Johnstone and Silverman (1997) described a level-dependent thresholding met
hod for extracting signals from correlated noise. The thresholds were chose
n to minimize a data based unbiased risk criterion. Here we show that in ce
rtain asymptotic models encompassing short and long range dependence, these
methods are simultaneously asymptotically minimax up to constants over a b
road range of Besov classes. We indicate the extension of the methods and r
esults to a class of linear inverse problems possessing a wavelet vaguelett
e decomposition.