Optimal statistical estimators of spectral density in \(L^ 2\) (Q1069641)

Spectral estimators in the following form of the smoothed periodograms \[ (1)\quad \hat f(\lambda)=(2\pi)^{-1}\sum_{| k| <N}\alpha_ kc_ ke^{-ik\lambda}\quad (c_ k=(N-k)^{-1}\sum X_{t+k}X_ t) \] of a stationary random sequence \(X=\{X_ t\}\) are investigated. Under a priori information \(X\in {\mathfrak X}\), where \({\mathfrak X}\) is some set of random sequences, minimax estimators \(f^*\) minimizing the maximal risk \[ R_ N(\hat f,{\mathfrak X})=\sup_{X\in {\mathfrak X}}E_ X\| \hat f- f\|^ 2_{L_ 2} \] over all \(\hat f\) of the form (1) are constructed for various sets \({\mathfrak X}\) given by restrictions on the covariance function and the fourth cumulant of the sequences X. The rate of convergence of the maximal risk to zero as \(N\to \infty\) is investigated and is compared with that of classical estimators. The best order of decrease of the maximal risk for these estimators is also established.