Comparing two nonparametric regression curves in the presence of long memory in covariates and errors (Q2174527)

Two stochastic processes $Y_{1,j}$ and $Y_{2,j}$, $j=1,2,\dots,n$ are observed, \[ Y_{1,j}=\mu_1(X_j+\sigma_1u_{1,j}), \] \[ Y_{2,j}=\mu_2(X_j +\sigma_2u_{2,j}). \] Here $\mu_1$ and $\mu_2$ are two real-valued regression functions, $\sigma_1$ and $\sigma_2$ are two positive numbers and the errors $u_{i ,j}=\displaystyle\sum_{k=0}^{\infty} b_{i ,k}{\epsilon}_{i,j-k}$, $i = 1,2$, where $\epsilon_{1,j}$ and $\epsilon_{2,j}$ are two independent sequences of i.i.d. standardized random variables. The common covariate process $X_j$ is assumed to be a stationary long memory moving average process. The stated problem is to test the null hypothesis $H_0:\mu_1(x)=\mu_2(x)$, for $x\in [a,b]$ against the two-sided alternative hypothesis: $H_a:\mu_1(x)-\mu_2(x) =\delta(x)\neq 0$ for some $x\in [a,b]$, based on data $X_j$, $Y_{1,j}$, $Y_{2,j}$, $j=1,2,\dots,n$, provided from the model. One introduces a process $U_n$ as a marked empirical process with the marks $D_j:=Y_{1,j}-Y_{2,j}$ and one bases tests on Kolmogorov-Smirnov-type functions. Under some quite special assumptions, results on estimators, asymptotic behaviors and Monte Carlo simulations are shown.