Eigenvector overlaps of random covariance matrices and their submatrices (Q6973504)

Let \(A\) be a deterministic rectangular \(M \times N\) matrix, \(M\geq N,\) and let \(B_{t}\) be a \(M\times N\) matrix containing independent Brownian motions. Let consider the matrix \(X_{t}= A+ N^{-1/2} B_{t}.\) For \(m<M, n<N,\) a very interesting problem is the comparison of the \(M\times N\) matrix \(\tilde{X}_{t},\) where \(\tilde{X}_{t}^{i j} = X_{t}^{i j}, 1\leq i\leq m, 1\leq j \leq n,\) and \( \tilde{X}_{t}^{i j}=0,\) otherwise, and the matrix \(X_{t}\) trough their singular vectors, assuming \(n\leq m.\)\N\NThe authors deal with the overlaps between the eigenvectors of the square matrices \(R_{t}= X_{t} ^{T}X_{t}, \, \tilde{R}_{t}= \tilde{X}_{t} ^{T} \tilde{X}_{t},\, L_{t}= X_{t} X_{t} ^{T} , \, \tilde{L}_{t}= \tilde{X}_{t}\tilde{X}_{t} ^{T},\) respectively. These are empirical covariance matrices of \(X_{t}\) and \(\tilde{X}_{t}\). More precisely, they focus the attention on the limiting behavior with those of the full matrix, when \( N/M ,\) \(m/M\) as well as \(n/N\) converge to fixed ratios. This means that these limits are considered in the bulk of both spectra, i.e., a comparison between the principal component analysis of \(X_{t}\) with \(\tilde{X}_{t} 's\) eigenvector by eigenvector, assuming the independent features which corresponds to the Marchenko-Pastur setup.\N\NNotice that in the context of symmetric Gaussian matrices in [\textit{E. Attal} and \textit{R. Allez}, J. Phys. A, Math. Theor. 58, No. 4, Article ID 045003, 36 p. (2025; Zbl 1557.60006)], explicit formulas for the limiting rescaled mean squared overlaps between the eigenvectors of a leading principal submatrix and those of the full matrix are deduced. The basic tool is based on the analysis of the eigenvector flow under the Dyson Brownian motion and the study of the dynamics of a specific resolvent. These results are confined for symmetric matrices and their principal minors. It is important to point out there is no trivial deterministic relation between the eigenvectors of a symmetric matrix and those of one of its principal leading submatrices.\N\NThe authors extend the above method to the singular vectors of rectangular Gaussian matrices, or equivalently, to the eigenvectors of Wishart matrices. By examining the dynamics of the singular vectors, some analogous results for the limiting overlaps in the macroscopic regime are deduced. The method makes use of the dynamics of the singular vectors and of specific resolvents when the matrix coefficients follow Brownian trajectories. Explicit forms for the limiting rescaled mean squared overlaps for right and left singular vectors in the bulk of both spectra, for any initial matrix \(A,\) are given. When \(A\) is the null matrix, this corresponds to the Marchenko-Pastur setup for covariance matrices, and the resulting formulas simplify into Cauchy-like functions. Notice that the time \(t\) is the variance of the noise added to \(A\) but it is also a way to derive dynamics leading to the main results.