Magnitude Is All You Need? Rethinking Phase in Quantum Encoding of Complex SAR Data
Research overturns assumption that quantum models need complex-valued data, showing simpler encoding yields 99.57% accuracy.
A new quantum machine learning (QML) study challenges fundamental assumptions about how to encode complex-valued data like Synthetic Aperture Radar (SAR) imagery. Researchers Sakthi Prabhu Gunasekar and Prasanna Kumar R systematically evaluated five quantum encoding strategies—including magnitude-only, joint complex, and I/Q-based approaches—on the MSTAR benchmark dataset for Automatic Target Recognition (ATR). Contrary to expectations that quantum models operating in complex Hilbert spaces would benefit from complex-valued data, their results showed magnitude-only encoding consistently outperformed all phase-aware methods in hybrid quantum-classical architectures, achieving 99.57% accuracy on 3-class tasks and 71.19% on 8-class tasks.
In hybrid models, phase-aware methods provided negligible or even negative improvements (~0%), suggesting classical components effectively compensate for missing phase information. However, the study revealed a crucial architectural dependency: in purely quantum architectures with only 184-224 trainable parameters and no classical components, phase information became essential, contributing up to 21.65% accuracy improvements. This finding demonstrates that phase utility isn't inherent to the data but depends critically on model architecture.
The research provides practical design guidelines for encoding complex-valued data in QML applications, highlighting the importance of encoding-architecture co-design. For practitioners working with current NISQ-era quantum hardware, this means simpler magnitude-only encoding may yield better results in hybrid systems, while pure quantum approaches require careful phase handling. The paper, currently under review for IEEE QCE 2026, represents a significant step toward more efficient quantum algorithms for real-world applications like defense, remote sensing, and image analysis.
- Magnitude-only encoding achieved 99.57% accuracy on 3-class SAR target recognition, outperforming all complex-valued strategies in hybrid models
- Phase information provided up to 21.65% improvement in purely quantum architectures but negligible benefits in hybrid systems
- Study reveals phase utility depends on architecture, not data, with hybrid models compensating via classical components
Why It Matters
Provides practical quantum encoding guidelines that could simplify implementation and improve performance for defense, remote sensing, and image analysis applications.