Oracle inequality for a statistical Raus-Gfrerer type rule
Provides theoretical guarantees for a parameter selection method in linear inverse problems, relevant to statisticians and practitioners in inverse problems.
The paper studies statistical linear inverse problems in Hilbert spaces, analyzing a variant of the Raus-Gfrerer parameter choice rule. It shows that this rule yields oracle inequalities leading to order optimal error bounds (up to logarithmic factors) for solutions with a self-similarity structure.
The authors study statistical linear inverse problems in Hilbert spaces. Approximate solutions are sought within a class of linear one-parameter regularization schemes, and the parameter choice is crucial to control the root mean squared error. Here a variant of the Raus{Gfrerer rule is analyzed, and it is shown that this parameter choice gives rise to error bounds in terms of oracle inequalities, which in turn provide order optimal error bounds (up to logarithmic factors). These bounds can only be established for solutions which obey a certain self-similarity structure. The proof of the main result relies on some auxiliary error analysis for linear inverse problems under general noise assumptions, and this may be interesting in its own.