acceptodds
Under review as a conference paper at ICLR 2027

RAF-Diff: Rational Adaptive Functional Diffusion for Time Series Analysis

Abstract

Time series analysis tasks such as forecasting, imputation, anomaly detection, and classification are crucial for applications spanning climate science, finance, retail, and cloud infrastructure. We present Rational Adaptive Functional Diffusion (), a conditional diffusion model that introduces two innovations: (i) a Rational Adaptive Correlation Projection (RACP) in the forward process and (ii) a Rational Denoising Transformer (R-DiT) architecture, specifically designed for time series modeling. Unlike traditional fixed Gaussian noise based forward process, our RACP module makes the forward process trainable and data-adaptive, capturing complex inter-feature correlations through learned rational function transformations. Our R-DiT architecture with its rational function layers provides superior expressiveness for capturing periodic time series signals. achieves superior performance across four fundamental time series analysis tasks, significantly outperforming existing prominent models in forecasting (eight datasets), imputation (six datasets), classification (ten datasets) and anomaly detection (five datasets). Comprehensive ablation studies across all tasks validate the utility of each component, demonstrating the model's robustness in diverse time series challenges.

Primary area:
learning on time series and dynamical systems
Keywords:
Diffusion Models, Time Series Analysis, Rational adaptive Functions

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Mb1xOpenReview ↗
Rating
3
Confidence
3
Informativeness
3(+1)
Presentation
3(+1)
Soundness
2
Read review
Paper summary

The paper proposes a conditional diffusion framework for four time-series tasks. Two components are introduced. RACP makes the forward process trainable by replacing its mean with a learned time-dependent projection, following a similar idea to CN-Diff. R-DiT replaces the MLP sublayer of a DiT block with a rational KAN layer using a Safe Padé denominator. The framework is evaluated on forecasting (8 datasets), imputation (6), classification (10, UEA) and anomaly detection (5), with ablations on both components and additional sensitivity, efficiency and significance analyses.

Critical strengths
  1. The empirical scope is unusually broad and is the main reason I lean toward acceptance. One framework is evaluated on 29 dataset–task combinations against 39 forecasting, 16 imputation, 13 classification and 18 anomaly baselines. Very few diffusion papers in this area attempt even two of the four tasks.

  2. The bookkeeping in the results tables is mostly sound. I replayed every rank in Tables 2, 3 and 5 against the printed values and found no rank that violates monotonicity, and the average ranks in Tables 3 and 5 are correct. The model has the best average rank on all four tasks.

  3. There is no target leakage in the sampling procedure. Algorithm 2 evaluates RACP on the predicted x^\hat{x}, not on the ground truth, which is the obvious failure mode for a learned forward process and the authors avoid it.

  4. The derivations in Eqs. 10–12 check out. I verified them independently and they are correct to machine precision.

  5. The model is small (0.5M parameters) and the code is released. A compact deterministic diffusion model that is competitive on four tasks is a useful artifact.

Critical weaknesses
  1. The two ablation tables contradict each other, so the contribution of RACP is never quantified. Table 7 reports Weather forecasting MSE of 0.6960 without RACP and 0.2930 with it, which reads as a 57.9% share. Table 8 reports the same comparison as 0.325 (DiT + MLP) and 0.308 (R-DiT + MLP) against 0.293 with RACP, a 1.5% to 4.9% share. The with-RACP rows agree exactly across the two tables (0.2930 and 0.293), so the conflict is localised to the no-RACP configuration. Table 2 corroborates the small number: the authors' own non-rational conditional diffusion reimplementation, CnDiff*, reaches 0.405 on ETTh1 against the full model's 0.4015, a 1.0% gap, while Table 7 implies a 57.7% gap on the same dataset. Either one table is wrong, or "w/o RACP" in Table 7 is not the same configuration as "MLP projection" in Table 8, in which case Table 7 does not isolate RACP at all. As it stands the second contribution is not supported by its own evidence.
  2. The objective as printed is not a valid variational bound, and nothing prevents the learned forward process from encoding the target. The reconstruction term in Eq. 3 sums log⁡pθ(x∣xt,c)\log p_\theta(x \mid x_t, c) over all tt rather than taking the t=0t=0 term, and the Appendix C.2 joint carries an additional product of p(x0∣xt)p(x_0 \mid x_t), so the expression is unnormalised. This matters more than a typo, because ϕ\phi is trained jointly with θ\theta on the same objective. An information-preserving projection with γ=1/αˉt\gamma = 1/\sqrt{\bar\alpha_t} is exactly representable in the Safe Padé basis, and it would cut the reconstruction loss while breaking the match with the inference prior N(c,I)N(c, I). The only term anchoring the forward process to that prior is the prior KL, which is weighted by αˉT\bar\alpha_T. At the paper's own settings that weight is 0.031 for T=100T=100 and about 10−1510^{-15} for T=1000T=1000. Nothing in the training constrains or monitors this. I am not claiming the model does exploit it, and the reported results suggest it does not, but a paper whose central mechanism is a trainable forward process needs either a proper bound or an explicit constraint and a diagnostic.
  3. The novelty framing overstates the gap on both components. The Abstract and Introduction describe existing forward processes as fixed and non-trainable, but the related work correctly credits CN-Diff [1] with non-linear learnable transformations and learnable conditions in the forward process for time-series forecasting. The Introduction's premise therefore contradicts the paper's own related work. Trainable or data-dependent forward processes are also well established outside time series, for example variational diffusion models [7], cold diffusion [8] and blurring diffusion [9], none of which is cited. On the architecture side, placing KAN layers in transformers is already done by KAT and the temporal KAN transformer, so the delta is the rational basis [4, 5, 6] inside a diffusion denoiser. The claim of being the first diffusion framework to be state of the art on all four tasks is also stronger than the evidence. TimeDiT [2], which the paper cites, already covers three of the four, and unified multi-task models such as UniTS (arXiv:2403.00131) cover all four in one model.

Minor issues

  • Table 6's statistic is never defined. The caption reports Mean, Median, Std, P95 and Max of some per-window quantity before and after the RACP projection, but the text uses the table to support the correlation claim in both directions, since Electricity rises from 0.5888 to 0.8800 while SMAP falls from 0.9482 to 0.7684.
  • In Table 2, eight models print 0.016 on Exchange and receive ranks 2,2,2,2,2,2,8,9. Seventeen columns across Tables 2 and 3 have this property, and the tie convention is not applied consistently even within one column.
  • Table 22's average rank of 2.2 reproduces exactly from the accuracy columns alone, but the surrounding text implies it covers accuracy, recall and precision. Over all three the value is 3.80.
Questions for authors

Please refer to weaknesses for details.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Mb1x

Weakness 1: The two ablation tables ⋯\cdots its own evidence.

Response: Tables 7 and 8 evaluate two different aspects of the proposed forward projection, rather than the same ablation configuration.

Table 7 evaluates the contribution of the learnable forward transformation relative to an identity mapping. Specifically, in the "w/o RACP" configuration, we replace the RACP transformation with an identity mapping, i.e., (RACPϕ(x,t)=x)(\mathrm{RACP}_{\phi}(x,t)=x) (DDPM), while retaining the remaining forward diffusion formulation. Thus, this experiment compares the proposed learnable rational transformation against a non-learnable identity transformation. On Weather, the forecasting MSE decreases from 0.6960 to 0.2930, while on ETTh1, it decreases from 0.9495 to 0.4015. These results quantify the performance improvement obtained by replacing the identity transformation with RACP.

Table 8 isolates the additional benefit of rational parameterization over a conventional learnable MLP projection. Unlike Table 7, both configurations retain a learnable forward transformation, differing only in its parameterization. With the R-DiT denoiser fixed, replacing MLP with RACP reduces Weather forecasting MSE from 0.308 to 0.293 (4.87%) and ETTm2 imputation MSE from 0.034 to 0.014 (58.82%). Similarly, SWaT anomaly detection F1 improves from 87.35 to 90.16, and PEMS-SF classification accuracy increases from 0.844 to 0.898. These results directly quantify the additional benefit of the proposed rational projection over the learnable MLP alternative.

Regarding CnDiff in Table 2, this baseline is distinct from the identity-mapping ablation in Table 7. CnDiff already employs a learnable nonlinear transformation in its forward process, whereas Table 7 uses a fixed identity transformation as the baseline. Consequently, the relatively small difference between CnDiff (0.405) and RAF-Diff (0.4015) on ETTh1 does not contradict the larger performance improvement observed relative to the identity mapping. Rather, these experiments evaluate different baselines: a fixed identity transformation, a learnable MLP projection, and an existing nonlinear transformation-based diffusion model.

Weakness 2: The objective as printed ⋯\cdots and a diagnostic.

We update Eq.(13) to its normalized version as follows. However, we emphasize that the loss term remains the same, only the way the joint is written changed.

Step 1: Non-Markovian forward process. From Eq. (1), the forward marginals are qϕ(xt∣x,c)=N(αˉtRACPϕ(x,t)+(1−αˉt)c  (1−αˉt)I)q_\phi(x^t \mid x, c) = \mathcal{N}\big(\sqrt{\bar\alpha_t}\mathrm{RACP}_\phi(x,t) + (1-\sqrt{\bar\alpha_t})c\; (1-\bar\alpha_t)I\big).

As shown in the corrected Eqs. (5)-(6), the transition qϕ(xt∣xt−1,x,c)q_\phi(x^t \mid x^{t-1}, x, c) depends on the data xx directly, through RACPϕ(x,t)−RACPϕ(x,t−1)\mathrm{RACP}_\phi(x,t) - \mathrm{RACP}_\phi(x,t-1).

The forward process is therefore non-Markovian, and following DDIM [1] we write it as qϕ(x1:T∣x,c)=qϕ(xT∣x,c)∏t=2Tqϕ(xt−1∣xt,x,c).q_\phi(x^{1:T} \mid x, c) = q_\phi(x^T \mid x, c)\prod_{t=2}^{T} q_\phi(x^{t-1} \mid x^t, x, c). Unlike in DDPM, where the data enters only the first forward step, here xx enters every forward transition.

Step 2: Generative joint in Normalized Form (Eq.~13).

For time series diffusion models, forecasting xθx_\theta has been found to be more effective than predicting ϵθ\epsilon_\theta [2,3]. Our denoiser therefore predicts the clean series x^θ(xt,t)\hat x_\theta(x^t, t) from every noisy state, and each reverse step is built on this prediction, pθ(xt−1∣xt,c)=qϕ(xt−1∣xt,x^θ(xt,t),c)p_\theta(x^{t-1} \mid x^t, c) = q_\phi\big(x^{t-1} \mid x^t, \hat x_\theta(x^t, t), c\big) (Eq.~16). Because xx enters every forward transition, the generative process mirrors this with a link back to the data at every step. Each such link generates its own auxiliary variable x~(t)\tilde x^{(t)}: $ p_\theta\big(x^{1:T}, \tilde x^{(1:T)} \mid c\big) = p(x^T \mid c)\prod_{t=2}^{T} p_\theta(x^{t-1} \mid x^t, c);\prod_{t=1}^{T} p_\theta\big(\tilde x^{(t)} \mid x^t, c\big), \qquad p_\theta\big(\tilde x^{(t)} \mid x^t, c\big) = \mathcal{N}\big(\hat x_\theta(x^t, t), \sigma^2 I\big). $

The reverse chain now ends at x1x^1 (since qϕ(x0∣x1,x,c)q_\phi(x^0 \mid x^1, x, c) has zero variance), and each auxiliary variable x~(t)\tilde x^{(t)} is generated exactly once by a normalized conditional. Integrating out x~(1),…,x~(T)\tilde x^{(1)}, \dots, \tilde x^{(T)}, then x1,…,xT−1x^1, \dots, x^{T-1}, then xTx^T gives one at every step, so the joint is normalized. In particular, no density is squared at t=1t = 1.

Step 3: Likelihood. During training, every auxiliary variable is observed to equal the data, x~(1)=⋯=x~(T)=x\tilde x^{(1)} = \dots = \tilde x^{(T)} = x. Our goal is to maximize pθ(x~(1:T)=x∣c)=∫pθ(x1:T,x~(1:T)=x∣c) dx1:T,p_\theta\big(\tilde x^{(1:T)} = x \mid c\big) = \int p_\theta\big(x^{1:T}, \tilde x^{(1:T)} = x \mid c\big)\, dx^{1:T}, which is intractable, so we maximize a variational lower bound on it.

AuthorsOpenReview ↗

Response to Reviewer Mb1x

Step 4: Variational bound, which is exactly Eq. (3).

Using qϕq_\phi as the variational posterior and applying Jensen's inequality,

log⁡pθ(x~(1:T)=x∣c)≥Eqϕ[log⁡pθ(x1:T,x~(1:T)=x∣c)qϕ(x1:T∣x,c)].\log p_\theta\big(\tilde x^{(1:T)} = x \mid c\big) \ge \mathbb{E}_{q_\phi}\left[\log \frac{p_\theta\big(x^{1:T}, \tilde x^{(1:T)} = x \mid c\big)}{q_\phi(x^{1:T} \mid x, c)}\right].

The numerator (Eq. 13) and the denominator (Step 1) are factorized over the same steps in the same direction, so the log-ratio splits term by term:

−log⁡pθ(x~(1:T)=x∣c)≤DKL(qϕ(xT∣x,c) ∥ p(xT∣c))⏟Lprior-\log p_\theta\big(\tilde x^{(1:T)} = x \mid c\big) \le \underbrace{D_{\mathrm{KL}}\big(q_\phi(x^T \mid x, c)\,\|\,p(x^T \mid c)\big)}_{\mathcal{L}_{\text{prior}}}

$ + \underbrace{\sum_{t=2}^{T} \mathbb{E}_{q_\phi} D_{\mathrm{KL}}\big(q_\phi(x^{t-1} \mid x^t, x, c),|,p_\theta(x^{t-1} \mid x^t, c)\big)}_{\mathcal{L}_{\text{diff}}}$

−∑t=1TEqϕlog⁡pθ(x∣xt,c)⏟Lrec.\underbrace{-\sum_{t=1}^{T} \mathbb{E}_{q_\phi}\log p_\theta(x \mid x^t, c)}_{\mathcal{L}_{\text{rec}}}.

This is exactly Eq. (3), with the diffusion sum indexed from t=2t = 2. The bound holds for every ϕ\phi. This structure is already established in state-space models [4]: over the diffusion steps, Eq. (13) has an initial state, Markov transitions and an emission at every step, and the bound above is the standard variational objective of deep state-space models

Relation to the standard data likelihood. The sampler outputs x~(1)\tilde x^{(1)}. By the chain rule,

log⁡pθ(x~(1:T)=x∣c)=log⁡pθ(x~(1)=x∣c)+log⁡pθ(x~(2:T)=x∣x~(1)=x,c).\log p_\theta\big(\tilde x^{(1:T)} = x \mid c\big) = \log p_\theta\big(\tilde x^{(1)} = x \mid c\big) + \log p_\theta\big(\tilde x^{(2:T)} = x \mid \tilde x^{(1)} = x, c\big).

The first term is the usual data likelihood. The second is a consistency term on the intermediate predictions, bounded above by −(T−1)d2log⁡(2πσ2)-(T-1)\tfrac{d}{2}\log(2\pi\sigma^2) for fixed-variance Gaussian emissions. Hence

−log⁡pθ(x~(1)=x∣c)≤LRAF-Diff−(T−1)d2log⁡(2πσ2)-\log p_\theta\big(\tilde x^{(1)} = x \mid c\big) \le \mathcal{L}_{\text{RAF-Diff}} - (T-1)\tfrac{d}{2}\log(2\pi\sigma^2)

so Eq. (3) also bounds the standard negative log-likelihood, up to a constant independent of ϕ\phi and θ\theta.

Target-information leakage and prior matching.

We agree that a learnable forward process could potentially retain target information in the final diffusion state. However, such behavior is penalized by the prior KL term. Specifically, for p(xT∣c)=N(c,I)p(x^T\mid c)=\mathcal{N}(c,I), the exact prior KL is

$ \mathcal{L}_{\mathrm{prior}} =D_{\mathrm{KL}}\left( q_\phi(x^T\mid x,c),|,p(x^T\mid c) \right)$

$\mathcal{L}_{\mathrm{prior}} =\frac{1}{2}\left[ |\delta_T(x,c)|_2^2 +d\left(-\bar{\alpha}_T-\log(1-\bar{\alpha}_T)\right) \right], $

where $ \delta_T(x,c)= \sqrt{\bar{\alpha}_T}\operatorname{RACP}_\phi(x,T) +\left(\sqrt{1-\bar{\alpha}_T}-1\right)c. $

Thus, the small value of αˉT\bar{\alpha}_T does not eliminate the penalty for a fixed displacement of the final-state mean from the inference prior. In particular, for the information-preserving construction RACP⁡ϕ(x,T)=x/αˉT\operatorname{RACP}_\phi(x,T)=x/\sqrt{\bar{\alpha}_T}, the mean-dependent KL becomes $ \frac{1}{2} \left| x+\left(\sqrt{1-\bar{\alpha}_T}-1\right)c \right|_2^2, $ which does not vanish as αˉT→0\bar{\alpha}_T\to 0. We emphasize that the prior KL penalizes, rather than strictly prohibits, target-information retention.

Empirical diagnostic.

To assess whether the trained model exhibits the proposed information-preserving behavior, we measured the final-state prior KL per dimension and the relative projection norm ∥RACP⁡ϕ(x,T)∥2/∥x∥2\|\operatorname{RACP}_\phi(x,T)\|_2/\|x\|_2. The results are summarized below.

DatasetPrior KL/dimProjection norm ratio
ETTh15.9×10−35.9\times10^{-3}0.1500.150
ETTm15.5×10−35.5\times10^{-3}0.0710.071
MSL2.1×10−52.1\times10^{-5}0.3900.390
SMAP3.2×10−53.2\times10^{-5}0.3300.330

The learned projection norm ratios (0.0710.071--0.3900.390) are substantially below the amplification factor 1/αˉT≈5.951/\sqrt{\bar{\alpha}_T}\approx 5.95 required by the specific construction identified by the reviewer. The small prior KL values further indicate that the learned final-state distributions remain close to the inference prior. Although these diagnostics cannot exclude every form of target-information leakage, they provide evidence against the particular amplification mechanism raised in the review.

AuthorsOpenReview ↗

Response to Reviewer Mb1x

Weakness 3: The novelty framing ⋯\cdots four in one model.

We emphasize that our novelty is not these ingredients (learnable nonlinear diffusion, rational functions) individually, but their specific parameterization and integration within conditional diffusion for multivariate time-series analysis (forecasting, classification, imputation and anomaly detection). The integration is not a "trivial engineering exercise", rather it is a grounded research contribution that advances the state of the art. We further clarify the distinction and provide matched evidence. We revised the Abstract, Introduction, and Related Work to clarify these distinctions and avoid overstating novelty.

1. Novelty of the forward process.

We acknowledge that modifying or learning the forward diffusion process is not new in itself. Variational Diffusion Models (VDM) optimize the noise schedule, Cold Diffusion considers alternative degradation operators such as blurring and masking, and Blurring Diffusion Models introduce structured blurring-based corruption. NDM and CN-Diff further demonstrate learnable nonlinear transformations, with CN-Diff applying such transformations to time-series forecasting. We will explicitly discuss these works and correct the Introduction's characterization of prior forward processes as universally fixed and non-trainable.

Our contribution is therefore not the introduction of a trainable forward process per se, but the

Rational Adaptive Correlation Projection (RACP): a learnable, time-dependent rational-function transformation within the conditional forward diffusion process, designed to capture inter-feature dependencies in multivariate time series. Unlike approaches primarily concerned with learning noise schedules or specifying alternative corruption operators, RACP directly parameterizes the data transformation through rational functions. Relative to CN-Diff, we investigate this rational parameterization and extend its application beyond forecasting to imputation, anomaly detection, and classification.

The benefit of this specific parameterization is evaluated in Table 8. With the R-DiT denoiser fixed, replacing the MLP forward projection with RACP improves Weather forecasting MSE from 0.308 to 0.293, ETTm2 imputation MSE from 0.034 to 0.014, SWaT anomaly detection F1 from 87.35 to 90.16, and PEMS-SF classification accuracy from 0.844 to 0.898. These results provide direct evidence for the benefit of rational projection over an MLP alternative in the evaluated configurations.

2. Novelty of the denoising architecture.

We agree with the reviewer that incorporating KAN layers into transformers is not new. In particular, KAT already introduces rational basis functions into transformer architectures, while temporal KAN-based transformers investigate such functional representations for time-series modeling.

We clarify that the contribution of R-DiT is not the invention of rational KAN layers or their use in transformers, but their integration into a diffusion denoising architecture for time-series analysis, together with timestep-dependent conditioning and the proposed RACP forward process.

Table 8 provides controlled evidence for R-DiT. With the projection type fixed, R-DiT improves upon conventional DiT across all four evaluated tasks, under both MLP and RACP projections.

To further examine whether a rational KAN-based transformer without diffusion can account for the performance of RAF-Diff, we additionally evaluated a non-diffusion Transformer equipped with rational KAN layers (KAT-style).

Table: Comparison between a non-diffusion rational KAN Transformer (KAT-style) and RAF-Diff across four tasks.

TaskDatasetKAT-styleRAF-Diff
Forecasting (MSE ↓\downarrow)ETTh10.6240.401
ETTm10.7640.338
Exchange0.0560.016
Imputation (MSE ↓\downarrow, 25% mask)ETTh10.1040.052
ETTm10.0260.021
Anomaly Detection (F1 ↑\uparrow)MSL83.5889.74
SMAP71.0391.84
Classification (Acc. ↑\uparrow)UWave0.8590.884

RAF-Diff outperforms this non-diffusion rational KAN Transformer across the evaluated datasets, including substantial improvements on ETTm1 forecasting (0.764 to 0.338) and SMAP anomaly detection (71.03 to 91.84).

We emphasize that this additional comparison evaluates the effectiveness of the complete RAF-Diff framework relative to a non-diffusion rational KAN Transformer, rather than isolating the effect of a single component. The component-level contributions of RACP and R-DiT are more directly evaluated through the controlled comparisons in Table 8.

Together, these experiments support the value of combining rational-function-based forward projection and denoising within a conditional diffusion framework.

Reviewer_JHGROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RAF-Diff combines a learned rational projection in the forward diffusion process with a denoising transformer that uses rational function layers. The approach aims to improve the modeling of feature dependencies and periodic structure in time series. The paper evaluates forecasting, imputation, classification, and anomaly detection, reporting competitive results alongside component ablations and additional analyses.

Critical strengths
  1. The evaluation covers four major time series tasks and includes diverse datasets, sensitivity studies, and computational comparisons.
  2. The crossed ablation in Table 8 provides useful evidence for the individual and combined benefits of the rational projection and denoiser.
Critical weaknesses
  1. The main architectural ingredients have close precedents: CN-Diff [1] already learns nonlinear transformations in forward diffusion, and KAT [2] introduces rational KAN layers into transformers. Although the paper acknowledges these works, its additional methodological contribution is primarily their combination and adaptation to time series tasks. This is a useful extension, but the advance over existing learned diffusion and rational network designs remains limited.
  2. Table 8 shows that the proposed components help, but does not test whether their benefits come from better modeling of periodicity or high frequency structure. The correlation analysis also interprets increases for forecasting and imputation, and decreases for anomaly detection, favorably without establishing that useful dependencies are better represented. The experiments therefore provide limited insight into when or why rational functions should be preferred.
  3. The claim that earlier diffusion frameworks address at most two task families overlooks TSDE [3], which evaluates all four considered here. Task coverage alone therefore provides a limited basis for the paper’s broader significance. For classification and anomaly detection, the absence of a comparable rational architecture trained without diffusion also leaves the added value of the diffusion formulation unclear.
  4. In Appendix C.2, multiplying reverse transition densities by additional reconstruction densities does not establish a normalized joint model, leaving the claimed variational interpretation unsupported as written. These reconstruction terms may still be useful as auxiliary losses, but their empirical value should be distinguished from the claim that the objective follows from the stated generative model.

Reference

[1] Rishi et al., "Conditional Diffusion Model with Nonlinear Data Transformation for Time Series Forecasting," ICML, 2025.

[2] Xingyi Yang and Xinchao Wang, "Kolmogorov-Arnold Transformer," ICLR, 2025.

[3] Senane et al., "Self-supervised learning of time series representation via diffusion process and imputation-interpolation-forecasting mask," KDD, 2024.

Questions for authors
  1. Which specific frequency and dependence patterns empirically favor rational functions, and to what extent do these patterns explain the observed performance gains?
  2. How much of the performance gain in classification and anomaly detection is retained if a comparable rational architecture is evaluated without diffusion?
  3. Please provide a properly normalized generative formulation that rigorously justifies the stated objective, and clarify which components are treated as heuristic or auxiliary losses.
6 comments
AuthorsOpenReview ↗

Response to Reviewer JHGR

Weakness 1: The main architectural ⋯\cdots network designs remains limited.

We emphasize that our novelty is not these ingredients (learnable nonlinear diffusion, rational functions) individually, but their specific parameterization and integration within conditional diffusion for multivariate time-series analysis (forecasting, classification, imputation and anomaly detection). The integration is not a "trivial engineering exercise", rather it is a grounded research contribution that advances the state of the art. We further clarify the distinction and provide matched evidence.

(1) RACP: Rational adaptive forward projection. Unlike CN-Diff's nonlinear transformation for forecasting, RACP introduces a time-dependent rational projection designed to model inter-feature dependencies across four tasks. Table~6 demonstrates task-dependent changes in mean pairwise feature correlations (Electricity: 0.589→0.8800.589 \rightarrow 0.880; SMAP: 0.948→0.7680.948 \rightarrow 0.768). The performance benefit of rational parameterization is isolated against a learnable MLP projection in Table 8.

(2) R-DiT: Rational diffusion denoising. While KAT uses rational functions in Transformers, R-DiT integrates them into a diffusion denoiser with timestep conditioning, enabling reconstruction across different noise levels. Table 8 compares DiT and R-DiT under identical projection types, directly evaluating the denoiser contribution.

Table A: Component-level ablation (Table 8 of the paper).

Task (metric)DatasetDiT + MLPDiT + RACPR-DiT + MLPR-DiT + RACP
Forecasting (MSE ↓\downarrow)Weather0.3250.3200.3080.293
Imputation (MSE ↓\downarrow)ETTm20.0350.0280.0340.014
Anomaly detection (F1 ↑\uparrow)SWaT80.8088.3387.3590.16
Classification (Acc. ↑\uparrow)PEMS-SF0.8280.8380.8440.898

With the denoiser fixed, RACP improves performance over MLP projection across all four evaluated tasks. Likewise, R-DiT improves over DiT under both projection types. These controlled comparisons show that each proposed component contributes performance improvements within the evaluated diffusion framework, rather than the gains arising solely from their combination.

(3) Additional comparison with a rational KAN Transformer without diffusion. To examine whether rational KAN layers alone account for the gains, we evaluated a non-diffusion Transformer equipped with rational KAN layers (KAT-style).

Table B: Additional KAT-style baseline comparison.

Task (metric)DatasetKAT-styleRAF-Diff
Forecasting (MSE ↓\downarrow)ETTh10.6240.401
Forecasting (MSE ↓\downarrow)ETTm10.7640.338
Forecasting (MSE ↓\downarrow)Exchange0.0560.016
Imputation, 25% mask (MSE ↓\downarrow)ETTh10.1040.052
Imputation, 25% mask (MSE ↓\downarrow)ETTm10.0260.021
Anomaly detection (F1 ↑\uparrow)MSL83.5889.74
Anomaly detection (F1 ↑\uparrow)SMAP71.0391.84
Classification (Acc. ↑\uparrow)Handwriting0.3770.382
Classification (Acc. ↑\uparrow)UWave0.8590.884

RAF-Diff outperforms the non-diffusion rational KAN Transformer on all evaluated datasets, notably on ETTm1 forecasting (MSE: 0.764→0.3380.764 \rightarrow 0.338) and SMAP anomaly detection (F1: 71.03→91.8471.03 \rightarrow 91.84). This comparison supports the effectiveness of the complete RAF-Diff framework, while Table A isolates the contributions of RACP and R-DiT within diffusion.

(4) General applicability. RAF-Diff uses a common conditional diffusion formulation across forecasting, imputation, anomaly detection, and classification, achieving average ranks of 1.5, 1.5, 2.4, and 1.9, respectively (Tables 2--5).

AuthorsOpenReview ↗

Response to Reviewer JHGR

Weakness 2: Table 8 shows ⋯\cdots preferred.

Question 1: Which specific frequency ⋯\cdots performance gains?

We conducted additional analyses to identify (i)\textit{(i)} which frequency patterns favor rational functions, (ii)\textit{(ii)} which nonlinear dependence patterns favor RACP, (iii)\textit{(iii)} how much these patterns contribute to performance gains, and (iv)\textit{(iv)} whether RACP better preserves useful dependencies.

(i) Frequency patterns and quantitative attribution.

By Parseval's theorem (the total energy of a signal in the time domain equals its total energy in the frequency domain), the MSE decomposes exactly into frequency bands. For each test window and channel, the power spectrum of the prediction error sums to the MSE:

Pk=∣FFT(y^−y)k∣2L2,∑kPk=MSE.P_k = \frac{\lvert \mathrm{FFT}(\hat{y} - y)_k \rvert^2}{L^2}, \qquad \sum_k P_k = \mathrm{MSE}.

We partition frequency bins into three bands based on their periods: long-period (>1>1 day), intermediate/daily (66--2424 hours), and high-frequency (<6<6 hours). Periods are converted to frequency bins using each dataset's sampling rate (Weather: 10 min; ETTh1: 1 h; ETTm1: 15 min) and horizon length HH (Weather: 672; ETTh1: 168; ETTm2: 192). In steps, the band edges are therefore 1 day = 144 steps for Weather, 24 for ETTh1 and 96 for ETTm1. Because the bands partition the reported MSE, the share of the gain in each band is an exact decomposition of the improvement.

A band holding most of the error would also receive most of an improvement that ignores frequency. We therefore report, for each band, both its share of the MSE gain and its selectivity:

selectivityb=share of the MSE gain in band bshare of the baseline MSE in band b.\text{selectivity}_b = \frac{\text{share of the MSE gain in band } b}{\text{share of the baseline MSE in band } b}.

A selectivity above 1 means the improvement is concentrated in that band beyond its size.

DatasetComparisonLong (greater 1 day): share / selectivityDaily (1/4–1 day): share / selectivityHigh (less than 1/4 day): share / selectivity
WeatherR-DiT vs DiT[80.9 ]% / 1.08[4.1 ]% / 0.33[14.9 ]% / 1.18
ETTm1R-DiT vs DiT[ 86.1]% / 1.22[ -1.4]% / -0.06[15.4 ]% / 1.04
ETTh1R-DiT vs DiT[ 70.0]% / 1.30[10.5 ]% / 0.76[19 ]% / 1.08
  • Long-period structure: Accounts for 7070--86%86\% of the net MSE improvement, with selectivity 1.081.08--1.301.30.
  • High-frequency structure: Contributes 14.914.9--19.4%19.4\%, with selectivity 1.041.04--1.181.18, consistent with the ability of rational functions to represent sharper variations.
  • Dominant daily/intermediate periodicity: Selectivity is below 11 in all comparisons, with a slight deterioration on ETTm1 (−1.4%-1.4\%). Thus, the main gains arise beyond the intermediate periodic band, particularly in long-period and high-frequency structure.

These shares provide an exact frequency-wise accounting of MSE improvements.

(ii) Task-dependent correlation patterns.

Table 6 shows that RACP changes the cross-channel correlation of the latent representation, and that the direction of the change depends on the task:

TaskDatasetMean corr. before RACPMean corr. after RACP
ForecastingElectricity0.5890.880 ↑
ImputationETTh10.5310.998 ↑
Anomaly detectionSMAP0.9480.768 ↓
Anomaly detectionMSL0.8930.578 ↓

These results demonstrate task-dependent changes in feature relationships. We do not assume that either increasing or decreasing correlation is inherently beneficial; we evaluate the usefulness of the learned dependencies directly in Part (iv).

(iii) Which nonlinear dependencies favor RACP?

To measure dependency patterns from the data, without any model , we compute the nonlinear gap on the training split (z-normalised): the mean distance correlation minus the mean absolute correlation over all channel pairs. Distance correlation detects any dependence, linear or not, so a positive gap means the channels are related in a nonlinear way.

DatasetNonlinear gapRACP projection gain
PEMS-SF+0.186+6.4% acc. (Table 8)
ETTm2+0.081− 59% MSE (Table 8)
Weather+0.053− 4.9% MSE (Table 8)
SWaT+0.018+3.2% F1 (Table 8)
MSL0.000+0.26 pts acc. (Table 7)
SMAP0.000−0.04 pts acc. (Table 7)

For the first four datasets, improvements compare R-DiT + RACP against R-DiT + MLP (Table 8). MSL and SMAP use the identity-mapping ablation from Table 7.

Datasets with larger nonlinear-dependence proxy values, particularly PEMS-SF and ETTm2, also exhibit substantial RACP improvements. In contrast, MSL and SMAP exhibit negligible proxy values and small performance changes. This is consistent with the hypothesis that rational projections are useful for nonlinear cross-feature dependencies.

AuthorsOpenReview ↗

Response to Reviewer JHGR

(iv) Does RACP represent useful dependencies?

To assess whether RACP captures meaningful and accurate inter-feature dependencies, we compare the cross-channel correlation matrix of the model's predictions, CpredC_{\mathrm{pred}}, with that of the ground truth, CtrueC_{\mathrm{true}}, using the Frobenius norm.

$ E_{\mathrm{corr}} = \left| C_{\mathrm{pred}}-C_{\mathrm{true}} \right|_F.$

Lower values indicate better agreement with the ground-truth linear dependence structure.

DatasetConfigurationCorrelation error ↓Mean abs. corr. (true)MSE ↓
Weather (21 ch.)R-DiT + MLP4.6080.531 (0.353)0.308
R-DiT + RACP2.1260.424 (0.353)0.293
ETTh1 (7 ch.)R-DiT + MLP1.1550.393 (0.262)0.428
R-DiT + RACP1.0180.333 (0.262)0.401
  • Weather: RACP reduces correlation-matrix error by 53.9% compared with MLP (4.608→2.1264.608\rightarrow2.126), while reducing MSE from 0.3080.308 to 0.2930.293.
  • ETTh1: Correlation error decreases by 11.9% (1.155→1.0181.155\rightarrow1.018), alongside an MSE reduction from 0.4280.428 to 0.4010.401.

Thus, RACP does not merely alter correlation magnitude; its predictions better match the ground-truth cross-channel correlation structure in both evaluated datasets. This directly supports improved preservation of useful linear dependencies, although it does not by itself establish improved nonlinear dependency recovery.

Summary. The new analyses identify three complementary findings: (i) long-period components explain 7070--86%86\% of the observed net forecasting MSE gains, with additional high-frequency improvements; (ii) larger nonlinear-dependence proxy values are associated with substantial RACP gains; and (iii) RACP improves agreement between predicted and ground-truth cross-channel correlations by 53.9%53.9\% on Weather and 11.9%11.9\% on ETTh1.

Together, these results provide quantitative evidence about where rational functions are advantageous and how their learned feature relationships relate to prediction quality.

We have added these new analyses in Appendix L.

Weakness 3: The claim that earlier ⋯\cdots formulation unclear.

Question 2: How much of ⋯\cdots without diffusion?

We thank the reviewer for pointing out TSDE; we will add it to Related Work and state our claim more precisely. TSDE does evaluate all four task families, but within a different paradigm: it is a self-supervised representation-learning method that uses diffusion (IIF masking) as a pretext task to pretrain general-purpose embeddings, which are then adapted to each task via fine-tuning or lightweight heads. RAF-Diff is a conditional generative diffusion model trained end-to-end in supervised method on each task's target, and its contributions lie in the diffusion formulation itself: the learnable rational forward process (RACP) and the rational denoiser (R-DiT).

Significance beyond task coverage. Our significance rests not on coverage alone but on:

  • a learnable rational forward process improves over MLP projection at a matched parameter count;
  • the rational denoiser R-DiT, which improves over DiT on all four tasks under both projections (Table~8);
  • consistently top average ranks across 29 datasets: 1.5 (forecasting), 1.5 (imputation), 1.9 (classification) and 2.4 (anomaly detection) (Tables~2--5).

Question 2: how much is retained without diffusion?

We evaluated a non-diffusion Transformer equipped with the same rational KAN layers (KAT-style), with hyperparameters tuned for each dataset.

TaskDatasetNo diffusionRAF-Diff
Anomaly det. (F1 ↑)MSL83.5889.74
SMAP71.0391.84
Classification (Acc ↑)Handwriting0.3770.382
UWave0.8590.884

Anomaly detection: The diffusion-based framework provides substantial performance improvements, with F1-score gains of 6.16 points on MSL and 20.81 points on SMAP compared with the non-diffusion rational architecture. Classification also benefits, with accuracy improvements of 0.5 percentage points on Handwriting and 2.5 percentage points on UWave. These results demonstrate the added value of the complete RAF-Diff framework, with particularly pronounced

Reviewer_4C9fOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes RAF-Diff framework, which introduces a learnable rational-function projection into the mean of a conditional forward diffusion process and uses a Rational Denoising Transformer for reconstruction. The framework is evaluated on forecasting, imputation, classification, and anomaly detection. The paper attributes its gains to adaptive transformations and rational-function expressivity, supported by component ablations.

Critical strengths
  1. The implementation demonstrates useful task breadth. Evaluation across four time-series tasks supports the framework’s practical versatility and provides evidence beyond a single-task demonstration.

  2. Table 8 crosses MLP/RACP projection with DiT/R-DiT denoising. This helps distinguish the contributions of the two modifications and supports their usefulness in the tested configurations, although capacity and compute matching remain important.

Critical weaknesses
  1. CN-Diff already uses learnable nonlinear transformations in conditional forward diffusion. NDM studies nonlinear diffusion transformations more generally, and KAT introduces rational-basis KAN layers into Transformers. RAF-Diff does not sufficiently identify the new principle beyond combining these ideas in a time-series framework. The broad redesign claims exceed the demonstrated distinction, and some forecasting differences from CN-Diff are small or tied. Matched comparisons are needed to establish a rational-specific advantage.

  2. Eq. (13) multiplies reverse-chain factors by additional conditional densities of the same x0x^0, which generally does not yield a normalized joint distribution; for T=1T=1, a conditional density is squared. Without a valid normalization or auxiliary-variable construction, the subsequent variational argument does not follow. Eq. (5) also splits a Gaussian noise coefficient into two terms while reusing the same noise, an invalid pathwise equality.

Questions for authors
  1. How is Eq. (13) normalized? Are its additional factors auxiliary variables, auxiliary losses, or part of the generative joint? Please relate the corrected formulation to the implemented loss and sampler, and clarify whether Eq. (5) requires independent noises and equality in distribution.

  2. What mechanism is new after accounting for CN-Diff and KAT? Which matched result isolates a rational-function advantage from differences in capacity, parameter count, or optimization?

  3. How are Table 5’s thresholds selected, is PA applied, and what validation/test label access is used? Were RAF-Diff and the baselines evaluated with the same preprocessing and model-selection rules?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 4C9f

Weakness 1: CN-Diff already uses ⋯\cdots to establish a rational-specific advantage.

Questions 2: What mechanism ⋯\cdots optimization?

Response: We emphasize that our novelty is not these ingredients (learnable nonlinear diffusion, rational functions) individually, but their specific parameterization and integration within conditional diffusion for multivariate time-series analysis (forecasting, classification, imputation and anomaly detection). The integration is not a "trivial engineering exercise", rather it is a grounded research contribution that advances the state of the art. As requested, we further clarify the distinction and provide matched evidence.

(i) Distinct methodological contribution.

RAF-Diff introduces rational functions at two stages: (i) RACP parameterizes a learnable, time-dependent forward transformation designed to capture inter-feature dependencies; and (ii) R-DiT integrates rational-function layers into a timestep-conditioned diffusion denoiser.

Unlike CN-Diff, which focuses on nonlinear conditional diffusion for forecasting, RAF-Diff investigates rational projection and denoising across four time-series tasks. NDM establishes learnable nonlinear diffusion for images, while KAT employs rational functions within Transformers without a diffusion process.

Thus, our contribution is the rational parameterization of both forward transformation and diffusion denoising, together with controlled evaluation of their individual and combined effects.

(ii) Matched rational vs MLP comparisons (Table 8).

To isolate the contribution of rational functions, we evaluate a 2×22\times2 ablation varying only the forward projection (MLP vs RACP) and denoiser (DiT vs R-DiT). Within each dataset, conditioning, hyperparameters, training protocol, and evaluation settings are held fixed. Total trainable parameters are shown in parentheses.

Task (metric)DatasetDiT + MLPDiT + RACPR-DiT + MLPR-DiT + RACP
Forecasting (MSE ↓\downarrow)Weather0.325 (8,950,417)0.320 (8,950,427)0.308 (8,950,437)0.293 (8,950,447)
Imputation (MSE ↓\downarrow)ETTm20.035 (3,433,374)0.028 (3,433,384)0.034 (3,433,414)0.014 (3,433,424)
Anomaly detection (F1 ↑\uparrow)SWaT80.80 (314,935)88.33 (314,945)87.35 (315,015)90.16 (315,025)
Classification (Acc. ↑\uparrow)PEMS-SF0.828 (13,425,166)0.838 (13,425,176)0.844 (13,425,186)0.898 (13,425,196)
  • Rational forward projection: Holding R-DiT fixed, replacing MLP with RACP improves Weather MSE (0.308→0.2930.308\rightarrow0.293), ETTm2 MSE (0.034→0.0140.034\rightarrow0.014), SWaT F1 (87.35→90.1687.35\rightarrow90.16), and PEMS-SF accuracy (0.844→0.8980.844\rightarrow0.898). Improvements also occur with DiT fixed.

  • Rational denoiser: Holding the projection fixed, R-DiT improves upon DiT under both MLP and RACP projections across all four evaluated tasks.

  • Capacity control: Across the four configurations, the maximum difference is 90 parameters (<0.03<0.03%). Thus, the improvements cannot be explained simply by a increase in parameter count.

(iii) Comparison with a non-diffusion rational KAN Transformer.

To investigate whether rational KAN layers alone account for RAF-Diff's performance, we additionally evaluated a non-diffusion Transformer equipped with rational KAN layers (KAT-style).

Task (metric)DatasetRational KAN TransformerRAF-Diff
Forecasting (MSE ↓\downarrow)ETTh10.6240.401
ETTm10.7640.338
Exchange0.0560.016
Imputation, 25% mask (MSE ↓\downarrow)ETTh10.1040.052
ETTm10.0260.021
Anomaly detection (F1 ↑\uparrow)MSL83.5889.74
SMAP71.0391.84
Classification (Acc. ↑\uparrow)Handwriting0.3770.382
UWave0.8590.884

RAF-Diff outperforms the non-diffusion rational KAN baseline on all evaluated datasets, notably on ETTm1 forecasting (0.764→0.3380.764\rightarrow0.338) and SMAP anomaly detection (71.03→91.8471.03\rightarrow91.84).

This comparison supports the effectiveness of the complete diffusion-based framework relative to a standalone rational KAN Transformer.

AuthorsOpenReview ↗

Response to Reviewer 4C9f

Weakness 2: Eq. (13) multiplies ⋯\cdots pathwise equality.

Question 1: How is Eq. (13) normalized? ⋯\cdots equality in distribution.

(a) Eq:(5)

For ease of understanding, the original derivation used a single noise term ϵ\epsilon in both components. We have now given it with explicit time indices on the noise. Following Understanding Diffusion Models: A Unified Perspective [1] (Eqs. 61-70), we use independent noise variables ϵt,ϵt−1∼iidN(0,I)\epsilon_t, \epsilon_{t-1} \overset{\text{iid}}{\sim} \mathcal{N}(0, I) and define

ϵˉt\bar{\epsilon}_t = 1−αtϵt+αt(1−αˉt−1)ϵt−11−αˉt\frac{\sqrt{1-\alpha_t} \epsilon_t + \sqrt{\alpha_t (1- \bar{\alpha}_{t-1})}\epsilon_{t-1}} {\sqrt{1-\bar{\alpha}_t}}

Since the sum of independent Gaussians is Gaussian with variance (1−αt)+αt(1−αˉt−1)=1−αˉt(1-\alpha_t) + \alpha_t(1-\bar{\alpha}_{t-1}) = 1-\bar{\alpha}_t, we have ϵˉt∼N(0,I)\bar{\epsilon}_t \sim \mathcal{N}(0, I). The updated Eq.(5), in which every equality holds pathwise, is

xt=αˉtRACPϕ(x,t)+(1−αˉt)c+1−αˉtϵˉtx^t = \sqrt{\bar{\alpha}_t}\mathrm{RACP}_\phi(x,t) + (1-\sqrt{\bar{\alpha}_t})c + \sqrt{1-\bar{\alpha}_t}\bar{\epsilon}_t

$x^t = \sqrt{\bar{\alpha}_t}\mathrm{RACP}_\phi(x,t) + (1-\sqrt{\bar{\alpha}_t})c + \sqrt{1-\alpha_t}\epsilon_t + \sqrt{\alpha_t}\sqrt{1-\bar{\alpha}_{t-1}}\epsilon_{t-1} $

$x^t = \sqrt{\bar{\alpha}_t}\mathrm{RACP}_\phi(x,t) + (1-\sqrt{\bar{\alpha}_t})c + \sqrt{1-\alpha_t}\epsilon_t $ +

αt(xt−1−αˉt−1 RACPϕ(x,t−1)−(1−αˉt−1) c)\sqrt{\alpha_t}\big(x^{t-1} - \sqrt{\bar{\alpha}_{t-1}}\,\mathrm{RACP}_\phi(x,t-1) - (1-\sqrt{\bar{\alpha}_{t-1}})\,c\big)

xt=αt xt−1+αˉt(RACPϕ(x,t)x^t = \sqrt{\alpha_t}\,x^{t-1} + \sqrt{\bar{\alpha}_t}\big(\mathrm{RACP}_\phi(x,t) -

RACPϕ(x,t−1))+(1−αt) c+1−αt ϵt\mathrm{RACP}_\phi(x,t-1)\big) + (1-\sqrt{\alpha_t})\,c + \sqrt{1-\alpha_t}\,\epsilon_t

where the third line uses xt−1=αˉt−1 RACPϕ(x,t−1)+(1−αˉt−1) c+1−αˉt−1 ϵt−1x^{t-1} = \sqrt{\bar{\alpha}_{t-1}}\,\mathrm{RACP}_\phi(x,t-1) + (1-\sqrt{\bar{\alpha}_{t-1}})\,c + \sqrt{1-\bar{\alpha}_{t-1}}\,\epsilon_{t-1}, a sample from Eq. (4) at step t−1t-1. Because ϵt\epsilon_t is independent of xt−1x^{t-1}, the last line gives exactly the transition in Eq. (6). The transition (6) and the marginal (4) are therefore consistent, and the posterior derivation (Eqs. 7-12) is unchanged.

The derivation requires independent noises ϵt,ϵt−1\epsilon_t, \epsilon_{t-1}. With the coupling ϵˉt\bar\epsilon_t defined above, every line of Eq. (5) holds pathwise, and the marginal statement xt∼qϕ(xt∣x,c)x^t \sim q_\phi(x^t \mid x, c) holds in distribution, since ϵˉt∼N(0,I)\bar\epsilon_t \sim \mathcal{N}(0, I).

(b) Eq. (13):

We update Eq.(13) to its normalized version as follows. However, we emphasize that the loss term remains the same, only the way the joint is written changed.

Step 1: Non-Markovian forward process. From Eq. (1), the forward marginals are qϕ(xt∣x,c)=N(αˉtRACPϕ(x,t)+(1−αˉt)c  (1−αˉt)I)q_\phi(x^t \mid x, c) = \mathcal{N}\big(\sqrt{\bar\alpha_t}\mathrm{RACP}_\phi(x,t) + (1-\sqrt{\bar\alpha_t})c\; (1-\bar\alpha_t)I\big).

As shown in the corrected Eqs. (5)-(6), the transition qϕ(xt∣xt−1,x,c)q_\phi(x^t \mid x^{t-1}, x, c) depends on the data xx directly, through RACPϕ(x,t)−RACPϕ(x,t−1)\mathrm{RACP}_\phi(x,t) - \mathrm{RACP}_\phi(x,t-1).

The forward process is therefore non-Markovian, and following DDIM [2] we write it as qϕ(x1:T∣x,c)=qϕ(xT∣x,c)∏t=2Tqϕ(xt−1∣xt,x,c).q_\phi(x^{1:T} \mid x, c) = q_\phi(x^T \mid x, c)\prod_{t=2}^{T} q_\phi(x^{t-1} \mid x^t, x, c). Unlike in DDPM, where the data enters only the first forward step, here xx enters every forward transition.

Step 2: Generative joint in Normalized Form (Eq.~13).

For time series diffusion models, forecasting xθx_\theta has been found to be more effective than predicting ϵθ\epsilon_\theta [3,4]. Our denoiser therefore predicts the clean series x^θ(xt,t)\hat x_\theta(x^t, t) from every noisy state, and each reverse step is built on this prediction, pθ(xt−1∣xt,c)=qϕ(xt−1∣xt,x^θ(xt,t),c)p_\theta(x^{t-1} \mid x^t, c) = q_\phi\big(x^{t-1} \mid x^t, \hat x_\theta(x^t, t), c\big) (Eq.~16). Because xx enters every forward transition, the generative process mirrors this with a link back to the data at every step. Each such link generates its own auxiliary variable x~(t)\tilde x^{(t)}: $ p_\theta\big(x^{1:T}, \tilde x^{(1:T)} \mid c\big) = p(x^T \mid c)\prod_{t=2}^{T} p_\theta(x^{t-1} \mid x^t, c);\prod_{t=1}^{T} p_\theta\big(\tilde x^{(t)} \mid x^t, c\big), \qquad p_\theta\big(\tilde x^{(t)} \mid x^t, c\big) = \mathcal{N}\big(\hat x_\theta(x^t, t), \sigma^2 I\big). $

The reverse chain now ends at x1x^1 (since qϕ(x0∣x1,x,c)q_\phi(x^0 \mid x^1, x, c) has zero variance), and each auxiliary variable x~(t)\tilde x^{(t)} is generated exactly once by a normalized conditional. Integrating out x~(1),…,x~(T)\tilde x^{(1)}, \dots, \tilde x^{(T)}, then x1,…,xT−1x^1, \dots, x^{T-1}, then xTx^T gives one at every step, so the joint is normalized. In particular, no density is squared at t=1t = 1.

Step 3: Likelihood. During training, every auxiliary variable is observed to equal the data, x~(1)=⋯=x~(T)=x\tilde x^{(1)} = \dots = \tilde x^{(T)} = x. Our goal is to maximize pθ(x~(1:T)=x∣c)=∫pθ(x1:T,x~(1:T)=x∣c) dx1:T,p_\theta\big(\tilde x^{(1:T)} = x \mid c\big) = \int p_\theta\big(x^{1:T}, \tilde x^{(1:T)} = x \mid c\big)\, dx^{1:T}, which is intractable, so we maximize a variational lower bound on it.

AuthorsOpenReview ↗

Response to Reviewer 4C9f

Step 4: Variational bound, which is exactly Eq. (3).

Using qϕq_\phi as the variational posterior and applying Jensen's inequality,

log⁡pθ(x~(1:T)=x∣c)≥Eqϕ[log⁡pθ(x1:T,x~(1:T)=x∣c)qϕ(x1:T∣x,c)].\log p_\theta\big(\tilde x^{(1:T)} = x \mid c\big) \ge \mathbb{E}_{q_\phi}\left[\log \frac{p_\theta\big(x^{1:T}, \tilde x^{(1:T)} = x \mid c\big)}{q_\phi(x^{1:T} \mid x, c)}\right].

The numerator (Eq. 13) and the denominator (Step 1) are factorized over the same steps in the same direction, so the log-ratio splits term by term:

−log⁡pθ(x~(1:T)=x∣c)≤DKL(qϕ(xT∣x,c) ∥ p(xT∣c))⏟Lprior-\log p_\theta\big(\tilde x^{(1:T)} = x \mid c\big) \le \underbrace{D_{\mathrm{KL}}\big(q_\phi(x^T \mid x, c)\,\|\,p(x^T \mid c)\big)}_{\mathcal{L}_{\text{prior}}}

$ + \underbrace{\sum_{t=2}^{T} \mathbb{E}_{q_\phi} D_{\mathrm{KL}}\big(q_\phi(x^{t-1} \mid x^t, x, c),|,p_\theta(x^{t-1} \mid x^t, c)\big)}_{\mathcal{L}_{\text{diff}}}$

−∑t=1TEqϕlog⁡pθ(x∣xt,c)⏟Lrec.\underbrace{-\sum_{t=1}^{T} \mathbb{E}_{q_\phi}\log p_\theta(x \mid x^t, c)}_{\mathcal{L}_{\text{rec}}}.

This is exactly Eq. (3), with the diffusion sum indexed from t=2t = 2. The bound holds for every ϕ\phi. This structure is already established in state-space models [6]: over the diffusion steps, Eq. (13) has an initial state, Markov transitions and an emission at every step, and the bound above is the standard variational objective of deep state-space models

Relation to the standard data likelihood. The sampler outputs x~(1)\tilde x^{(1)}. By the chain rule,

log⁡pθ(x~(1:T)=x∣c)=log⁡pθ(x~(1)=x∣c)+log⁡pθ(x~(2:T)=x∣x~(1)=x,c).\log p_\theta\big(\tilde x^{(1:T)} = x \mid c\big) = \log p_\theta\big(\tilde x^{(1)} = x \mid c\big) + \log p_\theta\big(\tilde x^{(2:T)} = x \mid \tilde x^{(1)} = x, c\big).

The first term is the usual data likelihood. The second is a consistency term on the intermediate predictions, bounded above by −(T−1)d2log⁡(2πσ2)-(T-1)\tfrac{d}{2}\log(2\pi\sigma^2) for fixed-variance Gaussian emissions. Hence

−log⁡pθ(x~(1)=x∣c)≤LRAF-Diff−(T−1)d2log⁡(2πσ2)-\log p_\theta\big(\tilde x^{(1)} = x \mid c\big) \le \mathcal{L}_{\text{RAF-Diff}} - (T-1)\tfrac{d}{2}\log(2\pi\sigma^2)

so Eq. (3) also bounds the standard negative log-likelihood, up to a constant independent of ϕ\phi and θ\theta.

Relation to the implemented loss and sampler.

ComponentRole in the joint (Eq. 13)Role in the implemented lossRole in the sampler
p(xT∣c)p(x^T \mid c)Initial stateLprior\mathcal{L}_{\text{prior}} (closed form, Eq. 17)xT∼N(c,I)x^T \sim \mathcal{N}(c, I)
pθ(xt−1∣xt,c)p_\theta(x^{t-1} \mid x^t, c), t=2..Tt = 2..TReverse transitionsLdiff\mathcal{L}_{\text{diff}} (closed form, Eq. 16)One reverse step per tt, using x^θ(xt,t)\hat x_\theta(x^t, t)
pθ(x~(t)∣xt,c)p_\theta(\tilde x^{(t)} \mid x^t, c), t=1..Tt = 1..TAuxiliary observed variables (emissions)Lrec\mathcal{L}_{\text{rec}}: 12σ2∥x−x^θ(xt,t)∥2\tfrac{1}{2\sigma^2}\lVert x - \hat x_\theta(x^t, t)\rVert^2 + constOnly x~(1)\tilde x^{(1)} is used, as the output x^θ(x1,1)\hat x_\theta(x^1, 1)

The implemented loss is therefore exactly the bound above, and the implemented sampler is ancestral sampling from the corrected Eq. (13): it draws xTx^T from the prior, applies the reverse transitions down to x1x^1, and returns the mean of the emission x~(1)\tilde x^{(1)}. The auxiliary variables x~(2:T)\tilde x^{(2:T)} are part of the generative joint and contribute to training, but are not needed at sampling time.

Question 3: How are Table 5’s ⋯\cdots rules?

All anomaly detection experiments, for RAF-Diff and for every baseline in Table 5, follow the standard TSLib protocol even validation/test label without modification. Following TSLib and Anomaly Transformer[5], the point-adjustment (PA) strategy is applied.


References:

[1] Luo, C. (2022). Understanding diffusion models: A unified perspective.

[2] Song, J., Meng, C., & Ermon, S. (2020). Denoising diffusion implicit models.

[3] Feng, S., Miao, C., Zhang, Z., & Zhao, P. (2024, March). Latent diffusion transformer for probabilistic time series forecasting.

[4] Shen, L., & Kwok, J. (2023, July). Non-autoregressive conditional diffusion models for time series prediction.

[5] Xu, J., Wu, H., Wang, J., & Long, M. (2021). Anomaly transformer: Time series anomaly detection with association discrepancy.

[6] Karl, M., Soelch, M., Bayer, J., & Van der Smagt, P. (2016). Deep variational bayes filters: Unsupervised learning of state space models from raw data.

Thanks for author's response and the additional experimental results, which address most of my concerns. Therefore, I am maintaining my positive rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Author Remarks

We sincerely thank all reviewers and chairs for their time and constructive feedback. We would like to summarize the key contributions of our work, the extensive experimental evaluation, and how we addressed the concerns raised during the rebuttal.

1. Methodological Novelty and Contributions

RAF-Diff introduces a conditional diffusion framework for four major time-series tasks: forecasting, imputation, classification, and anomaly detection. Its key contributions are:

  • Rational Adaptive Correlation Projection (RACP): A learnable, time-dependent rational-function transformation incorporated into the forward diffusion process to model complex inter-feature dependencies.

  • Rational Denoising Transformer (R-DiT): A diffusion denoiser incorporating rational-function layers with timestep conditioning to improve reconstruction across noise levels.

Integration of both components into a common diffusion architecture applicable across four tasks. Unlike existing methods such as CN-Diff and KAT, our contribution lies in the rational parameterization of both the forward transformation and diffusion denoiser, together with their controlled empirical validation.

2. Extensive Experimental Evaluation

We evaluated RAF-Diff across 29 dataset–task combinations, covering 8 forecasting, 6 imputation, 10 classification, and 5 anomaly detection datasets. Our evaluation includes extensive comparisons against existing methods, component-level ablations, sensitivity studies, efficiency comparisons, and statistical analyses. RAF-Diff achieves strong performance across all four tasks, with average ranks of 1.5 for forecasting, 1.5 for imputation, 1.9 for classification, and 2.4 for anomaly detection, demonstrating its effectiveness and broad applicability.

3. Comprehensive Response to Reviewer Concerns

During the rebuttal, we carefully addressed all major technical concerns raised by the reviewers, including:

  • Novelty and comparison with prior work: Clarified the distinctions from CN-Diff, KAT, TSDE, and other relevant methods, and refined our novelty claims.

  • Component contributions: Provided controlled, nearly parameter-matched comparisons between rational and MLP projections, and between DiT and R-DiT, demonstrating the individual effectiveness of both components.

  • Role of diffusion: Conducted additional comparisons against a non-diffusion rational KAN Transformer across all four tasks.

  • Frequency and feature dependencies: Performed new frequency-domain analyses, nonlinear-dependence measurements, and correlation-matrix evaluations to explain when and why rational functions improve performance.

  • Theoretical soundness: Corrected and clarified the normalized generative formulation, variational derivation, training objective, and sampling procedure, along with additional empirical diagnostics.

  • Experimental consistency: Clarified the differences between ablation settings, anomaly detection protocols, ranking conventions, and other evaluation details.

These responses included substantial additional theoretical analysis and new experiments beyond the original submission.

4. Reviewer Feedback After Rebuttal

We appreciate that the reviewers acknowledged our detailed responses:

  • Reviewer 4C9f explicitly stated that our additional results addressed most concerns and maintained a positive rating.
  • Reviewer Mb1x acknowledged our careful responses and increased both presentation and soundness scores while maintaining a positive overall rating.
  • Reviewer JHGR appreciated our thorough explanations without specifying further unresolved technical concerns.

We respect the reviewers’ assessments and appreciate their feedback. We believe that our comprehensive rebuttal has addressed the major questions raised and has substantially strengthened the technical and empirical justification of the proposed framework.

We believe RAF-Diff makes a meaningful methodological and empirical contribution to diffusion-based time-series modeling and merits favorable consideration.

Thank you for your time and thoughtful consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.