Adaptive Gaussian-Fourier Positional Encoding

• Adaptive Gaussian-Fourier positional encoding is a method that deterministically maps high-dimensional spatial data using a blend of Gaussian RBFs, cosine kernels, and fixed Fourier channels.
• It dynamically adapts kernel parameters based on input geometry, ensuring robust representation across varying scales and sampling densities.
• The approach achieves notable empirical performance improvements in tasks such as 3D classification and segmentation, as evidenced by increased ModelNet40 accuracy and ShapeNetPart mIoU.

Adaptive Gaussian-Fourier positional encoding is a deterministic, input-adaptive, and multi-modal feature mapping designed to capture local and global positional relationships in high-dimensional data such as images and point clouds. These encodings blend Gaussian radial basis functions (RBFs), cosine kernels, and, when applicable, fixed-frequency Fourier terms, with key parameters derived from geometric properties of the data itself. This family of methods includes both trainable variants in sequence and vision Transformers and parameter-free architectures for point-cloud classification and segmentation (Li et al., 2021, Saeid et al., 31 Jan 2026).

1. Mathematical Formulation of Adaptive Gaussian-Fourier Encodings

Adaptive Gaussian-Fourier encodings create a vectorized representation of spatial or geometric positions using a mixture of Gaussian and trigonometric kernels, with the degree of mixing dynamically determined by the global structure of the data.

For a set of $N$ 3D points $X = [x_1, ..., x_N] \in \mathbb{R}^{N \times 3}$, the adaptive positional code is constructed as:

$$H_{\rm pos} = \underbrace{H_{\rm adaptive}}_{\text{input-adaptive channel}} \;\Big\Vert\; \underbrace{H_{\rm Fourier}}_{\text{fixed-frequency channel (segmentation only)}}$$

Adaptive Channel: For a coordinate $x \in \mathbb{R}$ and anchor $v_m$:

• Gaussian RBF: $$\phi_{\rm RBF}(x, v_m) = \exp\left(-\frac{1}{2}\left(\frac{x - v_m}{\sigma_a + \epsilon}\right)^2\right)$$
• Cosine: $$\phi_{\cos}(x, v_m) = \cos\left(\frac{x - v_m}{\sigma_a + \epsilon}\right)$$
• Adaptive blending: $$\phi_{\rm adaptive}(x, v_m) = \lambda \phi_{\rm RBF}(x, v_m) + (1 - \lambda)\phi_{\cos}(x, v_m)$$

Here,

• $$\sigma_g = \tfrac{1}{3}\sum_{i=1}^3 \mathrm{Std}(X_{:,i})$$ is the mean per-axis standard deviation,
• $\sigma_a = \sigma_0 (1 + \sigma_g)$ (where $\sigma_0$ is a base bandwidth),
• $$\lambda = \mathrm{sigmoid}\left((\sigma_g - \tau)\kappa\right)$$ for task-specific hyperparameters $\tau$, $\kappa$.

Fixed-Frequency Channel (for 3D segmentation): For $L$ frequencies,

$$\phi_{\rm Fourier}(x) = [\sin(\frac{\beta x}{\omega_1}), \cos(\frac{\beta x}{\omega_1}), ..., \sin(\frac{\beta x}{\omega_L}), \cos(\frac{\beta x}{\omega_L})]$$

$\omega_j = \alpha^{j/L}$ for $j=1...L$, with $\beta$ a global scale.

The composite code $H_{\rm pos}$ is formed by concatenating all adaptive and Fourier codes across coordinates.

2. Construction and Adaptivity of the Encoding

The adaptivity arises through the direct dependence of bandwidth $\sigma_a$ and blend ratio $\lambda$ on $\sigma_g$, which measures the global scale of the input. As the spatial spread of the data changes (e.g., varying object scale or density), these parameters are recomputed, ensuring that the encoding adapts its locality and harmonic capacity to best resolve features at the appropriate scale.

For tightly clustered or small objects, $\sigma_g$ is small, leading to narrow kernels and a blend favoring the RBF component. For dispersed objects, larger $\sigma_g$ amplifies the cosine response and broadens the kernel, improving robustness to scale/density changes. Ablations confirm that adaptivity in $\sigma_a$ and $\lambda$ is crucial for maintaining peak performance across diverse datasets without task-specific retuning (Saeid et al., 31 Jan 2026).

3. Algorithmic Integration and Complexity

In non-parametric 3D architectures such as NPNet, these encodings are integrated within a hierarchical framework comprising farthest point sampling, $k$-nearest neighbor grouping, and pooling. At each stage, for each centroid and local neighborhood, features $H_{\mathcal N}$ (e.g., relative positions) and $H_{\rm pos}$ (adaptive/Fourier positional codes) are combined via elementwise multiplication:

$$\widetilde H_{\mathcal N} = (H_{\mathcal N} + H_{\rm pos}) \odot H_{\rm pos}$$

Subsequent mean and max pooling provide descriptors for each centroid.

The computational complexity per stage is $\mathcal{O}(N_t \cdot k \cdot d)$, where $N_t$ is the number of centroids, $k$ the neighborhood size, and $d$ the embedding dimension. Memory cost is $\mathcal{O}(N_t d + k d)$ (Saeid et al., 31 Jan 2026).

4. Robustness to Scale and Density Variations

A distinguishing property of adaptive Gaussian-Fourier positional encoding is its ability to maintain stable performance across drastic changes in input scale and sampling density. Standard encodings with fixed bandwidth often underperform outside a narrow parameter regime, either blurring detail at large scales or introducing aliasing at high resolution. By dynamically tying $\sigma_a$ and $\lambda$ to input statistics, these adaptive encodings avoid such degeneration, enabling accurate and stable representation over a broad range of object sizes and point densities. Empirically, adaptive selection yields peak accuracy of 85.45% on ModelNet40; fixing hyperparameters degrades performance outside of finely tuned settings (Saeid et al., 31 Jan 2026).

5. Application in Parametric and Non-Parametric Architectures

In parametric vision/language architectures, such as Transformers for image classification (e.g., ViT) or detection (e.g., DETR), adaptive Gaussian-Fourier encodings can be implemented using learnable Fourier features. Here, the frequencies are initialized from a zero-mean Gaussian (variance $\gamma^{-2}$, task-specific), and subsequently modulated through a multi-layer perceptron (MLP) before injection into attention layers. This configuration enables both flexibility and shift-invariance in the embedding, facilitating both convergence and generalization (Li et al., 2021).

In non-parametric designs exemplified by NPNet, the encoding is fully deterministic, and the blend between kernels is set by input statistics, with no learned parameters. This paradigm is compatible with template-matching, memory-based, and few-shot learning approaches, where weightless and adaptive representation is beneficial (Saeid et al., 31 Jan 2026).

6. Empirical Performance and Ablation Studies

On 3D object classification tasks (e.g., ModelNet40), NPNet using adaptive Gaussian-Fourier encoding achieves 85.45% accuracy, matching the top-performing non-parametric baselines. On part segmentation (ShapeNetPart), the combined adaptive and fixed-frequency encoding boosts instance mIoU from 70.4% to 73.56%. Ablation demonstrates that the adaptivity of $\sigma_a$ and $\lambda$ is essential; fixing either narrowly restricts the range of input scales for which the method is effective. Inclusion of fixed-frequency (global) Fourier channels is particularly impactful in segmentation, where global context must be captured for accurate part boundaries (Saeid et al., 31 Jan 2026).

In Transformer-based models, learnable Gaussian-Fourier encodings yield consistent gains across tasks. For instance, in ImageNet64 image generation, learnable Fourier+MLP converges ~20% faster and achieves a 0.03 bits/dim reduction compared to sinusoidal embeddings. In object detection, both accuracy and transfer robustness improve relative to baseline sine encodings (Li et al., 2021).

7. Comparative Overview

Method	Parametric/Non-parametric	Adaptivity	Empirical Peak Accuracy (ModelNet40)	Segmentation mIoU (ShapeNetPart)
NPNet (Adaptive Gaussian-Fourier)	Non-parametric	Yes (from input geometry)	85.45%	73.56%
Point-NN/Point-GN	Non-parametric	No	81.8% / 85.3%	70.4% (Point-NN)
Learnable Fourier+MLP (Transformer)	Parametric	Yes (learned)	See (Li et al., 2021) for per-task results	–

Adaptive Gaussian-Fourier positional encoding provides a theoretically grounded, empirically validated approach for achieving scale- and density-robust representations, functioning as a core module in both non-parametric and parametric models for vision and geometry-based tasks (Li et al., 2021, Saeid et al., 31 Jan 2026).