Abstract
Estimating the tangent spaces of a data manifold is a fundamental problem indata analysis. The standard approach, Local Principal Component Analysis(LPCA), struggles in high-noise settings due to a critical trade-off inchoosing the neighborhood size. Selecting an optimal size requires priorknowledge of the geometric and noise characteristics of the data that are oftenunavailable. In this paper, we propose a spectral method, Laplacian EigenvectorGradient Orthogonalization (LEGO), that utilizes the global structure of thedata to guide local tangent space estimation. Instead of relying solely onlocal neighborhoods, LEGO estimates the tangent space at each data point byorthogonalizing the gradients of low-frequency eigenvectors of the graphLaplacian. We provide two theoretical justifications of our method. First, adifferential geometric analysis on a tubular neighborhood of a manifold showsthat gradients of the low-frequency Laplacian eigenfunctions of the tube alignclosely with the manifold's tangent bundle, while an eigenfunction with highgradient in directions orthogonal to the manifold lie deeper in the spectrum.Second, a random matrix theoretic analysis also demonstrates that low-frequencyeigenvectors are robust to sub-Gaussian noise. Through comprehensiveexperiments, we demonstrate that LEGO yields tangent space estimates that aresignificantly more robust to noise than those from LPCA, resulting in markedimprovements in downstream tasks such as manifold learning, boundary detection,and local intrinsic dimension estimation.