Bounding suprema of canonical processes via convex hull (Q6150881)

The aim of the paper is to estimate the expected value of the supremum of the process \[ X_t=\sum_{i}t_iX_i, t=(t_1, \dots,t_n)\in T \subset \mathbf{R^n.} \] There is a long line of research devoted to bounding \(\mathbb{E}\sup_tX_t\) via the chaining method. However, chaining methods do not work well for heavy-tailed random variables. In this paper, the author investigates another approach based on the convex hull method. Upper bound is obtained easily, but the main question is how to reverse it, in the sense that for which processes and which sets \(T\) this upper bound can be inversed, in some sense? The main question is related to Talagrand conjectures about suprema of positive selector processes. Talagrand investigates the possibility of enclosing \(T\) into a solid convex hull, which is bigger than the convex hull. On the other hand, in the case under consideration, some regularity conditions on variables \(X_i\) are needed which is clearly not satisfied for nontrivial classes of selector processes. The author considers cases when the specified inversion is possible and the corresponding conditions, as well as examples when such inversion is impossible. For the entire collection see [Zbl 07730209].