{"id":"arxiv/2605.05043","url":"https://arxiv.org/abs/2605.05043","tags":["math.NA"],"score":0,"title":"Finding accurate eigenvalues and eigenvectors of positive semi-definite matrices given a subspace","domain":"math.NA","source":"arxiv","authors":["Yuji Nakatsukasa","Zheng Tang"],"reviews":0,"subject":"math.NA","summary":"We revisit a classical problem in numerical linear algebra: given an $k$-dimensional subspace $\\mathcal{Q}$ that approximates the leading eigenspace of an $n\\times n$ positive semi-definite matrix $A$, the goal is to extract high-accuracy eigenvalues. The Rayleigh-Ritz (RR) method is the standard algorithm for the task, which has been shown to be optimal in several ways (when $A$ is symmetric, not necessarily positive semi-definite $A\\succeq 0$). In this paper, we show that when $A \\succeq 0$, alternative methods can outperform RR, while having the same computational complexity, that is, the…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-05-06","reactions":{}}