Index-Set Fourier Series Features

• Index-Set Fourier Series Features (ISFSF) are explicit feature maps built from truncated multivariate Fourier series using sparse, structured index sets to represent periodic kernels.
• ISFSF deterministically selects frequencies to achieve faster convergence and reduce feature numbers by an order of magnitude compared to random Fourier approaches.
• The method is effective in scalable Gaussian process regression, ANOVA sensitivity analysis, and arithmetic Fourier geometry, demonstrating broad cross-disciplinary applications.

Index-Set Fourier Series Features (ISFSF) are explicit feature maps constructed from truncated multivariate Fourier series using sparse, structured index sets. Originally motivated by the need for scalable, accurate representations of periodic structure in high-dimensional kernels, ISFSF provides deterministic and tunable bases for machine learning and functional approximation, as well as connections to analytic number theory and spectral geometry. The concept of index sets—specific, mathematically defined subsets of multi-index frequency space—enables ISFSF to attain fast convergence and interpretable structure, in contrast to methods based on randomly sampled frequencies. Applications span scalable Gaussian process regression with periodic kernels (Tompkins et al., 2018), analysis of variable interactions via ANOVA-decompositions (Potts et al., 2024), and the study of arithmetic Fourier series generated by the prime numbers (Vartziotis, 12 Feb 2026).

1. Mathematical Foundations and Formal Definitions

ISFSF originates with the classical multivariate Fourier expansion of periodic or shift-invariant kernels. Any periodic kernel $$k\colon \mathbb{T}^D\times\mathbb{T}^D\to\mathbb{R}$$ on the $D$-torus admits the series

$$k(\mathbf{x},\mathbf{x}') = \sum_{\mathbf{n} \in \mathbb{Z}^D} c_{\mathbf{n}} e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} - \mathbf{x}' \rangle }$$

where $\mathbf{n}$ runs over integer multi-indices and $$\boldsymbol{\omega}_{\mathbf{n}} = \frac{2\pi}{T}\mathbf{n}$$. The coefficients $c_{\mathbf{n}}$ are obtained by integration against the kernel (Tompkins et al., 2018). To construct ISFSF, the infinite sum is truncated to a finite index set $\mathcal{I} \subset \mathbb{Z}^D$: $$k_{\mathcal{I}}(\mathbf{x}, \mathbf{x}') = \sum_{\mathbf{n} \in \mathcal{I}} c_{\mathbf{n}}\, e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} \rangle} e^{-i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x}' \rangle}$$ This yields an explicit feature embedding: $$\varphi_{\mathcal{I}}(\mathbf{x}) = \left[\sqrt{c_{\mathbf{n}}}\, e^{i \langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} \rangle}\right]_{\mathbf{n} \in \mathcal{I}} \in \mathbb{C}^{|\mathcal{I}|}$$ with $$k_{\mathcal{I}}(\mathbf{x}, \mathbf{x}') = \varphi_{\mathcal{I}}(\mathbf{x})^\mathsf{H} \varphi_{\mathcal{I}}(\mathbf{x}')$$.

For real-valued computation, complex features are split into sines and cosines, leveraging the symmetry $D$0.

In arithmetic contexts, such as prime-indexed Fourier geometry, the relevant index set is the set of all primes up to $D$1, and the construction involves explicit arithmetic weights (e.g., $D$2) (Vartziotis, 12 Feb 2026).

2. Index Set Design and Sparsification Principles

A central aspect of ISFSF is selection of the index set $D$3. This enables control over trade-offs in approximation quality, computational cost, and interpretability. Key constructions include:

• $D$4-ball (tensor-product) sets: $D$5, where $D$6 weights each dimension, and $D$7 sets the refinement.
• Energy-norm hyperbolic cross (ENHC) sets: These have the form

$D$8

with $D$9. As $$k(\mathbf{x},\mathbf{x}') = \sum_{\mathbf{n} \in \mathbb{Z}^D} c_{\mathbf{n}} e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} - \mathbf{x}' \rangle }$$0, this recovers tensor-product sets; as $$k(\mathbf{x},\mathbf{x}') = \sum_{\mathbf{n} \in \mathbb{Z}^D} c_{\mathbf{n}} e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} - \mathbf{x}' \rangle }$$1, the set becomes sparse in high $$k(\mathbf{x},\mathbf{x}') = \sum_{\mathbf{n} \in \mathbb{Z}^D} c_{\mathbf{n}} e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} - \mathbf{x}' \rangle }$$2 (Tompkins et al., 2018).

• Low-order ANOVA index sets: In sensitivity analysis and interpretability contexts, index sets are formed by all variable subsets $$k(\mathbf{x},\mathbf{x}') = \sum_{\mathbf{n} \in \mathbb{Z}^D} c_{\mathbf{n}} e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} - \mathbf{x}' \rangle }$$3 with $$k(\mathbf{x},\mathbf{x}') = \sum_{\mathbf{n} \in \mathbb{Z}^D} c_{\mathbf{n}} e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} - \mathbf{x}' \rangle }$$4 for small $$k(\mathbf{x},\mathbf{x}') = \sum_{\mathbf{n} \in \mathbb{Z}^D} c_{\mathbf{n}} e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} - \mathbf{x}' \rangle }$$5 (Potts et al., 2024).

Sparse index sets (notably ENHC) can reduce feature count from exponential-in-$$k(\mathbf{x},\mathbf{x}') = \sum_{\mathbf{n} \in \mathbb{Z}^D} c_{\mathbf{n}} e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} - \mathbf{x}' \rangle }$$6 to sub-exponential, while retaining high accuracy, enabling scaling to very high dimensions.

3. Feature Construction and Algorithmic Implementation

The construction of ISFSF proceeds as follows:

• Select $$k(\mathbf{x},\mathbf{x}') = \sum_{\mathbf{n} \in \mathbb{Z}^D} c_{\mathbf{n}} e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} - \mathbf{x}' \rangle }$$7 according to the desired sparsity/truncation strategy.
• Compute frequencies $$k(\mathbf{x},\mathbf{x}') = \sum_{\mathbf{n} \in \mathbb{Z}^D} c_{\mathbf{n}} e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} - \mathbf{x}' \rangle }$$8 and coefficients $$k(\mathbf{x},\mathbf{x}') = \sum_{\mathbf{n} \in \mathbb{Z}^D} c_{\mathbf{n}} e^{i\langle \boldsymbol{\omega}_{\mathbf{n}}, \mathbf{x} - \mathbf{x}' \rangle }$$9 (closed-form or via quadrature).
• For real-valued features, collect pairs $\mathbf{n}$0 to compute $\mathbf{n}$1 with weights $\mathbf{n}$2.
• The approximate kernel evaluation becomes a standard dot product in the ISFSF feature space.

Pseudo-code for index-set construction, feature map evaluation, and approximate kernel computation is provided in (Tompkins et al., 2018).

Computational complexity per data point is $\mathbf{n}$3 for feature evaluation, and $\mathbf{n}$4 for kernel computation. Index-set enumeration may be $\mathbf{n}$5 in worst-case but is mitigated by hyperbolic sparsification.

4. Theoretical Properties and Error Bounds

The error in approximating the true kernel by ISFSF vanishes as the index set grows. For tensor truncations with cutoff $\mathbf{n}$6, the tail bound is

$\mathbf{n}$7

and for the periodic squared exponential kernel, the truncation error $\mathbf{n}$8 decays exponentially in $\mathbf{n}$9 (Tompkins et al., 2018).

For ANOVA-based ISFSF decompositions, high-order interaction terms decay exponentially in their order $$\boldsymbol{\omega}_{\mathbf{n}} = \frac{2\pi}{T}\mathbf{n}$$0 for regular functions, controlling approximation error. The total error for the fitted predictor can be decomposed as

$$\boldsymbol{\omega}_{\mathbf{n}} = \frac{2\pi}{T}\mathbf{n}$$1

where each term can be made small by proper selection of $$\boldsymbol{\omega}_{\mathbf{n}} = \frac{2\pi}{T}\mathbf{n}$$2, number of features, and regularization (Potts et al., 2024).

5. Comparisons with Random and Quasi-Monte Carlo Fourier Features

ISFSF differs fundamentally from Random Fourier Features (RFF) and Quasi-Monte Carlo (QMC) approaches:

• ISFSF: Deterministically selects frequencies via the structural Fourier expansion of the kernel, yielding low-variance, tunable, and rapidly converging features. Enables explicit inclusion of anisotropy and control over harmonics.
• RFF: Employs Monte Carlo sampling of frequencies from the kernel’s spectral density, resulting in $$\boldsymbol{\omega}_{\mathbf{n}} = \frac{2\pi}{T}\mathbf{n}$$3 convergence and high variance when resolving sharp periodic structure or high-dimensional problems.
• QMC Fourier features: Use low-discrepancy sequences for improved convergence over pure RFF, but still lack deterministic, structure-aligned selection.

Empirical results demonstrate that ISFSF achieves equivalent or lower error with an order of magnitude fewer features than RFF and QMC methods, particularly for high-dimensional and strongly periodic data. For example, in regression on periodic textures, 49 ISFSF features (ENHC) matched the accuracy of approximately 793 RFF features (Tompkins et al., 2018).

6. Experimental and Applied Contexts

ISFSF has been implemented and evaluated in several domains:

• Periodic kernel approximation: ISFSF provides low-normalized Frobenius error in reconstructing large kernel matrices, with ENHC sets maintaining low error at high dimensions ($$\boldsymbol{\omega}_{\mathbf{n}} = \frac{2\pi}{T}\mathbf{n}$$4) (Tompkins et al., 2018).
• Large-scale regression: In Gaussian process regression for periodic textures, ISFSF features lead to lower RMSE and negative log loss compared to RFF and QMC baselines.
• Arithmetic Fourier geometry: Using index sets of primes and arithmetic weights $$\boldsymbol{\omega}_{\mathbf{n}} = \frac{2\pi}{T}\mathbf{n}$$5, planarly sampled ISFSF curves reveal experimental, possibly fractal geometry, with self-similar structure as $$\boldsymbol{\omega}_{\mathbf{n}} = \frac{2\pi}{T}\mathbf{n}$$6 grows. These findings are experimental; no explicit dimension results are proven (Vartziotis, 12 Feb 2026).
• ANOVA sensitivity analysis: ISFSF generalizes RFF models to explicit low-order interaction decompositions, enabling interpretable variable importance discovery and improved data efficiency (Potts et al., 2024).

7. Number-Theoretic and Spectral Connections

A distinct line arising from (Vartziotis, 12 Feb 2026) considers ISFSF generated by arithmetic structures. Let $$\boldsymbol{\omega}_{\mathbf{n}} = \frac{2\pi}{T}\mathbf{n}$$7 denote the sum of prime factors of $$\boldsymbol{\omega}_{\mathbf{n}} = \frac{2\pi}{T}\mathbf{n}$$8 (with multiplicity), $$\boldsymbol{\omega}_{\mathbf{n}} = \frac{2\pi}{T}\mathbf{n}$$9, and $c_{\mathbf{n}}$0 the $c_{\mathbf{n}}$1-adic valuation. Then

$c_{\mathbf{n}}$2

motivates the sparse, prime-indexed Fourier series

$c_{\mathbf{n}}$3

The $c_{\mathbf{n}}$4 norm satisfies

$c_{\mathbf{n}}$5

The continuous exponential $c_{\mathbf{n}}$6 is interpreted as the infinite-dimensional analogue of discrete Fourier polygons (eigenvectors of circulant Hermitian matrices), with sampled $c_{\mathbf{n}}$7 curves exhibiting experimentally observed fractal-like structure in the plane (Vartziotis, 12 Feb 2026).

• "Index Set Fourier Series Features for Approximating Multi-dimensional Periodic Kernels" (Tompkins et al., 2018)
• "ANOVA-boosting for Random Fourier Features" (Potts et al., 2024)
• "Fourier Series Generated by Additive Prime Factor Functions" (Vartziotis, 12 Feb 2026)