Tracy-Widom law for the extreme eigenvalues of large signal-plus-noise matrices (Q6137710)

Let \(\mathbf{R}\) be a deterministic, rectangular matrix, called ``signal matrix'', and \(\textbf{X}\) be a real random matrix with independent entries, called ``noise matrix''. The so called ``signal-plus-noise matrix model'' \(\mathbf{S} = \mathbf{R} + \mathbf{X}\) has been studied in the past, with applications in machine learning and signal processing. The authors prove that under mild assumptions on \(\mathbf{R}\) and \(\mathbf{S}\), most notably a tail assumption on the entries of \(\mathbf{X}\), the largest eigenvalue of \(\mathbf{S}\mathbf{S}^{*}\) follows the Tracy-Widom distribution of type 1 in the large dimension limit, as do matrices of the Gaussian orthogonal ensemble, see [\textit{C. A. Tracy} and \textit{H. Widom}, Commun. Math. Phys. 177, No. 3, 727--754 (1996; Zbl 0851.60101)]. Furthermore, it is shown that the tail condition is necessary to obtain the convergence to the Tracy-Widom distribution. The proof is done by analyzing the Stieltjes transform of the model. The authors first prove the result in the case of Gaussian noise, using in particular an interpolation procedure relating the signal-plus-noise matrix to a Wishart matrix. In this part, their approach is similar to the one of [\textit{J. O. Lee} and \textit{K. Schnelli}, Rev. Math. Phys. 27, No. 8, Article ID 1550018, 94 p. (2015; Zbl 1328.15051)]. A technical part of the paper is devoted to deriving the so-called optical theorems, which give control over the Green kernel of the random matrix.