We employ singular value decomposition (SVD) to study the eigenvalue spectra of random spin systems. By SVD, eigenvalue spectrum is decomposed into orthonormal modes $W_k$ with weight $\lambda_k$. We show that the scree plot ($\lambda_k$ with respect to $k$) in the ergodic phase contains two branches that both follow power-law $\lambda_k\sim k^{ -\alpha}$ but with different exponents $\alpha$. By evaluating $W_k$, it's verified the part of $\lambda_k $ with $k>k_{ \text{ Th}}$ is universal that follows random matrix theory, where $k_{ \text{ Th}}$ is related to the Thouless energy. We further demonstrate that $\alpha$ corresponds only to the exponential part of the level spacing distribution while being insensitive to the level repulsion, or equivalently the system's symmetry. Consequently, $\alpha$ gives an underestimation for the many-body localization transition point, which suggests a non-ergodic behavior that may be attributed to the Griffiths regime.
Categories: cond-mat.dis-nn