Multiscale scanning in inverse problems (Q1990595)

This paper introduces a novel method to detect active components of an unknown function of interest w.r.t. a prescribed dictionary \(\left\{\varphi_i\right\}_{i\in I}\) from indirect observations \(Y_j = T f\left(x_j\right) + \xi_j\). Here, \(T\) is a bounded linear operator acting between proper Hilbert spaces, \(x_j\) are deterministic sampling points, and \(\xi_j\) are independent errors. The method is based on a multiscale test statistic, which allows to test \(\left\langle \varphi_i, f\right\rangle = 0\) vs. \(\left|\left\langle \varphi_i, f\right\rangle\right|>0\) simultaneously over all subsets \(J \subset I\). The authors present a unified asymptotic theory for the considered global test statistic, which allows to calibrate the corresponding multiple hypothesis test universally and independent from the specific data set and reveals asymptotic minimax optimality. The method is then applied to an inverse problem from super-resolution fluorescence microscopy and its finite sample performance is investigated in a simulation study.