The paper tackles a classical problem in statistics: how well does a certain test statistic behave in the tails of its distribution when sample sizes are moderate rather than enormous? The specific statistic studied is a multivariate version of the self-normalized sum, which is closely related to Hotelling's T-squared test, a standard tool for testing whether a multivariate population mean is zero. The main result shows that the tail probabilities of this statistic closely match those of a chi distribution (the length of a standard normal random vector) even far out into the tails, as long as the underlying data has a finite third moment. This improves on earlier work that required a slightly stronger moment condition, closing a long-standing gap in the theory.
The technical challenge is that when you go far into the tails of a distribution, small errors in approximation can become relatively large, so getting sharp results requires careful control of every source of error. The authors handle this by first truncating the random vectors at a scale dictated by the third moment condition, which keeps the covariance matrix well-behaved. They then apply a "whitening" transformation to make the problem rotationally symmetric, adding a small ridge term (a regularization trick borrowed from machine learning and numerical linear algebra) to avoid problems when the sample covariance matrix is nearly singular. This converts the original problem into one about how probability mass is distributed across spherical caps, a more tractable geometric object.
The final steps use two classical mathematical tools in a clever sequence. First, an averaging argument reduces the multidimensional problem to a collection of one-dimensional problems, one for each direction in space. Second, a technique called Abel inversion, which is a way of recovering a radial distribution from its projections, reconstructs the full tail behavior from these directional results. At the end, deterministic comparison arguments remove the ridge regularization and the truncation, recovering the result for the original statistic. The geometric framing throughout the proof is what makes the argument cleaner and more flexible than purely algebraic approaches used in prior work.