Adaptive minimax testing in inverse Gaussian sequence space models

TL;DR

This paper derives optimal minimax testing radii in inverse Gaussian sequence models with noisy operator observations, establishing procedures that adapt to regularity with unavoidable logarithmic penalties.

Contribution

It introduces new minimax radii for testing in inverse Gaussian models with noisy operators, including adaptive procedures with proven optimality and unavoidable log-factor deterioration.

Findings

Minimax radii depend on noise levels and null hypothesis.

Adaptive tests achieve near-optimal radii with log-factor penalty.

Results apply to Sobolev spaces and ill-posed inverse problems.

Abstract

In the inverse Gaussian sequence space model with additional noisy observations of the operator, we derive nonasymptotic minimax radii of testing for ellipsoid-type alternatives simultaneously for both the signal detection problem (testing against zero) and the goodness-of-fit testing problem (testing against a prescribed sequence) without any regularity assumption on the null hypothesis. The radii are the maximum of two terms, each of which only depends on one of the noise levels. Interestingly, the term involving the noise level of the operator explicitly depends on the null hypothesis and vanishes in the signal detection case. The minimax radii are established by first showing a lower bound for arbitrary null hypotheses and noise levels. For the upper bound we consider two testing procedures, a direct test based on estimating the energy in the image space and an indirect test. Under mild assumptions, we prove that the testing radius of the indirect test achieves the lower bound, which shows the minimax optimality of the radius and the test. We highlight the assumptions under which the direct test also performs optimally. Furthermore, we apply a classical Bonferroni method for making both the indirect and the direct test adaptive with respect to the regularity of the alternative. The radii of the adaptive tests are deteriorated by an additional log-factor, which we show to be unavoidable. The results are illustrated considering Sobolev spaces and mildly or severely ill-posed inverse problems.