Pith. sign in

REVIEW 3 major objections 5 minor 35 references

PaCoDi claims that moving diffusion into the Fourier domain, with parallel real and imaginary denoisers plus an interactive correction branch, generates more accurate time series at half the attention FLOPs of temporal diffusion.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 02:39 UTC pith:2S4IUEBM

load-bearing objection Solid spectral diffusion paper with a real efficiency win, but the headline theory only covers the conditional process; the generative success rests on an unanalyzed correction heuristic. the 3 major comments →

arxiv 2602.17706 v2 pith:2S4IUEBM submitted 2026-02-10 cs.LG

Parallel Complex Diffusion for Scalable Time Series Generation

Rongyao Cai , Yuxi Wan , Kexin Zhang , Ming Jin , Zhiqiang Ge , Qingsong Wen , Yong Liu This is my paper
classification cs.LG MSC 68T0760H1062M10
keywords time series generationdiffusion modelsFourier domain diffusioncomplex-valued diffusionHermitian symmetryscore-based generative modelsattention efficiencyheteroscedastic noise
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the difficulty of diffusion-based time series generation is largely a problem of where the diffusion happens. Working in the raw time domain forces the denoiser to untangle dense local correlations, which the authors call the 'curse of entanglement' and which drives quadratic attention costs. PaCoDi moves the diffusion into the Fourier domain, where the noise becomes statistically orthogonal between real and imaginary components, and proves that the conditional reverse process factorizes into two independent real-valued chains. This legitimizes a parallel architecture of two real-valued denoisers, one for each quadrature component, with a Mahalanobis loss that accounts for heteroscedastic spectral noise, plus an interactive correction branch that re-introduces the phase-amplitude coupling discarded by the mean-field approximation. The result is a generator that the paper reports as more accurate than temporal diffusion baselines on several benchmarks while cutting attention FLOPs in half.

Core claim

PaCoDi's central discovery is that the Fourier transform can act as a diagonalizing operator for diffusion itself, not merely as a feature preprocessing step. The paper proves that spectral Gaussian noise has statistically independent real and imaginary parts (Theorem 3.1), that the forward transition factorizes into quadrature components, and that the reverse posterior conditioned on an initial boundary factors into independent real and imaginary marginals (Proposition 3.2). This factorization is the license to replace any complex-valued denoiser with two parallel real-valued networks, bypassing holomorphic constraints that would otherwise limit expressivity. Because real signals have Hermi

What carries the argument

The conditional reverse factorization (Proposition 3.2) is the key identity: it shows that, for a fixed initial spectrum, the reverse denoising posterior splits into independent real and imaginary components, which is what permits parallel quadrature networks. The Hermitian-symmetry compression (Section 5.3) is the efficiency mechanism: it discards the redundant negative-frequency half of the spectrum, turning a length-L sequence into two length-L/2 real branches and halving attention complexity. The Mean Field Theory approximation with the interactive correction branch (Eq. 12) is the mechanism that covers the gap between the decoupled theory and the entangled marginal needed for actual sam

Load-bearing premise

That the true marginal score of real data can be faithfully approximated by factorized real and imaginary branches plus a feedforward interactive correction—no error bound is proven for this step.

What would settle it

On a synthetic time series with a controlled phase-amplitude coupling (e.g., amplitude-modulated sinusoids where the coupling is known exactly), compute the exact marginal score numerically at an early diffusion step. If the MFT-plus-correction score estimator's output deviates from the true score beyond the noise floor, or if the full model's generation fidelity on such data stays at the level of the decoupled ablation, the paper's central approximation fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Time series diffusion can be made length-scalable: because attention runs on two sequences of length L/2 instead of one of length L, the quadratic cost drops by half with no added asymptotic overhead beyond an FFT.
  • The same Fourier-domain decoupling can be used for any real-valued signal (audio, sensor streams, images) since the Hermitian symmetry argument is generic, not specific to the tested datasets.
  • The heteroscedastic Mahalanobis loss is the correct objective for spectral-domain diffusion; training with a plain Euclidean loss would be biased by the non-isotropic noise covariance.
  • The continuous-time SDE formulation puts the discrete PaCoDi on the same footing as score-based generative models, so existing SDE sampling and guidance techniques can be imported.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the MFT-plus-correction approximation is as general as the paper suggests, the same parallel architecture could be reused for any domain with a natural unitary transform, such as wavelets or orthogonal polynomial bases, as long as the noise covariance is computed accordingly.
  • A testable extension is to apply PaCoDi to very long sequences (L > 1000) where the quadratic attention saving becomes the dominant factor; the paper's scaling analysis predicts an increasingly large gap in wall-clock time.
  • The decoupled ablation's collapse suggests that phase-amplitude coupling is the main thing the correction branch must learn; a simpler learned coupling, perhaps a low-rank cross-branch linear map, might recover most of the benefit at even lower cost.
  • The paper's theory predicts that the marginal score is entangled only through the data prior; this suggests a diagnostic: compare a dataset's phase-amplitude dependence (e.g., via mutual information between magnitude and phase of spectral coefficients) against PaCoDi's fidelity gain over the decoupled variant.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes PaCoDi, a spectral-domain diffusion model for time series generation. The forward diffusion is defined on the DFT of the signal, and the denoiser is split into two parallel real-valued networks for the real and imaginary parts of the spectrum, trained with a Mahalanobis heteroscedastic loss. Hermitian symmetry of real signals is exploited to halve the effective sequence length and attention FLOPs. The paper also gives a continuous-time SDE formulation via a 'Spectral Wiener Process', and claims equivalence between the discrete and continuous objectives. Experiments on conditional and unconditional generation benchmarks report consistent improvements over several baselines, with an ablation showing that the interactive correction branch is necessary for good fidelity.

Significance. The empirical results are strong and the idea of diagonalizing temporal dependencies via the Fourier transform is attractive. The covariance computations in Appendix A are a useful reference, and the 50% attention reduction from Hermitian compression is elegant and clearly stated. However, the paper's headline theoretical guarantee is proved only for transitions conditioned on a fixed initial spectrum X0, while the deployed generative model uses the marginal transition q(X_{t-1}|X_t). The bridge between the two is an MFT approximation plus an interactive correction heuristic, with no approximation bound or consistency analysis; Table 5 shows that the pure decoupled variant collapses. There is also a concrete inconsistency in the reverse SDE drift and a tautological 'equivalence' between discrete and continuous losses. If these are fixed or the claims appropriately narrowed, the work would be a valuable contribution. A code link and extensive appendices are included.

major comments (3)
  1. [Sec. 3.2-3.3, Eqs. (10)-(12)] Prop. 3.2 and Thm. 4.2 prove factorization only for transitions conditioned on a fixed boundary X0. The generative sampler uses q(X_{t-1}|X_t) = ∫ q(X_{t-1}|X_t,X0) q(X0|X_t) dX0, which Eq. (10) itself states is entangled. Eqs. (11)-(12) replace this marginal by two independent branches plus an interactive correction h(·); no error bound or consistency argument is given for h(·). Table 5 shows that without h(·) the decoupled model collapses (C-FID 0.886 vs 0.100 for PaCoDi), so the empirical gain comes from the correction mechanism rather than from the proven decoupling. This is the central theory-to-practice gap and should be addressed or the claims should be narrowed.
  2. [Sec. 4.1, App. E.1-E.2, Eqs. (15)-(19)] The Spectral Wiener Process W=F(w) has increments with real/imag covariance blockdiag(Σ_r, Σ_i), as derived in Appendix A. The forward SDE (15) therefore has diffusion matrix β·blockdiag(Σ_r, Σ_i) in the augmented real representation. The reverse SDE (18)-(19) uses g(t)g(t)^T = β I in the drift correction. The correct reverse drift is f − β blockdiag(Σ_r, Σ_i) ∇ log p_t, so Eqs. (19) are not the time reversal of Eq. (15). The factor 1/2 from Wirtinger calculus does not remove the missing Σ_r, Σ_i. Please derive the reverse SDE in the real augmented space and reconcile with the score-noise identity (21).
  3. [App. E.5.3, Eq. (115) and Sec. 4] The claimed equivalence between Lcont and Ldisc is obtained by setting λ(t)=λ_t(1−ᾱ_t). But λ_t is defined in Eq. (9) as (1−α_t)/(2α_t(1−ᾱ_t)), and λ(t) is defined in Sec. 4 as (1−α_t)/(2α_t). Choosing λ(t)=λ_t(1−ᾱ_t) makes the equality hold by construction; it is not a derived equivalence. Under the definitions actually used in the paper, the continuous and discrete objectives are not shown to be equivalent. The weighting terms should be stated consistently and the claim either proved or removed.
minor comments (5)
  1. [Abstract and Sec. 1] Phrases such as 'the complex diffusion process can be split into independent real and imaginary branches' should be qualified as conditional on the fixed initial spectrum X0; the unconditional/marginal process is not split.
  2. [Sec. 3.4] Liouville's theorem applies only to bounded entire functions on C; neural networks with non-holomorphic activations are not entire. The claimed dilemma between stability and nonlinearity for complex networks is overstated and should be rewritten.
  3. [Sec. 5.3, Eq. (25)] The Nyquist term (k=L/2) exists only for even L; spell out the odd-L case explicitly. Also clarify the normalization convention used for the DFT.
  4. [Table 3 caption] Define 'Cont.' and 'Disc.' (SDE and DDPM variants) in the caption; currently the reader must infer this from Sec. 6.3.
  5. [Appendix A] The full L-dimensional random vectors ε_r and ε_i are degenerate (zero variance at DC/Nyquist and Hermitian-symmetry constraints). The factorization p(E)=p(ε_r)p(ε_i) should be stated as holding on the supporting subspace, since the usual Lebesgue density does not exist for the degenerate components.

Circularity Check

1 steps flagged

Minor definitional equivalence in App. E.5.3; core derivation and experimental claims are otherwise self-contained.

specific steps
  1. self definitional [Appendix E.5.3, Eq. (115), 'Equivalence of Loss Functions between DDPM and SDE']
    "Furthermore, to align the weighting with standard diffusion models, we set the SDE weighting function λ(t) such that: λ(t)/(1−ᾱ_t)=λ_t =⇒ λ(t)=λ_t(1−ᾱ_t)."

    The appendix claims to prove that the continuous score-matching loss Lcont equals the discrete noise-prediction loss Ldisc. The only nontrivial step is defining λ(t) = λ_t(1−ᾱ_t), which exactly cancels the 1/(1−ᾱ_t) factor introduced by the score-noise substitution. Without this choice of weighting, the two objectives differ by a time-dependent factor, so the 'equivalence' is an identity by construction rather than a derived result. This is a minor, non-load-bearing circularity: the empirical comparisons in Tables 3–5 do not rest on this weighting identity.

full rationale

The central theoretical claims—the statistical orthogonality of spectral Gaussian noise (Theorem 3.1), the conditional reverse factorization for a fixed X0 (Proposition 3.2), and the parallel reverse SDE under the same conditioning (Theorem 4.2)—are derived from the linearity/unitarity of the DFT and Gaussian algebra, not from fitted constants or self-citations. The paper explicitly acknowledges that the actual generative transition uses the marginal q(X_{t-1}|X_t), which integrates over the coupled data prior (Eq. 10, App. B.2), and that the mean-field factorization plus Interactive Correction Branch is an approximation (Eqs. 11–12). This is a rigor/scope gap, not circularity. The fully decoupled ablation collapse in Table 5 is an empirical limitation that the paper itself reports, and it does not make the method's success a restatement of its inputs. Self-citations to surveys [9] and [19] are not load-bearing for the main derivation, and no uniqueness theorem or ansatz is imported from the authors' prior work. The only identified circular step is the continuous/discrete loss 'equivalence' in Appendix E.5.3, which is achieved by defining the SDE weight λ(t) to force equality; this is definitional but minor and does not affect the central empirical claims. Overall, the paper exhibits no significant load-bearing circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 1 invented entities

The paper introduces no fitted physical constants, but it relies on a diagonalization assumption for arbitrary time series, a mean-field factorization that contradicts real data coupling, a zero-mean/DC-discard assumption, and a questionable holomorphy claim. The only hand-chosen degree of freedom affecting a theoretical claim is the SDE loss weight.

free parameters (1)
  • Continuous-time loss weight lambda(t) = lambda_t * (1 - alpha_bar_t)
    Introduced in Appendix E.5.3 to make Lcont identically equal to Ldisc; chosen by hand rather than derived from an independent constraint.
axioms (4)
  • domain assumption The DFT diagonalizes temporal dependency structure, making spectral components statistically independent.
    Invoked throughout Section 5.1 to justify the frequency-native approach; holds exactly only for Gaussian stationary/circulant processes, not arbitrary time series.
  • ad hoc to paper Mean-field factorization p(X0) approximately equals p(R0)p(I0).
    Appendix B.2 and E.4 assume the data prior factorizes into real and imaginary marginals to obtain factorized generative transitions; the paper acknowledges this discards phase coupling and adds an interactive correction.
  • domain assumption Inputs are standard-normalized with zero mean, so the DC frequency can be discarded without information loss.
    Section 5.3, Eq. 24 discards X0; if data have nonzero mean, information is lost for unconditional generation.
  • ad hoc to paper Liouville's theorem applies to neural-network noise estimators, so complex-valued networks cannot be both stable and nonlinear.
    Section 3.4; this premise is inaccurate because non-holomorphic complex networks are standard and trainable via Wirtinger calculus, and no citation is provided.
invented entities (1)
  • Spectral Wiener Process W_t = F(w_t) no independent evidence
    purpose: Defines the stochastic driver of the continuous-time frequency SDEs and the reverse-time decoupled SDEs.
    This is a mathematical construction (Fourier transform of a standard Wiener process), not an unexplained physical entity; its covariance structure follows from the definition and is verifiable by simulation, but it is not an independent empirical postulate.

pith-pipeline@v1.3.0-alltime-deepseek · 32660 in / 17709 out tokens · 182351 ms · 2026-08-03T02:39:42.554977+00:00 · methodology

0 comments
read the original 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.

Figures

Figures reproduced from arXiv: 2602.17706 by Kexin Zhang, Ming Jin, Qingsong Wen, Rongyao Cai, Yong Liu, Yuxi Wan, Zhiqiang Ge.

Figure 1
Figure 1. Figure 1: Architecture evolution from temporal to parallel complex diffusion. DiT blocks are visualized in a simplified form [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Data Distribution Visualization on ETTh1 ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Computational Complexity Analysis latent manifold of various temporal dynamics. A critical finding, validated by our results, is the stability of PaCoDi across the se￾quence horizon. For example, on ETTh1, while Diffusion-TS shows a performance degradation in Discriminative scores at 𝐿 = 256, PaCoDi maintains high fidelity. This validates our theoretical claim that parallel quadrature decoupling effectivel… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 6 linked inside Pith

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. GPT-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  2. [2]

    Jonathan Crabbé, Nicolas Huynh, Jan Stanczuk, and Mihaela Van Der Schaar

  3. [3]

    Abhyuday Desai, Cynthia Freeman, Zuhui Wang, and Ian Beaver. 2021. Timevae: A variational auto-encoder for multivariate time series generation.arXiv preprint arXiv:2111.08095(2021)

    Pith/arXiv arXiv 2021

  4. [4]

    Yunfeng Ge, Jiawei Li, Yiji Zhao, Haomin Wen, Zhao Li, Meikang Qiu, Hongyan Li, Ming Jin, and Shirui Pan. 2025. T2S: High-resolution Time Series Generation with Text-to-Series Diffusion Models. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25. 5208–5216. Main Track

    2025

  5. [5]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783 (2024)

    Pith/arXiv arXiv 2024

  6. [6]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. InAdvances in Neural Information Processing Systems, Vol. 33. Curran Associates, Inc., 6840–6851

    2020

  7. [7]

    Jinwoo Kim, Max Beier, Petar Bevanda, Nayun Kim, and Seunghoon Hong. 2025. Sequence Modeling with Spectral Mean Flows.arXiv preprint arXiv:2510.15366 (2025)

    arXiv 2025

  8. [8]

    Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. 2018. Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks. InThe 41st International ACM SIGIR Conference on Research & Development in Information Retrieval(Ann Arbor, MI, USA)(SIGIR ’18). 95–104

    2018

  9. [9]

    Xu Liu, Taha Aksu, Juncheng Liu, Qingsong Wen, Yuxuan Liang, Caiming Xiong, Silvio Savarese, Doyen Sahoo, Junnan Li, and Chenghao Liu. 2025. Empowering Time Series Analysis with Synthetic Data: A Survey and Outlook in the Era of Foundation Models. arXiv:2503.11411 [cs.LG]

    Pith/arXiv arXiv 2025

  10. [10]

    Hao Ni, Lukasz Szpruch, Marc Sabate-Vidales, Baoren Xiao, Magnus Wiese, and Shujian Liao. 2022. Sig-wasserstein GANs for time series generation. In Proceedings of the Second ACM International Conference on AI in Finance (ICAIF ’21). Article 28, 8 pages

    2022

  11. [11]

    Jeha Paul, Bohlke-Schneider Michael, Mercado Pedro, Kapoor Shubham, Singh Nirwan Rajbir, Flunkert Valentin, Gasthaus Jan, and Januschowski Tim

  12. [12]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transform- ers. InProceedings of the IEEE/CVF international conference on computer vision. 4195–4205

    2023

  13. [13]

    Yurui Qian, Qi Cai, Yingwei Pan, Yehao Li, Ting Yao, Qibin Sun, and Tao Mei

  14. [14]

    Jensen, and Xueqi Cheng

    Zezhi Shao, Fei Wang, Yongjun Xu, Wei Wei, Chengqing Yu, Zhao Zhang, Di Yao, Tao Sun, Guangyin Jin, Xin Cao, Gao Cong, Christian S. Jensen, and Xueqi Cheng

  15. [15]

    Yang Song and Stefano Ermon. 2020. Improved Techniques for Training Score- Based Generative Models. InAdvances in Neural Information Processing Systems, Vol. 33. 12438–12448

    2020

  16. [16]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Boosting Diffusion Models with Moving Average Sampling in Frequency Domain. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 8911–8920

  17. [17]

    Haixin Wang, Jiashu Pan, Hao Wu, Fan Zhang, and Tailin Wu. 2025. FourierFlow: Frequency-aware Flow Matching for Generative Turbulence Modeling.arXiv preprint arXiv:2506.00862(2025)

    Pith/arXiv arXiv 2025

  18. [18]

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Fore- casting. InAdvances in Neural Information Processing Systems, Vol. 34. 22419– 22430

    2021

  19. [19]

    Yiyuan Yang, Ming Jin, Haomin Wen, Chaoli Zhang, Yuxuan Liang, Lintao Ma, Yi Wang, Chenghao Liu, Bin Yang, Zenglin Xu, Shirui Pan, and Qingsong Wen

  20. [20]

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Ste- fano Ermon, and Ben Poole. 2021. Score-Based Generative Modeling through Stochastic Differential Equations. InInternational Conference on Learning Repre- sentations

    2021

  21. [21]

    Xinyu Yuan and Yan Qiao. 2024. Diffusion-TS: Interpretable Diffusion for General Time Series Generation. InThe Twelfth International Conference on Learning Representations

    2024

  22. [22]

    Chen Zhao, Weiling Cai, Chenyu Dong, and Chengwei Hu. 2024. Wavelet- based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 8281–8291

    2024

  23. [23]

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond Efficient Transformer for Long Se- quence Time-Series Forecasting.Proceedings of the AAAI Conference on Artificial Intelligence35, 12 (2021), 11106–11115. Conference’17, July 2017, Washington, DC, USA Rongyao Cai et al. A Statistical Prope...

    2021

  24. [24]

    ACM Comput

    A Survey on Diffusion Models for Time Series and Spatio-Temporal Data. ACM Comput. Surv.(Dec. 2025)

    2025

  25. [25]

    Jinsung Yoon, Daniel Jarrett, and Mihaela van der Schaar. 2019. Time-series Generative Adversarial Networks. InAdvances in Neural Information Processing Systems, Vol. 32

    2019

  26. [29]

    The probability density function of the isotropic temporal noise is given by𝑝(𝝐)∝exp(− 1 2∥𝝐∥ 2)

    Preservation under Unitary Transformation:The normalized DFT operator can be represented as a unitary matrixU ∈C 𝐿×𝐿 satisfying UUH = I. The probability density function of the isotropic temporal noise is given by𝑝(𝝐)∝exp(− 1 2∥𝝐∥ 2). Since the unitary transform is an isometry, it preserves theℓ 2-norm: ∥E∥ 2 =∥U𝝐∥ 2 =𝝐 HUHU𝝐=∥𝝐∥ 2 (40) The covariance mat...

  27. [30]

    Linearity and Superposition:Consider the noise aggregation step in the diffusion chain. In the temporal domain, we have the well-known property: √𝛼𝑡 𝝐𝑡−1+√1−𝛼 𝑡 𝝐𝑡 =𝝐∗∼N(0,𝜎 2I) (given𝛼 𝑡+(1−𝛼 𝑡)=1)(41) Applying the linear operatorFto both side: F(√𝛼𝑡 𝝐𝑡−1+√1−𝛼 𝑡 𝝐𝑡)= √𝛼𝑡F(𝝐 𝑡−1)+√1−𝛼 𝑡F(𝝐 𝑡)= √𝛼𝑡E𝑡−1+√1−𝛼 𝑡E𝑡 (42) By linearity, the left-hand side equalsF...

  28. [31]

    The temporal additivity ensures Í𝑇 𝑡=1𝑤𝑡 𝝐𝑡 =√1− ¯𝛼𝑡 𝝐∗∼N(0,(1− ¯𝛼𝑡)I)

    Induction to Arbitrary Diffusion Step𝑡:By induction, the cumulative noise at step𝑡can be expressed as: Ecumulative = 𝑇∑︁ 𝑡=1 𝑤𝑡E𝑡 =F 𝑇∑︁ 𝑡=1 𝑤𝑡 𝝐𝑡 ! (43) where Í𝑇 𝑡=1𝑤 2 𝑡 =1− ¯𝛼𝑡 . The temporal additivity ensures Í𝑇 𝑡=1𝑤𝑡 𝝐𝑡 =√1− ¯𝛼𝑡 𝝐∗∼N(0,(1− ¯𝛼𝑡)I). Consequently: Ecumulative =√1− ¯𝛼𝑡F(𝝐∗)= √1− ¯𝛼𝑡E∗ (44) whereE∗ is a sample drawn from the identical sp...

    2017

  29. [32]

    Forward Transition𝑞(X𝑡|X𝑡−1):The forward diffusion step is defined asX𝑡 = √︁ 1−𝛽 𝑡X𝑡−1+ √︁ 𝛽𝑡E𝑡 . Since the complex noiseE=𝜺 𝑟+𝑗𝜺 𝑖 possesses statistically independent quadrature components (Theorem 3.1), the transition density naturally splits: 𝑞(X𝑡|X𝑡−1)=𝑞(R 𝑡|R𝑡−1)·𝑞(I 𝑡|I𝑡−1).(46)

  30. [33]

    When the boundary conditionX0 isfixed(i.e., , conditioned upon), the mean vector √ ¯𝛼𝑡X0 is a constant

    Conditional Marginal𝑞(X𝑡|X0):We must distinguish between the data correlation within X0 =R 0+𝑗I 0 and the stochastic evolution of the diffusion. When the boundary conditionX0 isfixed(i.e., , conditioned upon), the mean vector √ ¯𝛼𝑡X0 is a constant. The stochasticity is driven solely by the cumulative noise. As proven rigorously in Lemma C.1 (see below), t...

    2017

  31. [34]

    The Precision Matrix (Inverse Covariance):Extracting the quadratic termsx ⊤ 𝑡−1(·)x𝑡−1 : ˜𝚺 −1 𝑡 = 𝛼𝑡 1−𝛼 𝑡 𝚺−1+ 1 1− ¯𝛼𝑡−1 𝚺−1 = 𝛼𝑡 1−𝛼 𝑡 + 1 1− ¯𝛼𝑡−1 𝚺−1 = 𝛼𝑡(1− ¯𝛼𝑡−1)+(1−𝛼 𝑡) (1−𝛼 𝑡)(1− ¯𝛼𝑡−1) 𝚺−1 = 1− ¯𝛼𝑡 (1−𝛼 𝑡)(1− ¯𝛼𝑡−1) 𝚺−1. (64) Inverting this yields the reverse covariance: ˜𝚺𝑡 =(1−𝛼 𝑡)(1− ¯𝛼𝑡−1) 1− ¯𝛼𝑡 𝚺.(65) This result confirms that the shape ...

  32. [35]

    The Mean Vector:Extracting the linear terms−2 ˜𝝁⊤ 𝑡 ˜𝚺 −1 𝑡 x𝑡−1 involves the matrix𝚺−1: ˜𝝁⊤ 𝑡 ˜𝚺 −1 𝑡 = √𝛼𝑡 1−𝛼 𝑡 x⊤ 𝑡 𝚺−1+ √ ¯𝛼𝑡−1 1− ¯𝛼𝑡−1 x⊤ 0 𝚺−1.(66) Multiplying from the right by ˜𝚺𝑡 : ˜𝝁𝑡 = ˜𝚺𝑡 𝚺−1 √𝛼𝑡 1−𝛼 𝑡 x𝑡+ √ ¯𝛼𝑡−1 1− ¯𝛼𝑡−1 x0 = (1−𝛼 𝑡)(1− ¯𝛼𝑡−1) 1− ¯𝛼𝑡 𝚺𝚺−1 √𝛼𝑡 1−𝛼 𝑡 x𝑡+ √ ¯𝛼𝑡−1 1− ¯𝛼𝑡−1 x0 = √𝛼𝑡(1− ¯𝛼𝑡−1) 1− ¯𝛼𝑡 x𝑡+ √ ¯𝛼𝑡−1(1−𝛼 𝑡) 1− ¯𝛼𝑡 x0...

    2017

  33. [2021]

    PSA-GAN: Progressive self attention gans for synthetic time series.arXiv preprint arXiv:2108.00981(2021)

    Pith/arXiv arXiv 2021

  34. [2024]

    InProceedings of the 41st International Conference on Machine Learning(Vienna, Austria)(ICML’24)

    Time series diffusion in the frequency domain. InProceedings of the 41st International Conference on Machine Learning(Vienna, Austria)(ICML’24). Article 374, 32 pages

  35. [2025]

    on Knowl

    Exploring Progress in Multivariate Time Series Forecasting: Comprehensive Benchmarking and Heterogeneity Analysis.IEEE Trans. on Knowl. and Data Eng. 37, 1 (2025), 291–305

    2025