Ideal denoising for signals in sub-Gaussian noise (Q2470539)

The main subject of the paper is oracle based denoising. It is devoted to the extension and improvement of the already known algorithm. The study is concentrated on the incorporation of a wider class of noise vectors, maintaining the properties of this algorithm. A signal embedded in noise is under consideration. The paper presents the results of the development of estimators for such signals, obtained by thresholding coefficients in an orthonormal basis. The problem is considered when there is a library of such bases, from which it is possible to choose the optimal one. The authors are looking for estimators which are satisfying (with high probability) the well known oracle-type inequality for the given basis. Such estimators make use of knowledge of the signal and are of excellent quality, but they are unavailable in practice. After proper re-scaling, it can be argued that the considered type inequality is asymptotically optimal and the oracle risk decays in a best possible manner. Therefore, this type of inequality gives an a priori measure for the quality of the algorithm associated to the satisfying inequality estimators. An algorithm that allows the selection of an orthonormal basis from a library of bases had been already presented in the former publications. The main goal of this paper is to extend previous results to a wider class of noise vectors. The accepted assumption on the noise vector is that its coordinates are independent identically distributed Gaussian random variables. The authors are borrowing the noise set-up and related background results from the known publications, but they generalize the Gaussian hypothesis and only require the noise vector to satisfy a strict sub-Gaussian hypothesis. This set-up generalizes the obtained until now results. A particular theorem is proven that characterizes sub-Gaussian random variables and gives an indication for the wider scope of the study. For example, noise coordinates with the uniform distribution are included in the setting. The results are carefully obtained so that the algorithmic content of the previous results are preserved. In particular, the used thresholding parameters are the same as the ones used in the Gaussian case. The authors also take the opportunity to improve on the value of some of the key parameters appearing in the main inequality. Their proof follows the known one but uses a classical argument to derive a more specific, relative to the given noise vector, concentration inequality. The essence of the presented approach is in the way that the oracle risk is defined for the given basis. It is easy to see that the oracle risk appearing in the inequality uses very little knowledge about the noise distribution. Actually, it only uses the variances of the noise coordinates. In fact the obtained results show that the key aspect of the noise distribution is the existence of an exponential second moment inequality. This result only depends on the tail decay of the Gaussian distribution. The effect of a change of coordinates is also considered. These observations are leading into the wide class of strict sub-Gaussian random vectors as a natural class where results of the considered type can be proven. The paper is organized as follows: Section 2 summarizes the main known results. Section 3 describes the set-up of strict sub-Gaussian noise vectors. The main result, the proof of Theorem 3, is presented in Section 4. Section 5 briefly discusses some technical issues. Appendix A is provided, where some of the properties of the required sub-Gaussian and strict sub-Gaussian noise vectors are stated. Finally, for the reader's convenience, Appendix B is given, where some intermediate results needed along the way are depicted.