跳转至

多维空间中随机特征方法的谱收敛性

文章背景与核心概要

本文研究了多维目标函数下随机特征方法（Random Feature Method, RFM）的理论基础。作者证明了其在多种函数空间（包括索伯列夫空间、热夫雷空间、超解析空间和带限函数类）中的谱收敛性，确立了单一随机空间能够以从超指数级到代数级不等的速率同时逼近不同目标的特性。

此外，该研究将这些逼近界转化为多维二阶椭圆型边值问题和特征值问题的收敛性估计，并揭示了一个根本性的数值权衡：驱动高谱精度的底层机制，同时也导致了随机特征矩阵出现严重的病态性。

Spectral Convergence of Random Feature Method in Multiple Dimensions

元数据

  • arXiv ID: 2609.03401
  • 学科领域: 数值分析 (math.NA)、人工智能 (cs.AI)、机器学习 (cs.LG)、统计学理论 (math.ST)
  • 作者: Pingbing Ming, Hao Yu
  • 提交日期: 2026年9月3日
  • 篇幅: 48页，1个图表，2个表格

摘要 (Summary)

本文研究了针对多维目标的随机特征方法（RFM）的理论基础。作者证明了其在各种函数空间（包括索伯列夫空间、热夫雷空间、超解析空间和带限函数类）中的谱收敛性，确立了单一随机空间可以同时以超指数级到代数级的速率逼近目标。此外，该工作将这些逼近界转化为多维二阶椭圆型边值问题和特征值问题的收敛性估计，同时揭示了一个根本性的数值权衡：正是促成高谱精度的机制，同时引发了随机特征矩阵中严重的病态性。

This paper investigates the theoretical foundations of the Random Feature Method (RFM) for multidimensional targets. The authors prove spectral convergence across various function spaces—including Sobolev, Gevrey, ultra-analytic, and bandlimited classes—establishing that a single randomized space can simultaneously approximate targets with super-exponential to algebraic rates. Furthermore, the work translates these approximation bounds into convergence estimates for multidimensional second-order elliptic boundary value and eigenvalue problems, while uncovering a fundamental numerical trade-off: the very mechanisms driving high spectral accuracy also cause severe ill-conditioning in random feature matrices.

1. 谱收敛与逼近估计 (Spectral Convergence & Approximation Estimates)

作者证明了 RFM 在索伯列夫、热夫雷、超解析和带限类别中针对多维目标的谱收敛性。该分析在由核积分算子生成的插值尺度内建立了通用的高概率逼近估计。在仅由采样特征决定的单一事件上： * 一个随机空间可以逼近规定源球（source ball）中的每一个目标。 * 对于每个目标，单个系数向量所定义的逼近元能够在所有容许的误差范数下同时达到谱精度。 * 根据目标正则性的不同，对于正则性自适应的频率分布以及不断增长的频率窗口上的均匀分布，收敛速率从超指数级到代数级不等。

  1. Spectral Convergence & Approximation Estimates: The authors prove spectral convergence of the RFM for multidimensional targets in Sobolev, Gevrey, ultra-analytic, and bandlimited classes. The analysis establishes general high-probability approximation estimates within the interpolation scale generated by a kernel integral operator. On a single event determined exclusively by sampled features:
  2. One random space approximates every target in a prescribed source ball.
  3. For each target, a single coefficient vector defines an approximant attaining spectral accuracy simultaneously in all admissible error norms.
  4. Depending on target regularity, rates range from super-exponential to algebraic for both regularity-adapted frequency distributions and uniform distributions on growing frequency windows.

2. 离散化与椭圆型问题 (Discretization & Elliptic Problems)

针对强形式和弱形式的 RFM 离散化，本文均建立了抽象误差估计。这成功地将前述的逼近界转化为针对多维二阶椭圆型边值问题和特征值问题的严谨收敛性估计。

  1. Discretization & Elliptic Problems: Abstract error estimates are established for both strong- and weak-form RFM discretizations. This successfully converts the preceding approximation bounds into rigorous convergence estimates for multidimensional second-order elliptic boundary value and eigenvalue problems.

3. 矩阵条件数与病态性 (Matrix Conditioning & Ill-Conditioning)

对于随机特征矩阵（RFMtxs），本文证明了采用傅里叶特征时的超指数奇异值衰减和采用 \(\tanh\) 特征时的指数衰减，并给出了相应的条件数下界。这识别出一个统一的底层机制：正是带来高精度的同一谱逼近，同时导致了严重的矩阵病态性。

  1. Matrix Conditioning & Ill-Conditioning: For random feature matrices (RFMtxs), the paper proves super-exponential singular-value decay with Fourier features and exponential decay with \(\tanh\) features, alongside corresponding condition-number lower bounds. This identifies a unified underlying mechanism: the exact same spectral approximation that yields high accuracy simultaneously drives severe ill-conditioning.