🔬 physics

Neural-Network Error Mitigation for Quantum Teleportation and Superdense Coding

This paper demonstrates that lightweight neural networks can effectively mitigate structured noise in quantum teleportation and superdense coding protocols through classical post-processing, significantly improving accuracy in low-shot regimes while offering no benefit against unstructured, random noise.

Original authors: Hritik Chaudhary, Gokul K.C.

Published 2026-09-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hritik Chaudhary, Gokul K.C.

Original paper licensed under CC BY 4.0 (https://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 strange and counterintuitive world of quantum physics, information behaves in ways that defy our everyday experience. Two of the most famous tricks in this realm are quantum teleportation and superdense coding. Teleportation allows a particle's exact state to be transferred from one place to another without the particle itself traveling through the space in between, relying on a mysterious connection called entanglement and a small amount of classical data. Superdense coding does the reverse, allowing two pieces of information to be packed into a single particle and sent across a distance. Both tricks are fundamental to the future of secure communication and powerful computing. However, these delicate processes are incredibly fragile. In the real world, the equipment used to create and measure these particles is never perfect. Tiny imperfections in the machines, or "noise," can scramble the information, causing the teleportation to fail or the message to arrive garbled. Scientists have long sought ways to fix these errors, often by building more complex quantum machines, but a new approach asks if a simple computer program could clean up the mess after the fact.

Researchers Hritik Chaudhary and Gokul K.C. from Kathmandu University in Nepal investigated whether a small, artificial intelligence program could act as a digital filter to repair these damaged messages. They did not try to build a new quantum device or change the laws of physics. Instead, they simulated the entire process on a classical computer with perfect precision first, creating a baseline of what a flawless transmission looks like. Then, they introduced realistic errors into the simulation, mimicking the kinds of mistakes that happen in actual laboratories, such as miscalibrated sensors or biased measurement tools. Crucially, they set up the experiment so that the computer program acting as the "fixer" knew nothing about the errors. It was only shown the final, messy results of the measurements, just as a real human operator would see them, without being told what went wrong or how the machine was broken. The goal was to see if the program could learn to recognize the pattern of the mistakes and reverse them on its own.

The team tested this idea on both teleportation and superdense coding under different conditions. In one scenario involving superdense coding, they simulated a situation where the measurement tools were biased, meaning they were more likely to misread one type of signal than another. They found that when the number of measurements was very small—a common situation in real experiments where time or resources are limited—the trained computer program significantly outperformed a standard, unthinking method of guessing the most likely answer. Specifically, when the system was allowed only five measurements, the intelligent program correctly identified the message about five percentage points more often than the standard method. As the number of measurements increased, both methods improved and eventually reached near-perfect accuracy, but the advantage of the smart program was most valuable when data was scarce. This suggests that for tasks where gathering data is expensive or slow, a learned correction can make a tangible difference.

The results were even more striking in the teleportation experiment, where the researchers simulated a specific, consistent error: a systematic miscalibration of the equipment that rotated the information in a fixed, predictable way. In this case, the computer program learned to recognize the specific signature of this rotation and effectively "un-rotated" the data. By doing so, it raised the quality of the recovered information, known as fidelity, from about 93 percent to 99 percent. The improvement grew stronger as the researchers provided more measurements, because more data helped the program distinguish the consistent error from random noise. However, the researchers were careful to test the limits of this idea. In a control experiment, they replaced the consistent error with a completely random, unpredictable jumble of mistakes that changed every single time. In this chaotic scenario, the computer program offered no help at all; in fact, it performed slightly worse than doing nothing. This proved that the program's success was not magic, but a direct result of its ability to learn and exploit a repeating pattern in the noise.

The study concludes that this kind of lightweight, learned correction is a powerful tool, but only under specific circumstances. It shines when the errors are structured and repeatable, such as a machine that is consistently slightly off-kilter, and when the amount of data available is limited. It offers no benefit when the noise is random and chaotic, as there is no pattern for the program to learn. This finding provides a clear roadmap for the future of quantum technology. It suggests that rather than always trying to build more perfect, error-free quantum machines, engineers might be able to achieve better results by pairing their hardware with smart, simple software that knows how to clean up the specific, recurring mistakes of their particular devices. The work demonstrates that while we cannot eliminate all noise, we can teach machines to ignore the parts of the noise that follow a rule.

Technical Summary: Neural-Network Error Mitigation for Quantum Teleportation and Superdense Coding

Problem Statement
Quantum teleportation and superdense coding are foundational protocols for quantum communication, yet their performance degrades significantly under channel imperfections such as entanglement decoherence, gate miscalibration, and measurement readout bias. While classical post-processing offers a route to recovery without additional quantum resources, evaluating neural network (NN) based mitigation is often complicated by models that are granted privileged information—such as exact outcome probabilities or known noise parameters—that a real receiver would not possess. This paper addresses the question of whether a small neural network, acting solely as a classical post-processing layer with access only to empirical finite-shot measurement data, can effectively mitigate structured noise in these protocols.

Methodology
The study employs a strictly separated architecture where the quantum protocols are simulated exactly using tensor operations on complex qubit amplitudes, while the neural network operates exclusively as a post-processing decoder. The evaluation is conducted under "honest" conditions: the networks receive only empirical frequencies derived from a finite number of measurement shots and are never provided with the underlying noise parameters.

Two distinct scenarios are investigated:

Superdense Coding: The protocol is subjected to a biased Pauli channel (mimicking dephasing-dominated noise where $T_2 < T_1$ ) followed by an asymmetric readout confusion matrix. The network receives empirical outcome frequencies ( $\hat{p}$ ) and classifies the intended two-bit message. It is compared against a majority-vote baseline.
Quantum Teleportation: The protocol suffers from a systematic, fixed gate miscalibration (modeled as a deterministic rotation of the Bloch vector) combined with independent random jitter. The network receives a noisy Bloch vector estimate ( $\hat{r}$ ) derived from finite-shot tomography and attempts to reconstruct the clean state vector. It is compared against the uncorrected measured state.

Both models are small feed-forward networks trained in PyTorch. Crucially, a control experiment is performed where the systematic miscalibration in the teleportation scenario is replaced by a fully random, per-qubit rotation to test the limits of learning when noise lacks recurring structure.

Key Results
The results delineate a clear boundary for the efficacy of learned error mitigation:

Superdense Coding (Low-Shot Regime): In the presence of biased noise and asymmetric readout, the learned decoder outperforms the majority-vote baseline specifically in the low-shot regime. At five shots, the learned decoder achieves approximately 87.1% accuracy compared to 82.0% for the baseline (a ~5 percentage point gain). As the number of shots increases, both methods converge toward perfect accuracy, and the performance gap narrows. At a single shot, both methods perform identically, as no frequency information exists for the network to exploit beyond the single observed outcome.
Teleportation (Systematic Miscalibration): The network successfully learns to invert the systematic rotation, raising the mean state fidelity from approximately 0.93 (uncorrected) to 0.99. Unlike the superdense coding case, the benefit of the network increases with the number of shots. This is because higher shot counts reduce the sampling noise (tomographic noise), making the underlying systematic error more visible and easier for the network to learn and correct. The uncorrected fidelity plateaus near 0.93 regardless of shot count because the systematic error remains.
Control Experiment (Structureless Noise): When the systematic miscalibration is replaced by random, non-recurring noise, the neural network provides no benefit. In fact, it slightly degrades fidelity (by ~0.03) compared to the direct measurement. This confirms that the network cannot learn to correct noise that lacks a recurring structure.

Significance and Claims
The paper claims that its primary contribution is not a novel mitigation architecture, but a controlled characterization of when lightweight learned post-processing is worthwhile. The findings indicate that:

Learned correction is effective specifically in the low-shot, structured-noise regime. It excels when noise carries recurring patterns (such as gate miscalibration or biased readout) that can be learned from data, particularly when measurement resources are scarce.
Learned correction offers no advantage when noise is structureless (random) or when the noise parameters are already known to the system.
The approach is distinct from methods that assume access to noise models; it demonstrates that a receiver can improve performance using only the empirical data available in a real-world setting, provided the noise is not purely random.

The authors maintain a modest scope, noting that the noise models are software-based rather than hardware-characterized and that the experiments use single-qubit states. The results serve as a representative guide for where such techniques should be applied on physical devices, rather than a definitive proof of performance on specific hardware. Future work suggested includes extending these methods to multi-qubit codes and evaluating them on data from physical processors.

Technical Summary: Neural-Network Error Mitigation for Quantum Teleportation and Superdense Coding

More like this