acceptodds
Under review as a conference paper at ICLR 2027

PhyTM: A Physics-Informed Tri-Modal Transformer-Mamba Framework for Time Series Forecasting

Abstract

With the growing volume of time series data across real-world applications, accurate forecasting has become essential for understanding temporal dynamics and predicting future trends. However, existing approaches often overlook the physical processes underlying observed time series, resulting in limited incorporation of physical dynamics under noisy observations, weak temporal-physical awareness in attention-based components, and insufficient dynamical constraints in State-Space Models (SSMs) that may lead to spectral drift and structural distortion over long forecasting horizons. To address these challenges, PhyTM is proposed as a physics-informed tri-modal forecasting framework that incorporates physical priors into multimodal temporal modeling and improves long-term forecasting stability. First, a Dynamic Mode Decomposition (DMD)-based representation module is introduced to embed time series into a delayed state space and extract spectrally consistent dynamical modes for temporal evolution modeling. Second, a physics-informed dual-stream attention Transformer is designed for textual temporal representations by integrating semantic attention with DMD-derived physical priors to enhance temporal-physical awareness beyond conventional positional encoding. Third, a physics-structured SSM is developed for visual temporal representations to enforce physically consistent state transitions and patch-level evolution while maintaining linear-time efficiency. Finally, extensive experiments on long-term and short-term forecasting benchmarks verify the effectiveness of PhyTM, which achieves a 23.4% Mean Squared Error (MSE) reduction on Exchange at and a 12.7% Overall Weighted Average (OWA) reduction on M4-Others. Code and data are available at https://anonymous.4open.science/r/PhyTM-2026-E33D.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Physics-Informed Learning, Dynamic Mode Decomposition, Multimodal Learning, Transformer, Mamba

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_WXb7OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PhyTM, a physics-informed tri-modal framework for time-series forecasting. It integrates DMD-derived dynamical priors into Transformer attention for textual representations and structured Mamba state transitions for visual representations, alongside a numerical forecasting pathway. The three representations are fused for prediction. Experiments on eight long-term forecasting datasets and the M4 benchmark demonstrate competitive performance, supported by ablation studies, sensitivity analyses, and dynamical interpretability evaluations.

Critical strengths
  1. Dynamically structured modeling: Incorporating DMD-derived spectral priors into both attention and state-space transitions provides a coherent approach to introducing dynamical inductive biases, potentially improving the interpretability of temporal representations.

  2. Broad empirical evaluation: Experiments across eight long-term forecasting datasets and M4 demonstrate competitive results on several benchmarks, with ablation studies providing some evidence for the contributions of individual components.

Critical weaknesses
  1. Limited methodological novelty and theoretical justification: The framework primarily combines established techniques (DMD, Transformer, Mamba, and multimodal representations). The claimed physical consistency is insufficiently justified, as contraction of individual state transitions does not guarantee stability or physical fidelity of the overall forecasting model.

  2. Insufficient empirical validation: The ablations do not adequately isolate the benefits of physical priors from increased architectural complexity. Moreover, substantial underperformance on Electricity and Traffic raises concerns about generalizability, while computational efficiency and statistical significance remain insufficiently evaluated.

Questions for authors
  1. Theoretical validity : What justifies applying DMD-derived temporal dynamics to spatially serialized GAF patches? Can the authors establish that contractive state transitions meaningfully improve physical consistency or forecasting stability?

  2. Attribution of improvements: How can the authors distinguish the benefits of physical priors from increased model capacity? Would parameter-matched baselines and randomized spectral priors yield comparable improvements?

  3. Generalizability and efficiency: Why does PhyTM substantially underperform on Electricity and Traffic? Can the authors provide computational cost comparisons and multi-seed results to substantiate its practical advantages and performance reliability?

9 comments
AuthorsOpenReview ↗

Thank you for the thoughtful review. The response establishes the theoretical basis of DMD-guided state propagation and provides controlled experiments to evaluate its effectiveness and computational efficiency.

We thank the reviewer for the constructive comments. We have conducted additional theoretical analyses and controlled experiments to examine the methodological contribution and forecasting performance of PhyTM. We have also updated README.md in the anonymous repository to describe the additional experiments, implementation scripts, execution instructions, and result locations. The repository contains the original model implementation (PhyTM.zip), additional experimental code (rebuttal.zip), experimental results (results.zip), and datasets (data_long/ and data_short/). The implementation details and complete experimental results are available in the repository.

W1.

PhyTM uses shared DMD-derived dynamical parameters to guide two temporal operators. The same decay rates and oscillation frequencies are incorporated into the relative-time attention bias of Phys-Transformer and the rotation-decay state transitions of Phys-Mamba.

1. Theoretical analysis. Each continuous-time mode is represented by a rotation-decay block. Under bounded input forcing, if ∥Aˉt∥2≤q<1\|\bar A_t\|_2\leq q<1 and ∥Btut∥2≤M\|B_tu_t\|_2\leq M, the latent state satisfies ∥ht∥2≤qt∥h0∥2+M(1−qt)1−q\|h_t\|_2\leq q^t\|h_0\|_2+\frac{M(1-q^t)}{1-q}. This bound characterizes latent-state propagation under the stated assumptions.

2. Spectral-source replacement. We trained different spectral configurations under the same architecture using rebuttal/exp_q2_spectrum.py. Original DMD achieves the lowest MSE (0.368953), MAE (0.390558), and spectral distance (4.014608), outperforming Random Spectrum, FFT-derived Spectrum, Decay-only, Learnable Stable Oscillator, and Mismatched DMD.

3. Dynamical-pattern identification. We evaluated frozen encoders with linear probes using rebuttal/probe_q2.py. Fixed DMD achieves the highest Macro-F1 in three of four trend/fluctuation identification tasks. On ETTh1 Trend, it achieves 40.38%, compared with 33.19% for Random Spectrum.

4. Long-horizon validation. At H=720, Full PhyTM achieves a spectral distance of 4.181, compared with 4.588 for Standard Mamba and 4.731 without DMD priors. The results are reported in the supplementary experimental results.

W2.

1. Dynamical representation capacity. The effective ranks explaining 90% of variance are 4–10 for ETT, 27 for Electricity, and 51 for Traffic (Appendix L.1). DMD reconstruction error is positively correlated with relative forecasting error (r = 0.86, p = 0.006, Appendix L.2). We further trained PhyTM with different modal capacities using rebuttal/exp_q3_modes.py. Increasing K from 8 to 32 reduces Electricity MSE from 0.215 to 0.181 and Traffic MSE from 0.601 to 0.502. Complete results are presented in Appendix B.

2. Additional baseline comparisons. We extended the baseline evaluation with SimpleTM and TQNet across eight datasets and four forecasting horizons. PhyTM achieves MSE/MAE of 0.264/0.308 on ETTm2 and 0.269/0.372 on Exchange. It also achieves the lowest MAE on Weather (0.242) and ETTm1. On M4, PhyTM achieves the best overall MASE (1.462) and OWA (0.863). The complete comparisons are provided in Appendices D and I.

3. Computational efficiency. We profiled PhyTM using rebuttal/profile_q3.py on a trained ETTh1 checkpoint, with batch size 16, 5 warm-up iterations, and 30 timed repetitions.

MetricPhyTMAurora
Total Parameters38.854M210.8M
Trainable Parameters0.256M13.9M
Training Time/Epoch12.50 s45 s
Inference Latency/Batch35.47 ms82.56 ms
Peak Inference Memory209.88 MB1241 MB
Peak Training Memory243.23 MB3072 MB
MSE0.3693140.376
MAE0.3930990.397

PhyTM uses fewer parameters and requires less training time, inference time, and memory than Aurora, while achieving lower MSE and MAE. The complete comparison is presented in Appendix C.

Q1.

(1) GAF temporal correspondence. The GAF matrix encodes pairwise temporal relationships. Each 16 × 16 patch represents a local block of these relationships rather than a consecutive segment of raw observations. The raster-scan order preserves the structured ordering of temporal-block pairs, providing an ordered sequence for state propagation. Here, Δ_p=16 denotes the latent patch-transition scale, not the original sampling interval.

(2) Patch-order validation. We trained Phys-Mamba with original and randomized patch orders using rebuttal/exp_q1q2.py --group order. Randomizing patch order increases MSE from 0.368 to 0.392 on ETTh1, from 0.171 to 0.185 on ETTm2, from 0.163 to 0.176 on Weather, and from 0.082 to 0.090 on Exchange. These results are reported in Appendix A.

(3) State stability and forecasting. We evaluated state-transition stability using rebuttal/exp_q1q2.py --group stability. Negative Decay achieves a spectral radius of 0.982 and an N-step state norm of 0.614. At H=720, Full PhyTM achieves a spectral distance of 4.181, compared with 4.588 for Standard Mamba and 4.731 without DMD priors. The corresponding results are provided in the supplementary experimental results.

Q2.

(1) Parameter control. The parameter-matched comparison shows that Full PhyTM achieves lower MSE under identical total and trainable parameter counts. The experiment is implemented in rebuttal/exp_q1q2.py --group capacity.

(2) Spectral-source comparison. Original DMD achieves the lowest MSE (0.368953), MAE (0.390558), and spectral distance (4.014608) among the six evaluated spectral configurations. The experiment is implemented in rebuttal/exp_q2_spectrum.py.

(3) Dynamical identification. Fixed DMD achieves the highest Macro-F1 in three of four trend/fluctuation identification tasks. The frozen-encoder probes are implemented in rebuttal/probe_q2.py.

These comparisons show that Full PhyTM improves forecasting under matched parameter counts and that Original DMD outperforms alternative spectral configurations.

Q3.

(1) Dynamical representation capacity. We trained PhyTM with different DMD mode numbers using rebuttal/exp_q3_modes.py. Increasing K from 8 to 32 reduces Electricity MSE from 0.215 to 0.181 and Traffic MSE from 0.601 to 0.502. These results show improved forecasting with increased modal capacity.

(2) Forecasting performance. The additional baseline comparisons show that PhyTM achieves the lowest MSE and MAE on ETTm2 and Exchange, the lowest MAE on Weather and ETTm1, and the best overall MASE and OWA on M4. The complete comparisons are provided in Appendices D and I.

(3) Computational efficiency. We measured PhyTM using rebuttal/profile_q3.py. Compared with Aurora, PhyTM reduces total parameters from 210.8M to 38.854M, training time from 45 s to 12.50 s per epoch, and inference latency from 82.56 ms to 35.47 ms per batch. Peak inference memory decreases from 1241 MB to 209.88 MB, while peak training memory decreases from 3072 MB to 243.23 MB. PhyTM also achieves lower MSE/MAE of 0.369314/0.393099, compared with 0.376/0.397. The complete comparison is presented in Appendix C.

AuthorsOpenReview ↗

Theoretical Analysis and Empirical Validation of PhyTM

We sincerely thank the reviewer for the thoughtful review and constructive suggestions regarding the theoretical justification and empirical effectiveness of PhyTM. In response, additional theoretical analyses and controlled experiments have been conducted. The main improvements are summarized as follows.

First, validation of the contribution of DMD-derived dynamics. To determine whether forecasting improvements originate from DMD-derived information rather than generic spectral parameterization or increased model capacity, six spectral configurations and parameter-matched experiments were conducted. Original DMD achieves the lowest MSE of 0.368953. Under matched model capacity, Full PhyTM reduces MSE from 0.405 to 0.368 on ETTh1. These results demonstrate the independent contribution of DMD-derived dynamical priors. These experiments address the concern raised in Q2.

Second, validation of the theoretical basis of the proposed mechanisms. To examine the stability of DMD-guided state propagation, a bounded-input state-propagation analysis was added under the stated assumptions. Negative Decay achieves a spectral radius of 0.982. Additional patch-order experiments show that randomizing GAF patch ordering increases ETTh1 MSE from 0.368 to 0.392. These results support the stability of the proposed state transitions and the importance of structured patch ordering. These additions address the concern raised in Q1.

Third, evaluation of forecasting performance and generalization. To examine the effectiveness of PhyTM on datasets with different dynamical complexity, modal-capacity experiments were conducted. Increasing K from 8 to 32 reduces MSE from 0.215 to 0.181 on Electricity. Additional comparisons with SimpleTM and TQNet were also provided. Computational profiling shows that PhyTM reduces inference latency from 82.56 ms to 35.47 ms per batch compared with Aurora. These results demonstrate the forecasting effectiveness and computational efficiency of the proposed framework. These results address the concern raised in Q3.

Finally, the manuscript and anonymous repository were updated with additional implementation details, experimental scripts, results, datasets, and execution instructions.

Overall, these additional analyses and experiments provide further evidence for the theoretical motivation, methodological contribution, and practical effectiveness of PhyTM. We hope these results address your concerns and would greatly appreciate your further feedback and reconsideration of the updated evidence.

Thank you for your time and consideration.

Reviewer_k5LHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PhyTM, a physics-informed tri-modal Transformer-Mamba framework for time series forecasting, which aims to leverage dynamical information extracted from observed time series to guide multimodal representation learning. The model employs delay embedding and Dynamic Mode Decomposition (DMD) to extract decay rates and oscillation frequencies, which are incorporated as shared priors into the textual and visual branches. Specifically, the textual branch introduces physics-informed biases into Transformer attention, while the visual branch uses DMD-derived parameters to construct state transition matrices for Mamba. The numerical branch directly models the original time series, and the three representations are subsequently fused for forecasting. Experiments cover multiple long-term and short-term forecasting benchmarks, supported by ablation studies and dynamical visualizations. The paper reports competitive forecasting results, with particularly notable performance on the Exchange and M4 datasets.

Critical strengths
  1. One notable strength of this paper is that the dynamical information extracted by DMD is not simply concatenated as additional features. Instead, it is represented through decay rates and oscillation frequencies, which are incorporated into the attention bias of the textual path and the Mamba state transitions of the visual path, respectively. This allows the same set of dynamical parameters to play a structural role across different representation spaces. By directly embedding data-driven dynamical priors into the internal operators of the model, this design provides a more targeted approach than conventional frequency-domain feature fusion and establishes a degree of consistency across the proposed components.
  2. Another strength is the broad experimental coverage, which goes beyond reporting final forecasting accuracy. The paper includes long-term forecasting, M4 short-term forecasting, ablation studies of physics-informed components and the tri-path structure, hyperparameter sensitivity analysis, and dynamical visualizations. In particular, the temporal- and spectral-domain analyses of decay and oscillation modes provide some interpretability support for the proposed dynamical modeling assumptions. Furthermore, PhyTM achieves competitive results across multiple datasets and forecasting settings.
Critical weaknesses
  1. Insufficient validation of the independent contribution of DMD-derived physical priors. The paper uses DMD-derived decay rates and oscillation frequencies as physics-informed priors but lacks controlled comparisons with Fourier frequencies, random spectral parameters, and freely learnable parameters under the same architecture. Consequently, it remains unclear whether the performance improvements stem from DMD-derived dynamical priors or from general spectral parameterization and structured temporal biases. This weakens the evidence supporting the paper's core methodological contribution.
  2. Insufficient evidence for consistent competitiveness across datasets. According to the reported experimental results, PhyTM noticeably underperforms some competing methods on Electricity and Traffic. The paper attributes these performance gaps to higher dynamical complexity but provides insufficient experimental evidence to substantiate this explanation. Furthermore, the existing baseline comparisons do not sufficiently establish the overall competitive advantage of the proposed method.
Questions for authors
  1. The paper refers to the decay rates and oscillation frequencies extracted by DMD as physics-informed priors, but does not provide comparisons with random spectral parameters, Fourier frequencies, or freely learnable (α,ω)(\alpha, \omega) under the same model architecture. How can the authors demonstrate that the main performance gains specifically originate from the dynamical priors provided by DMD rather than from general spectral parameterization?
  2. The method enforces negative decay rates for all modes through α~k=−exp⁡(⋅)\tilde{\alpha}_k = -\exp(\cdot), thereby ensuring stable decay. Why is this constraint appropriate for all datasets and forecasting scenarios? How would allowing α~k\tilde{\alpha}_k to be zero or positive affect forecasting performance and the learned dynamical patterns?
  3. The paper employs shared DMD-derived dynamical priors to guide the textual and visual paths, while the numerical path processes the raw observations without direct physical constraints. How do the authors define and verify physical consistency in the final fused representation, given that the three paths are combined through feature concatenation and an MLP prediction head?
  4. The textual, GAF-based visual, and numerical representations are all derived from the same time series, while removing a path in the ablation studies also reduces model capacity. Have the authors conducted parameter-matched single-path or dual-path comparisons to establish whether the performance improvements arise from complementary representations rather than increased model capacity?
  5. PhyTM combines GPT-2, GAF transformation, Transformer, and Mamba components, potentially introducing substantial computational overhead. Can the authors report the total and trainable parameter counts, training time, inference latency, and memory consumption of the complete model, and clarify whether the additional computational costs are justified by the forecasting improvements?
  6. The main comparison omits several recently published long-term forecasting baselines, such as SimpleTM, TimeMixer++, TimeFilter, TQNet, and TimeBase. Given that PhyTM already underperforms some existing methods on Electricity and Traffic, how can the current comparisons substantiate its overall competitiveness? Does PhyTM retain a significant advantage over these more recent methods?
  7. The paper directly applies DMD spectral parameters estimated from the original time series to state transitions between GAF image patches. However, GAF images encode pairwise temporal relationships, and the patch scanning order does not directly correspond to the physical temporal evolution of the original sequence. Without an explicit temporal correspondence, how do the authors justify the physical validity of these state transitions rather than treating them as merely mathematically stable recurrences?
  8. The paper refers to the visual encoder as Phys-Mamba, but Eqs.~(9)--(12) only describe DMD-parameterized state transitions and linear recurrences, without explicitly showing Mamba's input-dependent selective mechanism. Does this module actually retain the core selectivity of Mamba, and what fundamentally distinguishes it from a conventional structured state-space model?
9 comments
AuthorsOpenReview ↗

Thank you for the thoughtful review. The response clarifies how DMD priors guide the multimodal architecture and provides additional experiments to verify their contribution to forecasting performance.

We thank the reviewer for the constructive comments. We have conducted additional experiments and updated the anonymous repository with the corresponding implementations and results. Specifically, README.md has been updated to describe the additional experiments, implementation scripts, execution instructions, and result locations. The repository contains the original model implementation (PhyTM.zip), additional experimental code (rebuttal.zip), experimental results (results.zip), and datasets (data_long/ and data_short/). The implementation details and complete experimental results are available in the repository.

W1.

PhyTM uses shared DMD-derived decay rates and oscillation frequencies to guide temporal attention in Phys-Transformer and state transitions in Phys-Mamba.

1. Spectral-source replacement. We trained six spectral configurations under the same architecture. Original DMD achieves the lowest MSE (0.368953), MAE (0.390558), and spectral distance (4.014608).

Spectral PriorMSE ↓MAE ↓Spectral Distance ↓
Random Spectrum0.3829900.4047384.017384
FFT-derived Spectrum0.3797180.3990464.026815
Decay-only0.3755940.3940314.022170
Learnable Stable Oscillator0.3738210.3931044.020774
Mismatched DMD0.3864120.4072254.041963
Original DMD0.3689530.3905584.014608

The experiments were implemented using rebuttal/exp_q2_spectrum.py and rebuttal/export_q2_spectrum.py.

2. Dynamical-pattern identification. We evaluated frozen encoders using linear probes. Fixed DMD achieves the highest Macro-F1 in three of four trend/fluctuation identification tasks, reaching 40.38% on ETTh1 Trend compared with 33.19% for Random Spectrum. The experiments use rebuttal/exp_q2_spectrum.py and rebuttal/probe_q2.py.

3. Parameter-matched comparison. Both the non-physical model and Full PhyTM contain 38.854M total parameters and 0.256M trainable parameters. Full PhyTM reduces MSE from 0.405 to 0.368 on ETTh1, from 0.188 to 0.171 on ETTm2, and from 0.175 to 0.163 on Weather. The experiment was executed using rebuttal/exp_q1q2.py --group capacity.

The spectral replacement and parameter-matched experiments show lower forecasting errors with DMD-derived dynamics under controlled model capacity. The corresponding analyses are presented in Section 4.5.

W2.

(1) Dynamical representation capacity. The effective ranks explaining 90% of variance are 4–10 for ETT, 27 for Electricity, and 51 for Traffic (Appendix L.1). DMD reconstruction error is positively correlated with relative forecasting error (r = 0.86, p = 0.006, Appendix L.2). Increasing K from 8 to 32 reduces Electricity MSE from 0.215 to 0.181 and Traffic MSE from 0.601 to 0.502. The experiment was executed using rebuttal/exp_q3_modes.py. Complete results are presented in Appendix B.

(2) Additional baseline comparisons. We extended the baseline evaluation with SimpleTM and TQNet across eight datasets and four forecasting horizons. PhyTM achieves MSE/MAE of 0.264/0.308 on ETTm2 and 0.269/0.372 on Exchange. It also achieves the lowest MAE on Weather (0.242). On M4, PhyTM obtains the best overall MASE (1.462) and OWA (0.863), reducing overall MASE by 8.3% compared with the best baseline. The complete comparisons are provided in Appendices D and I.

(3) Computational efficiency. We profiled PhyTM using rebuttal/profile_q3.py on a trained ETTh1 checkpoint, with batch size 16, 5 warm-up iterations, and 30 timed repetitions.

MetricPhyTMAurora
Total Parameters38.854M210.8M
Trainable Parameters0.256M13.9M
Training Time/Epoch12.50 s45 s
Inference Latency/Batch35.47 ms82.56 ms
Peak Inference Memory209.88 MB1241 MB
Peak Training Memory243.23 MB3072 MB
MSE0.3693140.376
MAE0.3930990.397

PhyTM uses fewer parameters and requires less training time, inference time, and memory than Aurora, while achieving lower MSE and MAE. The complete comparison is presented in Appendix C.

Q1.

We replaced the original spectrum with Random Spectrum, FFT-derived Spectrum, Decay-only, Learnable Stable Oscillator, and Mismatched DMD under the same architecture. Original DMD achieves the lowest MSE (0.368953), MAE (0.390558), and spectral distance (4.014608). Fixed DMD also achieves the highest Macro-F1 in three of four dynamical-pattern identification tasks. The spectral replacement experiment was implemented using rebuttal/exp_q2_spectrum.py, while the frozen-encoder identification experiment used rebuttal/probe_q2.py. The complete results are available in the anonymous repository.

Q2.

The negative-decay parameterization constrains the unforced rotation-decay transition. After discretization, its transition eigenvalues have magnitudes below one. Zero decay corresponds to neutral propagation, while positive effective growth can amplify the unforced state. We evaluated these configurations using rebuttal/exp_q1q2.py --group stability. Negative Decay achieves a spectral radius of 0.982 and an N-step state norm of 0.614, compared with 1.036 and 1.287 for Signed Growth/Decay. At H=720, Full PhyTM achieves a spectral distance of 4.181, compared with 4.588 for Standard Mamba and 4.731 without DMD priors. The corresponding results are provided in the supplementary experimental results.

Q3.

The three modalities contribute complementary information to the final forecasting representation. The textual and visual paths share DMD-derived dynamical priors, while the numerical path preserves the original observations. The tri-modal ablation shows that removing the textual path increases ETTh1 MSE from 0.368 to 0.420, removing the visual path increases it to 0.418, and removing the numerical path increases it to 0.415. These ablation results are presented in Appendix J.2. The corresponding configurations are implemented in run_benchmark.py using no_text, no_image, and no_num.

Q4.

We conducted two evaluations to examine representation contributions and model capacity. First, the tri-modal ablation examines the contribution of each representation. Removing the textual, visual, and numerical paths increases ETTh1 MSE from 0.368 to 0.420, 0.418, and 0.415, respectively (Appendix J.2). Second, the parameter-matched experiment controls total and trainable parameters at 38.854M and 0.256M. Full PhyTM reduces MSE from 0.405 to 0.368 on ETTh1, from 0.188 to 0.171 on ETTm2, and from 0.175 to 0.163 on Weather. The experiment was executed using rebuttal/exp_q1q2.py --group capacity.

Q5.

We measured the complete PhyTM model using rebuttal/profile_q3.py on a trained ETTh1 checkpoint, with batch size 16, 5 warm-up iterations, and 30 timed repetitions. PhyTM contains 38.854M total parameters and 0.256M trainable parameters, compared with 210.8M and 13.9M for Aurora. Training time is 12.50 s per epoch compared with 45 s, while inference latency is 35.47 ms per batch compared with 82.56 ms. Peak inference and training memory are 209.88 MB and 243.23 MB, compared with 1241 MB and 3072 MB. PhyTM also achieves lower MSE/MAE of 0.369314/0.393099, compared with 0.376/0.397 for Aurora. The complete comparison is provided in Appendix C.

Q6.

We extended the baseline comparison with SimpleTM and TQNet across eight long-term forecasting datasets and four prediction horizons. PhyTM achieves the lowest MSE and MAE on ETTm2 (0.264/0.308) and Exchange (0.269/0.372). It also achieves the lowest MAE on Weather (0.242) and ETTm1. On M4, PhyTM obtains the best overall MASE (1.462) and OWA (0.863). The complete comparisons are provided in Appendices D and I.

Q7.

The GAF matrix encodes pairwise temporal relationships. Each 16 × 16 patch represents a local block of these relationships rather than a consecutive segment of raw observations. The raster-scan order preserves the structured ordering of temporal-block pairs, providing an ordered sequence for state propagation. Here, Δ_p=16 denotes the latent patch-transition scale rather than the original sampling interval. We trained Phys-Mamba with original and randomized patch orders using rebuttal/exp_q1q2.py --group order. Randomizing patch order increases MSE from 0.368 to 0.392 on ETTh1, from 0.171 to 0.185 on ETTm2, from 0.163 to 0.176 on Weather, and from 0.082 to 0.090 on Exchange. These results are presented in Appendix A.

Q8.

Phys-Mamba retains the input-dependent selective operations of standard Mamba. Δ_t is generated through an input-dependent projection followed by Softplus. B_t and C_t are also generated from input-dependent projections, with rotation modulation through cos(ω_kΔ_t) and sin(ω_kΔ_t). Standard output gating is preserved. The DMD-derived decay component constrains the state matrix, while oscillatory information is incorporated through the time-varying B_t and C_t. The implementation is provided in models/image_modal.py and described in Section 3.3.2 and Eq. (12).

AuthorsOpenReview ↗

Validating DMD Priors and Tri-Modal Contributions in PhyTM

We sincerely thank the reviewer for the comprehensive review and constructive questions regarding the methodological contribution and empirical effectiveness of PhyTM. In response, additional controlled experiments and analyses have been conducted. The main improvements are summarized as follows.

First, validation of the contribution of DMD-derived dynamical priors. To distinguish the contribution of DMD from generic spectral parameterization and increased model capacity, six spectral configurations and parameter-matched experiments were conducted. Original DMD achieves the lowest MSE of 0.368953. Under matched model capacity, Full PhyTM reduces MSE from 0.405 to 0.368 on ETTh1. These results demonstrate the independent contribution of DMD-derived information beyond general spectral parameterization and model capacity. These experiments address the concern raised in Q1 and Q4.

Second, validation of the proposed modeling mechanisms. To examine the role of stability constraints and the contributions of different modalities, negative-decay comparisons and tri-modal ablations were conducted. Negative Decay achieves a spectral radius of 0.982. Removing the textual, visual, or numerical branch increases ETTh1 MSE from 0.368 to 0.420, 0.418, and 0.415, respectively. These results demonstrate the importance of stable state transitions and the complementary contributions of the three branches. The selective operations retained in Phys-Mamba were also clarified. These analyses address the concerns raised in Q2, Q3, Q7, and Q8.

Third, evaluation of forecasting performance and computational efficiency. To examine the competitiveness of PhyTM across different datasets, additional comparisons with SimpleTM and TQNet were conducted. On M4, PhyTM achieves an overall MASE of 1.462 and OWA of 0.863. Computational profiling against Aurora further shows that PhyTM reduces total parameters from 210.8M to 38.854M and inference latency from 82.56 ms to 35.47 ms per batch. These results demonstrate the forecasting competitiveness and computational efficiency of the proposed framework. These results address the concerns raised in Q5 and Q6.

Finally, the manuscript and anonymous repository were updated with additional implementation details, experimental scripts, results, datasets, and execution instructions.

Overall, these additional experiments provide further evidence for the contribution of DMD-derived dynamics and the complementary roles of the three modalities. We hope these results address your concerns and would greatly appreciate your further feedback and reconsideration of the updated evidence.

Thank you for your time and consideration.

Reviewer_XwgwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work presents a long‑horizon time‑series forecasting architecture integrating DMD‑estimated decay‑frequency priors into temporal attention and state‑space recurrence, instead of treating DMD merely as auxiliary features. It adopts serialized GAF patches for temporal modeling. Evaluations cover eight datasets including M4 with extensive ablations. The model yields competitive gains on partial benchmarks, yet the physical interpretation of DMD‑driven dynamics and practical reproducibility remain insufficiently validated.

Critical strengths

S1. Coherent architectural mechanism (Secs. 3.2–3.4). The same DMD‑derived decay/frequency parameters guide temporal attention and the visual state‑space recurrence instead of appearing only as appended features. This is a technically understandable integration strategy and its explicit parameterization (Eqs. 7 and 9–11) makes the central modeling hypothesis testable. It is the strongest reason to consider the paper for acceptance.

S2. Broad benchmarking with nontrivial positive results (Secs. 4.1–4.3; Tables 1–3, 7–10). The evaluation includes eight long‑horizon datasets, M4, hyperparameter studies, physics‑component ablations, and modality ablations. Exchange improves from 0.321 to 0.269 average MSE against the strongest listed comparator; M4‑Others OWA improves from 0.723 to 0.631. These effects suggest the architecture can be practically useful in some regimes, though they do not alone establish the proposed explanation.

Critical weaknesses

W1. The core physical interpretation is not causally isolated or adequately validated (Secs. 3.2–3.5; Appendix C.5; Tables 3 and 9). DMD provides a global linear spectral estimate from observations, not a verified governing law. Positive raw growth rates are reflected into negative rates before learning, so the stable prior can depart from the measured evolution. The visual state transition operates over serialized 2‑D GAF patches, for which Δp is explicitly not the original physical time interval; the physical meaning of patch‑to‑patch evolution is therefore unclear. Further, Eq. 11 guarantees only unforced block contraction, not reduced forecast drift. Existing ablations switch off entire modules but do not compare against parameter‑matched learned Fourier/relative‑position biases, random/shuffled DMD spectra, or generic learnable stable oscillator blocks. Consequently, the results cannot determine whether data‑derived physical structure—rather than a useful regularizer—is responsible for the gains. This weakens the paper’s primary scientific claim.

W2. Empirical support is uneven and the cost/reproducibility trade‑off is unresolved (Tables 1–2; Sec. 4; Appendices B–D). The model is substantially worse than the strongest listed comparator on Traffic (0.601 vs. 0.428 average MSE, approximately 40.4% higher) and Electricity (0.215 vs. 0.178, approximately 20.8% higher); it also does not lead M4 sMAPE (12.230 vs. 11.991). No repeated‑run dispersion or significance tests are provided. The full pipeline encodes each channel through GPT‑2, constructs an L×L GAF, and applies L×L attention, so the claim of linear‑time efficiency cannot be inferred for the end‑to‑end system from a linear state‑space update alone. End‑to‑end runtime, GPU memory, FLOPs, parameter count, and scaling with lookback length and channel count are absent. The global DMD fitting protocol for M4’s independent series, the precise truncation/selection rule, and several patch/SSM settings are also insufficiently detailed for reproduction. These omissions make the practical superiority and the generality of the conclusion uncertain.

Questions for authors

Q1 (method/protocol; assessment‑critical): Precisely how are the DMD reference spectra computed for each dataset, especially M4’s 100,000 independent series of different frequencies and lengths? Please specify truncation rank r, the rule for choosing K = 8 (including conjugate pairs), training‑only segmentation, handling of series boundaries, zero decay rates, epsilon_decay, and the exact leakage‑prevention protocol. Would per‑series/window DMD materially change the results?

Q2 (mechanism; assessment‑critical): How does Eq. 12 relate to an actual selective Mamba layer? Are the input‑dependent Δ, B, C, and gating operations retained? What are the GAF patch dimensions, serialization order, and Δp values, and how can adjacent 2‑D patches be interpreted as physical‑time transitions? Can the authors provide direct comparisons of forecast spectral drift or dynamical consistency, not just latent contraction?

9 comments
AuthorsOpenReview ↗

Thank you for the thoughtful review. The response validates the effectiveness of DMD-derived dynamics through controlled spectral comparisons and further examines PhyTM's forecasting performance across datasets.

We thank the reviewer for the constructive comments. We have conducted additional experiments and updated the anonymous repository with the corresponding implementations and results. Specifically, README.md has been updated to describe the additional experiments, implementation scripts, execution instructions, and result locations. The repository contains the original model implementation (PhyTM.zip), additional experimental code (rebuttal.zip), experimental results (results.zip), and datasets (data_long/ and data_short/). The implementation details and complete experimental results are available in the repository.

W1.

PhyTM uses DMD-derived decay rates and oscillation frequencies as shared structural priors to guide temporal attention in Phys-Transformer and state transitions in Phys-Mamba. We conducted additional experiments on dynamical-pattern identification, spectral replacement, parameter-matched comparison, state stability, and long-horizon forecasting.

ExperimentKey ResultsImplementation
Dynamics identificationFixed DMD achieves 40.38% vs. 33.19% on ETTh1 Trend and 37.08% vs. 35.71% on Exchange Fluctuation.probe_q2.py
Spectral replacementOriginal DMD achieves MSE 0.368953, MAE 0.390558, and spectral distance 4.014608.exp_q2_spectrum.py
Parameter-matched comparisonETTh1: 0.368 vs. 0.405. ETTm2: 0.171 vs. 0.188. Weather: 0.163 vs. 0.175.exp_q1q2.py
State stabilityNegative Decay achieves spectral radius 0.982 and N-step state norm 0.614.exp_q1q2.py
Long-horizon consistencyAt H=720: PhyTM 4.181, Standard Mamba 4.588, Without DMD 4.731.Existing experimental records

The parameter-matched models contain identical total parameters (38.854M) and trainable parameters (0.256M). Full PhyTM achieves lower forecasting errors under the same parameter budget. The spectral replacement results also show that Original DMD outperforms the alternative spectral configurations. The corresponding analyses are presented in Section 4.5 and the supplementary experimental results.

W2.

(1) Dynamical representation capacity. The effective ranks explaining 90% of variance are 4–10 for ETT, 27 for Electricity, and 51 for Traffic (Appendix L.1). DMD reconstruction error is positively correlated with relative forecasting error (r = 0.86, p = 0.006, Appendix L.2). We further trained PhyTM with different numbers of DMD modes using rebuttal/exp_q3_modes.py. Increasing K from 8 to 32 reduces Electricity MSE from 0.215 to 0.181 and Traffic MSE from 0.601 to 0.502. This comparison shows that increasing modal capacity improves forecasting on datasets with more complex dynamics. Complete results are provided in Appendix B.

(2) Additional baseline comparisons. We extended the baseline evaluation with SimpleTM and TQNet across eight datasets and four forecasting horizons. On ETTm2, PhyTM achieves MSE/MAE of 0.264/0.308, compared with 0.2750/0.3223 for SimpleTM and 0.2772/0.3225 for TQNet. On Exchange, PhyTM achieves 0.269/0.372, outperforming both methods. It also achieves the lowest MAE on Weather (0.242). On M4, PhyTM obtains the best overall MASE (1.462) and OWA (0.863), reducing overall MASE by 8.3% compared with the best baseline. Complete comparisons are provided in Appendices D and I.

(3) Computational efficiency. We profiled PhyTM using rebuttal/profile_q3.py on a trained ETTh1 checkpoint, with batch size 16, 5 warm-up iterations, and 30 timed repetitions.

MetricPhyTMAurora
Total Parameters38.854M210.8M
Trainable Parameters0.256M13.9M
Training Time/Epoch12.50 s45 s
Inference Latency/Batch35.47 ms82.56 ms
Peak Inference Memory209.88 MB1241 MB
Peak Training Memory243.23 MB3072 MB
MSE0.3693140.376
MAE0.3930990.397

PhyTM uses fewer parameters and requires less training time, inference time, and memory than Aurora, while achieving lower MSE and MAE. The complete comparison is presented in Appendix C.

Q1.

DMD reference spectra are estimated exclusively from training data and remain fixed during validation and testing (Appendix G.5). For M4, the 100,000 independent series are grouped into six sampling-frequency subsets.

ConfigurationSetting
Delay embeddingdemb=8d_{\mathrm{emb}}=8
Truncated SVD rankr=8r=8
Modal selectionK=8K=8 original eigenvalues
Conjugate pairsRetained separately, without merging
State dimension16

Snapshot pairs are constructed within individual series without crossing sequence boundaries or mixing frequencies. Decay rates and frequencies are aggregated using the median across the 512×C512\times C window-channel instances within each frequency subset. The sampling interval is Δt=1\Delta t=1. Eigenvalue magnitudes are lower-bounded by 10−610^{-6}, decay rates are clipped to [−5,5][-5,5], and angular frequencies are restricted to [−π,π][-\pi,\pi]. Positive growth rates are reflected to negative values of equal magnitude, while zero decay is assigned −ϵdecay-\epsilon_{\mathrm{decay}}. The resulting spectra remain fixed during validation and testing. The implementation is provided in models/physics_extractor.py.

Q2.

(1) Selective state-space mechanism. Phys-Mamba retains the input-dependent selective operations of standard Mamba. Δt\Delta_t is generated through an input-dependent projection followed by Softplus. BtB_t and CtC_t are also generated from input-dependent projections, with rotation modulation through cos⁡(ωkΔt)\cos(\omega_k\Delta_t) and sin⁡(ωkΔt)\sin(\omega_k\Delta_t). Standard output gating is preserved. The DMD-derived decay component constrains the state matrix, while oscillatory information is incorporated through the time-varying BtB_t and CtC_t. The implementation is provided in models/image_modal.py and described in Section 3.3.2 and Eq. (12).

(2) GAF construction and patch configuration. For an input sequence of length 96, the GAF transformation produces a 96×9696\times96 image. Non-overlapping 16×1616\times16 patches divide the image into a 6×66\times6 grid, yielding 36 tokens. The patches are serialized in row-major raster-scan order with learnable positional encoding. Here, Δp=16\Delta_p=16 denotes the latent patch-transition scale, while the actual selective step size Δt\Delta_t remains input-dependent. The raster-scan order preserves the structured ordering of temporal-block pairs, providing an ordered sequence for state propagation (Appendix F.2).

(3) Patch-order validation. We trained Standard Mamba and Phys-Mamba using original and randomized patch orders. Randomizing patch order increases MSE for both encoders across all four datasets. For Phys-Mamba, MSE increases from 0.368 to 0.392 on ETTh1, from 0.171 to 0.185 on ETTm2, from 0.163 to 0.176 on Weather, and from 0.082 to 0.090 on Exchange. The experiment was executed using rebuttal/exp_q1q2.py --group order. The results are presented in Appendix A.

AuthorsOpenReview ↗

Validating DMD-Derived Dynamics and Physical Mechanisms in PhyTM

We sincerely thank the reviewer for the detailed comments and constructive questions. In response to your concerns regarding DMD-derived dynamics and the physical interpretation of PhyTM, additional experiments and analyses have been conducted. The main improvements are summarized as follows.

First, validation of the independent contribution of DMD-derived dynamics. To determine whether forecasting improvements originate from DMD-derived information rather than generic spectral parameterization or increased model capacity, six spectral configurations and parameter-matched experiments were conducted. Original DMD achieves the lowest MSE of 0.368953. Under matched model capacity, Full PhyTM reduces MSE from 0.405 to 0.368 on ETTh1 and from 0.188 to 0.171 on ETTm2. These results demonstrate that DMD-derived dynamics provide informative priors beyond generic spectral parameterization and model capacity. These experiments address the concern raised in W1.

Second, validation of the proposed physical mechanisms. To examine the stability of DMD-guided state transitions and their effectiveness in long-term forecasting, negative-decay and long-horizon spectral experiments were conducted. Negative Decay achieves a spectral radius of 0.982, while Full PhyTM achieves a spectral distance of 4.181 at H=720, compared with 4.588 for Standard Mamba. The M4 DMD estimation protocol was also clarified, ensuring that reference spectra are estimated exclusively from training data. The input-dependent selective operations of Mamba are retained in Phys-Mamba. These results support the stability and implementation of the proposed mechanisms. These additions address the questions raised in Q1 and Q2.

Third, evaluation of forecasting performance and generalization. To examine the effectiveness of PhyTM on datasets with different dynamical complexity, modal-capacity experiments were conducted. Increasing K from 8 to 32 reduces MSE from 0.215 to 0.181 on Electricity. Additional comparisons with SimpleTM and TQNet were conducted. Computational profiling shows that PhyTM reduces inference latency from 82.56 ms to 35.47 ms per batch compared with Aurora. These results demonstrate the forecasting effectiveness and computational efficiency of PhyTM. These results address the concerns raised in W2.

Finally, the manuscript and anonymous repository were updated with the corresponding implementation details, experimental scripts, results, and execution instructions.

Overall, these additional experiments provide further evidence for the contribution of DMD-derived dynamics and clarify their role in the proposed modeling mechanisms. We hope these results address your concerns and would greatly appreciate your further feedback and reconsideration of the updated evidence.

Thank you for your time and consideration.

Comments on the paper

7 comments
AuthorsOpenReview ↗

We appreciate the Area Chair’s time and consideration. The rebuttal provides additional experimental evidence for the effectiveness of DMD-derived dynamics and clarifies the theoretical basis of PhyTM while further demonstrating its forecasting performance and computational efficiency.

The rebuttal substantially strengthens the methodological and empirical evidence of PhyTM through additional controlled experiments, extended baseline comparisons, and implementation clarifications. The anonymous repository has also been updated. Specifically, README.md documents the additional experiments and execution instructions. rebuttal.zip contains the experimental scripts, results.zip contains the experimental results, PhyTM.zip provides the original implementation, and data_long/ and data_short/ contain the datasets.

1. Validation of DMD-Derived Dynamics

Controlled spectral replacement and parameter-matched experiments directly examine the contribution of DMD-derived dynamical information. Under identical model sizes of 38.854M total parameters and 0.256M trainable parameters, Full PhyTM reduces MSE from 0.405 to 0.368 on ETTh1, from 0.188 to 0.171 on ETTm2, and from 0.175 to 0.163 on Weather. Among six spectral configurations, Original DMD achieves the lowest MSE (0.368953), MAE (0.390558), and spectral distance (4.014608). Fixed DMD also achieves the highest Macro-F1 in three of four dynamical-pattern identification tasks. These results establish the contribution of data-derived dynamical information beyond model capacity and generic spectral parameterization.

2. Theoretical Validation

The methodological explanations clarify how shared DMD-derived decay rates and oscillation frequencies guide temporal attention in Phys-Transformer and state transitions in Phys-Mamba. Phys-Mamba retains the input-dependent selective operations of standard Mamba, including Δ, B, C, and output gating. The GAF patch construction and raster-scan ordering are also clarified. Randomizing patch order increases MSE across all four datasets. Negative Decay achieves a spectral radius of 0.982 and an N-step state norm of 0.614. At H=720, Full PhyTM achieves a spectral distance of 4.181, compared with 4.588 for Standard Mamba and 4.731 without DMD priors.

3. Empirical Effectiveness

Additional dynamical analysis and modal-capacity experiments examine performance across datasets with different temporal complexity. Increasing K from 8 to 32 reduces Electricity MSE from 0.215 to 0.181 and Traffic MSE from 0.601 to 0.502. Extended baseline comparisons show that PhyTM achieves MSE/MAE of 0.264/0.308 on ETTm2 and 0.269/0.372 on Exchange. On M4, PhyTM obtains the best overall MASE (1.462) and OWA (0.863). Computational profiling against Aurora further demonstrates its efficiency. PhyTM contains 38.854M total parameters compared with 210.8M for Aurora. Training time decreases from 45 s to 12.50 s per epoch, inference latency from 82.56 ms to 35.47 ms per batch, and peak inference memory from 1241 MB to 209.88 MB. PhyTM also achieves lower MSE/MAE of 0.369314/0.393099 compared with 0.376/0.397 for Aurora.

PhyTM establishes a unified framework in which shared DMD-derived dynamics guide both temporal information selection and latent-state propagation. Its methodological contribution is supported by controlled spectral comparisons and parameter-matched experiments. Its forecasting effectiveness is further demonstrated through patch-order validation, modal-capacity analysis, and extended baseline comparisons. The integration of textual, visual, and numerical representations enables complementary information to contribute to forecasting. The additional code and experimental results are available in the updated anonymous repository.

AuthorsOpenReview ↗

Summary of Additional Experimental Validation for PhyTM

We appreciate the Area Chair's time and effort in overseeing the discussion of our submission. Following the reviewers' feedback, we have conducted additional experiments to further examine the effectiveness of the proposed method and clarify several aspects of its design. The main additions are summarized below.

First, validation of DMD-derived dynamics. To determine whether the forecasting improvements originate from DMD-derived information rather than generic spectral parameterization or increased model capacity, six spectral configurations and parameter-matched experiments were conducted. Full PhyTM reduces MSE from 0.405 to 0.368 on ETTh1 and from 0.188 to 0.171 on ETTm2. These results demonstrate the contribution of DMD-derived dynamical priors beyond model capacity.

Second, validation of the proposed modeling mechanisms. To examine how DMD-derived dynamics guide the Transformer and Mamba branches, additional experiments were conducted on state-transition stability and long-horizon spectral consistency. At H=720, Full PhyTM achieves a spectral distance of 4.181, compared with 4.588 for Standard Mamba. Patch-order experiments were also performed to examine the effect of structured GAF serialization.

Third, evaluation of forecasting performance and efficiency. To further evaluate the effectiveness of PhyTM on datasets with different dynamical complexity, modal-capacity experiments were conducted. Increasing K from 8 to 32 reduces MSE from 0.215 to 0.181 on Electricity and from 0.601 to 0.502 on Traffic. Additional baseline comparisons and computational profiling show that PhyTM reduces inference latency from 82.56 ms to 35.47 ms per batch compared with Aurora.

Finally, the manuscript and anonymous repository were updated with additional implementation details, experimental scripts, results, datasets, and execution instructions to facilitate verification and reproducibility.

Overall, these additional results provide further evidence for the contribution of DMD-derived dynamics and the effectiveness of the proposed framework. We sincerely appreciate your time and consideration and would welcome any further feedback during the discussion period.

AuthorsOpenReview ↗

Follow-up on Our Submission

Dear Area Chair, Program Chairs, and Reviewers,

We would like to invite the Area Chair, Program Chairs, and reviewers to further engage in the discussion of our submission. We appreciate your time and consideration and welcome any further questions or suggestions regarding our work.

In response to the concerns raised during the review, additional experiments and analyses have been conducted. First, six spectral configurations and parameter-matched experiments were conducted to validate the contribution of DMD-derived dynamics. Full PhyTM reduces MSE from 0.405 to 0.368 on ETTh1. Second, state-transition stability and long-horizon spectral consistency were further examined. At H=720, Full PhyTM achieves a spectral distance of 4.181, compared with 4.588 for Standard Mamba. Patch-order experiments also examine the effect of structured GAF serialization. Third, modal-capacity experiments show that increasing K from 8 to 32 reduces Traffic MSE from 0.601 to 0.502. Additional baseline comparisons and computational profiling further demonstrate the effectiveness and efficiency of PhyTM.

We hope these additional results provide a clearer understanding of the contributions and effectiveness of PhyTM. We would greatly appreciate further discussion and feedback from the Area Chair, Program Chairs, and reviewers. We would also be happy to provide any additional clarification if needed.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.