{"id":"259a4eab-de29-4ee2-a3b3-1a55d0e63b0d","arxiv_id":"2602.17706","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"PaCoDi generates time series by diffusing real and imaginary spectral components in parallel, cutting attention FLOPs roughly in half while improving benchmark scores.","lead":"This paper proposes a diffusion model for time series that adds and removes noise in the frequency domain rather than in the raw time domain, processing the real and imaginary spectral parts as parallel channels. The authors claim this cuts attention cost by about half and improves generation quality on several benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Conditional factorization (Prop. 3.2) does not justify the factorized generative score; the quality claim rests on the unproven MFT/interactive-correction approximation (Eq. 12), and Table 5's decoupled collapse shows the gap is material.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the theoretical factorization is conditional on X0, while the generative process needs the entangled marginal score. The paper's own Sec. 3.3 and App. B.2 acknowledge this, and the Table 5 ablation shows the decoupled variant collapses, confirming that the interactive correction is doing the essential work. This concern is significant because it limits the theoretical claim: the paper proves a conditional factorization but then relies on an unquantified approximation for the actual sampling distribution. However, the empirical evidence is strong, the interactive correction is a concrete and plausible mechanism, and the authors provide code. The appropriate verdict remains conditional: the method may be accepted if the theoretical framing is corrected and the interaction mechanism is better characterized. I therefore recommend no change to the reader's conditional verdict, while flagging that the central theoretical bridge is the least secure part of the argument.","tokens_in":32984,"tokens_out":16170,"duration_ms":170118,"concrete_test":"Construct a synthetic two-frequency process with known closed-form marginal density and tunable phase-amplitude coupling (e.g., z_t = R_t exp(iθ_t) with R_t depending on past θ_t). Numerically compute the true marginal score ∇ log p_t(X) by quadrature, then compare the L2 error of (a) the MFT factorized score (Eq. 11) and (b) PaCoDi's interactive score (Eq. 20) across a range of coupling strengths. If the interactive score does not substantially reduce error over the factorized score, or if PaCoDi's Context-FID degrades toward the 'Dec.' row of Table 5 as coupling strength grows, the MFT/interactive bridge is unsupported and the concern lands; if it tracks the true marginal score, the bridge holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core decoupling results (Thm 3.1, Prop 3.2, Thm 4.2) are conditioned on a fixed initial spectrum X0. The generative sampler, however, requires the marginal transition q(X_{t-1}|X_t), which integrates over the coupled data prior (Eq. 10 in Sec. 3.3). The paper's bridge is the Mean Field approximation (Eq. 11) plus the Interactive Correction Branch (Eq. 12). No error bound, consistency guarantee, or capacity argument is given for this replacement: it is an architectural heuristic. The fully decoupled variant in Table 5 ('Dec.') collapses on Sines (C-FID 0.886 vs. 0.100 for PaCoDi), demonstrating that the proven factorization alone is insufficient and that the accuracy advantage depends almost entirely on the interaction mechanism. The paper explicitly acknowledges the marginal coupling in Sec. 3.3 and App. B.2, but it does not analyze when h(·) is sufficient. The formulation is also in tension: if h(·) carries the full quadrature counterpart, the branches are not truly independent; if h(·) is weak, the decoupled collapse may recur. Thus the headline claim that the complex diffusion process 'can be split into independent real and imaginary branches' is not actually supported for the marginal generative process.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":33376,"tokens_out":11858,"duration_ms":121287,"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":[{"comment":"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.","section":"Sec. 3.2-3.3, Eqs. (10)-(12)"},{"comment":"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).","section":"Sec. 4.1, App. E.1-E.2, Eqs. (15)-(19)"},{"comment":"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.","section":"App. E.5.3, Eq. (115) and Sec. 4"}],"minor_comments":[{"comment":"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.","section":"Abstract and Sec. 1"},{"comment":"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.","section":"Sec. 3.4"},{"comment":"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.","section":"Sec. 5.3, Eq. (25)"},{"comment":"Define 'Cont.' and 'Disc.' (SDE and DDPM variants) in the caption; currently the reader must infer this from Sec. 6.3.","section":"Table 3 caption"},{"comment":"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.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has strong empirical results and an elegant architectural idea, but the theory currently overclaims relative to what is proved. The gap between conditional factorization and marginal generation is acknowledged but not analyzed; the reverse-SDE drift error is concrete and fixable; the loss equivalence is tautological. I recommend major revision. If the authors fix the SDE derivation, state the DDPM/SDE weightings consistently, and either analyze or carefully circumscribe the MFT/interactive-correction bridge, the paper could be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper is a clear, well-executed spectral diffusion method for time-series generation, and it delivers a real 50% attention FLOP saving via parallel real/imaginary branches plus Hermitian compression. The DFT noise covariance derivation is careful, the conditional reverse factorization is correct, and the empirical wins over Diffusion-TS and others are consistent across datasets and horizons. Code is public, which is a plus.\n\nThe main soft spot is the gap between the theory and the deployed architecture. Prop. 3.2 and Thm. 4.2 prove factorization for the transition conditioned on a fixed initial spectrum X0. Generation uses the marginal q(X_{t-1}|X_t), which integrates over the entangled data prior. The paper acknowledges this and bridges it with a mean-field approximation plus an interactive correction branch h(·), but there is no bound or analysis of when h(·) is sufficient. The ablation makes the gap concrete: the fully decoupled variant collapses on Sines (C-FID 0.886 vs 0.100), so the correction mechanism, not the factorization theorem, is doing the load-bearing work. That is not fatal — the method works — but the claims should be reworded to distinguish what is proven from what is heuristic.\n\nTwo smaller issues. The continuous/discrete loss \"equivalence\" in Appendix E.5.3 is achieved by defining lambda(t) = lambda_t(1-ᾱ_t); that is a scaling convention, not a derived equivalence. And the holomorphic/Liouville argument in Sec. 3.4 is a bit of a strawman — real-valued branches are a natural parameterization choice, not an escape from a genuine dilemma.\n\nNone of this undercuts the empirical or efficiency contributions. The paper is honest about the marginal coupling, the tables are detailed, and the results look reproducible. It deserves a serious referee. I would ask the authors to reframe the theory, ablate the correction branch more thoroughly, and soften the equivalence and holomorphy claims. I would cite it for the efficiency idea and would send it to review rather than desk reject.","headline":"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.","tokens_in":33817,"tokens_out":2426,"would_cite":true,"duration_ms":28058,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","60H10","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["time series generation","diffusion models","Fourier domain diffusion","complex-valued diffusion","Hermitian symmetry","score-based generative models","attention efficiency","heteroscedastic noise"],"falsifier":"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.","tokens_in":32908,"feed_emoji":"📈","tokens_out":5005,"duration_ms":47149,"temperature":0.7,"pith_summary":"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.","feed_headline":"Frequency-domain diffusion cuts attention cost in half for time series","feed_subtitle":"Parallel real and imaginary denoisers keep phase structure while shrinking the sequence the model must attend to.","key_machinery":"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","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Fourier diffusion halves attention FLOPs for time series","Parallel complex diffusion in spectral domain speeds time series gen","Diagonalize temporal dependence with Fourier diffusion","Spectrally native diffusion cuts time series attention cost","Real-valued denoisers in frequency domain slash FLOPs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Fourier diffusion halves attention FLOPs for time series","Parallel complex diffusion in spectral domain speeds time series gen","Diagonalize temporal dependence with Fourier diffusion","Spectrally native diffusion cuts time series attention cost","Real-valued denoisers in frequency domain slash FLOPs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1174,"prompt_tokens":788,"completion_tokens":386,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":310}},"tokens_in":532,"tokens_out":386,"duration_ms":4400,"temperature":1.0,"reasoning_tokens":310,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T02:39:42.554977+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}