Random matrix theory studies how the eigenvalues of large matrices behave when those matrices are built from many independent random numbers. A "Wigner matrix" is a symmetric random matrix where the entries are independent and identically distributed. A "spiked" Wigner matrix adds a small structured signal on top of this noise, which is relevant to problems like detecting a weak signal buried in high-dimensional data. A "linear spectral statistic" (LSS) is a summary of the eigenvalue distribution formed by applying some smooth function to all the eigenvalues and adding the results up. The central goal of this paper is to prove rigorous theorems describing the mean and fluctuations of these summaries, even when the individual matrix entries have been passed through a potentially rough, nonlinear transformation.
The main technical challenge is that classical results in this area typically require the transformation applied to the entries to be smooth or at least well-behaved in ways that allow standard calculus tools. This paper gets around that requirement using a clever three-step strategy. First, at each level of the analysis, the authors construct a simpler "stand-in" random variable that exactly matches the first four statistical moments of the actual transformed entry but is bounded and therefore easier to handle. Second, they apply a known LSS theorem to these well-behaved stand-ins. Third, they use a Fourier-based comparison estimate to transfer the result back to the original rough transformation without needing differentiability or other restrictive smoothness conditions. The only requirements on the transformation are that it be centered, have the right variance, and satisfy a mild stability condition on its fourth-order behavior under small shifts induced by the signal.
The paper works out in detail what the mean and variance of the LSS look like in the final answer. The mean has four interpretable pieces: a standard bulk contribution from the Wigner noise, a correction due to the rank-one signal, a correction from the zero-diagonal structure, and a correction from how the variance of the transformation varies. The variance of the fluctuations matches the standard formula for Wigner matrices, characterized by a single fourth-cumulant parameter of the transformation. The authors also carefully distinguish between statistics that track only the bulk of the eigenvalue spectrum and those that include the one outlier eigenvalue that separates from the bulk when the signal is strong enough, showing exactly how that outlier shifts the centering. Concrete examples and criteria for when the theorem applies or fails are provided throughout.