Optimal estimation and confidence sets for discontinuities in noisy, blurred regression functions

点击翻译按钮获取中文摘要

Nonparametric estimation of smooth univariate (1d) and bivariate (2d) regression functions is a very well-developed theory. However, in the presence of a discontinuity (jump) in a 1d regression, or of a discontinuity curve (edge) in a 2d image function, standard smoothing methods like local polynomial estimation do no longer work without modifications. Further, the discontinuity point and height in 1d as well as the location of the edge and its contrast in 2d may be of substantial independent scientific interest. Currently, no generally appropriate methods are available neither for the construction of confidence intervals for jumps in 1d regression curves which are observed under blurring, that is, convolution with a point-spread function, nor for uniform confidence sets for edges in 2d images. Developing and investigating confidence intervals and sets in these problems is therefore the first main aim of the project. For 2d image functions, often an additional blurring occurs which must be taken into account in reconstructions of the image itself as well as of edges contained in the image. The systematic treatment of deblurring (in addition to denoising) in edge estimation is therefore the second main aim of the project. Both the theoretical optimality properties of edge estimators and the construction and implementation of more practical recovering methods shall be investigated.