Parallel Complex Diffusion for Scalable Time Series Generation

AI summary

Overview

Research area: Generative machine learning — diffusion models for time series synthesis, with a focus on spectral (frequency-domain) diffusion theory and computational efficiency.

Technical level: Advanced. The paper leans heavily on complex analysis, Itô stochastic differential equations, Wirtinger calculus, and mean field theory. Readers need prior grounding in diffusion models (DDPM/score-based SDEs) and Fourier analysis.

Scope (one sentence): The paper proposes PaCoDi, a framework that defines the time series diffusion path natively in the complex frequency domain and sidesteps the mathematical barrier of complex-valued neural estimators by splitting them into two parallel real-valued branches.

What This Paper Is About

Time series data has strong temporal dependencies, so a standard diffusion model must learn a highly entangled noise/score estimator that aggregates information across an entire sequence to denoise a single point — a burden the authors call the "curse of entanglement." Moving diffusion into the frequency domain via the Discrete Fourier Transform (DFT) helps because spectral decomposition diagonalizes second-order dependence into roughly orthogonal modes, but constructing a diffusion path directly in the complex domain is blocked by the fact that complex-differentiable (holomorphic) neural networks are severely restricted. The goal of PaCoDi is to keep the diffusion path in the spectral domain while replacing the complex-valued estimator with two parallel real-valued estimators for the real and imaginary components.

Key Contributions

1. Rigorous spectral diffusion theory. The authors establish quadrature forward transitions and conditional reverse factorization, both grounded in a proof of the statistical orthogonality of complex (Hermitian) Gaussian noise.
2. A unified spectral SDE framework. They define a Spectral Wiener Process to describe heteroscedastic spectral Brownian motion, and prove the equivalence between temporal, discrete-frequency, and continuous-frequency diffusions.
3. Approximate factorization with correction. They introduce a Mean Field Theory (MFT) approximation to keep the parallel architecture trainable after the initial data state is marginalized, plus an interactive correction branch to compensate for the phase coherence weakened by that approximation.
4. Efficiency and empirical performance. They exploit Hermitian symmetry to compress the spectral state space, cutting attention-block computational complexity by 50%, and report experiments on unconditional and conditional time series generation against 5 baselines on 5 benchmarks.

Main Findings

• Statistical orthogonality of spectral noise (Theorem 3.1). The DFT of isotropic Gaussian noise splits into real and imaginary components that are jointly Gaussian with Cov(ε_r, ε_i) = 0_{L×L}. Their covariances Σ_r and Σ_i are heteroscedastic, determined by frequency indices and the Nyquist frequency. The induced noise law factorizes perfectly: p(ℰ) = p(ε_r) p(ε_i).

• Conditional reverse factorization (Proposition 3.2). When conditioned on a fixed initial spectral state 𝒳_0, the reverse posterior factorizes into two independent marginals: q(𝒳_{t-1}|𝒳_t, 𝒳_0) = q(ℛ_{t-1}|ℛ_t, ℛ_0) · q(ℐ_{t-1}|ℐ_t, ℐ_0). Under the standard variational bound, the ELBO theoretically splits into two independent terms, so the real and imaginary trajectories can be trained as mathematically separable tasks.

• Mahalanobis, not Euclidean, objective. Because the complex noise components are heteroscedastic rather than isotropic, minimizing the ELBO corresponds to minimizing a Mahalanobis distance ‖z‖²_{Σ^{-1}} = z^T Σ^{-1} z rather than a simple Euclidean norm, with ELBO-derived weighting λ_t = (1 - α_t) / (2 α_t (1 - ᾱ_t)).

• A holomorphic-constraint workaround. Modeling the estimator as a native complex-valued network would require satisfying the Cauchy-Riemann equations, and the authors note that Liouville's Theorem forces any bounded entire holomorphic function to be constant — a tension between boundedness and nonlinear expressivity. PaCoDi avoids this by keeping the diffusion path in the frequency domain while making only the estimator real-valued and parallel, so each branch can use standard non-holomorphic activations.

• Marginal coupling and the MFT fix. The true conditional posterior is factorized, but the marginal q(𝒳_{t-1}|𝒳_t) integrates over the coupled data prior q(𝒳_0) and is inherently entangled (the authors cite phase coherence as the reason). The MFT proxy p_θ(𝒳_{t-1}|𝒳_t) ≈ p_{θ_r}(ℛ_{t-1}|ℛ_t) · p_{θ_i}(ℐ_{t-1}|ℐ_t) restores a factorized, trainable form, and the interactive correction branch feeds each quadrature a lightweight projection h(·) of its counterpart.

• Continuous-time extension (Theorem 4.2). Taking the continuum limit of the discrete transition yields a Variance Preserving (VP) Itô SDE d𝒳 = -½β(t)𝒳 dt + √β(t) d𝒲, where 𝒲_t = ℱ(w_t) is the Fourier transform of a standard temporal Wiener process. Its increments have quadrature orthogonality (dw_r ⟂ dw_i) and complex Gaussianity with heteroscedastic covariances Σ_r dt and Σ_i dt. The reverse-time SDE separates into two parallel real-valued SDEs.

• Score-noise identity. Using Wirtinger calculus, the complex score decomposes as ∇_{𝒳̄} log p_{t|0}(𝒳) = ½(∇_ℛ log p_{t|0}(𝒳) + j ∇_ℐ log p_{t|0}(𝒳)), and for heteroscedastic complex noise the conditional scores relate to the cumulative injected noise via the precision matrix: ∇_ℛ log p_{t|0}(ℛ) = -Σ_r^{-1} ε_r / √(1-ᾱ_t), with the analogous imaginary expression.

• Geometric and stability guarantees. The normalized DFT is unitary (U U^H = I), so Euclidean distance is preserved (Parseval's equivalence: ‖x - x̂‖²₂ = ‖𝒳 - 𝒳̂‖²₂), meaning minimizing the ELBO in frequency domain is mathematically identical to minimizing it in the temporal domain. Unitary transformations also preserve the spectral radius of the Hessian, so the noise schedule β_t and learning rates need no recalibration.

• 50% attention FLOP reduction. Hermitian symmetry lets the model discard redundant negative frequencies, giving a compressed state 𝒳̃ = [𝒳_0, …, 𝒳_K]^T ∈ ℂ^K with K = ⌊L/2⌋. Linear/MLP layers stay at iso-FLOPs (two branches at L/2 each), but self-attention drops from O(L²) to 2 × O(L²/4) = ½ O(L² · C). The O(L log L) FFT/iFFT overhead is described as asymptotically negligible. The compressed-space noise variances become σ²/2 (1 + δ_{k,L/2}) for the real part and σ²/2 (1 - δ_{k,L/2}) for the imaginary part.

• Reported baselines. For conditional generation the visible baseline list is T2S (Ge et al., 2025), Diff-TS (Yuan and Qiao, 2024), and TimeVAE (Desai et al., 2021), plus two zero-shot foundation models evaluated for generalization. The abstract states experiments cover 5 SOTA baselines on 5 benchmarks. The specific dataset names, benchmark names, and numerical results are not present in the available paper content — the text is truncated mid-sentence during the baseline description — so no accuracy, FID, or discriminative-score values can be reported here.

Methodology in Plain English

The authors start from a standard observation: diffusion models learn to remove noise step by step, and that job is hard when data points are tightly correlated, as adjacent time steps in a time series are.

Their fix is a change of coordinates. Instead of adding noise and denoising in the time domain, they Fourier-transform the sequence and run the whole diffusion process in the frequency domain. In frequency space, a stationary signal's correlations are approximately decoupled into separate modes, so the estimator no longer has to untangle the whole sequence at every step.

The obstacle they run into is that frequency coefficients are complex numbers, and building a complex-valued neural network that is properly complex-differentiable would be highly restrictive. Their solution is to prove that the noise in frequency space splits cleanly into independent real and imaginary halves when the initial state is known. That lets them keep the complex-valued diffusion path but train two ordinary real-valued networks, one per half.

Because the true initial state is unknown during generation, the two halves do not actually stay independent — the data's phase structure couples them. So the authors add a mean-field approximation to preserve the two-branch design, and then bolt on a small "correction" pathway where each branch sees a compact projection of the other. They also show the same construction in continuous time using a stochastic differential equation driven by what they call a spectral Wiener process, and they exploit a symmetry of real-valued signals (Hermitian symmetry) to throw away half the frequency coefficients, which halves the cost of the attention layers.

Why This Matters

Impact on research. The paper argues that existing frequency-aware time series diffusion methods only modify the neural estimator while leaving the diffusion path in the temporal domain — the "second trajectory" in their taxonomy. PaCoDi claims to be frequency-native, and its contribution is as much theoretical as practical: a proof that the complex spectral diffusion path is internally consistent, a bridge between discrete DDPM and continuous SDE formulations in the spectral domain, and a way around the holomorphic constraint that has blocked complex-valued diffusion modeling. It also offers a closed-form complexity argument for why spectral diffusion can be cheaper than temporal diffusion, not just more accurate.

Real-world applications (as enabled by time series generation of this type):

• Synthetic sensor, industrial process, and control-system data, given the authors' affiliations in cyber-systems and control (Zhejiang University) and process mathematics (Southeast University).
• Financial and economic scenario simulation, where realistic synthetic series are needed without exposing proprietary data.
• Healthcare and wearable-record synthesis, including generating privacy-preserving physiological time series for model development.
• Data augmentation and imputation pipelines in forecasting, anomaly detection, and e-commerce demand modeling, where downstream models are data-starved.
• Educational and personalized-learning platforms, given the Squirrel Ai Learning affiliation, where learner-interaction traces are long multivariate sequences.

Industry relevance. The headline efficiency claim — 50% attention FLOP savings with no information loss, plus no need to retune the noise schedule or learning rate — targets the practical cost of training and serving long-sequence diffusion models. The code is released at https://github.com/RongyaoCai/PaCoDi. The paper is published in the Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '26), August 09–13, 2026, Jeju Island, Republic of Korea (DOI: 10.1145/3770855.3817791).

Future Directions

1. Closing the MFT approximation gap. The authors explicitly note that the naive mean-field factorization weakens the cross-quadrature dependencies required to preserve phase coherence. Whether the interactive correction branch fully recovers that coherence, or whether tighter variational bounds or a learned coupling term could do better, is left open.
2. Extending beyond the stationarity assumption. The diagonalization argument is stated for stationary processes, since the DFT natively diagonalizes the second-order covariance structure in that setting. Handling strongly non-stationary, trend-dominated, or regime-switching real-world series is an unresolved question.
3. Scaling and foundation-model use. Since two of the evaluated baselines are zero-shot foundation models, a natural next step is whether the spectral formulation carries over to pretrained or large-scale time series foundation models, and whether the L → L/2 compression compounds with other efficiency techniques.
4. Theoretical unification. The paper proves equivalence between temporal, discrete-frequency, and continuous-frequency diffusions (Appendix D.5) and defines the spectral Wiener process — extending this apparatus to other generative formulations such as flow matching or probability-flow ODEs in the spectral domain is a logical continuation.

Target Audience

This paper is best suited to graduate students and researchers working on generative models for time series, especially those interested in frequency-domain or spectral methods, stochastic differential equation formulations of diffusion, and the theory of complex-valued neural networks. Practitioners building diffusion-based synthetic data pipelines for long multivariate sequences will also find the complexity analysis and the 50% attention FLOP argument directly relevant. It presupposes comfort with DDPM/score-based SDE machinery, Fourier analysis, Itô calculus, and Wirtinger derivatives; readers without that background will find the theoretical sections dense.

Authors’ abstract

Diffusion models learn data distributions indirectly through denoising, making the difficulty of generative modeling closely tied to the dependency structure of data. For time series, strong temporal dependence forces the noise / score estimator to recover highly entangled cross-time relationships, leading to the curse of entanglement. We mitigate this burden by changing the topology of the diffusion space: the Discrete Fourier Transform (DFT) decomposes temporal dependencies into spectral modes, diagonalizing second-order dependency structure and better aligning the data manifold with isotropic Gaussian noise and homogeneous diffusion dynamics. However, existing frequency-aware diffusion methods mainly use the DFT to design estimator blocks under temporal DDPM/SDE frameworks, while frequency-native diffusion paths face a mathematical barrier from complex-valued dynamics. We propose PaCoDi (Parallel Complex Diffusion), a frequency-native diffusion framework that constructs the diffusion path in the spectral domain while replacing the complex-valued estimator with parallel real-valued estimators for real and imaginary components. Theoretically, we prove the statistical orthogonality of spectral Gaussian noise, establish quadrature forward transitions and conditional reverse factorization, and extend discrete PaCoDi to continuous-time spectral SDEs through a Spectral Wiener Process. We further introduce a Mean Field Theory approximation with an Interactive Correction Branch to handle marginal coupling, and exploit Hermitian symmetry to reduce 50% attention FLOPs without information loss. Extensive experiments on unconditional and conditional time series generation demonstrate superior generative quality and computational efficiency against 5 SOTA baselines in 5 benchmarks, respectively. Code is available at https://github.com/RongyaoCai/PaCoDi.