This paper is concerned with signal detection in Gaussian noise under quadratically convex orthosymmetric (QCO) constraints. Specifically the null hypothesis assumes no signal, whereas the alternative considers signal which is separated in Euclidean norm from zero, and belongs to the QCO constraint. Our main result establishes the minimax rate of the separation radius between the null and alternative purely in terms of the geometry of the QCO constraint -- we argue that the Kolmogorov widths of the constraint determine the critical radius. This is similar to the estimation problem with QCO constraints, which was first established by Donoho et al. (1990); however, as expected, the critical separation radius is smaller compared to the minimax optimal estimation rate. Thus signals may be detectable even when they cannot be reliably estimated.

翻译：暂无翻译