A geometric proof of multivariate self-normalized moderate deviations via ridge regularization

Abstract

Let $X,X_1,X_2,\ldots$ be i.i.d.\ random vectors in $\mathbb R^d$, where $d\ge1$ is fixed. Assume that $\mathbb{E} X=0$, $\mathbb{E} XX^\top=I_d$ and $\mathbb{E}\|X\|^3<\infty$. We prove a Cramér type moderate deviation theorem for a multivariate self-normalized radius statistic, that is, for every $a_n=o(n^{1/6})$, \[ \sup_{0\le x\le a_n} \left|\frac{\mathbb{P}(U_{n,+}\ge x)}{\mathbb{P}(χ_d\ge x)}-1\right|\to0, \] where $χ_d$ is the Euclidean norm of a standard normal vector in $\mathbb{R}^d$ and \[ U_{n,+}^2=S_n^\top A_n^+S_n, \qquad S_n=\sum_{j=1}^nX_j, \qquad A_n=\sum_{j=1}^nX_jX_j^\top. \] Here $A_n^+$ denotes the Moore--Penrose inverse. We further prove that the same result holds for the corresponding Hotelling $T^2$ statistic. This work closes the moment gap between the previously known finite $(3+δ)$ moment condition for Hotelling moderate deviations and the finite third moment boundary. Our proof adopts a geometric approach. Truncation at the finite-third-moment scale yields a truncated covariance matrix whose concentration error is negligible relative to the target \(χ_d\) tail. Uniform projected self-normalized moderate deviations are then averaged over directions and, after a whitening map with a ridge term, become a spherical cap transform of the radial distribution. A local Abel inversion recovers the radial tail, and deterministic comparison removes the ridge and the truncation.