The maximum of a Gaussian process with nonconstant variance: A sharp bound for the distribution tail (Q751036)

Let X(t) be a Gaussian process on [0,1] and \(d_ x\) the pseudometric defined by \(d_ X(s,t)=E((X(t)-X(s))^ 2)^{1/2}.\) For each \(\epsilon >0\), let \(N_ X(\epsilon)\) be the smallest number of \(\epsilon\)-balls which cover [0,1]. If the function \(Q_ X(x)=\int^{x/2}_{0}H(N_ X(\epsilon))d\epsilon\) with \(H(u)=\max (\sqrt{\log u},1)\) is finite, then it follows from a result of Dudley that X has a version with continuous sample paths. The authors assume that \(Q_ X(x)\) is finite and that \(d_ x(s,t)\leq \phi (| s-t|)\) for some continuous nondecreasing function \(\phi\) with \(\phi (0)=0\). Under the additional assumption that \(\sigma^ 2(t)=E(X^ 2(t))\) has a unique argmax, say \(\tau\), they obtain an improved bound for the behaviour of P(\(\max_{0\leq t\leq 1} X(t)\geq u)\) for large u. By means of examples they show that the bound is sharp for a large class of processes.