Limiting eigenvalue distribution of heavy-tailed Toeplitz matrices (Q6896728)

This paper studies the empirical spectral distribution of random symmetric Toeplitz matrices whose entries are i.i.d. and lie in the domain of attraction of an \(\alpha\)-stable law with \(0<\alpha<2\). After an appropriate normalization, the authors prove that the empirical spectral distribution converges weakly, as the matrix size tends to infinity, to a random symmetric probability measure on \(\mathbb R\). This probability measure arises as a numerical spectral measure of a linear operator. The authors further analyze properties of the limiting distribution. It is symmetric around \(0\) and subgaussian. Its support is almost surely bounded when \(0<\alpha<1\), but unbounded when \(1\le \alpha<2\).\N\NIn the proof the Toeplitz matrix is interpreted as the principal submatrix of a circulant matrix, which can be diagonalized. Several nice truncations arguments, allowed by the heavy-tailed entries, follow. I will finish by including a paragraph from the authors that describe the paper particularly well:\N\N``In this article, we establish the limiting eigenvalue distribution of the heavy-tailed Toeplitz matrix. The limiting distribution of the symmetric circulant matrices with heavy-tailed entries was studied in [\textit{A. Bose} et al., Stat. Probab. Lett. 81, No. 11, 1706--1716 (2011; Zbl 1229.15031)]. For circulants, the eigenvalues are just the discrete Fourier transform of the first row. The eigenvalues of Toeplitz matrices, however, lack such explicit representation. We employ the standard method of moments, not on the original matrix but the one obtained by conjugating it with the discrete Fourier transform matrix. Since we are working with heavy-tailed entries, we truncate them at an appropriate level so that the truncated random variables have all finite moments. Our key observation is that after the Fourier conjugation, the heavy-tailed Toeplitz matrix is well-approximated by a random band matrix where the width of the band depends on the approximation level. This allows us to find a limiting random operator on \(\ell^2(\mathbb Z)\), whose law only depends on \(\alpha\), and the eigenvalue limiting distribution is connected to the spectral measure of this operator. In contrast to the light-tailed case, the limiting measure for heavy-tailed Toeplitz turns out to be random. The source of this randomness is the magnitudes and locations of the largest entries of the heavy-tailed random variables. Interestingly, unlike the heavy-tailed Wigner matrix, the limiting measure is not heavy-tailed. In fact, it has a subgaussian tail almost surely. We hope that this limiting operator can further be analyzed to extract more information about the limiting measure. ''