Improvement in Autocorrelation Times Measured by the Master-Field Technique using Field Transformation HMC in 2+1 Domain Wall Fermion Simulations

Nobuyuki Matsumoto · 2025

Read Improvement in Autocorrelation Times Measured by the Master-Field Technique using Field Transformation HMC in 2+1 Domain Wall Fermion Simulations by Nobuyuki Matsumoto (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-2502-05452.txt.

Introduction

Shuhei Yamamoto and co-authors improve simulation times in lattice fermion studies with a field transformation method in a 2025 paper. Lattice runs slow down near critical points, a problem called critical slowing down.

The method joins standard Monte Carlo with a field shift based on a single smearing step that can be inverted. Tests used domain-wall fermions with a set gauge action. Autocorrelation times were read with a master-field style method from few configurations. Larger smearing values cut the longest times for smooth energy and charge measures. Local measures gained less. The result helps lattice teams plan cheaper runs with smarter update rules.

Catalog record

CREATORNobuyuki Matsumoto
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE19 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-05452

Browse the full library · About the Tower of Babel

Full text

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

Improvement in Autocorrelation Times Measured by the Master-Field Technique using Field Transformation HMC in 2+1 Domain Wall Fermion Simulations Nobuyuki Matsumoto Abstract The Field-Transformation Hybrid Monte-Carlo (FTHMC) algorithm potentially mitigates the issue of critical slowing down by combining the HMC with a field transformation, originally proposed by Lüscher and motivated as trivializing the theory. For the transformation, we use a single invertible discrete smearing step inspired by the Wilson flow but which resembles a Jacobian-computable generalisation of the stout smearing step. This is applied to a system with Iwasaki gauge fields and 2+1 Domain-Wall fermions. We have studied the effect of different smearing parameter values on autocorrelation times of Wilson-flowed energies with different flow time. We have found a reduction of exponential autocorrelation times for infra-red observables such as Wilson flowed energy densities and topological charge densities when a larger value of the smearing parameter is used. The autocorrelation times of local observables are computed using an approach akin to the master-field technique, allowing us to estimate the effect of the field transformation with different parameters based on a small number of configurations. MnLargeSymbols’164 MnLargeSymbols’171 [a]Shuhei Yamamoto 1 Introduction In Ref. [ 10 ] , Lüscher proposed an algorithm to mitigate the issue of critical slowing down using a transformation of gauge fields. In this algorithm, a gauge field, U , is substituted by another field, V , using a diffeomorphism on a space of gauge fields as U=\mathcal{F}_{t}(V) where t is a parameter of the transformation. Under this change of variable, Z=\int\mathcal{D}Ue^{-S(U)}=\int\mathcal{D}V\text{Det}[\mathcal{F}_{*}(V)]\,e^{-S(\mathcal{F}(V))}=\int\mathcal{D}V\,e^{-S_{FT}(V)} so that we have a new action for the field V , namely S_{FT}=S(\mathcal{F}_{t}(V))-\text{ln Det }\mathcal{F}_{t*}(V). In Ref. [ 10 ] , it is shown that there exists a perfect trivializing map \mathcal{F}_{t} such that \text{ln Det }\mathcal{F}_{t*}(V)\lvert_{t=1}=S(\mathcal{F}_{t}(V))+\text{constant}. For a perfect trivializing map, we have S_{FT}=0 for V , and so V is distributed uniformly. With a trivialized action, all modes of observables evolve at the same rate, reducing their autocorrelation times. These ideal maps can be obtained as a solution to \dot{U}_{t}=Z_{t}(U_{t})U_{t} where Z_{t} is a \mathfrak{su}(3) -field. In this study, we use only a single discrete step as an approximation to the perfect trivializing map [ 6 , 2 , 4 ] inspired by Wilson flow and stout smearing. The numerical integration in Wilson flow is replaced by a single the Euler step on each gauge link on the lattice: U(x,\mu)\to\mathcal{E}_{x,\mu}(y,\nu)=\begin{cases}e^{Z_{t}(U)(x,\mu)}U(x,\mu)&\text{if }(y,\nu)=(x,\mu)\\ U(y,\nu)&\text{otherwise.}\end{cases} (1) Here, \displaystyle Z_{t}(U_{t})(x,\mu) \displaystyle=\mathcal{P}(P_{\mu\nu}(x,\mu))\equiv\frac{1}{2}(P_{\mu\nu}(x,\mu)-P_{\mu\nu}(x,\mu)^{\dagger})-\frac{1}{6}\text{tr }[P_{\mu\nu}(x,\mu)-P_{\mu\nu}(x,\mu)^{\dagger}] \displaystyle P_{\mu\nu}(x,\mu) \displaystyle=\sum_{\nu\neq\pm\mu}\rho_{\mu,\nu}U(x,\nu)U(x+\hat{\nu},\mu)U(x+\hat{\mu},\nu)^{\dagger}U(x,\mu)^{\dagger}. (2) This gives a numerically cheap transformation which we might hope leads to a practical gain in computer time. The step size of the Euler step prescribed in Eq. 1 is absorbed into \rho_{\mu,\nu}\equiv\rho in Eq. 2 , and its effect on autocorrelation is investigated. To make this technique suitable for computer simulation, a single link update in Eq. 1 is performed at even and odd sites and in each direction \mu only once. All links of a given polarisation on even or odd sites are updated in parallel. Then, the overall link update consists of 8 independent updates and is more akin to familiar stout smearing. Compared to recent approaches seeking to completely trivialize the gauge theory and apply direct sampling schemes [ 7 ] , this approach remains more similar to Lüscher’s HMC based algorithm in structure while taking pains to minimize the cost of the field transformation even if the making this transformation more approximate. The motivation is that rather than seeking to map fields from the trivial distribution in the strong coupling limit, we use the local averaging implicit in link smearing to make a wavelength-dependent transformation (in particular active in the ultraviolet) from our integration variables to our gauge field. Because the integration variable momenta are Gaussian distributed, this wavelength-dependent transformation can rescale the changes made to the physical gauge field differentially between the ultraviolet and infrared degrees of freedom, potentially changing the relative rate of evolution and avoiding large forces arising in the ultraviolet that might otherwise limit the infrared evolution of the gauge field. 2 Simulation Details This numerical experiment is conducted on a lattice of size 32^{4} with \beta=2.37 and with 2+1 dynamical Domain-Wall fermions of mass m_{l}=0.0047,m_{s}=0.0186 . We have performed several simulations with different HMC parameters, such as different \rho values of 0.1, 0.112, 0.124, different gauge step sizes, \delta\tau_{G}=1/48,1/96 , and different fermion step sizes \delta\tau_{F}=1/24,1/16,1/12,1/8 . The nonzero value for \rho indicates the use of field transformation in the HMC algorithm and will be referred to as FTHMC. The algorithm with \rho=0 reduces to the usual HMC algorithm. In the following, we focus on the runs with different \rho and \delta\tau_{G} but with fixed \delta\tau_{F}=1/24 and trajectory length \tau=1 . Table 1 collects the total number of thermalized configurations from each run. These simulations are carried out on Frontier and Andes at Oak Ridge National Laboratory. \rho 0.0 0.1 0.112 0.124 \delta\tau_{G}=1/48 233 230 188 230 \delta\tau_{G}=1/96 401 232 229 229 \delta\tau_{G}=1/144 - 230 - - Table 1 : The number of trajectories for each ensemble after thermalization To see how reasonable the Monte Carlo (MC) chains are, the difference in the total simulation Hamiltonian between adjacent configurations in the chain, dH , has been computed. The averages of dH over MC chains are computed and summarized in Tab. 2 . We have checked that the Creutz relation \langle e^{-\Delta H}\rangle=1 is satisfied within statistics. \rho 0.0 0.1 0.112 0.124 \delta\tau_{G}=1/48 0.026(6) 0.006(6) 0.009(5) 0.03(1) \delta\tau_{G}=1/96 0.015(6) 0\pm 0.009 0.009(8) 0.017(7) \delta\tau_{G}=1/144 - 0\pm 0.008 - - Table 2 : \braket{dH} for different runs based on configurations with Metropolis step. \delta\tau_{F} is fixed to 1/48 . 3 Measurements As a first observable, we computed the plaquette. Figure 2 shows histories of plaquette values evaluated on each configuration from MC chains generated using the HMC and FTHMC algorithms. Both chains begin with a hot configuration. These MC chains are both observed to thermalize, as indicated by plaquette approaching to its expected value, which in this case is 0.6388238(37) , taken from Ref. [ 1 ] . Figure 1 : The plot of plaquette value vs. configuration index for HMC algorithm (orange line) and FTHMC algorithm (blue line). The horizontal red line indicates the expected value of 0.6388238(37) for this ensemble [ 1 ] . Figure 2 : Autocorrelation coefficient (ACC) as a function of t for Wilson-flowed energy E16, evaluated using the standard volume-averaged autocovariance \Gamma^{(V)}(t) . This conventional estimator of ACC exhibits a large error. We will introduce an improved local autocorrelation approach. Another set of observables measured in this study is Wilson flowed energy with different flow time \tau_{W}=4,8,12,16 . The flowed energy with longer flow time couples more strongly to slower mode and has a longer autocorrelation time, making it suitable for discerning different algorithms based on their effect on autocorrelation. Histories of some flowed energy on thermalized configurations from different MC chains are shown in Fig. 3 . As can be seen from these figures, the value of \tau_{W}^{2}E(\tau_{W}) fluctuates wildly. Determination of the central values as well as identification of repeated cycles are hard. This makes it difficult to calculate their autocorrelation times. 4 Autocorelation The autocorrelation of the flowed energy is considered in this study. When

Read the complete text