⚛️ quantum physics

Tensor network characterization and mitigation of readout errors

This paper introduces a scalable tensor-network framework that models readout errors as matrix product operators to efficiently characterize and mitigate correlated noise in near-term quantum processors, demonstrating superior accuracy over uncorrelated models across various tasks and system sizes.

Original authors: Yuchen Guo, Shuo Yang

Published 2026-06-25

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuchen Guo, Shuo Yang

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 are trying to listen to a very quiet conversation in a room full of static. In the world of quantum computers, this "conversation" is the data the computer produces, and the "static" is readout error.

When a quantum computer finishes a calculation, it has to translate its quantum state into a simple "0" or "1" (like a light switch being off or on). Unfortunately, the hardware isn't perfect. Sometimes a "0" gets misread as a "1," or vice versa. This is the readout error.

For a long time, scientists treated these errors like independent, random glitches. They assumed that if Qubit A made a mistake, it had nothing to do with Qubit B. It was like assuming that if one person in a crowded room sneezed, it didn't affect anyone else.

The Problem with the Old Way
In reality, quantum chips are crowded. The wires and controls are close together. If Qubit A sneezes (makes an error), it often causes Qubit B to sneeze too. This is called correlated error. The old "independent" models missed this, leading to inaccurate results, especially when scientists tried to measure complex patterns involving many qubits at once.

The New Solution: The "Tensor Network" Translator
The authors of this paper, Yuchen Guo and Shuo Yang, propose a new way to fix this using a mathematical tool called a Tensor Network (specifically, a Matrix Product Operator, or MPO).

Think of the old method as trying to fix a noisy phone call by asking each person to repeat their own sentence separately. It misses the context.

The new method is like hiring a super-smart translator who understands the whole conversation at once.

Learning the Noise: First, the team "trains" this translator. They feed it a bunch of known test data (calibration data) and let it learn the specific patterns of the noise. Instead of looking at errors one by one, the MPO looks at the "neighborhood" of qubits, understanding that if one is wrong, its neighbors are likely wrong too.
Efficiency: The beauty of this translator is that it's incredibly efficient. Even as the quantum computer grows from 10 qubits to 20 or more, the amount of work the translator needs to do only grows slowly (almost linearly), rather than exploding exponentially.

What They Tested It On
The team didn't just write theory; they tested this on a real superconducting quantum chip (the "Baihua" chip) and in massive computer simulations. Here is what they found:

Real Hardware: On the real chip, the new MPO model was much better at describing the actual noise than the old models. It caught the "sneezing neighbors" that the old models ignored.
Measuring Complex Things: When trying to measure "non-local" things (properties that depend on the relationship between distant qubits), the new method gave much more accurate answers.
Random Circuits: For tasks that involve generating random patterns (like testing if a quantum computer is truly powerful), the new method could clean up the data to reveal the true pattern hidden beneath the noise.
Error Correction: They even showed how this translator could be built directly into the "decoder" for quantum error correction. Instead of just fixing the final answer, it helps the computer decide the most likely correct path while it is decoding the message, leading to fewer mistakes in the final logical result.

The Bottom Line
This paper introduces a unified, scalable way to clean up the "static" in quantum computers. By using a smart, network-based translator that understands how errors spread between neighbors, they can extract much more reliable information from current noisy machines. This makes the data from today's quantum computers more trustworthy and paves the way for larger, more complex experiments in the future.

Technical Summary: Tensor Network Characterization and Mitigation of Readout Errors

Problem Statement

Readout errors remain a dominant noise source in near-term intermediate-scale quantum (NISQ) devices, directly distorting classical measurement outcomes and biasing tasks such as expectation-value estimation, tomography, and sampling benchmarks. Standard readout error mitigation (REM) approaches typically characterize a classical transition matrix $\Lambda$ that maps ideal distributions to noisy ones. However, full characterization of $\Lambda$ scales exponentially with system size ( $2^N \times 2^N$ ). To circumvent this, existing methods often assume independent single-qubit readout processes (uncorrelated models). These factorized models neglect spatial correlations induced by crosstalk, shared control lines, and hardware-specific mechanisms. Such correlations are often non-negligible in real devices, particularly for nonlocal observables and large-scale sampling tasks, limiting the performance of uncorrelated mitigation strategies.

Methodology

The authors propose a unified tensor-network framework that models the readout error matrix $\Lambda$ as a Matrix Product Operator (MPO). This representation naturally captures local and short-range correlations with polynomial complexity.

1. Characterization via MPO Learning

Representation: The readout error matrix $\Lambda_{x,y}$ (probability of outcome $x$ given true state $y$ ) is represented as $|\text{Tr}(M^{[1]}_{x_1,y_1} \cdots M^{[N]}_{x_N,y_N})|^2$ , where $M^{[k]}$ are local tensors with a finite bond dimension $\chi$ . This ensures the positivity of matrix elements.
Normalization: Unlike Matrix Product Unitaries (MPUs), the MPO model only requires the row sums to be unity ( $\sum_x \Lambda_{x,y} = 1$ ), a weaker condition suitable for classical stochastic processes.
Training: The MPO parameters are learned directly from calibration data (preparing and measuring product states) using a likelihood-based optimization. The loss function minimizes the negative log-likelihood of observed outcomes, augmented with a penalty term to enforce the normalization constraint. Gradients are computed efficiently via tensor network contraction.

2. Mitigation Strategies

The framework integrates the learned MPO model into various REM tasks:

Nonlocal Observables: The ideal expectation value $\langle O \rangle_{ideal}$ is estimated by contracting the observable (represented as an MPS) with the inverse of the readout error matrix ( $\Lambda^{-1}$ ). The inverse is approximated variationally using a single-layer MPO to minimize $\|\Lambda \Omega - I\|_F^2$ .
Global Sampling: To recover samples from the ideal distribution $P_{ideal}$ given noisy samples, the method learns an MPS model for $P_{ideal}$ by maximizing the likelihood of the noisy data under the transformation $P_{noisy} = \Lambda P_{ideal}$ .
Random Measurements: For tasks like classical shadows and learning-based tomography (e.g., using Locally Purified Density Operators, LPDOs), the readout error model is integrated directly into the optimization objective. Instead of post-processing samples, the likelihood function is modified to sum over the true states weighted by the readout error matrix $\Lambda$ .
Quantum Error Correction (QEC) Decoding: For 2D surface codes, the MPO model is generalized to a Projected Entangled-Pair Operator (PEPO). The readout error model is fused locally with the decoder's tensor network (PEPS) to perform joint inference over data errors and readout errors, effectively treating the syndrome readout as a soft-information input rather than hard-voted bits.

Key Results

Experimental Validation (Superconducting Processor)

Experiments were conducted on the Baihua superconducting quantum chip (Quafu platform) using 6-qubit chains.

Correlation Capture: The MPO model ( $\chi=4$ ) consistently outperformed the uncorrelated product model in characterizing the full error matrix $\Lambda_{Full}$ . The MPO model achieved a relative distance of $\approx 0.043$ from the full characterization, compared to $\approx 0.074$ for the uncorrelated model, demonstrating its ability to capture correlated noise missed by standard models.
Nonlocal Observables: In experiments on 7-qubit cluster states, the MPO-based REM significantly improved the estimation of nonlocal string order parameters, particularly for longer strings where errors are more pronounced.

Numerical Simulations

Scalability: Simulations up to $N=20$ qubits showed that the sample cost required to learn the MPO model scales nearly linearly with system size, a stark contrast to the exponential cost of full characterization.
Global Sampling: For random circuit sampling, the method improved the Cross-Entropy Benchmarking (XEB) fidelity. While generating new ideal samples required an MPS with exponentially growing bond dimension, the method successfully inferred the XEB fidelity directly from noisy samples using tensor contraction, eliminating readout bias without explicit sampling.
Random Measurements: In classical shadow estimation and LPDO-based tomography of a 1D XY model Gibbs state, the REM method significantly reduced bias in magnetization estimates and improved density matrix reconstruction compared to unmitigated methods.
QEC Decoding: Simulations on a distance- $d=7$ rotated surface code showed that the PEPO-fused decoder significantly reduced the logical error rate compared to standard decoders, especially in regimes with high readout flip rates. The method effectively utilized soft information from repeated measurements, outperforming majority-vote approaches.

Significance and Claims

The paper establishes tensor-network readout error mitigation as a scalable and versatile approach for noise-aware quantum data processing. Its primary contributions are:

Unified Framework: It provides a single MPO/PEPO formalism that handles correlated readout errors across diverse applications, from local observables to global sampling and QEC decoding.
Efficiency: The approach avoids the exponential overhead of full error characterization while capturing correlations that uncorrelated models miss. The sample complexity grows near-linearly with system size.
Practicality: Experimental results on real hardware confirm that correlated readout errors are significant and that the MPO model effectively mitigates them.
Integration: The framework is designed to be a building block that can be integrated into existing algorithms (like classical shadows and QEC decoders) rather than acting solely as a post-processing step.

The authors conclude that this method offers a promising route toward reliable extraction of physical information from both near-term and future fault-tolerant quantum devices, particularly by enabling the treatment of measurement crosstalk within a unified tensor-network inference problem.