Detecting relevant changes in the mean of nonstationary processes -- a mass excess approach (Q2284384)

The authors consider the location scale model $X_{i,n}=\mu(i/n)+\epsilon_{i,n}$ where $\epsilon_{i,n}:i=(1,\dots,n)_{n\in N}$ denotes a triangular array of centered random variables, $\mu:[0,1]\to R$ is the unknown mean function. For $c>0$, the set of all points $t\in[0,1]$ where the mean function differs from its original value at point 0 by an amount larger than $c$ is defined as the level set, \[ \mathcal{M}_c=t\in[0,1]:|{\mu} (t)-{\mu} (0) |>c. \] Define $T_c:=\lambda(\mathcal{M}_c)$ the corresponding excess measure, $\lambda$ being the Lebesgue measure. The aim is to investigate the hypothesis \[ H_0:T_c\leq\Delta\text{ versus }H_1:T_c>\Delta. \] The change is understood as relevant if $T_c$ is larger than the threshold $\Delta$. Because of practical purposes one investigates one-sided hypotheses ${H_0}^+$: ${T_c}^+=\lambda({\mathcal{M}_c}^+)\leq\Delta$ versus ${H_1}^+:{T_c}^+>\Delta$, ${H_0}^-:{T_c}^-=\lambda({\mathcal{M}_c}^-)\leq\Delta$ versus ${H_1}^-:{T_c}^->\Delta$. In practical applications the two parameters \(c\) and $\Delta$ have to be specified. The excess measures are estimated and their properties are studied. The approach used in the article is related to the sojourn time of a stochastic process. Asymptotic properties of the estimators are largely discussed as well as the bias correction and the bootstrap procedure. Some simulation results are shown. Some of the proofs have to be provided by a supplementary material, on \url{doi:10.1214/19-AOS1811SUPP}.