Testing appearance of polynomial trend (Q1979088)

A change-point problem with polynomial trend after the change-point is considered. The observations are supposed to be \(Y_i=g(i,k)+e_i\), where \(e_i\) are i.i.d. with \(E e_i=0\), \(E e_i^2=1\), \(E e_i^4<\infty\), \(g(i,k)=0\) if \(i\leq k\), \(g(i,k)=g(i,k,a_1,\dots,a_m) =a_1(i-k)/n+\dots + a_m((i-k)/n)^m\) for \(i=k+1,\dots,n\). Here \(a_l\) and \(k\) are unknown. The problem is to test the hypothesis \(H_0\): \(k=n\) (no change) against the alternative \(A\): \(k\in\{1,\dots,n-m\}\). The maximum of the process \[ \chi_n^2(k)=\min_{a_1,\dots,a_m} \sum_{j=1}^n(Y_j-g(j,a_1,\dots,a_m))^2. \] is considered as a test statistic. The article is devoted to the estimation of the distribution of \(S_n=\max_{1\leq k\leq (1-\alpha)n}\chi_n^2(k)\) as \(n\to\infty\) under \(H_0\). The main result is that \[ \operatorname{Pr}_{H_0}(S_n>u^2)\approx (1-F_m(u^2))+(-\ln(\alpha))C_m u^{m-1}\exp(-u^2/2),\qquad (n\to\infty, u\to\infty), \] where \(F_m\) denotes the \(\chi^2\)-distribution with \(m\) degrees of freedom, and \(C_m\) is a real constant (\(C_2=8/\sqrt{6\pi^3}\)).