# Limitations of Amplitude Encoding on Quantum Classification

**Authors:** Xin Wang, Yabo Wang, Bo Qi, Rebing Wu  
**arXiv:** [2503.01545](https://arxiv.org/abs/2503.01545)  
**Date:** 2025-03-03  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2503.01545v1/

## Key points

- Authors: Xin Wang, Yabo Wang, Bo Qi, Rebing Wu
- Topic: Limitations of Amplitude Encoding on Quantum Classification
- Quantum machine learning (QML) shows promise for complex problems
- Amplitude encoding is a common technique but has limitations
- Quantum states encoded via amplitude tend to concentrate towards specific state
- Concentration phenomenon creates loss barrier for quantum classifiers
- Loss function reaches lower bound hindering classification accuracy improvement
- Training error may paradoxically rise with increased training data volume
- Diminished gains in prediction accuracy on new data sets due to unexpected behavior
- Need for more efficient encoding techniques in QML to unlock full potential

## Summary

In their paper titled "Limitations of Amplitude Encoding on Quantum Classification," authors Xin Wang, Yabo Wang, Bo Qi, and Rebing Wu delve into the realm of quantum machine learning (QML) to explore the efficacy of amplitude encoding in quantum classification tasks. While QML holds promise for tackling complex real-world problems, the authors highlight a critical limitation associated with amplitude encoding - a widely used technique for encoding classical data into quantum states. The study reveals that under common data assumptions, quantum states encoded via amplitude encoding tend to concentrate towards a specific state. This concentration phenomenon poses a significant challenge for quantum classifiers by creating a loss barrier that cannot be overcome through optimization algorithms. In essence, the loss function reaches a lower bound that hinders further improvement in classification accuracy. Moreover, through numerical simulations, the authors uncover a counterintuitive outcome of amplitude encoding: as the volume of training data increases, the training error may paradoxically rise instead of decreasing. This unexpected behavior leads to diminished gains in prediction accuracy on new data sets and highlights the limitations of current encoding strategies in QML. Overall, these findings underscore the need for more efficient encoding techniques in quantum machine learning to unlock its full potential. By addressing these limitations and exploring alternative approaches to data representation, researchers can enhance the performance and applicability of quantum classifiers in practical and meaningful tasks.