arXiv:2609.24602
Random-matrix and transport frequencies in eigenstate spectral functions
Kadir Çeven, Rohit Patil, Marcos Rigol, Fabian Heidrich-Meisner
Abstract
The fact that generic isolated many-body quantum systems thermalize is understood using the eigenstate thermalization hypothesis (ETH). In recent years, there has been much interest in the behavior of the ETH spectral functions, which characterize the smooth dependence of the variance of the off-diagonal matrix elements of observables on the associated energy and frequency, and whose low-frequency part contains information about the long-time dynamics. In finite systems described by the ETH, the spectral functions are expected to exhibit plateaus below a characteristic frequency $ω^{}_{\mathrm{ETH}}$. In this regime, the statistics of the matrix elements of observables are expected to be described by random matrix theory. Related frequencies that have been studied in the literature are $ω^{}_{\mathrm{SFF}}$, which controls the onset of random-matrix behavior in the spectral form factor, and the transport frequencies $ω^{}_{\mathrm{tr}}$, which are derived from transport coefficients. However, a direct quantitative comparison of these frequencies is lacking. Using exact diagonalization, we conduct such a comparison for the spectral functions of current operators in clean and disordered quantum spin ladders with diffusive energy and spin transport. We find clear evidence for the expected low-frequency plateaus in the ETH spectral functions. For the accessible system sizes, $ω^{}_{\mathrm{SFF}}$ is consistent with the extent of the plateaus, while the transport frequencies are systematically larger and lie in the nonuniversal regime of the spectral functions. Our findings highlight the need to better understand the origin of these quantitative differences.
AI Summary
The paper asks how three finite-size frequency scales compare in thermalizing many-body systems: the onset of random-matrix behavior in the spectral form factor, $\omega_{\mathrm{SFF}}$; the extent of low-frequency plateaus in ETH spectral functions, $\omega_{\mathrm{ETH}}$; and transport frequencies inferred from diffusion constants. These scales are often called the Thouless frequency despite potentially different prefactors. Exact diagonalization computes energy- and spin-current spectral functions and spectral form factors for disordered XX ladders up to 18 spins and clean XXZ ladders up to 24 spins. Current autocorrelations, evaluated independently using dynamical quantum typicality and Krylov evolution, provide diffusion constants. The disordered calculations cover $W/J_\parallel=0.45,0.9,1.35$, within the diffusive regime. Both current spectral functions exhibit low-frequency plateaus with approximately size-independent heights consistent with $|f_{J^{\mathrm{ch}}}(E_\infty,0)|^2=D^{\mathrm{ch}}\chi^{\mathrm{ch}}/\pi$. Across the studied parameters, the numerical hierarchy is $\omega_{\mathrm{SFF}}\approx\omega_{\mathrm{ETH}}<\omega_{\mathrm{tr}}^{\mathrm E}<\omega_{\mathrm{tr}}^{\mathrm S}$, using $\omega_{\mathrm{tr}}^{\mathrm{ch}}=2\pi(2\pi/L)^2D^{\mathrm{ch}}$. Transport frequencies lie outside the plateaus, while spectral-form-factor frequencies are consistent with their endpoints. Averaging spectral form factors over symmetry sectors also resolves the random-matrix ramp in clean ladders. The comparison is limited by accessible sizes, residual finite-size effects, and filter- and threshold-dependent extraction of the spectral-form-factor timescale. No systematic criterion independently extracts $\omega_{\mathrm{ETH}}$ from the featureless plateau crossover. The origin of the quantitative differences, their nonuniversal prefactors, and dependence on the observable remain unresolved.
Error found: The documented clean spin-current spectral-function procedure uses matrix elements within fixed spin-reversal-parity sectors, where that current has identically zero matrix elements, so it cannot produce the reported nonzero plateau without an undocumented cross-parity calculation.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 13 pages, 9 figures