Weighted STSAFT-Modulation Spaces and Approximation by Phase-Adapted Neural Networks

We develop a transform-adapted approximation framework for shallow neural networks based on the short-time special affine Fourier transform (STSAFT). Replacing the short-time Fourier transform (STFT) in the definition of classical modulation spaces, we work with weighted STSAFT-modulation spaces, building on the existing SAFT-associated modulation-space framework, and use their chirp correspondence with classical weighted modulation spaces through the chirp factorization of the SAFT kernel. We construct an associated neural dictionary and establish an $N$-term Sobolev approximation bound of order $N^{-1/2}$ on bounded one-dimensional domains. A global approximation theorem is also obtained when the spatial translation parameter ranges over the whole real line; a compact-support argument explains why an unrestricted global conclusion cannot in general be inferred from a bounded translation range. For a fixed-scale chirp family $f_κ(x)=e^{iπκx^2}u(x)$, we combine the global estimate with a phase analysis to obtain the stronger phase-dependent upper-bound factor $\langleλ\rangle^n\langleκ-λ\rangle^{n+3/2-1/r}.$ For each fixed nonzero Schwartz envelope, matching the dictionary phase reduces the growth of this specified upper bound from order $|κ|^{n+3/2-1/r}$ to order $|κ|^n$, while preserving the $N^{-1/2}$ rate. The associated mismatch scale is stated as a comparison of upper bounds, not as a threshold for the optimal approximation error. Four numerical experiments examine model comparison, error decay with dictionary size, phase mismatch, and the obstruction caused by restricted translations. We discuss the effective number of fitted output coefficients, including the rank loss of the sine branch at zero phase, and distinguish the numerical diagnostics from the theoretical best-approximation estimates.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-12
DOI
https://doi.org/10.48550/arxiv.2611.16448
Primary Topic
Mathematical Analysis and Transform Methods
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Weighted STSAFT-Modulation Spaces and Approximation by Phase-Adapted Neural Networks

Bei Liu, Ming Hu, Bolin Liu, Jie Shen
arXiv (Cornell University)
Mathematical Analysis and Transform Methods
preprint

Weighted STSAFT-Modulation Spaces and Approximation by Phase-Adapted Neural Networks

Bei Liu, Ming Hu, Bolin Liu, Jie Shen
preprint en

Abstract

We develop a transform-adapted approximation framework for shallow neural networks based on the short-time special affine Fourier transform (STSAFT). Replacing the short-time Fourier transform (STFT) in the definition of classical modulation spaces, we work with weighted STSAFT-modulation spaces, building on the existing SAFT-associated modulation-space framework, and use their chirp correspondence with classical weighted modulation spaces through the chirp factorization of the SAFT kernel. We construct an associated neural dictionary and establish an $N$-term Sobolev approximation bound of order $N^{-1/2}$ on bounded one-dimensional domains. A global approximation theorem is also obtained when the spatial translation parameter ranges over the whole real line; a compact-support argument explains why an unrestricted global conclusion cannot in general be inferred from a bounded translation range. For a fixed-scale chirp family $f_κ(x)=e^{iπκx^2}u(x)$, we combine the global estimate with a phase analysis to obtain the stronger phase-dependent upper-bound factor $\langleλ\rangle^n\langleκ-λ\rangle^{n+3/2-1/r}.$ For each fixed nonzero Schwartz envelope, matching the dictionary phase reduces the growth of this specified upper bound from order $|κ|^{n+3/2-1/r}$ to order $|κ|^n$, while preserving the $N^{-1/2}$ rate. The associated mismatch scale is stated as a comparison of upper bounds, not as a threshold for the optimal approximation error. Four numerical experiments examine model comparison, error decay with dictionary size, phase mismatch, and the obstruction caused by restricted translations. We discuss the effective number of fitted output coefficients, including the rank loss of the sine branch at zero phase, and distinguish the numerical diagnostics from the theoretical best-approximation estimates.

arXiv (Cornell University)
Mathematical Analysis and Transform Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon