Fast Spectrum Estimation of Some Kernel Matrices

A researcher has introduced a mathematical framework for estimating the eigenvalues of certain large kernel matrices without the computational expense of constructing the full matrix, according to a revised paper posted to arXiv [1]. The work, authored by Mikhail Lepilov and last revised on 25 May 2026, targets kernel matrices that arise frequently in data science tasks such as classification [1][2]. These matrices are formed from large sets of independent observations, and knowing how their eigenvalues decay is critical for determining whether a low-rank approximation is practical [2]. The new method provides meaningful bounds for all eigenvalues of a matrix, avoiding the cost of building the full structure [1]. The framework applies specifically to kernels with quick decay away from the diagonal, used on uniformly-distributed points in Euclidean space of any dimension [2]. The paper proves the method's efficacy given certain bounds on the kernel function and supplies empirical evidence for its accuracy [1]. In the process, a general interlacing-type theorem for finite sets of numbers is also proven [2]. The submission history shows an initial version from 1 Nov 2024 at 639 KB, followed by the updated version at 372 KB [1]. The author indicates a potential application of the framework to studying the intrinsic dimension of data [2]. Kernel matrices are central to many modern algorithms. In Gaussian processes, for instance, the kernel matrix encodes covariance relationships between data points, and its eigenvalue spectrum directly affects the feasibility of computations as datasets grow [4]. Approximation methods for Gaussian processes often aim to retain accuracy while drastically reducing computation time, a goal aligned with avoiding full matrix construction [4]. The challenge of handling large matrices efficiently has parallels in other domains. The fast Fourier transform, described by Gilbert Strang as "the most important numerical algorithm of our lifetime," reduces the complexity of computing a discrete Fourier transform from O(n²) to O(n log n) by factorizing the matrix into sparse factors [3]. Similarly, compressed sensing exploits signal sparsity to recover information from far fewer samples than traditional theorems require, relying on mathematical properties like incoherence rather than brute-force measurement [5]. The arXiv paper is hosted alongside experimental projects from arXivLabs, a framework for community collaborators developing new features on the platform [1].