When searching for a weak signal in noisy data, statisticians often look at the global maximum of a random field. But the distribution of that maximum is notoriously hard to work with. The authors of this paper propose a different approach: instead of studying the maximum alone, they study the second maximum—the largest value of the field after the global maximum and its local gradient have been removed. This second maximum, together with the independent part of the Hessian at the maximum, turns out to carry enough information to characterize the distribution of the global maximum exactly.
Using an ad-hoc Kac-Rice formula, the authors derive the conditional distribution of the maximum given the second maximum and the independent Hessian part. From this, they construct a spacing test based on the gap between the two maxima. The test is exact, not asymptotic, and works for any Gaussian random field on a Riemannian manifold. When the covariance function is known only up to a scaling factor, they provide a Studentized version, the t-spacing test, which is perfectly calibrated under the null hypothesis and has high power for detecting sparse alternatives.
The method is demonstrated on three concrete problems: detecting sparse alternatives in Gaussian symmetric tensors, continuous sparse deconvolution, and two-layer neural networks with smooth rectifiers. A notable corollary is an exact, non-asymptotic test for the spiked tensor model that does not require prior knowledge of the noise level. Numerical experiments confirm the calibration and power of the proposed procedures.
The latest version of the paper also makes explicit the connection to the earlier Kac-Rice test of Taylor, Loftus, and Tibshirani, and adds a new section on the behavior of close maxima. This positions the spacing test as both a theoretical refinement and a practical tool for high-dimensional inference.