acceptodds
Under review as a conference paper at ICLR 2027

UFO: Learning Universal Frequency Representation with Neural Operator for Time Series Forecasting

Abstract

Frequency representations play a vital role in time series forecasting, with recent deep learning approaches increasingly modeling temporal dependencies in the Fourier basis. In parallel, operator learning frameworks such as the Fourier Neural Operator (FNO) demonstrate that complicated operators can be parameterized via linear transforms in frequency space coupled with local non-linearities. However, existing methods remain constrained by discretization errors and full-frequency mode representation, which inherently bottleneck their prediction capacity. To mitigate this limitation, we replace Fourier bases with adaptive basis functions and determine their rotation coefficients. Time translation is then converted into these coefficients acting on the basis functions. Mathematically, we prove that both the basis functions and the rotation coefficients can be universally approximated by neural networks. Grounded in this theoretical framework, we propose the Universal Frequency Operator (UFO) for long-term time-series forecasting (LSTF). Instead of relying on fixed Fourier or wavelet bases, UFO learns adaptive bases and the corresponding rotation coefficients, capturing temporal dynamics by combining them through complex multiplication. Extensive experiments on 8 real-world datasets demonstrate that UFO consistently outperforms state-of-the-art baselines. To demonstrate that UFO captures core temporal dynamics in the basis functions, we ablate the history input's high-frequency mode and observe that UFO consistently outperforms other methods across all high-frequency cutoff levels. Code is available at this repository: https://anonymous.4open.science/status/UFO-3D30.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series forecasting, Neural Operator, Frequency Representation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_JWFYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on frequency-based methods for time series forecasting. Existing approaches typically perform forecasting by transforming data into the Fourier domain via the Fast Fourier Transform (FFT); however, this paper argues that the use of fixed Fourier bases—which impose rigid frequency constraints—lacks flexibility. Furthermore, the Discrete Fourier Transform (DFT) and the practice of truncating high-frequency components introduce representation errors, while the FFT itself entails high computational complexity. To address these limitations, the paper proposes UFO, a lightweight forecasting architecture that eliminates the need for an explicit FFT by combining learnable latent representations of time series with input-dependent complex-phase rotations. Extensive experiments demonstrate the effectiveness of UFO.

Critical strengths

1.UFO constructs a lightweight time-series forecasting architecture by combining learnable latent representations with complex-valued phase rotations.

2.The paper seeks to provide a mathematical foundation for latent representation learning and temporal dynamic modeling, drawing upon the universal approximation theorem.

3.The model’s performance was evaluated through comparative forecasting experiments across eight real-world datasets, and its robustness was examined via high-frequency truncation experiments.

Critical weaknesses
  1. The authors argue that fixed Fourier bases, discretization errors, and high-frequency truncation limit existing methods. However, a fixed-frequency representation does not necessarily constrain predictive performance, nor does the Discrete Fourier Transform (DFT) inherently introduce reconstruction errors for discrete signals. Furthermore, works such as FreDF [1] and FilterNet [2] have already improved the utilization of frequency information through frequency fusion or filtering; the authors fail to adequately justify the shortcomings of existing approaches or the necessity of learning adaptive bases.

  2. In essence, UFO combines learnable latent representations with input-dependent complex-valued phase rotations—a specific architectural configuration of existing techniques. FRNet, for instance, already employs frequency-domain rotations to model temporal dynamics, while Koopman-related methods have also explored dynamic evolution within latent spaces. The paper does not sufficiently clarify the substantive innovations of UFO compared to these existing methods, and it notably lacks a direct experimental comparison with FRNet.

  3. Theorem 1 is essentially an application of the classical universal approximation theorem to finite-dimensional continuous mappings, while the finite-dimensional Fourier transform is, by definition, a linear mapping. Similarly, Theorem 2 relies primarily on the universal approximation properties of continuous functions. These conclusions neither prove that adaptive bases are superior to fixed Fourier bases, nor do they guarantee that the learned latent representations possess meaningful frequency characteristics or are better at capturing temporal dynamics.

  4. The paper’s logic proceeds from universal approximation capabilities to adaptive frequency representations, and subsequently to enhanced temporal dynamic modeling and predictive performance—yet these inferences do not necessarily hold. Moreover, both theorems rely on MLPs with universal approximation power, whereas the actual UFO architecture—specifically its dimension-lifting mapping, rotation coefficient generation, and projection modules—utilizes only single linear layers. Consequently, the theoretical results do not provide direct guarantees for the practical architecture employed.

  5. The paper lacks critical ablation and comparative experiments—such as comparing adaptive bases against fixed Fourier bases, complex-valued rotations against standard MLPs or linear mappings, and the full UFO model against a version with the rotation module removed. As a result, the current experimental evidence is insufficient to determine whether the observed performance gains stem from the adaptive frequency representation, the rotation mechanism, or simply the underlying linear mappings. Furthermore, high-frequency truncation experiments merely demonstrate that the model maintains strong predictive performance under various truncation conditions; they do not constitute direct proof that the model has successfully captured the so-called "core" temporal dynamics.

  6. Although the computational complexity of the FFT is O(Tlog⁡T)O(T \log T), the input length used in the paper’s experiments is only 96, meaning the actual computational overhead of the FFT is likely quite minimal. Moreover, the dense linear mappings employed by UFO entail computational costs that depend on both the input and latent dimensions. The paper’s assertion that the overall complexity is O(n+m)O(n+m) lacks a solid foundation, and the claim regarding its efficiency advantage is not substantiated by any comparison of actual runtime or computational overhead.

[1] Not All Frequencies Are Created Equal: Towards a Dynamic Fusion of Frequencies in Time-Series Forecasting. ACM MM 2024.

[2] FilterNet: Harnessing Frequency Filters for Time Series Forecasting. NeurIPS 2024.

Questions for authors

See Weaknesses

Reviewer_Y1DfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

UFO lifts the input window into a latent tensor BB with a linear layer, predicts a frequency matrix Ω\Omega from BB with another linear layer, multiplies BB by exp⁡(iΩ~Δt)\exp(i\tilde{\Omega}\Delta t), adds the result back to BB, and projects to the horizon with a third linear layer. Two theorems state that MLPs can approximate (i) a finite-dimensional Fourier encoding map and (ii) the function B⊙exp⁡(iΩ~Δt)B \odot \exp(i\tilde{\Omega}\Delta t). Experiments cover eight standard LTSF benchmarks with lookback 96, plus a low-pass input ablation on ETTh1.

Critical strengths
  1. Simple design. The model consists of three linear layers and an element-wise complex rotation, which makes it easy to implement and inspect.
  2. Broad benchmark coverage. Results are reported on eight common benchmarks against eleven baselines, and UFO is competitive on all of them.
Critical weaknesses
  1. The theory does not support the method.

    • Theorem 1 concerns the encoding map F=C∘F∘TF = C \circ \mathcal{F} \circ T, which is linear between finite-dimensional spaces. A single linear layer represents it exactly, so universal approximation adds nothing. The theorem also says nothing about the discretization error that motivates the paper.
    • Theorem 2 shows that an MLP can approximate the explicit smooth function B⊙exp⁡(iΩ~Δt)B \odot \exp(i\tilde{\Omega}\Delta t). However, UFO computes this function in closed form (Eq. 19) and never approximates it with an MLP, so the Taylor expansion in Eqs. 10–12 plays no role in the model.
    • Neither theorem shows that rotating a learned latent basis corresponds to a time shift. The shift property in Eq. 8 holds only for the Fourier basis.
  2. The "neural operator" and "frequency" framing is not reflected in the model.

    • Not a neural operator. PP, QQ, and rir_i are fixed-size linear layers over the length-96 window, so UFO is tied to a single discretization, like any MLP forecaster. There is no integral kernel and no resolution-transfer experiment.
    • The conjugate-pairing analogy fails. BB is real, so the paired rotation with [Ω;−Ω][\Omega; -\Omega] yields conjugate pairs only if the two halves of BB are equal. This breaks the analogy with real-signal spectra.
    • Functionally a small MLP. What remains is a linear–gate–linear network with sin/cos gating, which is close to a small MLP.
    • Key details are missing. The paper does not state the value of Δt\Delta t, the sizes kk and cc, or how the complex B^\hat{B} is mapped back to a real-valued forecast.
  3. The complexity claim does not hold up.

    • The O(n+m)O(n+m) bound treats k⋅ck \cdot c as a constant. In fact, the dense lifting and projection layers cost O(n⋅k⋅c)O(n \cdot k \cdot c) and O(k⋅c⋅m)O(k \cdot c \cdot m) per variable, which for typical widths exceeds the O(nlog⁡n)O(n \log n) of an FFT at n=96n = 96.
    • The paper reports no runtime, memory usage, or parameter counts to back the efficiency claim.
  4. Baselines are incomplete, and the gains are within noise on most datasets.

    • Missing baselines. The closest prior work is not compared. FRNet (KDD 2024), which the paper itself describes as rotating frequency-domain bases to model temporal increments, is absent. So are FITS (ICLR 2024) and FNO, on which the method is built. Strong recent forecasters are also missing, e.g., PatchTST, CycleNet, xPatch, TimeKAN, and TimeMixer++.
    • Small margins. Against the best baseline per dataset in Table 1, the MSE gap is at most 0.005 on five of eight datasets: ETTh2 0.373 vs. 0.378, ETTm1 0.383 vs. 0.384, ETTm2 0.274 vs. 0.275, Weather 0.243 vs. 0.244, and Traffic 0.427 vs. 0.428. On ECL, UFO is worse than TimeXer (0.178 vs. 0.170).
    • Unsupported significance claim. The paper claims significance at the 95% level but reports neither the number of runs, standard deviations, nor the test used.
    • Single lookback. Only lookback 96 is evaluated.
  5. The frequency ablation does not show what is claimed.

    • Unclear protocol. In Figure 2, the errors of all models drop sharply as the cutoff decreases (UFO's MSE falls from about 0.37 to 0.28). This would be unexpected if only the inputs were filtered, which suggests the targets were filtered as well.
    • No evidence for the claim either way. All five models follow the same trend with roughly constant gaps, so the experiment does not show that UFO "captures core temporal dynamics in the basis functions".
    • Narrow scope. The ablation uses only ETTh1 and, apparently, only H = 96.
  6. Presentation.

    • Figure 1 relies on decorative clip-art and omits the implementation details listed in point 2.
    • The panels in Figure 3 are too small to read.
    • Table 1 contains apparent errors: for example, DLinear's ETTh2 MAE at H = 192 (0.323) is lower than its own value at H = 96 (0.395).
    • There are typos ("discret", "Equation equation 9"), and in Theorem 2 the symbol Ω\Omega denotes both a matrix and a scalar bound.
Questions for authors
  1. What are the values of Δt\Delta t, kk, and cc? How is the complex B^\hat{B} mapped back to a real-valued forecast? What are UFO's parameter count and training time compared with FITS and DLinear?
  2. In Figure 2, are only the inputs low-pass filtered, or the targets as well?
  3. How does UFO compare with FRNet, FITS, and a plain linear–GELU–linear MLP of the same width?
  4. How many random seeds were used, and which statistical test supports the claim of significance at the 95% level?
Reviewer_awPHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Universal Frequency Operator (UFO), a new model for long-term time series forecasting. Instead of using fixed Fourier bases, UFO learns adaptive basis functions and rotation coefficients to capture temporal dynamics through complex multiplication. The authors also present theoretical analysis based on universal approximation. They evaluate UFO on eight benchmark datasets, showing competitive forecasting performance against several baselines.

Critical strengths
  1. The paper evaluates UFO on eight benchmark datasets and shows competitive forecasting performance against several baselines.
Critical weaknesses
  1. Figure 1 provides little useful information and does not clearly illustrate the key technical details of the proposed method.
  2. The baselines used in the experiments are somewhat outdated. The comparisons would be more convincing if the authors included recent SOTA models from ICLR 2026 and ICML 2026, such as Phaseformer [1], MixLinear [2], and DMANet [3].

References:

[1] PHASEFORMER: FROM PATCHES TO PHASES FOR EFFICIENT AND EFFECTIVE TIME SERIES FORECASTING (ICLR 2026)

[2] MIXLINEAR: EXTREME LOW RESOURCE MULTIVARIATE TIME SERIES FORECASTING WITH 0.1K PARAMETERS (ICLR 2026)

[3] Anti-Aliasing Matters: A Dynamic Network for Time Series Forecasting (ICML 2026)

Questions for authors
  1. In the third paragraph of the Introduction, the authors mention that DFT introduces discretization error and that frequency truncation further increases this error. I have two questions: (1) How much does the discretization error from DFT actually affect forecasting performance? I would expect the impact to be quite small. Could the authors provide some quantitative results to show how much it matters in practice? (2) Could the authors explain what they mean by the "general truncation operation"? Is this something done manually?
  2. The core idea of UFO is to model temporal dynamics through basis functions. How does this fundamentally differ from the previous work FBM [1], which also explores time series forecasting from a basis function perspective? Could the authors clarify the key differences and advantages of UFO over FBM?
  3. The authors claim that both the basis functions and rotation coefficients can be universally approximated by neural networks. However, this seems to follow naturally from the classical universal approximation theorem, so it is unclear what new theoretical insight these results provide. More importantly, Theorems 1 and 2 rely on the approximation capabilities of MLPs, while UFO uses single linear layers for lifting, rotation, and projection. Could the authors clarify how these theoretical results actually support the proposed architecture?

References:

[1] Rethinking Fourier Transform from A Basis Functions Perspective for Long-term Time Series Forecasting (NeurIPS 2024).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.