arxiv: 2605.02938 · v1 · submitted 2026-05-01 · 💻 cs.LG · cs.AI

PAMNet: Cycle-aware Phase-Amplitude Modulation Network for Multivariate Time Series Forecasting

Yingbo Zhou , Yutong Ye , Zhiwei Ling , Shuhao Li , Rui Qian , Jian Xiong , Li Sun , Dejing Dou This is my paper

Pith reviewed 2026-05-09 19:24 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords multivariate time series forecastingphase amplitude modulationperiodic pattern decompositioncycle-aware networktime series periodicitydeep learning forecastingphase-amplitude coupling

0 comments

The pith

PAMNet improves multivariate time series forecasting by explicitly separating phase and amplitude in periodic patterns with a dual-branch modulator.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to improve forecasting accuracy for data with strong periodic patterns by decomposing those patterns into separate phase and amplitude parts instead of relying on implicit extraction or ignoring their coupling. It introduces a dual-branch structure where cyclical embeddings handle phase-dependent mean shifts and another branch handles amplitude-driven variance changes, then fuses them lightly without heavy attention layers. This explicit approach is meant to capture interactions more directly while keeping the model efficient. A sympathetic reader would care because many practical forecasting tasks involve cyclical behavior where better separation of these aspects could yield more reliable predictions on real datasets.

Core claim

PAMNet explicitly decomposes periodic patterns into complementary phase and amplitude components. The phase branch employs cyclical embeddings to capture phase-dependent mean shifts, while the amplitude branch models intensity variations to adapt to changes in variance. A lightweight modulator with element-wise fusion combines these components to enable explicit modeling of their interactions without complex attention mechanisms, leading to state-of-the-art performance on twelve real-world datasets.

What carries the argument

Dual-branch modulator with cyclical embeddings for phase positioning and learnable embeddings for amplitude modulation, combined via element-wise fusion.

If this is right

Explicit phase-amplitude decoupling captures periodic interactions more directly than implicit or coupled approaches.
The method delivers state-of-the-art forecasting accuracy on diverse real-world multivariate datasets.
Lightweight element-wise fusion provides an efficient alternative to attention-heavy architectures for cyclical time series.
The decomposition offers a new perspective that treats phase shifts and amplitude variations as separately adaptable factors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same separation could extend to other cyclical signals such as audio waveforms or biological rhythms where phase and amplitude are physically distinct.
Integrating the modulator with trend or residual components might yield hybrid models that handle non-stationary series more robustly.
Controlled tests on synthetic series with known phase-amplitude relationships would provide a direct check on whether the decoupling recovers the ground-truth factors.
The reduced reliance on attention suggests possible use in edge devices that require fast, low-memory periodic forecasting.

Load-bearing premise

Periodic components have an intrinsic phase-amplitude coupling that can be explicitly decomposed and modulated by learnable embeddings without losing critical interactions or causing overfitting.

What would settle it

Ablation experiments on the twelve datasets where disabling either the phase branch or the amplitude branch fails to increase forecasting error compared to the full model, or where non-separating baselines match or exceed PAMNet accuracy.

Figures

Figures reproduced from arXiv: 2605.02938 by Dejing Dou, Jian Xiong, Li Sun, Rui Qian, Shuhao Li, Yingbo Zhou, Yutong Ye, Zhiwei Ling.

**Figure 1.** Figure 1: Periodic non-stationarity in time series data. The evolution of the cycle mean and cycle standard deviation for channel OT in ETTh1 reveals a consistent periodic pattern in its statistical properties, confirming periodic non-stationarity rather than random drift. plicitly decomposes periodic patterns into complementary phase and amplitude components. Its core innovation lies in a dual-stream modulation tha… view at source ↗

**Figure 2.** Figure 2: Overview of PAMNet, a novel framework inspired by signal modulation. The architecture begins by generating adaptive phase EP and amplitude EA carrier signals from a cycle index t. These carriers then modulate the encoded time series EX within two parallel paths, which explicitly capture cyclical dynamics through a phase-amplitude decoupling modulator. The resulting representations are fused and passed thro… view at source ↗

**Figure 3.** Figure 3: Model efficiency comparison under the input-96-predict-96 setting. All results are obtained by using the official model configurations with the same batch sizes: 32 and 16 for the ETTm1 and Traffic datasets, respectively [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Ablation performance (MSE) on the ECL dataset with varying cycle length c. All results are averaged across forecasting horizons H ∈ {96, 192, 336, 720}. Cycle Length Sensitivity. We systematically vary the hyperparameter cycle length c = {23, 24, 96, 168, 192, 336, 720} to assess its impact on both PAMNet and a leading cyclical baseline, TQNet. As shown in 4, optimal performance for both models occurs nea… view at source ↗

**Figure 5.** Figure 5: Impact of varying lookback lengths on ECL. We further assess PAMNet’s adaptability across the following three generalization scenarios. Impact of Input Length. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 7.** Figure 7: Visualization of the modulated representations on ECL. Phase-Amplitude Weights [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗

**Figure 6.** Figure 6: The learned phase and amplitude weights on ETTh1. -10 0 10 20 t-SNE Component 1 -15 -10 -5 0 5 10 15 20 t-SNE Component 2 t-SNE: Feature Representations EX: Input Embeddings MP: Phase-modulated Features MA: Amplitude-modulated Features MX: Modulated Representations EP: Phase Embeddings EA: Amplitude Embeddings [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

**Figure 9.** Figure 9: The modulated representations on ETTm1. -150 -100 -50 0 50 100 150 t-SNE Component 1 -150 -100 -50 0 50 100 150 t-SNE Component 2 t-SNE: Feature Representations EX: Input Embeddings MP: Phase-modulated Features MA: Amplitude-modulated Features MX: Modulated Representations EP: Phase Embeddings EA: Amplitude Embeddings [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗

**Figure 8.** Figure 8: The learned phase and amplitude weights on ETTm1. As shown in Figures 8, 9, and 10, these visualizations collectively validate PAMNet’s core design: cycle-aware embeddings encode periodic dynamics, dual-path modulation enhances feature discriminability, and complementary fusion yields semantically rich temporal representations. By bridging signal processing principles with deep learning, PAMNet provides… view at source ↗

read the original abstract

Reliable periodic patterns serve as a fundamental basis for accurate multivariate time series forecasting. However, existing methods either implicitly extract periodicity through complex model architectures (e.g., Transformers) with high computational overhead or overlook the intrinsic phase-amplitude coupling when modeling periodic components explicitly. To address these issues, we propose a novel Cycle-aware Phase-Amplitude Modulation Network (PAMNet) that explicitly decomposes periodic patterns into complementary phase and amplitude components. The core innovation lies in its dual-branch modulator, featuring dedicated learnable embeddings for phase positioning and amplitude modulation. The phase branch employs cyclical embeddings to capture phase-dependent mean shifts, while the amplitude branch models intensity variations to adapt to changes in variance. A lightweight modulator with element-wise fusion efficiently combines these components, enabling explicit modeling of their interactions without complex attention mechanisms. Extensive experiments on twelve real-world datasets demonstrate that our method achieves state-of-the-art performance through its novel phase-amplitude decoupling mechanism, offering a new perspective for cyclical modeling in time series forecasting.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PAMNet adds an explicit phase-amplitude split with cyclical embeddings for periodic time series, but the SOTA claim still needs isolation from capacity and baseline effects.

read the letter

The main thing to know is that this paper puts forward a dual-branch modulator with dedicated learnable embeddings: one branch uses cyclical embeddings to track phase-dependent mean shifts, the other handles amplitude for variance changes, and they fuse the two with simple element-wise operations. It positions this against both heavy implicit models like Transformers and basic explicit decompositions that ignore the coupling between phase and amplitude. That split is the concrete new piece, and the motivation for keeping it lightweight while targeting periodicity is reasonable on its face. The architecture description itself is clear and focused, which is a plus for anyone who has dealt with seasonal patterns in real forecasting tasks. It avoids piling on attention layers and tries to make the periodic handling more direct. The soft spots sit mostly in the empirical grounding. The abstract asserts state-of-the-art results on twelve datasets thanks to the phase-amplitude mechanism, yet supplies no numbers, no baseline tables, no ablation on the split itself, and no statistical checks. Without those, it is difficult to tell whether the reported gains come from the explicit decoupling or simply from added parameters and the overall design. The stress-test point about independent modulation possibly missing residual interactions holds until the paper shows controls that rule it out. If the full experiments include proper ablations and hold up under scrutiny, the idea gains traction; otherwise the central claim stays under-supported. This is the sort of paper that would interest applied time-series people working on traffic, energy, or sensor data who want something more interpretable than another Transformer variant. A reader already thinking about explicit cycle modeling could pick up usable architectural ideas. It deserves a serious referee because the problem framing is coherent and the proposed fix is specific, even if the current write-up would need substantial strengthening on the results side before publication. I would send it to review with a request for detailed ablations and comparisons.

Referee Report

2 major / 1 minor

Summary. The paper proposes PAMNet, a Cycle-aware Phase-Amplitude Modulation Network for multivariate time series forecasting. It explicitly decomposes periodic patterns into complementary phase and amplitude components using a dual-branch modulator with dedicated learnable cyclical embeddings for phase positioning (capturing mean shifts) and amplitude modulation (capturing variance changes), combined via lightweight element-wise fusion. The central claim is that this mechanism achieves state-of-the-art performance on twelve real-world datasets while avoiding the overhead of implicit periodicity extraction in complex architectures like Transformers and the coupling oversights of prior explicit methods.

Significance. If the empirical results and mechanism isolation hold, the work offers a lightweight, interpretable alternative for cyclical modeling in time series that could reduce computational costs relative to attention-heavy models while providing explicit control over phase-amplitude interactions. Strengths include the parameter-efficient design and focus on a previously underexplored decomposition, but significance hinges on demonstrating that gains stem from the decoupling rather than capacity increases.

major comments (2)

[Abstract and Experiments] Abstract and Experiments section: The SOTA claim on twelve datasets lacks any reported quantitative metrics, baseline tables, statistical tests, or ablation results in the provided abstract, and the full manuscript does not isolate the contribution of the dual-branch phase-amplitude modulator from overall architecture capacity or added learnable embeddings. This is load-bearing for the novelty and mechanism claims, as improvements could arise from implicit effects rather than explicit decoupling.
[§3 and §4] §3 (Method) and §4 (Experiments): No derivation or controlled comparison shows why independent phase (cyclical mean-shift) and amplitude (variance) embeddings with element-wise fusion suffice to capture arbitrary periodic coupling without residual interactions that attention-based models exploit; the premise that this avoids overfitting to dataset-specific periodicity remains untested against variants that retain implicit coupling.

minor comments (1)

[§3] The description of the modulator fusion could include an explicit equation for the element-wise combination to clarify how phase and amplitude interact.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will incorporate revisions to strengthen the presentation of results and mechanism validation.

read point-by-point responses

Referee: The SOTA claim on twelve datasets lacks any reported quantitative metrics, baseline tables, statistical tests, or ablation results in the provided abstract, and the full manuscript does not isolate the contribution of the dual-branch phase-amplitude modulator from overall architecture capacity or added learnable embeddings.

Authors: We agree the abstract would benefit from explicit metrics. In the revision we will update the abstract to report average relative improvements (e.g., 5–12% MSE reduction) over strong baselines across the twelve datasets. The experiments section already contains ablation tables comparing the full PAMNet against phase-only, amplitude-only, and embedding-ablated variants. To more rigorously isolate the dual-branch mechanism from capacity effects, we will add new controlled ablations that match parameter budgets exactly (by widening competing baselines) and report statistical significance via paired t-tests over five random seeds. revision: yes
Referee: No derivation or controlled comparison shows why independent phase (cyclical mean-shift) and amplitude (variance) embeddings with element-wise fusion suffice to capture arbitrary periodic coupling without residual interactions that attention-based models exploit; the premise that this avoids overfitting to dataset-specific periodicity remains untested against variants that retain implicit coupling.

Authors: Section 3 motivates the design by noting that phase and amplitude are complementary and largely separable in periodic signals; element-wise fusion provides an efficient interaction term without attention overhead. While a formal proof of sufficiency for arbitrary couplings is not provided (the approach is empirical), we will add in the revision two new controlled experiments: (i) a coupled-embedding variant that concatenates phase and amplitude before modulation, and (ii) a lightweight attention-based coupling baseline, both evaluated on the same twelve datasets with identical training protocols. These will directly test whether explicit decoupling reduces overfitting to dataset-specific periodicities relative to implicit-coupling alternatives. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper proposes an empirical neural architecture (PAMNet) for multivariate time series forecasting and validates it via experiments on twelve real-world datasets. No mathematical derivation, first-principles result, or prediction is claimed that reduces to the inputs by construction. The core mechanism (dual-branch phase-amplitude modulator with learnable embeddings) is an architectural design choice whose value is assessed through standard held-out test performance, not through any self-referential fitting or renaming. No uniqueness theorems, self-citations as load-bearing premises, or ansatzes smuggled via prior work appear in the provided text. This is a standard ML modeling paper whose claims rest on external empirical benchmarks rather than internal definitional loops.

Axiom & Free-Parameter Ledger

2 free parameters · 0 axioms · 1 invented entities

The model introduces learnable embeddings whose values are determined during training on the target datasets; no external physical constants or pre-derived quantities are invoked. The approach assumes standard neural network optimization can discover useful phase and amplitude representations.

free parameters (2)

learnable phase embeddings
Dedicated embeddings in the phase branch that are optimized to capture phase-dependent mean shifts.
learnable amplitude embeddings
Dedicated embeddings in the amplitude branch that are optimized to model intensity variations.

invented entities (1)

dual-branch phase-amplitude modulator no independent evidence
purpose: To explicitly combine phase positioning and amplitude modulation via element-wise fusion
Core architectural component introduced to enable interaction modeling without attention mechanisms.

pith-pipeline@v0.9.0 · 5489 in / 1226 out tokens · 43187 ms · 2026-05-09T19:24:44.495978+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

36 extracted references

[1]

Proceedings of the 2023 7th International Conference on Advances in Artificial Intelligence (ICAAI) , pages =

Kristofers Volkovs and Evalds Urtans and Vairis Caune , title =. Proceedings of the 2023 7th International Conference on Advances in Artificial Intelligence (ICAAI) , pages =

2023
[2]

Proceedings of the 33rd

Peng Zhu and Yuante Li and Yifan Hu and Qinyuan Liu and Dawei Cheng and Yuqi Liang , title =. Proceedings of the 33rd
[3]

Wanneng Shu and Ken Cai and Neal Naixue Xiong , title =
[4]

Liu Sheng and Joseph Dunbar , title =

Mohammad Amin Morid and Olivia R. Liu Sheng and Joseph Dunbar , title =
[5]

Nguyen and Phanwadee Sinthong and Jayant Kalagnanam , title =

Yuqi Nie and Nam H. Nguyen and Phanwadee Sinthong and Jayant Kalagnanam , title =. The Eleventh International Conference on Learning Representations (ICLR) , year =
[6]

Advances in Neural Information Processing Systems (NeurIPS) , pages =

Haixu Wu and Jiehui Xu and Jianmin Wang and Mingsheng Long , title =. Advances in Neural Information Processing Systems (NeurIPS) , pages =
[7]

The Twelfth International Conference on Learning Representations (ICLR) , year =

Yong Liu and Tengge Hu and Haoran Zhang and Haixu Wu and Shiyu Wang and Lintao Ma and Mingsheng Long , title =. The Twelfth International Conference on Learning Representations (ICLR) , year =
[8]

International Conference on Machine Learning, (ICML) , volume =

Tian Zhou and Ziqing Ma and Qingsong Wen and Xue Wang and Liang Sun and Rong Jin , title =. International Conference on Machine Learning, (ICML) , volume =
[9]

Guoqi Yu and Jing Zou and Xiaowei Hu and Angelica I. Avil. Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling , booktitle =
[10]

The Eleventh International Conference on Learning Representations (ICLR) , year =

Yunhao Zhang and Junchi Yan , title =. The Eleventh International Conference on Learning Representations (ICLR) , year =
[11]

Thirty-Seventh

Ailing Zeng and Muxi Chen and Lei Zhang and Qiang Xu , title =. Thirty-Seventh
[12]

Periodicity Decoupling Framework for Long-term Series Forecasting , booktitle =

Tao Dai and Beiliang Wu and Peiyuan Liu and Naiqi Li and Jigang Bao and Yong Jiang and Shu. Periodicity Decoupling Framework for Long-term Series Forecasting , booktitle =
[13]

Advances in Neural Information Processing Systems (NeurIPS) , year =

Minhao Liu and Ailing Zeng and Muxi Chen and Zhijian Xu and Qiuxia Lai and Lingna Ma and Qiang Xu , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =
[14]

Forty-first International Conference on Machine Learning (ICML) , year =

Shengsheng Lin and Weiwei Lin and Wentai Wu and Haojun Chen and Junjie Yang , title =. Forty-first International Conference on Machine Learning (ICML) , year =
[15]

Advances in Neural Information Processing Systems (NeurIPS) , year =

Shengsheng Lin and Weiwei Lin and Xinyi Hu and Wentai Wu and Ruichao Mo and Haocheng Zhong , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =
[16]

Forty-second International Conference on Machine Learning (ICML) , year=

Temporal Query Network for Efficient Multivariate Time Series Forecasting , author=. Forty-second International Conference on Machine Learning (ICML) , year=
[17]

International Conference on Machine Learning (ICML) , year=

TimeBridge: Non-Stationarity Matters for Long-term Time Series Forecasting , author=. International Conference on Machine Learning (ICML) , year=
[18]

The Eleventh International Conference on Learning Representations (ICLR) , year =

Haixu Wu and Tengge Hu and Yong Liu and Hang Zhou and Jianmin Wang and Mingsheng Long , title =. The Eleventh International Conference on Learning Representations (ICLR) , year =
[19]

Zhang and Jun Zhou , title =

Shiyu Wang and Haixu Wu and Xiaoming Shi and Tengge Hu and Huakun Luo and Lintao Ma and James Y. Zhang and Jun Zhou , title =. The Twelfth International Conference on Learning Representations (ICLR) , year =
[20]

Thirty-Fifth

Haoyi Zhou and Shanghang Zhang and Jieqi Peng and Shuai Zhang and Jianxin Li and Hui Xiong and Wancai Zhang , title =. Thirty-Fifth
[21]

Seeger and Jan Gasthaus and Lorenzo Stella and Yuyang Wang and Tim Januschowski , title =

Syama Sundar Rangapuram and Matthias W. Seeger and Jan Gasthaus and Lorenzo Stella and Yuyang Wang and Tim Januschowski , title =. Advances in Neural Information Processing Systems (NeurIPS) , pages =
[22]

arXiv , volume =

Shengsheng Lin and Weiwei Lin and Wentai Wu and Feiyu Zhao and Ruichao Mo and Haotong Zhang , title =. arXiv , volume =
[23]

The Twelfth International Conference on Learning Representations (ICLR) , year =

Donghao Luo and Xue Wang , title =. The Twelfth International Conference on Learning Representations (ICLR) , year =
[24]

Tsang and Hui Xiong , title =

Jinliang Deng and Feiyang Ye and Du Yin and Xuan Song and Ivor W. Tsang and Hui Xiong , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =
[25]

Ihor Javorskyj and Roman Yuzefovych and Ivan Matsko and Zbigniew Zakrzewski and Jacek Majewski , title =. Digit. Signal Process. , volume =
[26]

Neural Networks , volume =

Stefan Elfwing and Eiji Uchibe and Kenji Doya , title =. Neural Networks , volume =
[27]

1968 , publisher=

Communication systems , author=. 1968 , publisher=

1968
[28]

PyTorch: An Imperative Style, High-Performance Deep Learning Library , booktitle =

Adam Paszke and Sam Gross and Francisco Massa and Adam Lerer and James Bradbury and Gregory Chanan and Trevor Killeen and Zeming Lin and Natalia Gimelshein and Luca Antiga and Alban Desmaison and Andreas K. PyTorch: An Imperative Style, High-Performance Deep Learning Library , booktitle =
[29]

The Thirteenth International Conference on Learning Representations (ICLR) , year =

Hao Wang and Lichen Pan and Yuan Shen and Zhichao Chen and Degui Yang and Yifei Yang and Sen Zhang and Xinggao Liu and Haoxuan Li and Dacheng Tao , title =. The Thirteenth International Conference on Learning Representations (ICLR) , year =
[30]

Kingma and Jimmy Ba , title =

Diederik P. Kingma and Jimmy Ba , title =. 3rd International Conference on Learning Representations (ICLR) , year =
[31]

The Association for the Advancement of Artificial Intelligence (AAAI) , pages =

Yulong Wang and Yushuo Liu and Xiaoyi Duan and Kai Wang , title =. The Association for the Advancement of Artificial Intelligence (AAAI) , pages =
[32]

The Association for the Advancement of Artificial Intelligence (AAAI) , pages =

Jingru Fei and Kun Yi and Wei Fan and Qi Zhang and Zhendong Niu , title =. The Association for the Advancement of Artificial Intelligence (AAAI) , pages =
[33]

Proceedings of the 31st

Pengfei Zhou and Yunlong Liu and Junli Liang and Qi Song and Xiangyang Li , title =. Proceedings of the 31st
[34]

Advances in Neural Information Processing Systems (NeurIPS) , year =

Yuxuan Wang and Haixu Wu and Jiaxiang Dong and Guo Qin and Haoran Zhang and Yong Liu and Yunzhong Qiu and Jianmin Wang and Mingsheng Long , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =
[35]

Journal of Machine Learning Research , year =

Laurens van der Maaten and Geoffrey Hinton , title =. Journal of Machine Learning Research , year =
[36]

2026 , booktitle =

PAMod: Modeling Cyclical Shifts via Phase-Amplitude Modulation for Non-stationary Time Series Forecasting , author =. 2026 , booktitle =

2026