acceptodds
Under review as a conference paper at ICLR 2027

IsoKoop: The Cycle Comes From Outside the Window

10 views

Abstract

In long-horizon multivariate forecasting, a look-back window shorter than the cycle it must forecast does not fix that cycle's length: two cycles of different length can agree on the whole window, so the length is absent from it, not hard to fit. Per-window normalisation turns that observation into an exact interface, one that binds published models too: moved in front of its normaliser, FreqCycle's cycle table costs that model up to 210.6% under a gain. A forecaster that standardises each window by its own mean and scale is scale-equivariant for every inner network, and a profile supplied from outside the window keeps that guarantee when subtracted in raw units only if it is a constant level, but at any shape once it enters in the normalised coordinates or is rescaled by the window's scale. The ratio of the dataset cycle to the look-back, fixed by the protocol, says what the window cannot hold, and IsoKoop supplies it. Against eight forecasters retrained in one pipeline on thirteen public benchmarks, four horizons and three seeds, IsoKoop attains the best mean rank in MSE and MAE (2.34 against 3.77 and 2.14 against 3.65) and the lowest MSE in 34 of the 52 settings (2.24 against 3.99 and 38 of 56 with our TrafficC panel). On the 883-variate network it uses a fourteenth of EMAformer's parameters, the runner-up in mean rank. IsoKoop's supply is a rank-R prototype read at the sample index, zero-initialised so that its profile, though not its period, is learned, and an optional level anchor. In our ablations, and in the baselines with no new run, the supply helps most where the cycle exceeds the look-back, on panels that here are also the sensor networks, so ratio and panel type are not separated; on the real shifts we tested, these readings lower the error only on Solar at H=96, less where the scale moved most, and raise it on PEMS04 and PEMS08. The same lens audits the benchmarks: on Exchange no model beats persistence, many variates change scale from training to test, and nine drifting sensors decide Traffic's ranking.

Primary area:
learning on time series and dynamical systems
Keywords:
multivariate time series forecasting, instance normalization, periodicity, Koopman operator

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

This paper propose IsoKoop to combine different modules, including a low-rank phase-indexed prototype, spectral lifting, gated transport, with mean-field coupling, and a robust time-frequency. It aims to improve the multivariate time-series forecasting performance leveraging the interation between external periodic information and per-window normalization. The proposed method is validated against 8 baselines on various datasets.

Critical strengths

Demostrate strong performance against different baselines over several datasets.

Critical weaknesses
  1. The authors claim one of its contributions is a criterion, not a module. I acknowledge the theoretical observation about period non-identifiability and the potential value of explicit periodic information. However, it remains unclear whether the paper has established a sufficiently validated, generalizable criterion for guiding practical model design based on this observation. Stage I contains protoype, level anchor, spectral lifting and structural codes. The performance gains from the state-construction stage or from the full model over a simplified baseline cannot be attributed solely to the explicit provision of periodic information. Also, all P/L>1 data is from PEMS, and the performance gains of the proposed modules could come from other characteristics of the PEMS data rather than the P/L ratio.

  2. Several benchmark improvements are small, and the architecture contains dataset-dependent choices. The comparison does not establish how much of the performance advantage comes from the proposed mechanisms rather than configuration selection.

Questions for authors

Does “window-only” exclude information encoded in parameters learned from the full training split? If not, what precisely do Design Remarks 4–5 imply about a trained forecaster’s need for explicit period or phase inputs, beyond the impossibility of universally identifying periods over the stated function class?

What search spaces and validation budgets were used for IsoKoop and the baselines? Were baseline recipes checked for adequate convergence under the common 20-epoch cap? Please distinguish published settings from any settings modified in the shared implementation.

Do you have controlled results varying L within the same dataset across the relevant P/L regimes while isolating the prototype or phase information? Such evidence would materially strengthen the central interpretation

1 comment

Thank you. We narrowed the claim and added mechanism controls.

  1. What “window-only” and P/L imply. “Window-only” describes inference inputs. Parameters can encode periodic information learned from the full training split. Remarks 4–5 establish ambiguity over the stated periodic-signal class, not a requirement that every trained forecaster receive explicit period or phase inputs. Without bounded candidate periods, any finite window admits longer competing continuations. With periods bounded by P_max, L≥P+P_max is sufficient for identification; L≥2P is the special case P_max=P. Here P is an exact minimal period. In the experiments it is a calendar reference for noisy data. The revised Introduction and Appendix K state these conditions. P/L describes window coverage, not a general threshold for forecasting gains. Period ambiguity also does not imply that phase is always unidentifiable from a short window.

  2. Within-dataset controls, including counterexamples. We trained the calendar profile Q on/off at each L with seeds 2027–2029, keeping the remaining recipe and train/validation/test forecast origins fixed within each dataset and horizon. At H=96, Table 15 gives these relative reductions in three-seed mean MSE:

Dataset (P)L valuesMSE reduction (%)
PEMS04 (288)168/288/576+4.14/+1.52/−0.56
PEMS03 (288)168/288/576−4.57/−8.52/−19.75
PEMS08 (288)168/288/576+8.88/+7.39/+1.95
Solar (144)72/144/288+33.94/−9.82/+5.42
Traffic (168)84/168/336+0.32/−0.11/+1.27
Electricity (168)84/168/336−4.59/−0.84/−1.44

On Solar at L72, reductions are 26.61%/27.97% at H192/336; every seed improves at all three horizons. Thus short-window gains extend beyond PEMS, while the negative and non-monotonic results rule out a universal P/L criterion. Solar retains the anchor and disables shrink; Traffic/Electricity retain calendar inputs in both arms. Q-off can still use learned priors. Varying L changes history and L-dependent projections. At fixed PEMS04 L168/H96, a phase-only code improves on Q-off by 1.48%, and the low-rank profile improves on that code by 2.70%, in every seed (Table 17).

  1. Configuration selection and small gains. A smaller structural preset gives PEMS04 MSE 0.10231 with 5.05M parameters, versus 0.10163 with 19.52M. On Solar it gives 0.20744 versus 0.18742 for a supplemental reference: both arms disable shrink, while the preset also disables the anchor and changes profile rank and width. Each uses its dataset's trainer. Multiple fields change, so this is neither an isolated capacity ablation nor the submitted Solar configuration (Appendix L.2).

  2. Search scope and recipe changes. Search budgets were unequal. IsoKoop's traceable round-6/7 records contain 547 successful L168 runs covering 364 dataset–stage–arm candidates, including stage references; 301 candidates used H96/seed2027 only. Tables 18–19 list widths, trainer, profile/rank/period/gain/router, anchor, coupling, shrink, loss and gate alternatives. Non-capacity selection requires ≥2 shared validation cells, mean relative gain ≥1.4%/√cells and paired t>1; capacity uses single-cell validation ranking with a 1.4% replacement threshold. These are heuristics, not significance tests. Earlier lifting, readout and Weather-loss searches are excluded; test-set profile diagnostics also informed shrinkage exploration.

Limited baseline checks used learning-rate multipliers {0.5,1,2} and width multipliers {1,2}, capped at 1024: 44 challenger runs across EMAformer/SOFTS on Electricity/PEMS04 and FreEformer on PEMS04, plus 12 SOFTS/PEMS04 confirmation runs. Selection used validation MSE at H96/336, seed2027. Some controls used CUDA and challengers ROCm. The submitted baselines use recorded official-repository recipes (PatchTST: Time-Series-Library port), adapted to L168, cap20 and generally instance normalization; iTransformer/FreDF on PEMS08 retain their exception. PEMS horizons differ from published 12/24/48/96 tables. Missing recipes inherit documented neighboring datasets; FreqCycle's PEMS recipe derives from Electricity. Appendix M records these adaptations; not all historical upstream revisions are pinned.

  1. Does cap20 explain the advantage? On PEMS04 L168/H96, independently trained three-seed cap20/30 runs give FreEformer MSE 0.11195/0.11121 and FreqCycle 0.19344/0.19304. IsoKoop at cap20 gives 0.10163, 8.62% below FreEformer-30, and beats both cap30 baselines in every seed (Table 20). The split, train-only scaler and 2,728 test origins are shared. Sizes are 19.52M/14.25M/0.125M parameters, with different recipes and batch sizes. All runs reached their caps. FreEformer's best validation epochs at cap30 were 30/29/30, and its cosine schedule depends on the cap. Equal epochs do not establish equal compute or convergence across the benchmark.

We welcome any further questions.

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

This paper studies long-horizon multivariate forecasting, focusing on cases where the input window is shorter than the underlying cycle. The authors analyze how seasonal information interacts with instance normalization and scale equivariance, and propose IsoKoop, which combines a learned cycle prototype, spectral representations, gated transport, and a time-frequency loss. Experiments on 13 public benchmarks report competitive results against eight forecasting models, supported by ablation studies and additional analyses.

Critical strengths
  1. Strong empirical results. The paper evaluates IsoKoop on 13 public benchmarks against eight forecasting models. The reported results are competitive across different prediction horizons, with particularly strong performance on several high-dimensional datasets.

  2. Useful analysis of normalization and seasonal information. The paper provides a mathematical analysis of how the placement of a seasonal profile affects scale equivariance. This offers a useful perspective on the interaction between normalization and periodic modeling, beyond the proposed architecture itself.

Critical weaknesses
  1. Limited novelty over existing work. Several key components have close precedents. CycleNet already uses learned periodic patterns and residual forecasting, while AsyTO employs a similar low-rank, phase-indexed prototype. The equivariance of normalization wrappers for arbitrary networks has also been studied by Saied and Fleuret (2026). Proposition 1(ii) offers a more specific result concerning the placement of external profiles, but the paper does not clearly establish what substantial methodological advance the overall architecture provides beyond these existing ideas.
  2. Insufficient evidence connecting the theoretical analysis to the empirical improvements. Several theoretical arguments do not directly apply to the implemented model. The reported configuration uses a raw-unit cycle prototype without a general scale-equivariance guarantee, while some datasets use frequency-band filters rather than spectral derivatives. More importantly, the claimed benefit of external cycle information is not sufficiently isolated. All main experiments use the same input length, and the datasets with longer cycles are also PEMS sensor networks. The EMAformer–iTransformer comparison also changes more than phase information. Although the reported forecasting results are strong, it remains unclear whether the proposed theoretical explanations account for these improvements.
Questions for authors
  1. Can the authors evaluate different input window lengths on the same dataset, particularly across the point where L becomes longer than P? This would help determine whether the benefit of external cycle information is related to P/L rather than dataset characteristics.

  2. Section 4.3 uses the comparison between EMAformer and iTransformer to support the benefit of phase information. However, their released implementations differ in architecture and training configurations. Can the authors provide a controlled comparison that isolates the effect of phase information?

  3. In layers/isokoop/objective.py, the frequency-domain loss uses g.detach().mean(dim=1) in the reported configuration. This prevents the frequency-domain term from directly updating the order-weight gate. Is this intentional, and how does removing the stop-gradient operation affect the results?

2 comments

Thank you for recognizing the empirical results and normalization analysis. Your concerns call for a clearer distinction between the placement theorem and explanations of accuracy. We give the requested controls below. MSEs are means over seeds 2027–2029 at L=168, H=96 unless specified. Paired arms share forecast origins. Paper references refer to the revised manuscript.

Novelty and the closest methods. We credit CycleNet for learned cycles and residual forecasting, AsyTO for low-rank phase-indexed profiles, and Saied and Fleuret (2026) for arbitrary-backbone normalization equivariance (Section 2). Our analysis asks how a fixed external profile changes that guarantee. With per-variate mean/std normalization and positive scales, Proposition 1(ii) shows that raw-profile subtraction and restoration preserve scale equivariance for every inner network if and only if the history and future profiles are the same constant per-variate level. The placement switch makes this distinction testable within a fixed forecasting architecture.

AsyTO predicts each variate from its own history. IsoKoop uses spectral lifting and gated transport with cross-variate mean-field coupling. Replacing only the transport by a residual MLP, retaining coupling and readout, gives PEMS04 MSE 0.11972 versus 0.10163 for transport, a 15.11% lower mean with improvement in every seed. Parameter counts differ by 0.008%. Solar favors the MLP, 0.18952 versus 0.20314 (Table 16, Appendix L.2).

The pinned official CycleNet core, adapted to the supplemental data/origin protocol, gives PEMS04 MSE 0.14925 (MLP) and 0.17567 (Linear), versus IsoKoop's 0.10163 (Appendix L.2). Sizes are 0.224M/0.105M/19.52M and epoch caps 30/30/20, respectively. This comparison has unequal capacity and training budgets; it measures accuracy at these recipes rather than establishing novelty by itself.

Theory and implementation. The main benchmark's nonconstant raw profile has no general backbone-independent scale-equivariance guarantee. The gain test checks the placement prediction (Table 8, Appendix C); it does not establish an accuracy advantage. Learned band-filtered channels also differ from exact spectral derivatives (Section 3.2, Appendix I). On PEMS04, band gives MSE 0.10163 and the derivative-multiplier variant 0.10281 (Appendix L.2). Both retain the equalizer, so this tests representations rather than the exact-derivative assumptions.

Q1: Within-dataset L sweeps. Table 15 (Appendix L.1) compares profile Q on/off across L < P, L = P and L > P on six datasets, keeping each pair's remaining recipe fixed and all L's forecast origins common within a dataset/horizon. At H=96, PEMS04's MSE reductions relative to Q-off are 4.14%, 1.52%, −0.56% for L=168/288/576 (P=288). On Solar (P=144), they are 33.94%, −9.82%, 5.42% for L=72/144/288. At L=72, Solar also improves by 26.61%/27.97% at H=192/336, in every seed. PEMS03 and Electricity have negative effects, and Solar is non-monotonic.

Solar sweeps retain the anchor and disable shrink. Traffic/Electricity retain calendar inputs in both arms. Q-off removes the profile, while trained weights or other inputs can still supply periodic information. Changing L also changes history and L-dependent projections. Thus P/L describes coverage, not a universal accuracy threshold. Appendix K states the signal-class identification conditions; these experiments use calendar-reference periods for noisy data.

Q2: Isolating phase. Both arms use the same iTransformer and phase-token module, identical parameter/token counts and training recipe, with a 10-epoch cap. Only true versus constant phase changes (Table 16):

DatasetTrue phaseConstant phase
PEMS040.203340.20755
Solar0.194090.19452

PEMS04 improves by 2.03%, in every seed. Solar's small mean gain has mixed seed directions. Section 4.3 now uses this control; the EMAformer–iTransformer gap remains a comparison between methods.

Q3: Stop-gradient. Detach predates this rebuttal; the available records do not establish whether it was originally intentional. It blocks the frequency-loss gradient through reused time-gate weights; prediction-error gradients remain, and the temporal objective trains the gate. Removing it reopens the path through the gate and its inputs (Table 16):

DatasetWith detachWithout detach
PEMS040.101630.10213
Solar0.203140.19040

Solar transport and detach controls retain shrink and anchor, unlike the shrink-off Solar L sweep. Removing detach slightly worsens every PEMS04 seed. On Solar, one seed drives the lower mean while the other two worsen; neither choice is uniformly better.

Thank you for the rebuttal and the additional experiments. I appreciate the authors' efforts, and the new results help clarify some of my concerns. However, I do not think they change my overall assessment of the paper. I will therefore keep my original score.

Reviewer_M6eDOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes IsoKoop, a multivariate time-series forecasting framework motivated by the difficulty of identifying periodic patterns when the observation window is shorter than the underlying cycle. The authors introduce a low-rank, phase-indexed prototype to supply periodic information beyond the input window and analyze how prototype placement affects scale equivariance under per-window normalization. The framework combines spectral state construction, gated Koopman-style transport with mean-field coupling, and a training objective incorporating temporal and frequency-domain errors.

Experiments across 13 public datasets, four forecasting horizons, and eight baselines show that IsoKoop achieves the best average ranking in both MSE and MAE, with the lowest MSE in 34 of 52 settings. The paper also presents component ablations and analyses of prototype placement, scale robustness, and dataset distribution shifts.

Critical strengths
  1. The paper addresses a relevant limitation of short-lookback forecasting: when the observation window is shorter than the underlying cycle, it may not contain enough information to uniquely identify the periodic pattern. The authors use the ratio (P/L) to motivate when information from outside the window could be useful. The low-rank prototype is a reasonable way to incorporate this information without introducing a large number of additional parameters.

  2. The analysis of scale equivariance is one of the stronger parts of the paper. The authors show that per-window normalization preserves scale equivariance regardless of the inner forecasting network, under the stated conditions. They also show why subtracting a fixed periodic profile before normalization can break this property, while placing it in normalized coordinates or rescaling it can preserve the guarantee. The experiments comparing these placements are particularly useful because they show that preserving scale equivariance does not necessarily lead to better forecasting performance under real distribution shifts.

  3. IsoKoop achieves the best average MSE and MAE rankings across 13 datasets and four prediction horizons, with the lowest MSE in 34 of 52 settings. The comparison includes eight baselines retrained under a common evaluation pipeline, which makes the results easier to assess. The authors also examine the contributions of different stages through factorial ablations and investigate the effects of prototype removal, mean-field coupling, and scale changes. These experiments provide a useful picture of where the method performs well and where individual components contribute less consistently.

Critical weaknesses
  1. One of the paper's central arguments is that external periodic information becomes particularly important when the underlying cycle length (P) exceeds the observation window (L). However, all experiments use a fixed lookback length of (L=168), and the datasets with (P>L) are also high-dimensional sensor networks. As the authors acknowledge, this makes it difficult to separate the effect of the proposed ratio from dataset-specific characteristics, such as cross-variable dependencies and periodic strength. The reported comparisons therefore provide indirect evidence rather than a controlled validation of the criterion. A more convincing experiment would vary the lookback length within the same dataset, particularly across (L<P), (L=P), and (L>P), and measure how the benefit of the external prototype changes. Without this analysis, it remains unclear whether (P/L) is a reliable indicator of when external periodic information improves forecasting.

  2. The authors evaluate the removal of Stage II and compare the proposed transport with attention-based alternatives. However, Stage II combines state-dependent diagonal gating, dense mixing, mean-field coupling, and spectral-entropy-based readout. Removing the entire stage therefore changes several mechanisms simultaneously. Although the mean-field coupling is separately ablated, the experiments do not establish whether the Koopman-inspired parameterization itself is responsible for the improvements. In particular, it remains unclear whether a conventional nonlinear network with comparable capacity could achieve similar results when the same mean-field coupling and readout are retained. A matched-capacity residual MLP replacing only the gated transport would provide a more direct assessment of whether the proposed Koopman-style formulation offers a meaningful advantage.

  3. The proposed prototype represents periodic patterns through a shared temporal basis and variable-specific coefficients. This implicitly assumes that the relevant periodic behavior across variables can be captured in a common low-dimensional subspace. While the authors select the prototype rank through validation and provide prototype-removal ablations, these experiments do not establish whether the low-rank factorization is preferable to simpler phase-dependent representations. In particular, an independent per-variable seasonal profile or a matched-capacity phase embedding could potentially provide similar information without imposing the same shared structure. Since the prototype's contribution is relatively small or inconsistent on some datasets, a comparison against these alternatives would help determine whether the proposed factorization is important or whether the improvement primarily comes from supplying additional phase information.

  4. IsoKoop obtains the period from the dataset protocol and retrieves the corresponding prototype using the absolute sample index. This assumes that the relevant periodicity and phase are known and remain sufficiently consistent between training and testing. Although this is reasonable for datasets with regular calendar-based patterns, real-world time series may contain changing cycle lengths, phase shifts, overlapping periodicities, or irregular events. The paper does not directly examine sensitivity to an incorrectly specified period or phase, making it difficult to assess how robust the proposed external periodic information is outside settings with reliable cycle metadata. Evaluating performance under controlled period misspecification and phase shifts would help clarify the practical limitations of this design.

Questions for authors
  1. Since all experiments use (L=168) and datasets with (P>L) are also sensor networks, what evidence supports (P/L) as the main factor behind the observed improvements rather than dataset-specific characteristics?
  2. Given that Stage II combines several mechanisms and the mean-field coupling contributes substantially to performance, how do the authors distinguish the contribution of the Koopman-style transport from the other components?
  3. What evidence supports the necessity of the shared low-rank factorization rather than simply providing phase-dependent information? How do the authors interpret the relatively small prototype-removal effects on some datasets?
  4. How are the predefined period and absolute phase determined for each dataset, and what assumptions are made about their consistency between training and testing?
  5. Since the learned prototype mainly captures per-sensor levels on Traffic and TrafficC, and its removal improves Solar performance, to what extent do the results support its intended role as a periodic representation?
1 comment

Thank you for suggesting these controls and recognizing the normalization analysis. Q denotes the cycle profile. Unless stated otherwise, mechanism controls use L168/H96 and average seeds 2027–2029. Percentages compare mean MSEs. Single-checkpoint interventions are identified. Table numbers refer to the revised manuscript.

Q1 / W1: Does P/L explain the benefit beyond dataset characteristics?

Within-dataset controls show short-window gains beyond PEMS. On Solar (P=144, L=72), MSE falls by 33.94%, 26.61%, and 27.97% at H=96/192/336, with every seed improving at each horizon. These controls retain the anchor and disable shrink. On PEMS04 at H=96, the gain falls from 4.14% to 1.52% to -0.56% at L=168/288/576. All lookbacks share train/validation/test forecast origins within each dataset and horizon.

Table 15 reports all complete comparisons across six datasets: PEMS03 and Electricity have negative gains, and Solar and PEMS08 are not always monotonic. Traffic H192/L336 lacks a complete three-seed Q-off arm and is marked incomplete. Thus P/L describes coverage, but does not alone predict benefit. The revised contribution and Section 4.3 state this distinction. Q-off can retain learned periodic priors; Traffic/Electricity retain calendar inputs in both arms. Changing L also changes history and L-dependent parameters.

Q2 / W2: Is the transport useful beyond mean-field coupling and capacity?

We replaced only the within-token transport with a residual MLP, retaining mean-field coupling and readout. On PEMS04 (L168/H96), our transport lowers MSE from 0.119717 to 0.101628 (15.11%), improving in every seed. The independently trained MLP and transport have 19,518,106 and 19,519,642 parameters, a 0.008% difference. Solar with shrink and anchor retained favors the MLP: 0.189524 versus 0.203137 (Table 16). The transport-specific advantage is therefore setting-dependent.

Q3 / W3: Is shared low rank necessary, or is supplying phase sufficient?

On PEMS04 (L168/H96), Q-off, phase-only, low-rank Q, and independent per-variable dense Q give MSEs of 0.106020, 0.104451, 0.101628, and 0.102007. Phase-only improves on Q-off by 1.48%; low-rank Q improves on phase-only by 2.70%, both in every seed. Phase-only adds a code to existing tokens and has 488 more parameters than low-rank Q; their representations differ.

Low-rank and dense Q have close mean errors using 4,760/88,416 profile parameters (19.52M/19.60M total). Solar with shrink off and anchor retained favors dense Q: 0.176126 versus 0.187417. Table 17 and Appendix L.2 support low rank as a compact representation, rather than a necessary or uniformly optimal way to use phase.

Q4 / W4: How are period and phase defined, and what if they are wrong?

Table 11 reports calendar-reference periods: daily P=288 for 5-minute PEMS and P=144 for 10-minute Solar; weekly P=168 for hourly Traffic/Electricity. Phase is the global forecast-origin index modulo P, continued across splits. It uses time indices, assuming regular sampling and consistent alignment.

On PEMS04 (L168/H96), retraining with incorrect P=259/317 gives MSE 0.106847/0.107068, versus 0.101628 at P=288; table size and phase indexing also change. Separately, test-time offsets of 0/36/72/144/216 steps on fixed trained checkpoints give 0.101628/0.508481/0.633080/0.905281/0.509487. Zero offset reproduces the aligned result; other offsets show strong sensitivity to misalignment. Appendix L.3 states this assumption. Unknown or drifting periods, multiple periods and irregular sampling remain outside these checks.

Q5: What do Traffic's level-dominated Q and Solar's removal results mean?

Independently retrained Traffic Q-on/off models are nearly tied at 0.377072/0.376692 on 3,413 ordinary L168 test windows. The H96 lookback sweep instead uses 3,077 common Lmax336 origins. For one raw-profile checkpoint with instance normalization, centre coupling and no profile gain/router/shrink, 83.85% of Q's squared energy is constant across phase. Removing its per-sensor cycle mean leaves predictions allclose in FP32 (atol=rtol=1e-5) on all 3,413 windows. Raw/centered Q both give MSE ≈0.372915; constant-only/zero Q give ≈0.591536. Thus this checkpoint uses the temporal component despite its smaller energy. Constant-only versus zero fails FP32 allclose but passes in FP64 (maximum difference 4.06e-9). This inference intervention is separate from Q-off retraining and covers neither TrafficC nor training gradients (Appendix L.3).

On Solar at L168/H96, Q-off, Q-on without shrink, and Q-on with shrink give 0.176404/0.187417/0.203137, with anchor retained. Shrink alone cannot explain the degradation. Together with the short-window gains, these results establish a conditional benefit for the periodic profile.

Please let us know if a particular mechanism or control needs further clarification.

Comments on the paper

1 comment
Area_Chair_1jgqOpenReview ↗

Reminder: Reviewer–Author Discussion

Thank you to the reviewers who have already engaged in the discussion and provided valuable feedback. As the discussion deadline is approaching, I would kindly encourage the remaining reviewers to review the authors' responses and share their follow-up comments, particularly on whether their concerns have been adequately addressed.

Authors are also encouraged to respond promptly to any further questions.

Thank you all for your time and contributions to a constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.