| name | quantum-diffusion-time-series-generation |
| description | Quantum generative diffusion model for real-world time series (QDiffusion-TS) - replaces feed-forward layers in diffusion transformers with QNNs, achieves ~1000x parameter reduction, 44% better Wasserstein distance, 71% forecasting improvement |
| tags | ["quantum","diffusion-model","time-series","generative-ai","qnn","hybrid-quantum-classical","finance","synthetic-data","parameter-efficient","wasserstein"] |
QDiffusion-TS: Quantum Generative Diffusion for Time Series
Paper Summary
Title: Quantum Generative Diffusion Model for Real-World Time Series
arXiv: 2606.27561
Authors: Jack Waller, Filippo Caruso, Dimitrios Makris, Rajagopal Nilavalan, Xing Liang
Date: June 25, 2026
Categories: cs.LG, quant-ph
Core Innovation
QDiffusion-TS is the first quantum generative diffusion model for time series synthesis. It extends a classical diffusion architecture by replacing feed-forward components within the denoising transformer with quantum neural networks, yielding a hybrid quantum transformer that reduces trainable parameters by nearly three orders of magnitude while outperforming classical baselines.
Key Results
Parameter Efficiency
- ~1000x parameter reduction in replaced components compared to classical counterparts
- Maintains or exceeds classical model performance despite massive compression
Generation Quality
- 44% reduction in Wasserstein distance vs classical model on Apple and Amazon stock data
- Synthetic data more accurately reproduces real distribution statistics
Downstream Task Performance
- Up to 71% improvement in RMSE for forecasting tasks when augmenting training data with quantum-generated synthetic data
- Validated on real financial time series
Architecture
Hybrid Quantum Transformer
- Diffusion backbone: Standard diffusion model for time series
- QNN substitution: Feed-forward layers replaced with parameterized quantum circuits
- Quantum encoding: Time series features mapped to quantum states via amplitude/angle encoding
- Measurement readout: Quantum measurement outcomes feed back into classical pipeline
Training Pipeline
- Train classical diffusion model first as reference
- Replace selected layers with QNN components
- Joint fine-tuning of quantum-classical interface
- Validate on real quantum hardware (IQM processor)
When to Use
- Time series data augmentation for forecasting
- Parameter-constrained generative modeling
- Financial data synthesis
- Any diffusion-based pipeline needing parameter reduction
- Hybrid quantum-classical ML experimentation
Implementation Notes
- Shallow quantum circuits sufficient (compatible with NISQ)
- Amplitude encoding for real-valued time series features
- Gradient-based optimization through quantum circuits
- Real hardware validation recommended over simulation-only
Pitfalls
- Quantum noise affects training stability on real hardware
- Encoding strategy must match data distribution
- Barren plateau risk with deep circuits
- Classical baseline still strong - quantum advantage is parameter-efficiency, not raw accuracy
Related Skills
quantum-generative-diffusion-medical - Quantum diffusion for medical imagingquantum-reservoir-finance - Quantum reservoir computing for financial time serieshybrid-quantum-ml-timeseries-forecasting - Hybrid quantum ML for time series
References
- arXiv:2606.27561 (June 25, 2026)
