Displacement structure and eigenvalue bounds for multiband prolate matrices

Abstract

We obtain explicit non-asymptotic bounds for the eigenvalue transition of multiband prolate matrices. Let $S\subset[-1/2,1/2]$ be a finite union of pairwise disjoint intervals, and let $λ_0\ge\cdots\geλ_{N-1}$ be the eigenvalues of $B_{N,S}$. If $ν$ is the number of boundary points of $S$ on the frequency torus, then, for every $0<\eps<1/2$, \[ \#\{k:\eps<λ_k<1-\eps\} \le ν\ceil{\frac{\log(4N)}{π^2} \log\!\left(\frac{4}{\eps(1-\eps)}\right)}. \] Compared with the multiband estimate of Zhu and Wakin~\cite{wakin2017approximating}, this improves the dependence on $\eps$ from order $1/[\eps(1-\eps)]$ to logarithmic dependence. We also prove a second bound in which the main logarithmic term depends on the time--bandwidth product $N|S|$ rather than only on $N$. We also compare $S$ with frequency sets having fewer boundary points. Using a classical large-sieve estimate, we take into account not only the total size of the change in the frequency set, but also how that change is distributed. For sets with many small gaps, this can give a bound growing like $\log N$ for fixed $\eps$, even when the bounds based only on the total size of the gaps grow at least linearly in $N$. We also obtain exponential bounds for individual eigenvalues and use them to estimate the error in approximating sampled multiband signals by leading eigenvectors. A discretization argument gives corresponding results for the continuous problem. Numerical examples compare the bounds and examine how the locations of the frequency intervals affect the eigenvalues.