Abstract:Understanding the statistical fluctuations of lattice observables over the gauge configurations is important both theoretically and practically. It provides a physical insight into tackling the famous signal-to-noise problem and the sign problem, and inspires new thoughts in developing methods to improve the signal of lattice calculations. Among many efforts, exploring the relationship between the real part and imaginary part of lattice numerical result is a new method to understand lattice signal and error, because both the real part and imaginary part come from the same sample of gauge field and their distributions on the gauge sample are related in principle. Specifically, by analyzing the distributions of the real part and imaginary part of quenched lattice two-point function with high statistics and non-zero momentum, this work proposes a possible quantitative formula connecting these two distributions as \begin{document}$R(x)=\displaystyle\int {\rm{d}}y S(y-x) \left[I(y) K(U_y)\right]$\end{document}, where \begin{document}$R(x)$\end{document} denotes the real-part distribution, \begin{document}$I(x)$\end{document} the imaginary-part distribution, \begin{document}$S(x)$\end{document} the underlying signal distribution and \begin{document}$K(U_x)$\end{document} a kernel function of the gauge field. This theoretical assumption has universal validity because the kernel function contains the gauge field information that determines all the distributions. The formula is numerically verified by calculating the non-trivial statistical correlations of the real part and the kernel-function-modified imaginary part under the further assumption of the kernel function. It is found that the most naïve guess of \begin{document}$K(U_x)=1$\end{document} does not work, which leads to no statistically significant correlation. Meanwhile, the assumption that \begin{document}$K(U_x)$\end{document} is only a sign function works well, giving rise to \begin{document}$\sim70\%$\end{document} correlation. Then, through the process of adding random distortions to the absolute values of the imaginary part, it is found that even a slight distortion, of around 1% could result in a significant reduction in the correlation between the real part and imaginary part down to less than 50% or lower. This essentially proves that the observed \begin{document}$\sim70\%$\end{document} correlation is highly non-trivial and the hypothesis that \begin{document}$K(U_x)$\end{document} is a sign function captures at least some of the physical mechanisms behind the scenes. Employing this correlation, the variance of lattice results can be improved by around 40%. It is not a significant improvement in practice; however, this study offers an innovative strategy to understand the source of statistical uncertainties in lattice QCD and to improve the signal-to-noise ratio in lattice calculation. Further research on the ability to use machine learning on various more accurate lattice data will hopefully give better instructions and constraint on the form of the kernel function.

Bayesian interpretation of Backus-Gilbert methods

Bayesian solution to the inverse problem and its relation to Backus-Gilbert methods

Bayesian Inference for Contemporary Lattice Quantum Field Theory

Spectral densities from Euclidean lattice correlators via the Mellin transform

A Bayesian approach to chiral extrapolations

Bayesian Inference with Gaussian Processes for the Determination of Parton Distribution Functions

Bayesian and Maximum Likelihood Estimation for Gaussian Processes on an Incomplete Lattice

Progress on the spectroscopy of lattice gauge theories using spectral densities

Teaching to extract spectral densities from lattice correlators to a broad audience of learning-machines

Meson spectroscopy from spectral densities in lattice gauge theories

Imaginary-part Distribution and Signal Improvement of Lattice Quantum Chromodynamics Data

Extraction of lattice QCD spectral densities from an ensemble of trained machines

Density of States of the lattice Schwinger model

Multigrid Monte Carlo Revisited: Theory and Bayesian Inference

The quenched glueball spectrum from smeared spectral densities

Combining Bayesian reconstruction entropy with maximum entropy method for analytic continuations of matrix-valued Green's functions

Lattice $B$-field correlators for heavy quarks

Entanglement entropy from non-equilibrium lattice simulations

Bayesian approach to superstatistics