Quantum Generative Diffusion Model for Real-World Time Series

Technical Summary: Quantum Generative Diffusion Model for Real-World Time Series

Problem Statement
Generative models have achieved significant success in data synthesis, particularly in image generation, but recent advancements rely heavily on increasing model scale, leading to substantial computational costs and energy consumption. While generative models have been applied to time series data, accurately modeling complex, high-dimensional distributions—especially in financial contexts characterized by non-stationarity, heavy-tailed distributions, and volatility clustering—remains challenging. Existing approaches, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and diffusion models, often require billions of parameters to achieve state-of-the-art performance. The paper addresses the need for a more parameter-efficient generative framework that can replicate the statistical properties of real-world financial time series without the prohibitive computational overhead of large-scale classical models.

Methodology
The authors propose QDiffusion-TS, the first quantum generative diffusion model specifically designed for time series synthesis. The framework is built upon the classical Diffusion-TS architecture, which utilizes an encoder-decoder transformer to capture both short-term dynamics and long-range dependencies via self-attention mechanisms.

The core innovation lies in the integration of Quantum Neural Networks (QNNs) into the denoising process. Specifically, the feed-forward neural network (FFNN) components within the transformer's encoder and decoder blocks are entirely replaced by parameterized quantum circuits (PQCs).

• Architecture: The model employs a hybrid quantum transformer. Input data is encoded into a quantum state (using amplitude encoding in simulation and angle encoding on hardware), processed through a variational circuit, and measured to produce classical outputs.
• Parameter Reduction: This substitution drastically reduces the number of trainable parameters. Each replaced FFNN component is reduced from 33,088 parameters to just 36, representing a reduction of nearly three orders of magnitude per component. The total model parameter count decreases from 589,030 to 461,254 (an overall reduction of approximately 20%).
• Training and Inference: The model is trained to predict the original data sample from a noisy state, utilizing a loss function that combines standard reconstruction error with a Fourier-based term to preserve temporal frequency characteristics. While training is conducted on a classical simulator (using the Pennylane library), the inference stage is validated on the IQM Emerald quantum processor.

Key Contributions

1. First QDM for Time Series: The paper introduces QDiffusion-TS, marking the first application of a quantum diffusion model to the synthesis of time series data, extending the paradigm beyond image generation.
2. Hybrid Quantum-Transformer Architecture: The work demonstrates a practical method for integrating QNNs into modern deep learning architectures, specifically replacing FFNNs in transformers to create a "quantum transformer" for denoising.
3. Parameter Efficiency: The study highlights that quantum components can achieve comparable or superior performance with significantly fewer parameters (nearly three orders of magnitude fewer in replaced components), offering a path toward more efficient generative modeling.
4. Hardware Validation: The approach is validated not only in simulation but also on real quantum hardware (IQM Emerald), demonstrating feasibility within the constraints of near-term devices.

Results
The model was evaluated on financial time series data from Apple and Amazon, focusing on the reproduction of log-return distributions and downstream forecasting tasks.

• Statistical Fidelity: QDiffusion-TS generated synthetic data that more accurately reproduced the real data distributions compared to its classical counterpart. The model reduced the Wasserstein distance by approximately 44% on average across features. The improvement was particularly notable in the reproduction of variance (63% more accurate on average) and the overall shape of the distribution.
• Downstream Forecasting: When synthetic data generated by QDiffusion-TS was used to augment training datasets for a BiLSTM forecasting task, predictive performance improved by up to 71% in RMSE compared to a baseline trained solely on real data. The quantum-generated data achieved peak forecasting accuracy comparable to classical-generated data.
• Hardware Performance: Execution on the IQM Emerald device closely matched simulated performance and, in some metrics, marginally outperformed the simulation (reducing Wasserstein distance by ~89% relative to the classical model on hardware). The results suggest that hardware noise did not degrade performance and may have introduced beneficial stochasticity.
• Data Regimes: While the quantum model generally achieved lower statistical distances across various training set sizes, the advantage was not exclusively confined to low-data regimes, suggesting the improvement stems from better capture of statistical structure rather than just sample efficiency.

Significance and Claims
The paper claims that QDiffusion-TS establishes a practical framework for more efficient and scalable data-driven generative modeling. The primary significance lies in demonstrating that quantum-enhanced architectures can consistently match and frequently surpass classical performance while utilizing substantially fewer parameters.

The authors argue that the high-dimensional Hilbert space of quantum systems allows for the representation of complex feature correlations with fewer parameters, and entanglement operations enable efficient modeling of strongly correlated features. Furthermore, the successful validation on real quantum hardware indicates that current devices are capable of supporting such generative tasks. The work suggests that integrating QNNs into large-scale architectures is a viable route toward hybrid models where quantum components play progressively larger roles, potentially mitigating the computational costs associated with scaling classical generative models. The paper remains modest regarding the specific mechanisms of quantum advantage in low-data regimes, noting that the hybrid nature of the model (retaining significant classical components) may limit the realization of specific low-data advantages in this context.