Abstract:Inhomogeneous random matrices with a non-trivial variance profile determined by a symmetric stochastic matrix and with independent symmetrically sub-Gaussian entries up to Hermitian symmetry, include many prominent examples, such as remarkable sparse Wigner matrices and random band matrices in dimension d, and have been of great interest recently. The maximum of entry variances as a natural proxy of sparsity is both a key feature and a significant obstacle. In this paper, we study low-rank additive perturbations of such random matrices, and establish a sharp BBP phase transition for extreme eigenvalues at the level of law of large numbers. Under suitable conditions on the variance profile and the finite-rank perturbation, we also establish the fluctuations of spectral outliers that may be characterized by the general inhomogeneous random matrices. This reveals the strong non-universality phenomena, which may depend on eigenvectors, sparsity or geometric structure. Our proofs rely on ribbon graph expansions, upper bounds for diagram functions, large moment estimates and counting of typical diagrams.
What problem does this paper attempt to address?
This paper aims to study the eigenvalue behavior of inhomogeneous random matrices (IRM) under low - rank additive perturbations, especially the BBP phase - transition phenomenon of extreme eigenvalues and their fluctuation characteristics. Specifically:
1. **Research Background**:
- Inhomogeneous random matrices refer to random matrices with non - trivial variance distributions. These matrices are very important in many practical applications, such as sparse Wigner matrices, random banded matrices, etc.
- The maximum entry variance of such matrices, as a natural proxy for sparsity, is both a key feature and a significant obstacle.
2. **Research Questions**:
- **Question 1**: Under the given conditions, will the extreme eigenvalues of the deformed inhomogeneous random matrix \(X_N\) undergo BBP phase - transition?
- **Question 2**: If BBP phase - transition occurs, what are the fluctuation characteristics of the extreme eigenvalues?
3. **Main Results**:
- **BBP Phase - transition**: The paper proves that under appropriate variance distributions and finite - rank perturbation conditions, the extreme eigenvalues of \(X_N\) will undergo BBP phase - transition at the level of the law of large numbers. Specifically, for the \(j\)-th largest eigenvalue \(\lambda_j(X_N)\), when \(a_j\leq1\), \(\lambda_j(X_N)\) converges almost surely to 2; when \(a_j > 1\), \(\lambda_j(X_N)\) converges almost surely to \(\frac{a_j + 1}{a_j}\).
- **Fluctuation Characteristics**: In the supercritical case (i.e., \(a > 1\)), the paper further proves the fluctuation characteristics of the first \(q\) largest eigenvalues and gives the expressions for their weak convergence to the eigenvalues of a \(q\times q\) matrix \(Z_\beta\). This matrix \(Z_\beta\) is composed of three independent random matrices \(H_{\text{ID}}\), \(H_{\text{Gaussian}}\) and \(H_{\text{Diag}}\).
4. **Methods and Structures**:
- The paper adopts a technique called "ribbon graph expansion" to deal with large powers of the deformed inhomogeneous random matrix. This method is widely used in the study of classical GOE/GUE ensembles but is less common in deformed models.
- Through a series of upper - bound estimations and term - by - term limits of graph functions, the paper proves the above results. The specific steps include:
1. **Graph Function Expansion**: Use the Wick formula to decompose the joint moment into a sum of graph functions.
2. **Upper - bound Estimation**: Perform upper - bound estimation on each graph function to obtain the dominating function.
3. **Term - by - term Limit**: Apply the dominated convergence theorem to take limits term by term.
5. **Applications**:
- The paper discusses several specific inhomogeneous random matrix models, including random banded matrices, random matrices with d - regular variance distributions, and inhomogeneous information - plus - noise models.
Through these studies, the paper not only reveals the complex phase - transition phenomena of inhomogeneous random matrices under low - rank perturbations but also provides new tools and methods for understanding the spectral properties of these matrices.