Super-resolution from noisy data (Q485201): Difference between revisions

The article is devoted to super-resolution techniques, which recover high-resolution information from coarse-scale data. There are many applications of these techniques -- in optical imaging, astronomy, medical imaging, spectroscopy, radar, geophysics, etc. In [\textit{E. J. Candès} and \textit{C. Fernandez-Granda}, Commun. Pure Appl. Math. 67, No. 6, 906--956 (2014; Zbl 1350.94011)], the authors considered the problem of recovering superpositions of point sources in a noiseless setting. The present article deals with a setting where the data are contaminated with noise. In this case, it is impossible to achieve infinite precision. The authors give a characterization of the estimation error as a function of the noise level and of the resolution desired. It is shown that it is possible to super-resolve point sources from noisy data with high precision via complex optimization. As it is outlined in the abstract, assume that we only have information about the spectrum of an object in the low-frequency band \(\left[-f_{\text{lo}},f_{\text{lo}}\right]\) and need to obtain a higher resolution estimate by extrapolating the spectrum up to a frequency \(f_{\text{hi}}>f_{\text{lo}}\). It is shown that if the sources are separated by \(2/f_{\text{lo}}\), then solving a convex program produces a stable estimate in the sense that the approximation error between the higher-resolution reconstruction and the truth is proportional to the noise level times the square of the super-resolution factor \(f_{\text{hi}}/f_{\text{lo}}\). The main result of the article is essentially an estimate of a high-resolution error under this separation condition. The article should be interesting for specialists in Signal Processing, Applied Harmonic Analysis, Physics, Medicine, and many other areas of science.