Wavelet smoothing of evolutionary spectra by nonlinear thresholding (Q1815715)

The contents of the paper are as follows: A two-dimensional periodic multiresolution and a collocation wavelet transform is recalled in Section 2. Then, in Section 3, we review the concept of evolutionary spectra of a locally stationary process, as described by \textit{R. Dahlhaus} [Fitting time series models to nonstationary processes. Beiträge z. Statistik 4, Univ. Heidelberg (1993)]. Moreover, we introduce the short-time periodogram as a localized periodogram over segments of the original time series. Section 4, the main part, treats wavelet estimation of the evolutionary spectrum by nonlinear thresholding of the empirical coefficients, i.e., the coefficients of the localized periodogram estimates. We first investigate their statistical behavior and estimate their tail probability to end up with an appropriate threshold. Our main theorem shows that the resulting threshold estimator achieves a near-optimal rate of convergence of the \(L_2\)-risk, i.e., the integrated mean-squared error between estimate and true spectrum. At this place, we restrict to rigorously derive the results for Hölder function classes only, mainly in order to simplify the proofs. However, we were able to generalize our main results to more interesting functions of lower inhomogeneously distributed regularity. With this, basically continuous functions of bounded variation are included. Finally, Section 5 deals with applications and simulations.