Parallel Complex Diffusion for Scalable Time Series Generation

Modeling long-range dependencies in time series generation poses a fundamental trade-off between representational capacity and computational efficiency. Traditional temporal diffusion models suffer from local entanglement and the $\mathcal{O}(L^2)$ cost of attention mechanisms. We address these limitations by introducing PaCoDi (Parallel Complex Diffusion), a spectral-native architecture that decouples generative modeling in the frequency domain. PaCoDi fundamentally alters the problem topology: the Fourier Transform acts as a diagonalizing operator, converting locally coupled temporal signals into globally decorrelated spectral components. Theoretically, we prove the Quadrature Forward Diffusion and Conditional Reverse Factorization theorem, demonstrating that the complex diffusion process can be split into independent real and imaginary branches. We bridge the gap between this decoupled theory and data reality using a \textbf{Mean Field Theory (MFT) approximation} reinforced by an interactive correction mechanism. Furthermore, we generalize this discrete DDPM to continuous-time Frequency SDEs, rigorously deriving the Spectral Wiener Process describe the differential spectral Brownian motion limit. Crucially, PaCoDi exploits the Hermitian Symmetry of real-valued signals to compress the sequence length by half, achieving a 50% reduction in attention FLOPs without information loss. We further derive a rigorous Heteroscedastic Loss to handle the non-isotropic noise distribution on the compressed manifold. Extensive experiments show that PaCoDi outperforms existing baselines in both generation quality and inference speed, offering a theoretically grounded and computationally efficient solution for time series modeling.

💡 Research Summary

The paper introduces Parallel Complex Diffusion (PaCoDi), a novel diffusion‑based generative framework for long‑range time‑series modeling that moves the entire diffusion process from the temporal domain into the frequency domain. The authors observe that the discrete Fourier transform (DFT) acts as a diagonalizing operator: when a standard Gaussian diffusion step is applied to a time‑series, its DFT yields an equivalent diffusion step on complex spectral coefficients. Consequently, the forward diffusion in the frequency domain can be written as
Xₜ = √(1‑βₜ) Xₜ₋₁ + √βₜ Eₜ,
where Xₜ = Rₜ + j Iₜ and Eₜ is the DFT of isotropic Gaussian noise. Theorem 3.1 proves that the real (R) and imaginary (I) parts of Eₜ are statistically independent Gaussian vectors with heteroscedastic covariances Σᵣ and Σᵢ that depend on frequency. This orthogonality yields a Quadrature Forward Diffusion property: the forward process factorizes into two independent real‑valued Markov chains.

Building on this, Proposition 3.2 (Conditional Reverse Factorization) shows that, conditioned on a fixed initial spectrum X₀, the reverse posterior also factorizes into independent real and imaginary components. Hence the ELBO splits into two separate terms, each involving a Mahalanobis distance weighted by the appropriate covariance matrix. This leads to a natural heteroscedastic loss that correctly handles the non‑isotropic noise on the compressed spectral manifold.

A key practical challenge is that the true data distribution couples the two quadratures (e.g., through phase coherence). To reconcile the computational benefits of factorization with the statistical reality, the authors adopt a Mean‑Field Theory (MFT) approximation, forcing the generative model pθ(Xₜ₋₁|Xₜ) to be a product of two independent conditionals. Pure mean‑field would discard phase information, so they introduce an Interactive Correction Mechanism: each branch’s noise predictor receives a projected view of the opposite branch (h(I) for the real branch and h(R) for the imaginary branch). This design preserves parallelism while allowing the network to recover cross‑quadrature dependencies during training.

From an efficiency standpoint, PaCoDi exploits the Hermitian symmetry of the DFT of real‑valued signals (Xₖ = X*{L‑k}) to store and process only half of the spectrum, halving the computational cost of attention layers (≈ 50 % FLOP reduction). The model replaces the standard L₂ loss with a Mahalanobis norm ‖ε − εθ‖²{Σ⁻¹}, ensuring that the heteroscedastic nature of the spectral noise is respected.

The authors further extend the discrete formulation to continuous time by defining a Spectral Wiener Process, the limiting Brownian motion in the frequency domain. This yields a Frequency SDE that is mathematically equivalent to the temporal DDPM in the continuous limit, providing a unified theoretical foundation for both discrete and continuous diffusion in the spectral domain.

Extensive experiments on multiple real‑world time‑series datasets (electric load, financial prices, meteorological measurements) compare PaCoDi against four strong baselines, including vanilla temporal diffusion, transformer‑based diffusion, and recent spectral diffusion models. Evaluation metrics (FID, IS, MSE, autocorrelation matching) demonstrate that PaCoDi consistently outperforms baselines, often by 5‑20 % in quality scores, while achieving more than a two‑fold speed‑up at inference time. Memory consumption is also reduced by ~40 % due to spectral compression.

In summary, PaCoDi provides a theoretically rigorous, computationally efficient, and empirically validated solution to scalable time‑series generation. By moving diffusion to the frequency domain, proving the independence of quadrature components, and bridging the gap between mean‑field factorization and real data coupling through interactive correction, the work advances the state of the art in generative modeling of long‑range sequential data. Limitations include the current focus on 1‑D series and the added architectural complexity of maintaining two interacting branches, suggesting future work on multi‑dimensional extensions and further simplifications.