⚛️ quantum physics

Lindblad Multiproduct Formulas

This paper introduces Lindblad Multiproduct Formulas, a quantum error mitigation technique utilizing two-dimensional tensor networks and loop-corrected belief propagation that offers computational advantages over direct expectation value calculations, achieving a 5.6× speedup on GPUs and demonstrating effectiveness on a 65-qubit IBM quantum processor.

Original authors: Niall F. Robertson, Andrea D'Urbano, Anton Dekusar, Max Rossmannek, Eric D. Switzer, James R. Garrison, Nicolas Lorente, Igor Pasichnyk, Paolo D'Alberto, Muhammad Osama, Zachary Streeter, Yasuko Ecker

Published 2026-09-07

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Niall F. Robertson, Andrea D'Urbano, Anton Dekusar, Max Rossmannek, Eric D. Switzer, James R. Garrison, Nicolas Lorente, Igor Pasichnyk, Paolo D'Alberto, Muhammad Osama, Zachary Streeter, Yasuko Eckert, Constantinos Evangelinos, Sergiy Zhuk

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 quest to build useful quantum computers, scientists face a persistent obstacle: noise. Unlike classical computers, which process information with near-perfect reliability, quantum machines are incredibly fragile. Their delicate states of matter, known as qubits, are easily disturbed by their environment, causing calculations to drift away from the truth. This noise is not just a minor glitch; it is the primary reason why current quantum computers struggle to solve problems that classical machines can handle easily. To get useful answers, researchers have developed techniques to "clean up" the noisy data these machines produce, often by running the same calculation multiple times with slightly different settings and then mathematically combining the results to guess what the perfect, noise-free answer would have been. However, as these calculations grow larger and more complex, the amount of data required to clean them up becomes overwhelming, often demanding more time and resources than the quantum computer itself can provide.

A team of researchers has now introduced a new approach to this problem, one that cleverly combines the power of quantum hardware with advanced classical simulation techniques. They call their method Lindblad Multiproduct Formulas. Instead of relying solely on the quantum computer to generate data or solely on classical computers to simulate the entire system, they split the work. The quantum machine measures the noisy outcomes, while a sophisticated classical algorithm calculates the specific weights needed to combine those noisy measurements into a single, accurate prediction. The researchers tested this hybrid method on a model of a "discrete time crystal," a strange state of matter that oscillates in time without losing energy, using a quantum processor with 65 qubits. Their results show that by using this technique, they could recover a much clearer picture of the system's behavior than either the quantum computer or the classical computer could achieve on their own, even when the classical simulations were simplified to run faster.

The core of this new method lies in how it handles the relationship between the noisy quantum machine and the classical computer. Usually, when scientists try to simulate a complex quantum system on a classical computer, they must simplify the math to make it manageable. These simplifications introduce their own errors, known as truncation errors. If the system is too complex, these errors become so large that the classical simulation is useless. The researchers realized, however, that the specific mathematical quantities needed to fix the quantum computer's errors might be much easier to calculate than the final answer itself. They used a technique involving two-dimensional networks of mathematical objects, which they contracted using a method called loop-corrected belief propagation. This allowed them to calculate the necessary correction factors with high precision, even though calculating the full, noise-free answer directly would have been too difficult for their classical computers.

To make this work, the team needed to generate a set of noisy measurements from the quantum computer that varied in a predictable way. Traditionally, this is done by artificially amplifying the noise, a process that requires running the experiment many, many times to get a reliable average. This is slow and expensive. The researchers instead introduced a technique they call "ergodic amplification." Rather than turning up the noise, they slightly tweaked the parameters of the quantum circuit itself—specifically, the strength of the interactions between the qubits. This change made the system behave as if it were noisier, without actually increasing the noise or requiring extra time on the quantum machine. This allowed them to gather the necessary data points much faster, making the entire process significantly more efficient.

The researchers tested their workflow on a specific model of a two-dimensional discrete time crystal, a system that had been studied before but is difficult to simulate accurately due to its complexity. They ran their experiments on a real quantum computer named ibm_basquecountry, which features 65 qubits arranged in a specific heavy-hexagonal pattern. They compared three different ways of getting the answer: a purely classical simulation that was simplified to run quickly, a purely classical simulation that was run with extreme precision (acting as a reference), and their new hybrid method. The results were striking. The simplified classical simulation failed to predict the correct behavior, even when the researchers tried to mathematically correct for its errors. The raw data from the quantum computer was also inaccurate due to noise. However, the hybrid method, which combined the noisy quantum data with the classical correction factors, produced a result that closely matched the high-precision reference simulation.

A key finding of the study is that the errors introduced by the simplified classical calculations did not ruin the final result. This might seem counterintuitive: if the classical part of the calculation is imperfect, why does the final answer come out so clean? The researchers explain that the errors in the correction factors tend to cancel each other out when they are applied to the noisy quantum data. Furthermore, the specific way the errors behave means that even if the correction factors are slightly off, the final combined answer remains stable. This allows the method to work even when the classical computer is not powerful enough to simulate the whole system perfectly. The team also found that by running the classical part of their calculation on a graphics processing unit, they could speed up the process by a factor of 5.6, making the method even more practical for larger problems.

The study also highlights the importance of knowing how confident one should be in the results. The researchers developed a way to estimate the error bars, or the range of uncertainty, for their final answer. They found that their method provided a reliable estimate of the true value, with error bars that were small enough to be useful. In contrast, other methods that rely on fitting noisy data to a curve often struggle to provide such reliable estimates, especially when the data is sparse or the noise is complex. By combining the quantum measurements with the classical calculations, the researchers were able to produce a result that was not only more accurate but also came with a clear measure of its own reliability.

While the method showed great promise for this specific type of quantum system, the researchers are careful to note that it may not work for every possible quantum circuit. The success of the technique depends on finding the right way to tweak the circuit parameters to generate the necessary data points. If the tweaks are too extreme, the mathematical corrections become unstable, and the error bars grow too large to be useful. For the specific model they tested, the researchers found a set of tweaks that worked perfectly, but they acknowledge that applying this to a completely different type of circuit would require finding a new set of rules. They suggest that for some circuits, the traditional method of amplifying noise might still be necessary, even though it is slower.

The work represents a significant step forward in the field of quantum error mitigation, demonstrating that the most effective path forward may not be to make quantum computers perfect or to make classical computers infinitely powerful, but to find a way for them to work together. By using classical computers to do the heavy lifting of calculating the correction factors, and quantum computers to provide the raw data that is otherwise impossible to simulate, the researchers have created a workflow that is greater than the sum of its parts. This approach opens the door to studying complex quantum systems that were previously out of reach, offering a glimpse of a future where quantum computers can solve real-world problems with the help of smart, classical partners.

The researchers conclude that their method, Lindblad Multiproduct Formulas, offers a robust way to extract accurate information from noisy quantum devices. It is a testament to the idea that in the age of quantum computing, the most powerful tool may be the ability to blend the strengths of different computational approaches. As quantum hardware continues to improve, techniques like this will be essential for unlocking the full potential of these machines, allowing scientists to explore the strange and wonderful behaviors of the quantum world with a clarity that was previously impossible.

Technical Summary: Lindblad Multiproduct Formulas

Problem Statement

Estimating expectation values of observable operators in the presence of noise is a central challenge in quantum computing. While techniques like Probabilistic Error Cancellation (PEC) and Zero Noise Extrapolation (ZNE) exist, they face significant scaling hurdles. PEC requires an exponentially increasing number of samples for deeper circuits, while ZNE often suffers from bias in fitting expectation values against noise rescaling parameters. Furthermore, purely classical simulation techniques, particularly those based on tensor networks, struggle to accurately simulate large, two-dimensional quantum systems due to the rapid growth of entanglement and the lack of a canonical form for truncation in two dimensions.

The authors investigate whether a hybrid error mitigation workflow can be designed that leverages state-of-the-art high-dimensional tensor network techniques. The core hypothesis is that while calculating the exact expectation values of observables for a given problem may be classically intractable, the specific quantities required to mitigate errors (i.e., the coefficients for a linear combination of noisy measurements) may be computationally cheaper to evaluate using tensor networks.

Methodology: Lindblad Multiproduct Formulas (L-MPF)

The paper introduces Lindblad Multiproduct Formulas (L-MPF), a technique that combines quantum hardware measurements with classical tensor network calculations to estimate noiseless expectation values. The workflow proceeds in three stages:

Noise Characterization and Data Acquisition (QPU):
- A noise model is learned from the quantum device (specifically ibm_basquecountry), yielding Pauli-Lindblad noise rates $\lambda_a$ .
- Noisy expectation values $\langle O \rangle_{\lambda_i}$ are measured on the quantum hardware.
- To generate these noisy values, the authors employ two methods:
  - Probabilistic Error Amplification (PEA): A standard technique that amplifies noise rates via sampling.
  - Ergodic Amplification (New Method): A technique that perturbs the parameters of the system's Hamiltonian (specifically the single-qubit gate parameters) to increase ergodicity, thereby mimicking the effect of increased noise without additional sampling overhead.
Classical Coefficient Calculation (CPU/GPU):
- The goal is to find coefficients $c_i$ such that the linear combination $\sum c_i \langle O \rangle_{\lambda_i}$ approximates the noiseless value $\langle O \rangle_{\lambda_0}$ (where $\lambda_0 = 0$ ).
- This is formulated as minimizing a cost function $F(c)$ , which represents the projection error of the noisy observables onto the target noiseless observable.
- The cost function depends on overlaps $F_{ij} = \langle\langle O | \Lambda^T_{\lambda_i} \Lambda_{\lambda_j} | O \rangle\rangle$ , where $\Lambda_\lambda$ is the vectorized Lindblad evolution operator.
- These overlaps are evaluated using two-dimensional tensor networks contracted via loop-corrected belief propagation. This approach avoids the need for a specific "middle-out" contraction path or the storage of high-dimensional Matrix Product Operators (MPOs), which would incur prohibitive memory costs.
- Clifford Rescaling: The workflow incorporates Clifford rescaling techniques to determine a global damping factor $d$ , imposing the constraint $\sum c_i = d$ .
- Error Estimation: The method outputs an estimated error bar by combining the projection error and the tensor network truncation error.
Mitigation:
- The final noiseless estimate is computed as $\langle O \rangle_{L-MPF} = \sum c_i \langle O \rangle_{\lambda_i}$ .
- Results are refined via bond dimension extrapolation to estimate the infinite bond dimension limit.

Key Contributions

L-MPF Algorithm: A new error mitigation scheme that uses 2D tensor networks to compute coefficients for a linear combination of noisy quantum measurements, rather than simulating the full state evolution.
Ergodic Amplification: A novel method to generate noisy expectation values by tuning Hamiltonian parameters (specifically the spin-flip coupling strength $\epsilon$ ) rather than amplifying noise rates via sampling. This significantly reduces the sampling cost compared to PEA.
Loop-Corrected Belief Propagation: The direct application of this state-of-the-art 2D tensor network technique to error mitigation, bypassing the need for specialized contraction paths (like middle-out) that are difficult to scale in two dimensions.
Error Analysis: Theoretical and empirical demonstration that the error on the mitigated observable can be smaller than the error of a purely classical tensor network simulation of the same observable. This is because tensor network errors in the coefficients enter the final result as second-order terms (products of coefficient errors and hardware noise errors).
Hardware Implementation: Successful implementation on 65 qubits arranged in a $3\times3$ heavy-hexagonal topology on the ibm_basquecountry quantum computer.

Results

The method was applied to a model of two-dimensional discrete time crystals (DTCs), a system previously studied in [1].

Accuracy: The L-MPF workflow produced estimates for the noiseless observable that were more accurate than those obtained from either the noisy quantum computer alone or purely classical tensor network simulations (even those with low bond dimensions).
Convergence: In convergence analysis, the L-MPF mitigated results converged to the high-bond-dimension classical reference faster than purely classical low-bond-dimension simulations, even when the classical simulations were extrapolated to the zero-truncation-error limit.
Performance: A GPU implementation of the classical part of the workflow achieved speedups of up to 5.6× compared to a single CPU processor, with speedups increasing as the number of Floquet cycles and truncation thresholds increased.
Ergodic vs. PEA: While PEA was tested in the appendix, the authors found that ergodic amplification was particularly well-suited for the DTC model, offering comparable accuracy with significantly reduced sampling costs.

Significance and Claims

The paper claims that L-MPF represents a promising direction for scalable quantum error mitigation, particularly for two-dimensional systems where classical simulation is difficult. The significance lies in the observation that the "hard" part of the problem (calculating the exact expectation value) does not need to be solved classically; instead, the "easier" part (calculating the correction coefficients) can be solved efficiently using tensor networks.

The authors emphasize that this approach allows for the direct incorporation of advancements in classical tensor network methods into quantum error mitigation workflows. They note that while the method requires careful consideration of the specific circuit to ensure error bars remain practical, it successfully mitigated errors in a non-trivial 2D system where purely classical methods struggled to match the accuracy of the hybrid approach. The work suggests that for certain classes of problems, the truncation error from the tensor network component of the mitigation workflow does not degrade the final result as severely as it would in a purely classical simulation.