acceptodds
Under review as a conference paper at ICLR 2027

Revisiting Noise Design in Time Series Diffusion Models: Adaptive Schedule with Temporal Prior

Abstract

Diffusion models have demonstrated remarkable generative capabilities and been increasingly adopted in time series analysis. Recent research primarily focus on developing effective architectures and training protocols to improve performance on downstream tasks. Despite the critical role of noise in diffusion models, it has received comparatively limited attention, hindering time series diffusion models from fully realizing their potential. In this work, we propose a principled approach to design Adaptive Noise for Time Series diffusion models (ANTS) based on the inherent and unique properties of data. Our core insight is that an efficient noise schedule should distribute corruption process to maintain uniform information degradation across diffusion steps, eliminating redundant noise levels that contribute little learnable signal. It improves the precision of noise estimation and facilitates better traceability of diffusion. Furthermore, we incorporate informative statistic priors to capture temporal dependencies and enhance sample fidelity during reverse process. Our method incurs only marginal additional offline computational cost and can be seamlessly integrated into frameworks based on standard diffusion processes. Experiments on generation and forecasting tasks across various datasets demonstrate that ANTS improves generative quality and predictive performance over baselines, validating its effectiveness for time series diffusion models.

Primary area:
learning on time series and dynamical systems
Keywords:
time series generative models, diffusion models, noise scheduling, time series generation, time series forecasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wQnsOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes ANTS, an Adaptive Noise for Time Series diffusion models with experiments showing time series generation and forecasting. An offline method, Integrated Absolute Autocorrelation Time (IAAT) is used based on dataset statistics to design the noise schedule for time series noising and stated as an efficient noise schedule that distribute corruption process to maintain uniform information degradation across diffusion steps, which eliminates redundant noise levels that contribute to little learnable signal. Also prior of the diffusion is also estimated using dataset statistics. Experiments related to time series generation and forecasting are shown with ablations.

Critical strengths
  1. The proposed method is model/architecture agnostic and can be computed offline with dataset characteristics.
  2. The ablations with multiple component variations (Figure 4) and Raw, Lag1AC, VarAC, comparisons (Table 3, Appendix D.3) are useful and provided further insights.
Critical weaknesses
  1. Proposed components are of very marginal or of no effect on performance. The ablation results of Figure 4 clearly shows that in several experimental settings, removing adaptive schedule, temporal prior, or Gaussian process components leads to very less change in performance. Although certain specific settings shows some improvement (eg: ETTh1-DiffusionTS, Wind) but it is not consistent across models and datasets (eg: ECL dataset for forecasting and ETTh1-PaDTS for generation).
  2. Tables 6 and 7 do not convincingly demonstrate the superiority of the proposed IAAT, alternatives such as Lag1AC and VarAC achieve comparable or better results in numerous settings.
  3. Core idea of ANTS (adaptive noise scheduling based on time-series statistics) has substantially overlap with a well established paper ANT (NeurIPS 2024) and another idea (controlling diffusion schedules through uniform distributional changes) is very much similar to CRS (Arxiv 2024 and TMLR 2026) which makes the proposed work, a incremental of existing methods.
Questions for authors
  1. Can authors provide direct experimental comparison over ANT (NeurIPS 2024) and Constant Rate Scheduling (CRS) under identical settings ?
  2. Proposition 1 establishes monotonic IAAT degradation under temporally independent Gaussian noise, but ANTS uses GP-correlated noise. Does IAAT still decrease monotonically under this prior? If not, how are Eq. 9 and Theorem 1 justified?
  3. What concrete advantages does IAAT offer over Lag1AC and VarAC, and why should it be preferred ?
  4. Are the hyper-parameters and denoising steps same for Baseline and Baseline + ANTS ?
5 comments
AuthorsOpenReview ↗

Response to Reviewer wQns

Dear Reviewer wQns,

We sincerely appreciate your thorough review and valuable comments. We believe that the main concerns may stem from experiment results and distinctions from prior work. We respond to each point below to clarify core contributions of our method.


Response to W1:

Thank you for raising this point. We clarify that the ablation results indicate that each component is necessary and contributes to the overall performance of ANTS. Although the gains may appear modest in a few cases from figure, the actual improvements are consistent across experiments and can be substantial for particular settings. Since ANTS is designed to be a architecture agnostic method, each component that potentially play a critical role is essential to achieving broad effectiveness.

We understand that the differing scales may obscure some effects. We provide the numerical ablation results in the table below. It shows the contribution of each component more clearly and indicate that full ANTS performs best among variants.

DatasetModelANTSw/o Adaptive ScheduleIAAT-Focusedw/ Zerow/o Temporal Priorw/o GP
ETTh1DiffusionTS1.2471.3032.1791.7271.2891.679
PaDTS1.5201.6741.5671.5711.7641.557
WindDiffusionTS0.5590.7181.0290.7780.6680.589
PaDTS0.3900.8330.5560.5650.7620.441
ECLSSSD0.3640.3920.3680.3850.3900.382
TimeDiff0.7390.7620.7420.7560.7470.749

Response to W2 and Q3:

Thank you for your feedback. We prefer IAAT since it achieves optimal overall performance across a broad range of settings compared with Lag1AC and VarAC. IAAT captures autocorrelation across all lags and reflects finer temporal dependence than Lag1AC and VarAC. Tables 3, 6 and 7 support its strong average performance. In particular, IAAT ranks among the top-2 in 42 of 48 settings, compared with 26 for Lag1AC and 22 for VarAC. It supports IAAT as a generally effective statistic.

Besides, as we have discussed in Section 5.4, there is no universally optimal statistic for measuring time series corruption. We emphasize that the advantage of IAAT is its broad robustness rather than being best for each individual setting. More generally, our main contribution lies in the method for adaptively constructing schedules based on time series statistics. As demonstrated by experiments in Appendix D.3, it exhibits strong robustness to different statistical measures. Although we adopt IAAT in the current implementation, the framework can be flexibly adapted to other suitable measures.

AuthorsOpenReview ↗

Response to Reviewer wQns

Response to W3 and Q1:

Thank you for raising this important issue. Our core contribution is an efficient schedule constructed directly from unique characteristics of corrupted time series which is underexplored. (1) ANT selects the best schedule from a manually specified candidate space. Its performance is limited by the space and incurs the cost of evaluating. By contrast, ANTS directly resamples the captured trajectory to construct target schedule. In this sense, ANTS represents the optimal solution that ANT seeks to approximate. (2) CRS introduces uniform distributional change but primarily focus on image. Besides,it incurs additional model training and computation overhead, whereas ANTS estimates the IAAT trajectory and constructs schedule offline. We previously discussed these points briefly in Related Work Section. We have revised it and added Appendix D.4 to more clearly highlight how ANTS differs from ANT and CRS.

While no publicly code is available to reproduce CRS, we provide a fair comparison between ANTS and ANT using the same diffusion backbone, prior and diffusion steps. The noise function and temperature search space are consistent with official ANT implementation. As shown in the table below, ANTS outperforms ANT in generation quality across datasets and metrics.

DatasetModelContext-FID Score (↓\downarrow)Correlational Score (↓\downarrow)Discriminative Score (↓\downarrow)
ANTS\textbf{ANTS}ANT\textbf{ANT}ANTS\textbf{ANTS}ANT\textbf{ANT}ANTS\textbf{ANTS}ANT\textbf{ANT}
ETTh1DiffusionTS1.247‾±0.095\underline{1.247}_{\pm 0.095}1.324±0.023{1.324}_{\pm 0.023}0.092‾±0.006\underline{0.092}_{\pm 0.006}0.096±0.005{0.096}_{\pm 0.005}0.163‾±0.008\underline{0.163}_{\pm 0.008}0.179±0.018{0.179}_{\pm 0.018}
PaD-TS1.520‾±0.083\underline{1.520}_{\pm 0.083}1.617±0.080{1.617}_{\pm 0.080}0.095‾±0.010\underline{0.095}_{\pm 0.010}0.104±0.006{0.104}_{\pm 0.006}0.170‾±0.011\underline{0.170}_{\pm 0.011}0.172±0.009{0.172}_{\pm 0.009}
WindDiffusionTS0.559‾±0.028\underline{0.559}_{\pm 0.028}0.721±0.046{0.721}_{\pm 0.046}0.072‾±0.005\underline{0.072}_{\pm 0.005}0.076±0.013{0.076}_{\pm 0.013}0.125‾±0.012\underline{0.125}_{\pm 0.012}0.132±0.019{0.132}_{\pm 0.019}
PaD-TS0.390‾±0.037\underline{0.390}_{\pm 0.037}0.484±0.041{0.484}_{\pm 0.041}0.043‾±0.003\underline{0.043}_{\pm 0.003}0.052±0.007{0.052}_{\pm 0.007}0.052‾±0.027\underline{0.052}_{\pm 0.027}0.072±0.042{0.072}_{\pm 0.042}

Response to Q2:

Thank you for careful review which helps us identify missing theoretical derivation. The monotonic degradation conclusion in Proposition 1 remains applicable to GP-correlated noise. We agree that Proposition 1 alone is insufficient to establish the conclusion. Therefore, we have extended theoretical derivation and revised the manuscript accordingly as below:

Specifically, letting X1X_1 denote the GP prior, we obtain the corresponding autocorrelation:

w_\tau=\frac{\alpha_\tau^2\operatorname{Var}(X_0^t)}{\alpha_\tau^2\operatorname{Var}(X_0^t)+(1-\alpha_\tau^2)\operatorname{Var}(X_1^t)}.$$ Accordingly, $$\psi_{\mathrm{IAAT}}(X_\tau)=1+2\sum_{k=1}^{\infty}\left|w_\tau\rho_k(0)+(1-w_\tau)\rho_k(1)\right|.$$ Since time series data typically have larger aggregate autocorrelation than GP prior, we suppose that $\sum_{k=1}^{T-1}\rho_k(0)>\sum_{k=1}^{T-1}\rho_k(1)$. This makes the derivative of convex $\psi_{\mathrm{IAAT}}$ positive. Thus, as $\alpha_\tau$ and $w_\tau$ decrease during forward diffusion, IAAT decreases monotonically toward $\psi_{\mathrm{IAAT}}(X_1)$. Our empirical IAAT trajectories show consistency with this result. Besides, the theoretical terminal boundary is not $1$ under GP-correlated noise. In practice, we use empirically measured values at endpoints. Thank you for pointing out this error and we have revised Eq. (9) accordingly. --- **Response to Q4:** For each *Baseline* and *Baseline + ANTS* variant, we keep the backbone hyper-parameters and number of denoising steps identical for fair comparison. We only modify the noise schedule and prior, as verified by our submitted anonymous code. --- We thank you again for your valuable comments, which have helped us strengthen the reliability and rigor of the paper. All revisions have been incorporated into the updated manuscript and highlighted in blue. We sincerely hope our detailed explanations could address your concerns. If you have any further questions or concerns about the method, we remain fully available to discuss them as thoroughly as possible. If the revisions meet your expectations, we would greatly appreciate your consideration of updating your score accordingly to support our work.
Reviewer_wQnsOpenReview ↗

Response to Author rebuttal

I appreciate the authors' effort in giving rebuttal, some of my concerns got resolved. But some concerns still remain open,

  1. ANT vs ANTS comparison (New comparison shown by authors in rebuttal) shows that many of the improvements are within one standard deviation, so the improvement is marginal, which does not outweigh the methodological overlap.
  2. The ablation table reports no standard deviations. Using Table 1’s std as reference, multiple variants on ETTh1 are within one std of full ANTS: w/o Adaptive Schedule and w/o Temporal Prior on DiffusionTS, and IAAT-Focused and w/o GP on PaD-TS.
  3. Section 5.3 still states "As evidenced by Figure 4, both w/o Adaptive Schedule and IAAT-Focused variants consistently lead to substantial performance degradation", from the table authors provided in rebuttal several differences are within std (e.g., 1.247 vs. 1.303 on ETTh1/DiffusionTS), which contradicts the statement "substantial performance degradation".
  4. The response on IAAT is not convincing. Lag1AC and VarAC still achieve comparable or better performance in many settings.

Considering the rebuttal, I maintain my score.

Reviewer_YFLiOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ANTS, a noise-design approach for time-series diffusion models that constructs a dataset-specific schedule from the measured evolution of integrated absolute autocorrelation time (IAAT). It combines this schedule with a complementary schedule and introduces a Gaussian prior incorporating data-dependent scales and Gaussian-process temporal correlations. Across the reported generation and forecasting experiments, integrating ANTS into several diffusion backbones improves most evaluated metrics, with some exceptions.

Critical strengths
  1. Adapting the corruption process to temporal statistics is a well-motivated direction that complements improvements to diffusion architectures.

  2. The method improves most reported generation and forecasting results across several diffusion backbones, supporting its practical usefulness.

Critical weaknesses
  1. The correspondence between the theoretical analysis and the final method could be stated more clearly. Appendix C.1 assumes independent white noise, while Algorithm 1 uses a GP-correlated prior. The minimax result in Appendix C.2 concerns the unmixed schedule, whereas the final algorithm combines it with a complementary schedule. Clarifying the applicable assumptions, the IAAT normalization, and which conclusions extend to the final construction would help substantiate the uniform-degradation rationale.

  2. The comparative evidence for replacing candidate-based scheduling with direct construction could be strengthened. The experiments and ablations support improvements over the original backbones, but do not directly compare with ANT, which also adapts schedules using IAAT. The zero-terminal-SNR control does not evaluate ANT’s full selection procedure. A focused comparison under matched settings would help establish the practical benefit of the proposed scheduling construction.

Questions for authors
  1. Which theoretical conclusions apply to the final method with the GP prior and mixed schedule, and which describe the idealized construction? In particular, could you clarify the space in which IAAT is computed, the endpoint normalization under correlated noise, and the scope of the minimax result after schedule mixing? An explanation of the actual procedure and its applicable assumptions would help distinguish the guarantees for the final design from the theoretical insights motivating its components.

  2. How does the proposed schedule compare with ANT in one representative setting using the same backbone, prior treatment, and a matched sampling budget? The existing variant without the temporal prior could provide a common setting for isolating the scheduling difference. A focused comparison of sample or forecast quality and offline schedule-construction cost would help identify the practical value of direct construction, without extending the comparison to every model and dataset.

2 comments
AuthorsOpenReview ↗

Response to Reviewer YFLi

Dear Reviewer YFLi,

We sincerely appreciate your constructive suggestions and recognition of our method as a novel complement to existing diffusion architectures. Your concerns center on the connection between the theoretical results and the final method, as well as experiment further highlighting ANTS's advantage. We provide a point-by-point response below.


Response to W1 and Q1:

Thank you for this important feedback which helps us identify missing derivation and necessary clarification. To distinguish theoretical analysis from final construction, we clarify the scope of each result:

  • Monotonic IAAT degradation remains applicable to GP-correlated prior. We agree that Appendix C.1 alone does not establish the result for GP-correlated noise. Therefore, we have extended the theoretical derivation to demonstrate its applicability to our final method. We have revised the manuscript accordingly, as detailed below. Specifically, letting X1X_1 denote the GP prior, we obtain the corresponding autocorrelation: w_\tau=\frac{\alpha_\tau^2\operatorname{Var}(X_0^t)}{\alpha_\tau^2\operatorname{Var}(X_0^t)+(1-\alpha_\tau^2)\operatorname{Var}(X_1^t)}.$$ Accordingly, $$\psi_{\mathrm{IAAT}}(X_\tau)=1+2\sum_{k=1}^{\infty}\left|w_\tau\rho_k(0)+(1-w_\tau)\rho_k(1)\right|.$$ Since time series data typically have larger aggregate autocorrelation than GP prior, we suppose that $\sum_{k=1}^{T-1}\rho_k(0)>\sum_{k=1}^{T-1}\rho_k(1)$. This makes the derivative of convex $\psi_{\mathrm{IAAT}}$ positive. Thus, as $\alpha_\tau$ and $w_\tau$ decrease during forward diffusion, IAAT decreases monotonically toward $\psi_{\mathrm{IAAT}}(X_1)$. Our empirical IAAT trajectories show consistency with this result. We have added the derivation to the manuscript.
  • Appendix C.2 theoretically supports the idealized schedule that forms a key component of the final method. Minimax result in Theorem 1 guides our idealized construction of an adaptive schedule with approximately linear degradation. However, as discussed at the end of Section 4.3, the idealized setting focusing solely on IAAT may overemphasize low noise levels while giving insufficient attention to high noise levels. Therefore, we combine it with a complementary schedule the final construction. The ablation in Section 5.3 supports the practical benefit of mixing strategy.

Regarding the implementation details in Q1:

  • We compute IAAT in original time series space by simulating corruption at different noise levels with temporal-aware prior. It is consistent with presentation in the main text and the implementation in our submitted anonymous code.

  • With a GP-correlated prior, the theoretical terminal boundary is not 11, while the IAAT of original time series varies across datasets. Therefore, we use empirically captured values at endpoints for normalization. Thank you for identifying the error in Eq. (9) and we have corrected it in the revised manuscript.

  • The initial unmixed adaptive schedule equalizes adjacent IAAT degradation and has minimax property. The complementary schedule is designed for a different consideration, so the final mixed schedule does not retain that exact guarantee.

AuthorsOpenReview ↗

Response to Reviewer YFLi

Response to W2 and Q2:

Thank you for raising this important point. The comparison indicates that ANTS outperforms ANT in overall performance. We use the same diffusion backbone, prior and diffusion steps for comparison. The noise function and temperature search space are consistent with official implementation of ANT. As shown in the table below, ANTS outperforms ANT in generation quality across datasets and metrics. ANT selects from predefined candidate schedules, whereas ANTS directly resamples the target trajectory. In this sense, the adaptive schedule of ANTS represents the optimal solution that ANT seeks to approximate.

ANTS also has lower offline schedule construction cost than ANT. ANTS estimates only 11 trajectory on a linearly spaced noise grid and obtains the adaptive schedule by interpolation, whereas ANT scores the schedules in predefined set with NN candidates. The detailed computation cost on single schedule across different scales is illustrated in Figure 6 of manuscript.

DatasetModelContext-FID Score (↓\downarrow)Correlational Score (↓\downarrow)Discriminative Score (↓\downarrow)
ANTS\textbf{ANTS}ANT\textbf{ANT}ANTS\textbf{ANTS}ANT\textbf{ANT}ANTS\textbf{ANTS}ANT\textbf{ANT}
ETTh1DiffusionTS1.247‾±0.095\underline{1.247}_{\pm 0.095}1.324±0.023{1.324}_{\pm 0.023}0.092‾±0.006\underline{0.092}_{\pm 0.006}0.096±0.005{0.096}_{\pm 0.005}0.163‾±0.008\underline{0.163}_{\pm 0.008}0.179±0.018{0.179}_{\pm 0.018}
PaD-TS1.520‾±0.083\underline{1.520}_{\pm 0.083}1.617±0.080{1.617}_{\pm 0.080}0.095‾±0.010\underline{0.095}_{\pm 0.010}0.104±0.006{0.104}_{\pm 0.006}0.170‾±0.011\underline{0.170}_{\pm 0.011}0.172±0.009{0.172}_{\pm 0.009}
WindDiffusionTS0.559‾±0.028\underline{0.559}_{\pm 0.028}0.721±0.046{0.721}_{\pm 0.046}0.072‾±0.005\underline{0.072}_{\pm 0.005}0.076±0.013{0.076}_{\pm 0.013}0.125‾±0.012\underline{0.125}_{\pm 0.012}0.132±0.019{0.132}_{\pm 0.019}
PaD-TS0.390‾±0.037\underline{0.390}_{\pm 0.037}0.484±0.041{0.484}_{\pm 0.041}0.043‾±0.003\underline{0.043}_{\pm 0.003}0.052±0.007{0.052}_{\pm 0.007}0.052‾±0.027\underline{0.052}_{\pm 0.027}0.072±0.042{0.072}_{\pm 0.042}

We thank you again for your valuable feedback, which has helped improve the clarity and rigor of our work. All revisions discussed above have been incorporated into the corresponding passages of the updated manuscript and highlighted in blue.

We genuinely hope that our clarification helps alleviate your concerns and provide a clearer understanding of ANTS. If you have any remaining questions, we would be glad to discuss and address them as thoroughly as possible.

If all concerns are satisfactorily resolved, we would greatly appreciate your consideration of updating score in support of our paper.

Reviewer_ymiNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose a method, ANTS, that improves the noise schedule for (DDPM-based) time series diffusion methods by incorporating characteristics from the data. Specifically, the noise levels are chosen as the quantiles of a CDF defined from a quantity related to the zero-frequency spectral density of the time series under diffusion. They show that this, rather heuristic, selection serves as a plug-in for distinct baseline diffusion models, helping them achieve better performance in time series generation and forecasting.

Critical strengths

The critical strength is a rather clever way to assign a noise schedule for the diffusion process that considers the inherent correlations of the time series at all lags. The community usually adopts the arbitrary (linear or) cosine schedule. The authors show that this improves performance in distintict tasks with no significant overhead. Besides that, the authors also go beyond considering the typical standard normal noise prior for the reverse diffusion and adopt a Gaussian process that also has information about the data correlations.

Critical weaknesses

The authors introduce a "discrepancy" \mathcal{D}(\tau, \tau+\delta \tau) between diffused distributions, from which they base their main arguments. It is further approximated with a "constant rate" in Eq (5). In Eq (9) the authors defined \alpha_\tau as the quantiles of F_X(\tau) but never show whether that is consistent with the ODE in Eq (7). My main concern with this is that the evolution of the "diffused distribution" is well known, both forward and backwards (e.g. Fokker-Planck equation; cf. EDM paper by Karras et al) so any statement of discrete differences or "discrepancies" should be shown consistent with it.

Besides this, the presentation of the paper should be improved:

  • Frequent misses of articles in the grammar, the "discrepancy" if vaguely introduced and only given a concrete definition until Eq (8).
  • There are several imprecise statements in the paper: e.g. "\alpha is introduced to control the corruption and make diffusion process toward desired direction." What is such "desired direction"? ... "To explicitly reflect the dependence of measurement on schedule ..." What measurement? This hasn't been mentioned in the narrative before; etc.
  • The reference to quantum annealing is anecdotal. There is no clear mapping to the problem under study.
Questions for authors
  1. How does the "discrepancy" and your assumed "constant rate" of change relates to the Fokker-planck equation and the assumptions behind it?
  2. In Algorithm 1, line ~670, you show an infinite sum over absolute correlations at different lags to obtain IAAT. In practice, your time series are windowed, so how many lags effectively enter?
  3. Having IAAT as the main statistics naturally separate the analysis into short-memory and long-memory time series. Did you consider how your proposed noise schedule affects the performance in these regimes?
2 comments
AuthorsOpenReview ↗

Response to Reviewer ymiN

Dear Reviewer ymiN,

We sincerely thank you for your insightful comments and constructive review, and for recognizing the novelty of our method. We understand that several concerns may stem from descriptions and transitions between sections that could be improved. Below, we respond to each point in turn in the hope of addressing your concerns.


Response to W1 and Q1:

Thank you for pointing out the ambiguity. We clarify that the objective represented by ODE in Eq. (7) is consistent with the discrete schedule ατ\alpha_\tau derived in Eq. (9). Following the manuscript, we use the change in statistic ψIAAT\psi_{\mathrm{IAAT}} as a proxy for the discrepancy D\mathcal{D} between diffused distributions along the path from time series to noise. FXF_X denotes the normalized cumulative IAAT degradation over [0,1][0,1]. Thus, the signal coefficient corresponding to the quantile FX−1(τ)F_X^{-1}(\tau) is ατ=1−FX−1(τ)\alpha_\tau=1-F_X^{-1}(\tau), equivalently FX(1−ατ)=τF_X(1-\alpha_\tau)=\tau. With FX′(1−ατ)=v(ατ)/ZF_X^{\prime}(1-\alpha_\tau)=v(\alpha_\tau)/Z and c=−Zc=-Z, differentiation gives

dατdτ=−Z⋅v(ατ)−1,\frac{d\alpha_\tau}{d\tau}=-Z\cdot v(\alpha_\tau)^{-1},

as required by Eq. (7). At discrete steps τ=i/Ts\tau=i/T_s, equal change of FXF_X yield equal IAAT degradation serving as proxy for discrepancy D\mathcal{D} between adjacent steps of adaptive schedule.

Under the forward SDE, the decreasing ατ\alpha_\tau induces nonnegative noise intensity βτ=−2 dlog⁡ατ/dτ\beta_\tau=-2\,d\log\alpha_\tau/d\tau, and its marginals satisfy the corresponding Fokker–Planck equation. The “discrepancy” D(τ,τ+Δτ)\mathcal{D}(\tau,\tau+\Delta\tau) of ANTS tracks the change in scalar IAAT along these marginals. Our design criterion makes that change approximately uniform without imposing additional evolution law on the full distribution.


Response to W2:

Thank you for your suggestions on improving presentation. Below, we clarify the specific points you raised and have revised the corresponding passages in Section 4.1 of manuscript:

  • We introduce the concept of "discrepancy" in Section 4.1 and subsequently give its concrete form of IAAT in Section 4.2. Section 4.1 of paper is primarily intended to present the general derivation and motivation of our method. We have added an additional note and explanation to make the connection between two sections and Eq. (8) easier to follow.

  • We have carefully reviewed and revised the manuscript for imprecise statements. In the mentioned examples, (1) “desired direction” refers to approximately uniform decrease in D\mathcal{D} as original data approach terminal distribution, while (2) “measurement” refers to discrepancy D\mathcal{D} which is instantiated by IAAT statistic of diffused time series in our method.

  • The adiabatic theorem in quantum annealing provides additional motivation for our uniform evolution schedule. As discussed in the manuscript, an interpolation parameter controls path between initial and target functions in quantum annealing, while ατ\alpha_\tau plays an analogous scheduling role along the diffusion path. Minimizing the maximum local inverse traceability rate in annealing is analogous to minimizing the maximum IAAT degradation rate v(ατ)dατ/dτv(\alpha_\tau)d\alpha_\tau/d\tau in our setting, with the aim of facilitating recovery in reverse diffusion. It provides appropriate analogy for motivation rather than merely anecdotal reference. We have improved the statement of mapping relationship in revision.

AuthorsOpenReview ↗

Response to Reviewer ymiN

Response to Q2:

We include all lags within finite time series window. For a window of length TT, the empirical IAAT sums over k=1,…,T−1k=1,\ldots,T-1. It could capture more comprehensive and fine-grained temporal dependence than limited lag set. We have corrected Algorithm 1 to reflect the implementation.


Response to Q3:

Thank you for this constructive suggestion. We provide a dataset-level analysis comparing IAAT of original time series which effectively characterizes memory persistence, with the performance gains of ANTS over baselines. Energy and Wind have substantially higher values are classified as longer-memory datasets, whereas ETTh1 and KDDCup have lower values are classified as shorter-memory datasets. As shown by the averaged improvements below, the longer-memory time series generally achieve larger gains from our method.

A possible explanation is that long-memory series contain informative correlations across many lags, the conventional noise schedule may cause a more severely aggregate loss of temporal dependence within a few diffusion steps. These abrupt changes leave model fewer intermediate states to learn to recover long-range patterns. By tracking autocorrelation changes across all lags through IAAT, ANTS allocates more steps where these dependencies degrade most rapidly, making the recovery more gradual and easier to learn. Short-memory series concentrate most of dependence at nearby lags, leaving less structure to benefit from the adjustment and resulting in smaller gains. We have added the analysis to the Appendix D.5.

MetricEnergy (long)Wind (long)ETTh1 (short)KDDCup (short)
Context-FID Score31.93%36.99%9.37%21.02%
Discriminative Score11.01%32.75%6.55%9.06%

We thank you again for your valuable suggestions and feedback, which have improved the rigor and readability of the paper. All revisions have been incorporated into the updated manuscript and highlighted in blue.

We hope that these revisions adequately address the points you raised. If you have any further concerns about the presentation or suggestions for improvement, we remain fully available for your feedback and be glad to discuss them.

If these revisions meet your expectations, we would greatly appreciate your reconsidering in the score to support our paper.

Comments on the paper

7 comments

Thank you for your careful review and prompt reply. We have revised the manuscript to clarify the raised points.

  • In Section 4.1, D\mathcal{D} was intended as a general measure of change between diffused time series states rather than a separately defined distance between marginal distributions. In our implementation, we instantiate D\mathcal{D} using IAAT as a proxy for changes in temporal property, corresponding to your interpretation (2). Considering that previous description may cause misunderstanding, we have fixed the definition and checked consistent usage throughout the manuscript.

  • We have added the results to the manuscript. The averaged IAAT over 55 calculations are Energy (54.76±0.0254.76\pm0.02), Wind (51.03±0.0151.03\pm0.01), KDDCup (35.02±0.0135.02\pm0.01) and ETTh1 (23.52±0.0223.52\pm0.02). Wind and Energy exhibit substantially higher IAAT values and are classified as the long-memory group, whereas ETTh1 and KDDCup form the short-memory group. We have also corrected typos caused by compiler in Table 9.

Thank you again for your careful and valuable feedback. We hope these clarifications fully resolve your concerns and meet your expectations.

Thank you for your careful review and prompt reply! We have revised Algorithm 1 and Eq. (9) to ensure that these expressions are consistent with our implementation. Below, we provide a detailed clarification of equal IAAT degradation between adjacent steps in the unmixed schedule.

  • Let ψIAAT(τ)\psi_{\mathrm{IAAT}}(\tau) denote the dataset specific IAAT. Using the empirically measured GP prior endpoint, we define the normalized cumulative IAAT degradation as F_X(\tau)=\frac{\psi_{\mathrm{IAAT}}(0)-\psi_{\mathrm{IAAT}}(\tau)}{Z}.$$ Since $F_X$ is monotonic and range within $[0,1]$, we allocate an equal fraction of the total IAAT degradation to each of the $T_s$ steps. Specifically, the unmixed schedule sets $\tau_i=F_X^{-1}(i/T_s)$ and $\alpha_i=1-\tau_i$ for $i=\{0,1,\ldots,T_s\}$. Thus, $F_X(\tau_i)=i/T_s$ and $F_X(\tau_i)-F_X(\tau_{i-1})=1/T_s$. Consequently, every adjacent pair exhibits the same IAAT degradation: $$\begin{aligned} \psi_{\mathrm{IAAT}}(\tau_{i-1})-\psi_{\mathrm{IAAT}}(\tau_i) =Z\bigl(F_X(\tau_i)-F_X(\tau_{i-1})\bigr) =\frac{Z}{T_s}. \end{aligned}$$

The equality holds for the empirically measured trajectory. We sincerely hope this clarification fully addresses your concern.

Thank you for the detailed responses and the revisions to the manuscript. After reviewing the updated paper, I find that the additional experiments and the clarifications regarding the meaning of discrepancy have addressed most of my initial concerns.

Overall assessment:

The paper, in its current form, still lack some clarity anchors, like an operational definition of "traceability". This is mentioned several times but never defined. Furthermore, the authors dropped the original "intuition" relating discrepancy to distributional change in the diffused time series. This favors the paper clarity but leaves open a theoretical connection that was never explored. Despite this, the method is novel and interesting. It expands the knowledge on recent efforts to explore the noise design of diffusion methods applied to time series. The experiments convincingly demonstrate its effectiveness. So, I change my recommendation to weak accept.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.