Improving Spectral Resolution from Real-time Evolution for Correlated Systems

Ta Tang, Chunjing Jia, Brian Moritz, Thomas P. Devereaux · 2025

Read Improving Spectral Resolution from Real-time Evolution for Correlated Systems by Ta Tang, Chunjing Jia, Brian Moritz, Thomas P. Devereaux (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-15539.txt.

Introduction

Ta Tang leads a team that sharpens blurry spectra from short simulations. For strongly linked materials, mock runs over time can predict how electrons leave a sample. Short runs and small samples blur both the energy and the direction detail. The team wanted finer detail without paying for much longer runs.

They propose a cheap learning method that extends short time and short range results. It uses a linear model that predicts later values from earlier ones. They apply it to two models of linked electrons, one far costlier than the other. The sharper spectra reveal traits tied to electrons coupled with lattice shakes, traits the blur had hidden. The method adds reach at low cost.

Catalog record

CREATORTa Tang, Chunjing Jia, Brian Moritz, Thomas P. Devereaux
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-15539

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Improving Spectral Resolution from Real-time Evolution for Correlated Systems Ta Tang, Chunjing Jia, Brian Moritz, Thomas P. Devereaux Abstract The quality of numerically simulated spectra using real-time evolution methods for strongly correlated systems is affected by both the length of simulation time and the system size, limiting resolution in both frequency and momentum. In this work, we propose a computationally cheap, linear autoregressive machine learning-based framework to extend short-time and distance results over a wider range. We demonstrate the proposed method to extend the lesser Green’s function for both the Hubbard model and the much more computationally challenging Hubbard-extended Holstein model. This technique significantly improves both the frequency and momentum resolution of the single-particle removal spectrum 𝒜 ⁡ ( k , ω ) \mathcal{A}(k,\omega) , allowing observation of otherwise obscured spectral features due to electron-phonon coupling. I Introduction Machine learning has emerged as a promising tool in condensed matter and materials physics, enabling new insights across a broad range of quantum many-body problems. Among the most pressing challenges in this field is the development of numerical algorithms capable of tackling strongly correlated systems at larger system sizes, in higher dimensions, and over longer time scales without sacrificing the accurate treatment of long-range entanglement and strong correlation effects. In recent years, neural-network quantum states (NQS) have been introduced as a variational ansatz for many-body wavefunctions, offering a route to simulate model Hamiltonians such as the Hubbard and Heisenberg models Carleo and Troyer 2017 ; Sharir et al. 2020 . These wavefunction-based approaches have the potential to move beyond the quasi-one-dimensional regime traditionally accessible to density matrix renormalization group (DMRG) methods; however, the growth of entanglement entropy of the wave functions Vidal et al. 2003 ; Vidal 2003 ; Vidal 2004 leads to a growing computational cost with longer times and larger system sizes. In this work, we propose a different type of novel machine learning-based framework that utilizes a linear autoregressive model to directly learn from spectral functions, or more precisely correlation functions \chi(r,t) , which may have a reduced complexity and present certain patterns that could be utilized to make predictions about values at later times and longer distances. This bypasses the need to represent the many-body wavefunction itself and accurately account for the growing entanglement entropy. The framework offers a new pathway to reach larger length and time scales that enable the study of dynamical correlations in strongly correlated systems with unprecedented efficiency and scalability. By leveraging machine learning for strongly correlated systems in a novel way, our approach opens the door to addressing longstanding challenges in quantum magnetism and high-temperature superconductivity that have remained beyond the reach of conventional numerical techniques. Dynamical quantities \chi(q,\omega) are essential for characterizing excitations in strongly correlated systems and can be used as a bridge between results from modeling and those from experimental measurements. \chi(q,\omega) can be computed via its Lehmann representation, for example by using the exact diagonalization (ED) method with Lanczos iterations to obtain dynamical quantities Dagotto 1994 . While ED is unbiased and versatile, it is limited to very small cluster sizes. A more powerful method for one-dimensional (1D), or quasi-1D systems, is density matrix renormalization group (DMRG) White 1992 ; White 1993 . Ground state and finite-temperature approaches White 2009 based on DMRG can be easily applied to very large systems ( \sim a few hundred sites). Extensions such as the correction vector method Kühner and White 1999 and variational approaches Jeckelmann 2002 have been developed to successfully capture dynamical properties, denoted as dynamical DMRG (DDMRG) methods, for much larger systems compared to ED. However, DDMRG requires separate converged simulations for each point in (q,\omega) -space and can be computationally inefficient if the goal is to obtain the entire spectra. An alternative method to calculate \chi(q,\omega) is to first compute the time-dependent correlator \chi(r,t)=\left<\hat{O}^{\dagger}_{r_{0}+r}(t_{0}+t)\hat{O}_{r_{0}}(t_{0})\right> and then apply Fourier transforms. Time-dependent DMRG (tDMRG) Vidal 2003 ; White and Feiguin 2004 ; Paeckel et al. 2019 is an extension of DMRG, making real-time evolution possible. This leads to a very efficient method to evaluate \chi(q,\omega) , where the full spectra can be obtained by a single tDMRG simulation. However, computational complexity grows rapidly as the length of time and cluster size increase, leading to low-frequency and -momentum resolution, especially for more numerically challenging problems, e.g. those involving multiple degrees of freedom or extended interactions. Previous works based on time series predictions White and Affleck 2008 ; Tian and White 2021 to improve the resolution have only extended the correlator near the center of the system along the time direction. Here we propose a method to extend the time-dependent correlator along both the time and space directions based on the pattern of short-time and short-distance results on a small cluster, simultaneously improving both the frequency and momentum resolution of the spectra. We first demonstrate this method on the single-particle removal spectra for the 1D Hubbard model. We then turn to a much more computationally expensive problem – the single-particle removal spectra of the 1D Hubbard-extended Holstein model with extended electron-phonon ( el-ph ) coupling. Our method provides high resolution for the simulated spectra in both cases, and highlights phonon-induced spectral features that are not apparent in the raw simulated spectra because of the low resolution. Figure 1: Comparison between predicted and simulated lesser Green’s function. The simulated data is obtained on a L=120 half-filled Hubbard chain using tDMRG. The total time simulated is 20[t_{h}^{-1}] . (a) Predicted Green’s function. The first half of the simulated data 10[t_{h}^{-1}] is used for training (indicated by the gray area). The data above the gray area is predicted. (b) Difference between predicted and simulated lesser Green’s function. Note that the color bar scale in (b) is 100 times smaller than that in (a). In (c) and (d), we plot the real and imaginary part of the predicted \mathcal{G}^{<}_{p}(r,t) (red and blue dashed lines) against the simulated \mathcal{G}^{<}_{s}(r,t) (black solid lines) at the center of the chain r=60 and at the boundary of the training data light cone r=32 . The gray areas indicate data used in training. II Method Computationally, using a method like tDMRG is very expensive, scaling like O(LTM^{3}) , where L is the system length measured in the number of sites, T is the simulated time measured in time steps, and M is the DMRG bond dimension, which also increases with L and T . Previous works have tried time series predictions, where for each fixed position r , the future time step \chi(r,t=n) is obtained by using data from previous d_{t} time steps, e.g. using a linear autoregressive model for linear prediction (LP) White and Affleck 2008 \mathcal{\chi}(r,t)=\sum_{j=1}^{d_{t}}\theta^{r}_{j}\mathcal{\chi}(r,t-j), (1) or a more sophisticated recursion method Tian and White 2021 . The process can be repeated to obtain future time step data at t=n+2,\,n+3,\,\dots . This can extend \chi(r,t) along the time dimension, improving the frequency resolution. An obvious limitation of this method is that it can only extend \mathcal{\chi}(r,t) along the time direction near the center portion of the chain since there is not enough data near the boundary to obtain \theta^{r} parameters and make good predictions. Therefore, it only improves the frequency resolution but not the momentum resolution. Additionally, the accuracy

Read the complete text