Fourier Neural Filter (FNF)

• Fourier Neural Filter is a neural module that parameterizes frequency-domain filters to enable adaptive, efficient, and global signal transformations.
• It leverages compact neural architectures like MLPs to replace traditional kernels, facilitating plug-and-play integration in images, audio, and time series.
• FNF achieves remarkable parameter efficiency by decoupling filter size from learnable parameters, supporting extensive receptive fields with reduced complexity.

A Fourier Neural Filter (FNF) is a general class of neural module that parameterizes and applies filters in the frequency domain to enable efficient, expressive, and content-adaptive transformations of input signals, whether in time series, images, or audio. Instead of directly learning spatial or temporal kernels, an FNF represents the filtering operation via neural parameterizations (such as small MLPs or implicit neural representations) in the frequency domain, combining the inductive biases of classical Fourier analysis with the adaptability and nonlinearity of neural networks. This framework supports plug-and-play application across modalities, facilitating large (potentially infinite) receptive fields and efficient global operations with a compact parameter footprint (Grabinski et al., 2023, Xu et al., 10 Jun 2025, Kim et al., 2024, Verma, 2023).

1. Foundational Formulation and Mathematical Principles

At the core of FNF models is the convolution theorem, which states that convolution in the time or spatial domain is equivalent to element-wise multiplication in the frequency domain: $$g \circledast k = \mathcal{F}^{-1}\left(\mathcal{F}(g) \odot \mathcal{F}(k)\right)$$ where $g(x)$ is the input (e.g., feature map or signal), $k(x)$ is the convolution kernel, and $\mathcal{F}()$ denotes the Fourier transform, with $\odot$ as element-wise multiplication. The Fourier Neural Filter replaces the explicit kernel $k(x)$ with a frequency-domain function, typically parameterized by an implicit neural function or compact neural module. For a 2D spatial convolution: $$F_\theta: \mathbb{R}^2 \to \mathbb{C}^C,\quad m(\omega_x, \omega_y) = F_\theta(\omega_x, \omega_y)$$ where each frequency $(\omega_x, \omega_y)$ maps to a channel-specific complex multiplier. In 1D/time-series, this reduces to $F_\theta: \mathbb{R} \to \mathbb{C}^D$ for D channels (Grabinski et al., 2023, Xu et al., 10 Jun 2025).

Transformations are implemented as follows:

1. Fourier-transform input ($\mathcal{F}(g)$)
2. Obtain frequency multipliers using a compact neural module ($F_\theta(\omega)$)
3. Perform point-wise multiplication in frequency domain
4. Inverse Fourier-transform to return to native domain

This allows the effective filter size to be decoupled from the parameter count and provides global context with complexity $O(N^2 \log N)$ or $O(N \log N)$ (1D) (Grabinski et al., 2023, Xu et al., 10 Jun 2025).

2. Architectural Variants and Instantiations

The FNF concept is instantiated in several forms, tuned to particular modalities:

• Neural Implicit Filter Function (NIFF): Parameterizes the Fourier spectrum as an implicit neural function (MLP or stack of $1 \times 1$ convolutions) over frequency grids. Empowers CNNs to learn filters potentially as large as the input dimension, retaining compact parameterization and direct compatibility with standard training pipelines (Grabinski et al., 2023).
• Input-Dependent and Gated FNF: As proposed in long-term time-series forecasting, FNF layers extend Fourier Neural Operators (FNO) by gating the global (frequency) response with a local, input-dependent gate, providing both global (spectral) and local (instance-specific) context (Xu et al., 10 Jun 2025):

$$(K v)(x) = T\left[ G(v)(x) \odot \mathcal{F}^{-1} \left( R_\phi \cdot \mathcal{F}(H(v)) \right)(x) \right]$$

Here, $T,G,H$ are learned linear maps, $R_\phi$ is a learned frequency-domain filter, and gating is complex-valued.

• Implicit Neural Fourier Filter (INFF): In time-series modeling with Neural Fourier Modelling (NFM), the INFF module generates learnable frequency-domain filters using a small SIREN MLP, parameterized as a function of spectral embeddings generated from the input. The frequency-domain filter coefficients are then applied multiplicatively to transformed features, enabling adaptive, length-agnostic filtering (Kim et al., 2024).
• Learnable DFT/STFT Front End FNF: In audio applications, FNF directly replaces the DFT/STFT bank with learnable filterbanks realized as fully-connected layers or convolutional (FIR) filters. Adaptive gating networks can route inputs to different expert filterbanks (Verma, 2023).

Variant	Param. Form	Domain	Notable Property
NIFF (Grabinski et al., 2023)	Implicit MLP (2D)	Images/CNNs	Infinitely large kernel, plug-in
INFF (Kim et al., 2024)	SIREN MLP (1D)	Time-Series	Instance/mode adaptive, compact
Input-Gated FNF	Linear + gate (1D)	Time-Series	Input-dependent kernel, gating
Learnable DFT [2308]	Matrix/Conv Layer	Audio	Content-adaptive, windowed

3. Implementation and Computational Properties

FNF modules universally exploit the efficiency of FFT/IFFT for the global operations:

• Parameter efficiency: Filter size (receptive field) decoupled from learnable parameter count, e.g., a handful of 1×1 layers suffice regardless of image/time-series length (Grabinski et al., 2023).
• Computational complexity: Frequency-domain filtering reduces the cost for large kernels from $O(N^4)$ (spatial) to $O(N^2 \log N)$ for images ($N \times N$) or $O(N \log N)$ for 1D signals (Grabinski et al., 2023, Xu et al., 10 Jun 2025, Kim et al., 2024).
• Memory: Stores only compact neural weights and current frequency maps ($O(N^2)$ memory).
• Extensions: Zero-padding, crops, and coordinate grid normalization enable adaptation to linear vs. circular convolutions and different input sizes.
• Training: Losses typically combine standard task objectives (e.g., cross-entropy, L1/MSE) with, in some variants, explicit frequency-domain spectral loss components for spectral accuracy (Kim et al., 2024).
• Plug-in capability: FNF modules are direct replacements for convolutions or, in time series, self-attention or other global mixers.

4. Empirical Findings and Application Domains

The operational flexibility of FNFs has been empirically validated in multiple domains:

• Image classification: NIFF-enabled CNNs for ImageNet-1k show that filters, even with access to kernels as large as $56\times56$, empirically localize most energy within $9\times9$ windows, with dominant components often $3\times3$ or $5\times5$ (Grabinski et al., 2023).
• Time-series forecasting: In multivariate forecasting (energy, weather, traffic), FNF architectures with dual-branch designs (DBD) consistently achieve lower MAE and MSE than Transformer, FNO, and MLP baselines across 11 datasets. Ablations highlight the benefit of parallel branch modeling (Xu et al., 10 Jun 2025).
• Audio signal processing: Learned filters specialize as onset detectors, harmonic (comb) filters, windowed sinusoids, and content-adaptive filterbanks, surpassing classical DFT approaches in pitch and timbre recognition tasks (Verma, 2023).
• Compactness: In NFM, FNF-based models require fewer than 40K parameters, with robust performance generalizing to unseen sampling rates and modalities (Kim et al., 2024).

5. Theoretical Analyses and Inductive Biases

The FNF framework embeds several key inductive biases and theoretical properties:

• Translation invariance: Fourier-based kernels are naturally translation-invariant, inheriting stationarity assumptions from classical signal processing (Xu et al., 10 Jun 2025).
• Expressive power: Input-dependent and gated variants (e.g., time-series FNF) strictly expand the class of learnable integral operators relative to fixed-kernel FNOs, without increasing computational complexity (Xu et al., 10 Jun 2025).
• Efficient global context: Global filtering is realized at $O(N \log N)$ cost, contrasting favorably with $O(N^2)$ attention mechanisms.
• Spectral adaptivity: Modules like INFF enable both instance-level and mode-level adaptivity (using implicit neural representations), providing spectral selectivity and flexibility beyond fixed DFT bases (Kim et al., 2024, Verma, 2023).
• Information bottleneck: Dual-branch designs for spatio-temporal data optimize parallel mutual information objectives, enhancing gradient flow and representational capacity beyond unified or sequential approaches (Xu et al., 10 Jun 2025).

6. Implementation Considerations and Practical Recommendations

Several specific implementation strategies and caveats have been highlighted:

• Precompute normalized frequency grids for FFT input (Grabinski et al., 2023).
• Maintain compact neural module depth (2–4 layers) to avoid overfitting (Grabinski et al., 2023).
• Employ GPU-accelerated FFT libraries for efficiency (Grabinski et al., 2023, Kim et al., 2024).
• For hybrid or sparse data, non-uniform FFT or interpolation may be adopted (Xu et al., 10 Jun 2025).
• Use batch normalization/activation in the native (spatial/time) domain, only filtering in Fourier domain (Grabinski et al., 2023).
• Zero-padding can approximate linear convolution when needed, though circular convolution is natural (and empirically advantageous) for many FFT-based approaches (Grabinski et al., 2023).
• For length-agnostic models (e.g., INFF/NFM), instance normalization and frequency extrapolation combine to make models robust to unseen test-time sampling rates (Kim et al., 2024).

7. Limitations, Empirical Observations, and Outlook

• FFT/IFFT overhead: May introduce computational latency unless properly fused (Grabinski et al., 2023).
• Circular convolution effects: Wrap-around artifacts are possible but typically minor (Grabinski et al., 2023).
• Receptive field utilization: Despite infinite theoretical receptive field, filters often remain local in energy (Grabinski et al., 2023).
• Small signal settings: For very short inputs (e.g., CIFAR-10), large filters confer little benefit and may degrade performance (Grabinski et al., 2023).
• Implementation complexity: FNF logic and gradient flows are more complex than typical time- or spatial-domain convolutions (Grabinski et al., 2023, Xu et al., 10 Jun 2025).
• Parameterization trade-offs: Small, implicit models risk underfitting on highly structured or irregular data (Kim et al., 2024).

A plausible implication is that FNFs provide a promising global signal-processing backbone for domains where parameter efficiency and principled inductive biases are critical, with demonstrated superiority over classical attention and kernel methods in long-range and multi-modal contexts. Potential extensions include hybrid FNF–attention networks, non-uniform domain generalizations, and use in generative or anomaly detection tasks (Xu et al., 10 Jun 2025, Kim et al., 2024).