Fast localization of eigenfunctions via smoothed potentials (Q2063155)

The authors are interested in the problem of predicting localized eigenfunctions \(\phi\) of Schrödinger operators of the form \(-\Delta + V\) on bounded domains in \(\mathbb{R}^d\), where the potential \(V\) is assumed to be rapidly varying. They build on a recent work of the third author [Commun. Partial Differ. Equations 46, No. 7, 1262--1279 (2021; Zbl 1487.35200)] who introduced an alternative to the landscape function approach of \textit{M. Filoche} and \textit{S. Mayboroda} [``Universal mechanism for Anderson and weak localization'', Proc. Natl. Acad. Sci. USA, 109, No. 37, 14761--14766 (2012; \url{doi:10.1073/pnas.1120432109})], who use the minima of the function \(1/u\), where \(u\) solves \((-\Delta + V)u=1\) (plus boundary conditions), to predict the location of localized eigenfunctions very well. The alternative approach introduced by the third author comes from observing that the eigenvalue equation can be rewritten in a slightly smoothed form, \[ -\Delta \phi(x) + (V \ast k_t)(x)\phi(x) = \lambda\phi(x) + \text{error} (x,t) \] for a specially chosen kernel \(k_t\), a ``universal convolution kernel'' \[ k_t(x) = \frac{1}{t} \int_0^t \frac{\text{exp}(-\|x\|^2/4s)}{(4\pi s)^{d/2}}\,ds, \] such that the error can be controlled in terms of \(\phi\) and \(\|V\|_\infty\) only (for each fixed \(x\) and for the correct power of \(t\)). An analysis shows that the function \(k_t \ast V\) should then (for suitable values of \(t\)) act as a rapidly computable landscape function, with very similar behaviour to the one of Filoche and Mayboroda. In the paper under review, the authors give an alternative representation of \(k_t\) well-suited to numerical computations, as it reduces the problem of calculating \(k_t \ast V\) to two applications of the Fast Fourier Transform. They present the corresponding algorithm, and also sketch how it can be extended to similar operators such as fractional Laplacians \(-\Delta^\alpha + V\) and Bilaplacians \((-\Delta)^2+V\). They also conduct a numerical analysis in a prototypical setting, showing among other things that their landscape function does in fact approximate the one of Filoche and Mayboroda [loc. cit.] quite well.