentially fewer resources and significantly reduced acquisition time compared to traditional extensible Lindblad tomography.

Original authors: Rune Thinggaard Birke, Johann Bock Severin, Malthe A. Marciniak, Emil Hogedal, Andreas Nylander, Irshad Ahmad, Amr Osman, Janka Biznárová, Marcus Rommel, Anita Fadavi Roudsari, Jonas Bylander, Giovann

Published 2026-02-17

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Rune Thinggaard Birke, Johann Bock Severin, Malthe A. Marciniak, Emil Hogedal, Andreas Nylander, Irshad Ahmad, Amr Osman, Janka Biznárová, Marcus Rommel, Anita Fadavi Roudsari, Jonas Bylander, Giovanna Tancredi, Daniel Stilck França, Albert Werner, Christopher W. Warren, Jacob Hastrup, Svend Krøjer, Morten Kjaergaard

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

Imagine you have a very complex, high-tech kitchen (a quantum computer) with five ovens (qubits). You want to know exactly how these ovens behave when you aren't cooking anything in them—how they lose heat, how they vibrate, and how they accidentally influence each other. This is called characterizing the "idle" state of the processor.

If you try to measure every single possible way these ovens could interact using traditional methods, it would take you 58 hours of non-stop testing just to get a clear picture. That's like trying to map every single grain of sand on a beach by picking them up one by one.

This paper introduces a new, clever shortcut called Shadow Lindblad Tomography (SLT). Here is how it works, explained simply:

The Problem: The "Exhaustive Search"

The old way of doing this (called Extensible Lindblad Tomography or ELT) is like a detective who insists on interviewing every single person in a city to find out who stole a cookie.

The Method: They prepare the ovens in every specific, known state, wait a tiny bit, and measure them. Then they repeat this for every possible combination.
The Result: It's accurate, but it takes forever. As you add more ovens (qubits), the time required explodes exponentially. For a 5-oven kitchen, it's manageable but slow. For a 50-oven kitchen, it would take 5 years.

The Solution: The "Shadow" Detective

The new method (SLT) is like a detective who uses random sampling and smart guessing. Instead of interviewing everyone, they interview a random group of people and use a mathematical trick to reconstruct the whole story.

Here is the analogy:

The Old Way (ELT): You shine a flashlight on the kitchen from every single angle, one by one, to see every shadow.
The New Way (SLT): You turn on a strobe light that flashes randomly. You take a few snapshots. Even though the light is random, if you take enough snapshots and use a computer algorithm to "stitch" them together, you can reconstruct the 3D shape of the kitchen perfectly.

How They Proved It Works

The researchers tested this on a real quantum chip with 5 qubits:

The Small Test (1 Qubit): They compared the old method and the new method on a single qubit. The results were identical. The new method was just as accurate but used fewer "snapshots."
The Medium Test (3 Qubits): They did the same for three qubits. Again, the new method matched the old one perfectly, but it required significantly less data.
The Big Test (5 Qubits): This is where the magic happened.
- They used the new "Shadow" method to map the entire 5-qubit processor.
- Time taken: Only 9 hours.
- Time saved: They estimated the old method would have taken 58 hours.
- The Catch: The new method relies on a physical assumption: that the ovens mostly only affect their immediate neighbors (local interactions), not the whole kitchen at once. The researchers proved this assumption is true for their hardware.

Why This Matters

Think of the "Shadow" method as a smart filter.

In the past, to learn how a quantum computer behaves, you had to measure everything.
Now, because we know quantum computers usually have "local" behavior (qubits mostly talk to their neighbors), the Shadow method can ignore the impossible interactions and focus only on the likely ones.
It recycles random data. Instead of throwing away a measurement after using it once, the algorithm uses that same random data point to estimate dozens of different properties simultaneously.

The Bottom Line

This paper proves that we don't need to measure everything to understand a quantum computer. By using randomized measurements and smart math (Classical Shadows), we can learn the "personality" of a quantum processor 6 times faster than before.

This is a huge step forward. It means that as quantum computers grow from 5 qubits to 50 or 500, we won't be stuck waiting years to calibrate them. We can use this "Shadow" technique to keep them running smoothly, making the dream of powerful quantum computers much more practical.

1. Problem Statement

The performance of solid-state quantum processors is degraded by spurious couplings and decoherence. To mitigate these issues, precise calibration and characterization of the "idling" processor (its dynamics when no gates are applied) are essential.

The Bottleneck: Traditional tomographic recovery of time-independent Lindblad dynamics (describing both Hamiltonian evolution and dissipation) scales exponentially with the number of qubits ( $n$ ). For a 5-qubit system, a full, model-agnostic reconstruction requires estimating over a million parameters, making it experimentally intractable (requiring months of data acquisition).
The Challenge: How to efficiently learn the Lindblad generator (Hamiltonian $H$ and jump operators $\{L_k\}$ ) for larger systems without assuming a specific noise model that might miss critical physics, while avoiding the exponential resource cost of standard tomography.

2. Methodology

The authors experimentally benchmark two protocols on a superconducting transmon processor: Extensible Lindblad Tomography (ELT) and Shadow Lindblad Tomography (SLT). Both aim to learn the Lindblad master equation (LME) governing the system's idle evolution.

A. Theoretical Framework

The system dynamics are modeled by the LME:
$\dot{\rho} = \mathcal{L}(\rho) = -i[H, \rho] + \sum_k \left( L_k \rho L_k^\dagger - \frac{1}{2}\{L_k^\dagger L_k, \rho\} \right)$
The goal is to estimate the coefficients of the Hamiltonian and dissipators expanded in the Pauli basis. The protocol relies on measuring the instantaneous rate of change of expectation values at $t=0$ :
$\frac{d}{dt}E_i(0) = \text{Tr}(B_i \mathcal{L}(A_i)) = \sum_j p_j \text{Tr}(B_i \mathcal{L}_j(A_i))$
where $p_j$ are the unknown Lindblad parameters.

B. Protocol 1: Extensible Lindblad Tomography (ELT)

Approach: A deterministic scheme using a complete set of product-state preparations (Pauli eigenstates) and Pauli measurements.
Scaling: Without assumptions, it scales exponentially. However, under $k$ -locality assumptions (e.g., dissipation is only on-site, interactions are at most 2-body), the number of required configurations scales polynomially.
Role in Paper: Used as the baseline for comparison. It provides a "ground truth" (within experimental uncertainty) for smaller subsystems (1 and 3 qubits).

C. Protocol 2: Shadow Lindblad Tomography (SLT)

Approach: A randomized scheme based on Classical Shadows.
- Instead of fixed states, it uses random Pauli preparations and random Pauli measurement bases.
- It recycles randomized data to estimate many Pauli Transfer Matrix (PTM) elements simultaneously via classical post-processing.
Key Innovation: The authors demonstrate that for low-weight Pauli operators (relevant under $k$ $k$ -locality assumptions), the estimator is bounded for every single shot.
- Unlike standard shadow tomography which often requires "median-of-means" estimators to handle heavy-tailed distributions, SLT allows for standard Gaussian error propagation and $\chi^2$ -fitting.
- This simplifies uncertainty quantification and makes the protocol compatible with standard regression tools.
Scaling: Under $k$ -locality, the number of randomizations ( $N_{SLT}$ ) required to estimate parameters with high confidence scales logarithmically with system size ( $O(\log n)$ ), rather than polynomially or exponentially.

3. Key Contributions

Experimental Benchmarking: First experimental demonstration of SLT for Lindblad dynamics on a superconducting processor, comparing it directly against the deterministic ELT baseline.
Validation of Locality Assumptions: Verified on 1- and 3-qubit subsystems that the dynamics are dominated by 1-local and 2-local terms, justifying the use of SLT for larger systems.
Statistical Simplification: Proved and demonstrated that for bounded- $k$ regimes, SLT estimators have bounded variance, allowing the use of standard $\chi^2$ -regression and Gaussian error propagation instead of complex robust estimators.
Scalability Demonstration: Successfully applied SLT to a 5-qubit processor, recovering all single-qubit dissipation and two-qubit coupling parameters in a feasible timeframe, whereas ELT would have been prohibitively expensive.

4. Experimental Results

A. Single-Qubit Verification

SLT and ELT results agreed within uncertainties.
SLT converged to ELT accuracy with $\sim 3,000$ randomizations.
The extracted parameters (detuning, dephasing, relaxation) matched known device physics and standard $T_1$ measurements.

B. Three-Qubit Benchmark

Goal: Reconstruct all 4,032 parameters of a 3-qubit LME without structural assumptions.
Result: SLT agreed with ELT within a few kHz using only a fraction of the shots required by ELT.
Observation: 3-local (three-body) interactions were found to be negligible (within error bars), confirming that a 2-local model is sufficient. The residual standard deviation saturated at a lower relative shot count for 3 qubits than for 1 qubit, consistent with $k$ -local scaling.

C. Five-Qubit Full Reconstruction

Setup: Applied SLT to the full 5-qubit processor assuming a restricted model (on-site dissipation + pairwise coherent interactions).
Efficiency:
- SLT: Completed in 9 hours of acquisition time using 25 million randomizations.
- ELT Estimate: The equivalent ELT experiment (even with 2-local assumptions) would require ~160 million runs, estimated at 58 hours.
- Unrestricted ELT: Would require ~162 days.
Findings:
- Single-qubit parameters were $\sim 100\times$ larger than two-qubit interaction terms, consistent with weakly coupled transmon arrays.
- Residual static couplings were minimal (few kHz).
- All extracted parameters were physically interpretable and consistent with single-qubit ELT results.

5. Significance and Outlook

Practical Scalability: The work demonstrates that randomized shadow protocols can be practically implemented to learn quantum processor dynamics as qubit counts increase, moving from intractable exponential scaling to feasible logarithmic scaling under realistic physical assumptions.
Metrology Tool: SLT provides a scalable tool for device-level metrology, yielding interpretable parameters tied to fabrication and layout, which is crucial for diagnosing noise and optimizing control.
Statistical Robustness: By avoiding median-of-means estimators and utilizing standard Gaussian statistics, the method lowers the barrier to entry for experimentalists and ensures transparent uncertainty quantification.
Future Impact: The authors estimate that for a 50-qubit system, ELT would take ~5 years to characterize idle noise, whereas SLT could do it in ~22 hours. This positions SLT as a critical subroutine for characterizing large-scale superconducting processors and analog quantum simulators.

Conclusion

The paper successfully validates Shadow Lindblad Tomography as a superior alternative to deterministic methods for characterizing open quantum system dynamics. By leveraging randomized measurements and physically motivated locality assumptions, SLT achieves exponential speed-ups in resource efficiency while maintaining high accuracy and compatibility with standard statistical analysis tools. This paves the way for the characterization of quantum processors with tens to hundreds of qubits.