acceptodds
Under review as a conference paper at ICLR 2027

Harnessing Oscillator Dynamics for High-Frequency Time Series Representation

Abstract

High-quality representation of high-frequency components remains challenging in multivariate time series forecasting (TSF), due to their noisy and non-stationary patterns. Inspired by coupled oscillator dynamics, we introduce a Kuramoto oscillator-based high-frequency refiner (KOR), which maps high-frequency signals into a multivariate oscillator system and extracts structured latent responses from coupling-constrained phase evolution. Its restoring dynamics suppress rapid micro-fluctuations, while the observed phase-locking and slipping regimes admit a local Adler-type interpretation through effective frequency mismatch. To control the residual energy processed by KOR and complement spectral details not fully retained by its macroscopic readout, we further develop KOR-centered frequency modeling framework (KFM), consisting of soft-gated frequency decomposition and spectral fusion. Extensive experiments demonstrate that KFM improves the large majority of evaluated backbone–dataset settings and achieves competitive performance. Code will be open sourced upon acceptance.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, high-frequency representation, Kuramoto oscillator system

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

The paper proposes KOR for representing high-frequency components in multivariate time series forecasting and integrates it into the KFM framework through soft-gated frequency decomposition and spectral fusion. Experiments on nine public datasets evaluate forecasting performance across multiple backbones, alongside module ablations, replacement experiments, robustness analysis, and computational comparisons.

Critical strengths
  1. The experiments cover nine datasets and six forecasting backbones, providing broad evidence of the framework’s effectiveness.
  2. Ablation, module-replacement, and robustness studies examine the contributions of KOR and the other KFM components.
Critical weaknesses
  1. The framework combines oscillator dynamics, frozen random projections, frequency decomposition, and spectral fusion. The specific contribution of Kuramoto dynamics relative to generic fixed random-feature representations is not fully isolated.
  2. The dynamical interpretation is supported by theoretical analysis, but the connection between the proposed phase dynamics and the observed forecasting improvements could be established more directly.
Questions for authors

See the above weakness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer YEqA (Question 1)

Q1. Contribution of Kuramoto dynamics beyond fixed random features

The oscillator dynamics, frozen random projections, frequency decomposition, and spectral fusion named in the question all belong to KFM but play different roles. KOR is the core contribution that adapts Kuramoto dynamics to time-series forecasting. It maps high-frequency input to oscillator excitation, evolves within-channel and across-channel coupled phases, and reads out the driven–free phase difference. The frozen random projections are internal to KOR and serve only to map the input to oscillator excitation and align the phase-difference readout with the output feature dimension. The soft gate in Sec. 3.1 routes input between frequency branches, while the spectral fusion in Sec. 3.3 supplies residual spectral information. The main-text ablation in Table 5 shows the largest overall degradation when KOR is removed. On ETTh1, removing KOR raises MSE by 3.47% relative to the full model, compared with 1.16% for removing soft gating and 0.23% for removing spectral fusion. On ETTm1, removing KOR raises MSE by 6.25%, removing spectral fusion raises it by 2.60%, and removing soft gating slightly lowers it by 0.26%. These ablations distinguish the contributions of the three modules within KFM and identify KOR as the primary contributor to forecasting accuracy.

To test whether fixed random features can replace KOR, we added a stateless Fixed Random Features control to the Appendix I.3 module comparison. It retains the high-frequency input, frozen random projections, output feature dimension, spectral fusion, and forecasting head, while replacing iterative oscillator phase evolution with fixed random nonlinear features. We also compare a double-zero-coupling KOR control. It retains the input drive, independent oscillator phase evolution, and driven–free phase-difference readout, while removing inter-oscillator coupling. Each entry below reports MSE / MAE.

High-frequency moduleETTh1ETTm1WeatherElectricityTrafficIllness
Fixed random features0.451 / 0.4470.412 / 0.4090.256 / 0.2830.199 / 0.2970.613 / 0.3282.033 / 0.940
Double-zero-coupling KOR0.441 / 0.4400.396 / 0.3990.254 / 0.2810.194 / 0.2900.613 / 0.3271.981 / 0.918
Full KOR0.432 / 0.4280.384 / 0.3880.248 / 0.2720.186 / 0.2840.603 / 0.3241.945 / 0.880

Full KOR has lower MSE and MAE than fixed random features on all six datasets. Relative to that control, it reduces MSE by 4.21% on ETTh1 and 6.53% on Electricity. Full KOR also outperforms double-zero-coupling KOR on both metrics across all six datasets. These results show that frozen random features alone do not explain KOR's forecasting gains and that inter-oscillator coupled phase evolution provides an additional contribution.

AuthorsOpenReview ↗

Response to Reviewer YEqA (Question 2)

Q2. Direct link between phase dynamics and forecasting

In KOR, the high-frequency input first becomes the external oscillator drive DD. The local Adler analysis shows how DD changes an oscillator's effective frequency mismatch relative to the coupling field. Under locally quasi-stationary conditions, the relation between mismatch and coupling strength determines local phase locking or slipping. KOR then forms its forecasting representation from the driven–free phase difference, Hmacro=S−1∑ssin⁡(θdriven,(s)−θfree,(s))H^{macro}=S^{-1}\sum_s\sin(\theta^{driven,(s)}-\theta^{free,(s)}). This gives a specific path from input drive through phase response to forecast. We therefore measured changes in the phase readout used by the model alongside changes in forecast error.

On the same Weather test batch, we changed only the projected oscillator drive while keeping the trained model, future targets, and other forecasting branches fixed. Appendix I.1, Table 17 provides the full intervention results. The paired measurements are summarized below. Relative drive change is ∥ΔD∥/∥D∥\|\Delta D\|/\|D\|, total energy change is relative to the original drive, and rHr_H is the mean relative change in the driven–free phase readout.

Drive interventionRelative drive changeTotal energy changePhase-readout change rHr_HMSE (change)
Original drive0.00000.00%0.0000.164 (—)
Prominent high-frequency attenuation0.0946−10.47%0.2320.179 (+9.15%)
Matched weak high-frequency attenuation0.0946−2.68%0.0830.167 (+1.83%)
Matched random high-frequency attenuation0.0946−6.54%0.1370.171 (+4.27%)
Prominent/weak high-frequency energy swap0.69930.00%0.1870.182 (+10.98%)
Spectral inversion (bins 1 to 13)1.27850.00%0.2440.180 (+9.76%)
Gaussian noise, 20 dB0.1000+1.02%0.0320.166 (+1.22%)
Gaussian noise, 10 dB0.3162+10.07%0.0480.167 (+1.83%)

At the same relative drive change of 0.0946, attenuating prominent high-frequency components changes the phase readout by 0.232 and increases MSE by 9.15%. Attenuating weak components yields 0.083 and 1.83%, while random high-frequency attenuation yields 0.137 and 4.27%. Comparably sized 20 dB Gaussian noise yields 0.032 and 1.22%. Because the intervention enters only through the KOR drive and the other branches remain fixed, its effect on the forecast passes through KOR's phase readout. In this batch, stronger forecast sensitivity to the selected prominent drive components accompanies a larger phase-readout response. The interventions match relative drive-change magnitude, but remove different amounts of energy, so this comparison does not isolate phase organization as the sole cause of the error differences.

The coupling mechanism has complementary phase-level evidence. On the same active ETTh1 oscillator edges, the driven locked-edge fraction is 16.10% for full KOR and 7.79% when both coupling strengths are zero, while MSE is 0.432 and 0.441, respectively. Coupling thus changes local phase organization alongside forecasting accuracy. These observations support the proposed local phase mechanism without implying global synchronization or establishing locking as the sole source of the forecasting gain.

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

This paper proposes a Kuramoto Oscillator-based high-frequency Refiner (KOR) for multivariate time series forecasting. KOR maps high-frequency signals into a coupled oscillator system and extracts representations from the phase differences between driven and free oscillator dynamics. The authors provide a theoretical analysis of the system's restoring dynamics and phase-locking/slipping behaviors. Building on KOR, they develop the KFM framework, which integrates soft-gated frequency decomposition and spectral fusion with existing forecasting backbones. Experiments across nine datasets and six backbones demonstrate widespread performance improvements, supported by ablation studies, module replacement experiments, and robustness analyses.

Critical strengths
  1. The paper introduces an approach that leverages Kuramoto oscillator dynamics to model high-frequency components in time series forecasting. In particular, the use of coupled oscillator dynamics and phase-difference responses offers an interesting alternative to conventional neural-network-based high-frequency encoders. It opens up a promising method for incorporating dynamical systems into time series representation learning.

  2. The experimental evaluation is comprehensive, covering multiple datasets and forecasting backbones. The proposed framework achieves improvements in the vast majority of evaluated settings, suggesting good generalizability across different model architectures.

Critical weaknesses
  1. Insufficient theoretical justification for the connection between Kuramoto dynamics and forecasting improvements. Although the paper provides theoretical analyses of coupling-induced restoring dynamics, noise attenuation, and phase-locking/slipping behavior, the connection between these dynamical properties and improved forecasting performance remains insufficiently established. In particular, the analysis explains how certain perturbations can be attenuated under local assumptions, but it does not clearly establish why such attenuation would preserve informative high-frequency structures while suppressing irrelevant noise. Therefore, the theoretical analysis provides useful intuition about the oscillator dynamics but does not fully explain the observed forecasting improvements.

  2. Several key design choices are heuristic, and the necessity of Kuramoto-specific dynamics is not sufficiently justified. The construction of the oscillator system, the use of frozen random projections, and the time-averaged phase-difference readout are primarily motivated by intuition rather than a clear task-specific justification. While the ablation and module replacement experiments demonstrate that KOR is effective compared with conventional neural encoders, they do not sufficiently isolate the benefits of Kuramoto-specific nonlinear coupling from those of more general dynamical filtering mechanisms. For example, a simpler stable dynamical system or leaky integrator may exhibit similar noise-attenuation properties. Additional comparisons with such alternatives would help clarify whether the performance gains arise specifically from Kuramoto dynamics or from more general properties of dynamical representations.

Questions for authors

Q1. Connection between Kuramoto dynamics and forecasting performance. The paper provides theoretical analyses of coupling-induced restoring dynamics, noise attenuation, and phase-locking/slipping behavior. However, the connection between these dynamical properties and the observed forecasting improvements remains somewhat unclear to me. Could the authors further clarify why these properties are expected to preserve predictive high-frequency information rather than simply suppress high-frequency variations? In particular, is there any theoretical or empirical evidence directly linking the proposed dynamical mechanisms to improved forecasting performance?

Q2. Necessity of Kuramoto-specific dynamics. The experimental results demonstrate the effectiveness of KOR compared with conventional neural high-frequency encoders. However, it remains unclear whether the improvements arise specifically from Kuramoto's nonlinear coupling dynamics or from more general properties of dynamical filtering. Could the authors provide further justification for adopting Kuramoto dynamics over simpler alternatives, such as stable linear dynamical systems or leaky integrators? Any additional evidence that helps distinguish these mechanisms would strengthen the methodological contribution.

Q3. Reproducibility of the experimental results. The consistent improvements across diverse datasets and backbones are encouraging, although many of the reported gains are relatively modest (often around 3–4%). While I appreciate the implementation details already provided in the appendix, I would welcome more comprehensive experimental configurations, particularly the hyperparameter settings and selection procedures, to facilitate independent reproduction.

2 comments
AuthorsOpenReview ↗

Response to Reviewer JoHo (Question 1)

We thank the reviewer for these questions.

Q1. From phase dynamics to forecasting

We clarify that KOR's phase representation is intended to respond primarily to prominent, high-energy high-frequency structure in the projected excitation, while reducing low-energy high-frequency perturbations in inter-oscillator disagreement modes and approximately Gaussian noise, rather than simply suppressing high-frequency variation. In the paper, Eq. (15) describes the driven–free phase difference as the joint result of the external drive and coupling-induced restoring force. Eq. (17) shows that a stable mode's response scales with drive amplitude while its response to faster fluctuations decreases with frequency. Appendix D.3 bounds its variance under Gaussian white noise. KOR then reads out the finite-time phase difference in Eq. (6). Thus, prominent excitation can still produce a readable response, while the effect of low-energy high-frequency perturbations is limited.

To test whether this phase response relates to forecasting, we intervened on the projected oscillator drive in the same Weather test batch, holding the trained model, future targets, and other forecasting branches fixed. Appendix I.1, Table 17 reports the paired results:

Drive interventionRelative drive changeTotal energy changeHigh-frequency energy changeHigh-frequency energy sharePhase-readout change rHr_HMSE (change)MAE (change)
Original drive0.00000.00%0.00%67.93%0.0000.164 (—)0.208 (—)
Prominent high-frequency attenuation0.0946−10.47%−15.41%64.18%0.2320.179 (+9.15%)0.226 (+8.65%)
Matched weak high-frequency attenuation0.0946−2.68%−3.95%67.04%0.0830.167 (+1.83%)0.213 (+2.40%)
Matched random high-frequency attenuation0.0946−6.54%−9.63%65.68%0.1370.171 (+4.27%)0.217 (+4.33%)
Prominent/weak high-frequency energy swap0.69930.00%0.00%67.93%0.1870.182 (+10.98%)0.224 (+7.69%)
Spectral inversion (bins 1 to 13)1.27850.00%−59.08%27.80%0.2440.180 (+9.76%)0.224 (+7.69%)
Gaussian noise, 20 dB0.1000+1.02%+0.74%67.74%0.0320.166 (+1.22%)0.209 (+0.48%)
Gaussian noise, 10 dB0.3162+10.07%+7.35%66.25%0.0480.167 (+1.83%)0.210 (+0.96%)

At the same relative drive change of 0.0946, attenuating prominent high-frequency components changes the phase readout by 0.232 and increases MSE by 9.15%. Attenuating weak high-frequency components yields only 0.083 and 1.83%. Comparably sized 20 dB Gaussian noise yields 0.032 and 1.22%. The paired changes in phase readout and forecast error directly show that prominent high-frequency drive structure contributes to forecasting, whereas the tested noise has a smaller effect.

AuthorsOpenReview ↗

Response to Reviewer JoHo (Questions 2 3)

Q2. Necessity of Kuramoto dynamics

To test whether the gains arise from generic dynamical filtering, we replaced KOR with a leaky integrator and a stable linear dynamical system while keeping the rest of KFM unchanged. We also added an ablation that removes inter-oscillator coupling to examine its contribution to forecasting.

DatasetFull KORLeaky integratorStable linear systemBoth couplings zero
ETTh10.432 / 0.4280.442 / 0.4390.443 / 0.4410.441 / 0.440
ETTm10.384 / 0.3880.404 / 0.4090.401 / 0.4040.396 / 0.399
Weather0.248 / 0.2720.253 / 0.2790.251 / 0.2810.254 / 0.281
Electricity0.186 / 0.2840.193 / 0.2930.195 / 0.2930.194 / 0.290
Traffic0.603 / 0.3240.611 / 0.3270.613 / 0.3280.613 / 0.327
Illness1.945 / 0.8801.988 / 0.9401.992 / 0.9511.981 / 0.918

Each entry reports MSE / MAE. Full KOR outperforms all three controls on both metrics across the six datasets. On ETTh1, its MSE is 2.26%, 2.48%, and 2.04% lower than that of the leaky integrator, stable linear system, and zero-coupling model, respectively. On Electricity, the corresponding reductions are 3.63%, 4.62%, and 4.12%.

KOR differs through the within-channel and across-channel sinusoidal phase couplings in Eqs. (9)–(10). Prominent high-frequency excitation can alter the local effective frequency mismatch and the resulting locking or slipping behavior, which the driven–free phase-difference readout in Eq. (6) captures over finite time (Appendices C.3–C.4). Coupling-induced restoring dynamics also attenuate low-energy high-frequency perturbations in inter-oscillator disagreement modes and limit noise responses in stable modes (Appendices B.2, B.4, and D.3). The dynamical replacements and coupling ablation together show that generic stable filtering alone does not explain the forecasting gains.

Q3. Reproducibility

Given space constraints, Appendix A.2, Table 6 reports the selected dataset-wise configurations for the representative TSMixer backbone. Each row covers all four prediction horizons for that dataset.

DatasetLook-back LLVirtual channels VVOscillators NNSegment length PPDrive scale γ\gammaBatch size
ETTh196641928532
ETTh296641928632
ETTm1966419212432
ETTm2966412812832
Weather96641288832
Traffic9612819216532
Electricity966419212532
Exchange966412812532
Illness1046412813516

We searched over a discrete grid of candidate values: V,N∈{32,64,128,192}V,N\in\{32,64,128,192\}, P∈{6,8,10,12,14,16}P\in\{6,8,10,12,14,16\}, γ∈{1.0,2.0,…,10.0}\gamma\in\{1.0,2.0,\ldots,10.0\}, and batch size ∈{16,32,64,128}\in\{16,32,64,128\}. Candidate configurations were compared on the validation set. We selected one configuration per dataset and used it for all four prediction horizons, as reported in Table 6. The look-back length LL was fixed by the paired experimental protocol rather than searched.

Reviewer_Vt6iOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper enhances neural network time-series forecasting by passing rapid fluctuations through a network of interconnected rhythmic units, known as a Kuramoto oscillator network. While this setup delivers solid practical results, it remains unclear whether the performance gains come from the units synchronizing with one another or simply from general non-linear filtering and pattern extraction. The strongest contributions are the consistent experimental improvements across benchmarks and the creative use of fixed dynamical systems to pull features from raw data.

Critical strengths
  1. Creative mathematical framework: Using coupled oscillator dynamics as a fixed non-linear feature extractor—particularly by comparing how the system behaves with external input versus running freely—is a fresh, engaging concept.

  2. Convincing component tests: Removing the oscillator module consistently degrades performance, and the method outperforms standard neural alternatives like multilayer perceptrons (standard dense layers) and convolutional neural networks (sliding filter layers).

Critical weaknesses
  1. Familiar core concept: Using Kuramoto oscillator networks for reservoir computing (using fixed physical or dynamical simulations to process data) has been explored in prior literature. The manuscript does not adequately discuss or contrast itself against these existing studies.
  2. Unverified claim about fast signals: The paper argues that the model preserves predictable, rapid fluctuations while filtering out random noise. However, the evaluation only measures total forecasting error across the entire time series, offering no direct proof that high-frequency patterns specifically are being isolated and retained.
Questions for authors
  1. Novelty relative to prior work: Could you clarify what distinguishes your Kuramoto oscillator representation from earlier research on Kuramoto-based reservoir computing? In particular, what is the core contribution beyond using a fixed oscillator network for feature extraction?

  2. True source of performance gains: Is there concrete evidence that the improvements come specifically from the oscillators coupling and synchronizing with one another, rather than from general non-linear feature extraction or smoothing? How should the ablation experiments be interpreted in light of this?

  3. Validating fast-signal retention: The paper claims this architecture preserves useful high-frequency information while suppressing noise, yet the benchmarks focus on overall forecasting error. Do you have experimental evidence showing that the improvements stem directly from better modeling of high-frequency components?

  4. Theoretical derivation versus actual code: Could you clarify how the theoretical synchronization model (the Adler equation) connects to the equations implemented in your code? The model uses simple additive inputs, whereas the mathematical derivation assumes inputs that depend on the oscillator's phase. Under what conditions is this approximation valid?

  5. Simulation time limits: In some experimental setups, the oscillator system runs for only three integration steps. Is three steps long enough for true synchronization or phase-locking to emerge, or is the module functioning more like a brief, static non-linear transformation?

  6. Statistical reporting in Table 15: Could you clarify how the standard deviations across forecasting horizons were calculated in Table 15? When aggregating across the same random seeds, several of the averaged standard deviations appear mathematically inconsistent with the individual horizon values.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Vt6i (Questions 1 2)

We thank the reviewer for the questions.

Q1. Relation to Kuramoto reservoir computing

Although our work and the two cited studies [1, 2] all use Kuramoto oscillators, a fixed oscillator network alone is not our claimed novelty. Our contribution is an input-conditioned, coupled phase representation of historical high-frequency structure for multivariate TSF. The differences are as follows.

  • Role in the model. Chiba et al. [1] use order parameters from a Kuramoto reservoir to approximate a target function, while de Jong et al. [2] train a readout and feed it back to reproduce target dynamics autonomously. KOR instead uses driven–free phase differences as an intermediate high-frequency time series representation.
  • Input and interactions. Chiba et al. [1] apply the same additive input to all oscillators with all-to-all coupling. In their main setup, de Jong et al. [2] apply phase-dependent input derived from the target system with all-to-all coupling. KOR instead maps segmented historical high-frequency residuals into oscillator-specific additive drives and uses frequency-proximity-based coupling within and across virtual channels.

We have clarified these distinctions and added both studies to the revised Related Work section.

References

[1] Hayato Chiba, Koichi Taniguchi, and Takuma Sumi. Reservoir computing with the Kuramoto model. arXiv:2407.16172, 2024.

[2] Thomas Geert de Jong, Hirofumi Notsu, and Kohei Nakajima. Harnessing omnipresent oscillator networks as computational resource. arXiv:2502.04818, 2025.

Q2. Gains from coupling and synchronization

Coupling and alternative explanations: In Table 5 of the revised paper, setting Kintra=Kinter=0K_{intra}=K_{inter}=0 removes oscillator interactions while retaining the same input drive, independent phase updates, nonlinear driven–free phase readout, and forecasting framework. The full model achieves lower MSE and MAE than this control on all six datasets. To test whether generic nonlinear feature extraction or stable filtering could explain the gain seen in the coupling ablation, Appendix I.3, Table 19 compares KOR with fixed random nonlinear features, MLP and CNN encoders, a leaky integrator, and a stable linear dynamical system. Full KOR has lower MSE and MAE than each on all six datasets. For example, ETTh1 MSE is 0.432 with KOR, versus 0.451 with fixed random features, 0.442 with the leaky integrator, and 0.443 with the stable linear system. Table 5 isolates the contribution of oscillator interactions within KOR, while Table 19 tests whether alternative feature encoders or stable filtering can account for the same forecasting gains.

Local synchronization: To examine whether coupling also changes local phase organization, we compare the full and double-zero-coupling models on the same adjacency-defined edges. Free-edge and driven-edge coherence measure the mean cos⁡(θi−θj)\cos(\theta_i-\theta_j) without and with the external drive, respectively, so higher values indicate closer phase alignment on those edges. The driven locked-edge fraction measures how many edges maintain a stable relative phase across the observation window, providing a temporal measure of local locking in addition to coherence.

DatasetModelMSE / MAEFree-edge coherenceDriven-edge coherenceDriven locked edges
ETTh1Full KOR0.432 / 0.4280.620.4916.10%
ETTh1Kintra=Kinter=0K_{intra}=K_{inter}=00.441 / 0.4400.440.367.79%
WeatherFull KOR0.248 / 0.2720.590.5314.69%
WeatherKintra=Kinter=0K_{intra}=K_{inter}=00.254 / 0.2810.470.406.95%
ElectricityFull KOR0.186 / 0.2840.660.6019.13%
ElectricityKintra=Kinter=0K_{intra}=K_{inter}=00.194 / 0.2900.520.488.07%

On all three datasets, coupling improves forecast error while increasing driven-edge coherence and the fraction of locally locked edges. Together, the coupling ablation and matched-edge phase measurements provide empirical support for local inter-oscillator phase organization as a contributor to the forecasting gains. These measurements extend the full-model phase observations in Appendix D.1 and agree with the local restoring-force analysis in Appendix B.2 and stable-mode reduction in Appendix D.2. Local locking reflects prominent components of the projected excitation and does not require global synchronization.

AuthorsOpenReview ↗

Response to Reviewer Vt6i (Questions 3 4)

Q3. Role of High-frequency and noise suppression

We tested whether prominent high-frequency components of the historical input excitation contribute to KOR's phase representation and downstream forecasting, and how the representation and forecast respond to added noise. On the same Weather test batch, we intervened on the oscillator drive spectrum while keeping the trained model, future targets, and other forecasting branches fixed. The merged Table 17 of the revised paper reports both forecasting errors and changes in KOR's driven versus free phase readout under these paired interventions.

Drive interventionRelative change in DDTotal energy changeHigh-frequency energy changeHigh-frequency energy sharerHr_HMSE (change)MAE (change)
Original drive0.00000.00%0.00%67.93%0.0000.164 (—)0.208 (—)
Prominent high-frequency attenuation0.0946−10.47%−15.41%64.18%0.2320.179 (+9.15%)0.226 (+8.65%)
Matched weak high-frequency attenuation0.0946−2.68%−3.95%67.04%0.0830.167 (+1.83%)0.213 (+2.40%)
Matched random high-frequency attenuation0.0946−6.54%−9.63%65.68%0.1370.171 (+4.27%)0.217 (+4.33%)
Prominent/weak high-frequency energy swap0.69930.00%0.00%67.93%0.1870.182 (+10.98%)0.224 (+7.69%)
Spectral inversion (bins 1 to 13)1.27850.00%−59.08%27.80%0.2440.180 (+9.76%)0.224 (+7.69%)
Gaussian noise, 20 dB0.1000+1.02%+0.74%67.74%0.0320.166 (+1.22%)0.209 (+0.48%)
Gaussian noise, 10 dB0.3162+10.07%+7.35%66.25%0.0480.167 (+1.83%)0.210 (+0.96%)

Relative drive change is ∥D′−D∥2/∥D∥2\|D'-D\|_2/\|D\|_2. Energy changes and the error changes in parentheses are relative to the original drive. rHr_H is the relative change in KOR's driven versus free phase readout.

At the same relative drive-change magnitude, attenuating prominent high-frequency components causes a much larger phase-readout change and forecasting degradation than attenuating weak or randomly selected high-frequency components. A comparably sized Gaussian-noise perturbation has a much smaller effect on both measures. Together with the energy-swap and spectral-inversion results, these interventions support that prominent high-frequency structure in the historical projected excitation is reflected in KOR's phase readout and contributes to forecasting, while the tested Gaussian-noise perturbations have smaller effects on both measures. Appendices B.2 and D.2 show that, on locally cohesive trajectories, coupling contracts stable disagreement modes and bounds their noise variance, while Table 17 shows that prominent high-frequency excitation still produces a forecast-relevant phase response.

Q4. Adler equation and the implemented drive

Thank you for pointing out the difference between the derivation and the implementation. In the implemented oscillator update, Eq. (8), the external input is the additive term γDv,n(s)\gamma D_{v,n}^{(s)}. The original derivation instead wrote this input as γDsin⁡(ΘD−θv,n)\gamma D\sin(\Theta_D-\theta_{v,n}), which does not follow from the implemented equation. We have removed that assumption from the revised Sec. 3.2 and Appendices C.2–C.3.

Only the two coupling terms in Eqs. (9)–(10) depend on oscillator phase. Combining them as Keffsin⁡(Φeff−θv,n)K_{eff}\sin(\Phi_{eff}-\theta_{v,n}) and defining ϕ=Φeff−θv,n\phi=\Phi_{eff}-\theta_{v,n} gives, directly from the continuous-time form of the implemented update, ϕ˙=(Φ˙eff−ωv,n−γDv,n(t))−Keffsin⁡ϕ.\dot\phi=\bigl(\dot\Phi_{eff}-\omega_{v,n}-\gamma D_{v,n}(t)\bigr)-K_{eff}\sin\phi. The sine term therefore comes from the implemented couplings, while the additive input enters the instantaneous frequency mismatch. The revised derivation does not approximate the additive input by a phase-dependent force. The resulting Adler locking criterion applies locally when Keff>0K_{eff}>0 and both KeffK_{eff} and the instantaneous frequency mismatch vary slowly over an interval resolved by the discrete updates.

AuthorsOpenReview ↗

Response to Reviewer Vt6i (Questions 5 6)

Q5. Number of integration steps

There may be a misunderstanding about the number of updates. The main experiments do not use only three steps, since the largest segment length is P=16P=16 when L=96L=96 and P=32P=32 when L=336L=336. Thus, Table 1 uses at least 6 oscillator updates and Table 2 at least 11. Tables 3 and 4 follow the corresponding backbone settings, with at least 6 updates for L=96L=96 and 11 for L=336L=336. These settings are clarified beside Figure 4 in the revised paper.

A count of three would result from pairing L=96L=96 with P=32P=32, but these values are not paired in the main experiments. Figure 4 includes larger PP values to show how performance changes as the number of updates decreases. These are not the main experimental settings. Table 10 uses L=336L=336, so even its largest P=64P=64 gives 6 updates, not three. We therefore describe the main-experiment behavior as a finite-time coupled phase response. In Appendix C.4, representative driven–free phase-difference trajectories approach locally stable behavior within the plotted 21 updates, while Appendix D.1 reports increased free-state coherence and stable relative phases for a subset of driven oscillators. These observations support the emergence of local phase organization within finite integration windows.

Q6. Standard deviations

We thank the reviewer for catching the incorrect avg-row standard deviations in Table 15. These were reporting errors, and we have corrected them in Table 16 of the revised paper. For each of the same five seeds, we average its four horizon errors, ar=14∑h=14er,ha_r=\frac14\sum_{h=1}^{4}e_{r,h}, and report mean⁡r(ar)±SD⁡r(ar)\operatorname{mean}_r(a_r)\pm\operatorname{SD}_r(a_r), using the sample standard deviation. The corrected values are 0.002 for Weather/PatchTST MSE, 0.002 for Weather/PatchTST+KFM MAE, 0.002 for Electricity/PatchTST MSE, and 0.003 for Electricity/PatchTST MAE. The aggregation rule is now stated in that table's caption.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.