Abstract:We consider $N\times N$ self-adjoint Gaussian random matrices defined by an arbitrary deterministic sparsity pattern with $d$ nonzero entries per row. We show that such random matrices exhibit a canonical localization-delocalization transition near the edge of the spectrum: when $d\gg\log N$ the random matrix possesses a delocalized approximate top eigenvector, while when $d\ll\log N$ any approximate top eigenvector is localized. The key feature of this phenomenon is that it is universal with respect to the sparsity pattern, in contrast to the delocalization properties of exact eigenvectors which are sensitive to the specific sparsity pattern of the random matrix.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the localization - delocalization transition phenomenon of eigenvectors near the spectral edge in non - uniform random matrices. Specifically, researchers are concerned with whether the approximate maximum eigenvectors of the matrix are localized or delocalized when the number of non - zero elements \(d\) in each row is in different proportions to the logarithm \(\log N\) of the matrix size \(N\).
### Background and Problem Description of the Paper
In many fields such as mathematical physics, computer science, and combinatorics, understanding whether the eigenvectors of large random matrices are localized or delocalized is an important research topic. A delocalized eigenvector means that its mass is roughly evenly distributed over all coordinates, while a localized eigenvector is mainly concentrated on a few coordinates.
The paper discusses two extreme cases:
- **Completely Delocalized**: For example, the classical Gaussian Orthogonal Ensemble (GOE), whose eigenvectors are uniformly distributed on the unit sphere and are therefore always delocalized.
- **Completely Localized**: For example, a diagonal matrix, whose eigenvectors are concentrated on a single coordinate.
To interpolate between these two extremes, researchers consider models with different degrees of sparsity, such as random band matrices. These matrices exhibit a phase transition from localization to delocalization near the spectral edge, depending on the bandwidth.
### Research Object
This paper studies a class of \(N\times N\) self - adjoint Gaussian random matrices defined by an arbitrary deterministic sparse pattern. These matrices are characterized by having \(d\) non - zero elements in each row. Researchers are interested in the behavior of these random matrices near the spectral edge, especially whether their approximate maximum eigenvectors are localized or delocalized.
### Main Findings
The main findings of the paper are:
- When \(d\gg\log N\), the random matrix has a delocalized approximate maximum eigenvector.
- When \(d\ll\log N\), any approximate maximum eigenvector is localized.
The key to this phenomenon is that it is universal for sparse patterns, in contrast to the fact that the delocalization property of exact eigenvectors is sensitive to specific sparse patterns.
### Definitions and Results
The paper introduces some key definitions:
- **\((L,\kappa)\)-Delocalized**: A vector \(v\in S^{N - 1}\) is said to be \((L,\kappa)\)-delocalized if for any set \(A\subseteq[N]\) with size not exceeding \(L\), \(\sum_{j\in A}v_j^2\leq\kappa^2\) is satisfied.
- **Approximate Maximum Eigenvector**: If a vector \(v\in S^{N - 1}\) satisfies \(\|Xv\|^2\geq(1 - \epsilon)\|X\|\), then \(v\) is called a \((1 - \epsilon)\)-approximate maximum eigenvector.
The main result of the paper is Theorem 1.6, which describes the localization - delocalization transition:
- **Localization Region**: When \(d\ll\log N\), any \((1 - \epsilon)\)-approximate maximum eigenvector is localized.
- **Delocalization Region**: When \(d\gg\log N\), there exists a \((1 - o(1))\)-approximate maximum eigenvector that is delocalized.
### Conclusion
Through the above research, the paper reveals a new localization - delocalization phenomenon, which is universal for sparse patterns, thus providing a new perspective for understanding the behavior of eigenvectors of non - uniform random matrices.