Fourier Accelerated HMC in the S U ( N ) × S U ( N ) SU(N)\times SU(N) Principal Chiral Model
* · 2025
Read Fourier Accelerated HMC in the S U ( N ) × S U ( N ) SU(N)\times SU(N) Principal Chiral Model by * (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-2503-20898.txt.
Introduction
Roger Horsley, Pablo Morande, and Brian Pendleton run Fourier-accelerated simulations of a two-dimensional chiral model that mimics key traits of quantum chromodynamics. Their paper from twenty twenty-five uses hybrid Monte Carlo with a Fourier boost. The model, a product of two SU(N) groups, shares asymptotic freedom and mass generation with the real strong force. Group sizes from small to mid are covered.
Even at modest correlation lengths, the boost cuts autocorrelation times by about tenfold. Extra compute cost stays slight. The edge shrinks as group size grows, likely since the group space widens. Mass spectra and scaling match the exact Bethe ansatz solution. The study gives lattice teams a tested speed lever for chiral models, with a clear note on where it fades.
Catalog record
| CREATOR | * |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 18 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-20898
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Fourier Accelerated HMC in the S U ( N ) × S U ( N ) SU(N)\times SU(N) Principal Chiral Model * Abstract We report the results of Fourier-accelerated HMC simulations of 2D S U ( N ) × S U ( N ) SU(N)\times SU(N) principal chiral models for N = 2, 3, 4, 6, 9. These models share several key properties with 4D QCD, such as asymptotic freedom and dynamical mass generation. Even for modest correlation lengths, we find integrated autocorrelation times are decreased by an order of magnitude relative to standard HMC, with relatively little computational overhead. Our results also suggest that the relative advantage of Fourier Acceleration over traditional HMC decreases as N increases, possibly due to the enlarged group space associated with larger N. Our Monte Carlo results agree with the mass spectra and continuum scaling behaviour predicted by the exact solution obtained using the Bethe ansatz. [a]Brian Pendleton 1 Introduction In lattice simulations, critical slowing down refers to the significant increase in run time required to obtain results as the lattice spacing, a , is reduced. This phenomenon arises because our algorithms generate Monte Carlo configurations which become highly correlated in the a\to 0 limit. In practice, the degree of autocorrelation in the configurations is measured through the integrated autocorrelation time (IAT), \tau_{IAT} , which can be interpreted as the number of configurations between independent samples. In the case of Hybrid Monte Carlo (HMC) [ 1 ] , critical slowing down emerges because the algorithm treats both long- and short-wavelength modes equally. The short wavelength modes evolve faster in the simulation and set a strict upper limit on the integration step size for accurate integration. This means the more physical long wavelength modes, which evolve more slowly, barely evolve in each molecular dynamics step, resulting in highly correlated configurations. Consider a free theory, the equations of motion lead to the dispersion relation \omega^{2}(p)=p^{2}+m^{2} [ 2 ] , which shows that modes with higher (shorter) momentum (wavelength) evolve faster than those with lower (longer) ones. Generally, the maximum integration step size dt is inversely proportional to the maximum possible frequency [ 2 ] . Therefore, the number of steps required for the mode with the lowest momentum to complete a whole cycle is given by: \frac{T}{dt}=\frac{\omega(p_{max})}{\omega(p_{min})}=\sqrt{\frac{p^{2}_{max}+m^{2}}{p^{2}_{min}+m^{2}}} (1) Critical slowing down occurs when this ratio diverges as a\to 0 as more modes are included in the simulation. One possible way to circumvent this issue is Fourier Acceleration (FA), in which the dynamics are modified such that the slow modes are accelerated and the fast ones are slowed down by adding a momentum-dependent "mass" to the system. Ideally, this would make the evolution speed of the modes independent of the momentum. To achieve this, the Hamiltonian is modified by introducing the inverse kernel of the action in the distribution of the momenta. Returning to the free theory, this modification yields a new dispersion relation of \omega^{2}(p)=\frac{p^{2}+m^{2}}{p^{2}+M^{2}} [ 2 ] where M is a tunable parameter included in the kernel introduced in the momenta distribution. Choosing M=m makes the evolution rate independent of the mode’s momentum, completely avoiding critical slowing down. While being exemplified through the free theory, the idea of modifying the dynamics of the system to change the evolution rates of the different modes can, in principle, be applied to asymptotically free theories (like QCD) [ 4 , 5 , 3 ] in which the problematic high momentum modes enter the simulation as asymptotically free modes. 2 The SU(N) x SU(N) Principal Chiral Model In the continuum, the 2D SU(N)\times SU(N) model is defined via the Euclidean Lagrangian: \displaystyle\mathcal{L}=\frac{1}{T}\,\operatorname{Tr}\,{\partial_{\mu}U\partial_{\mu}U^{\dagger}} \displaystyle U\in SU(N) (2) On the lattice, the model is defined via the action: \displaystyle S=-\beta N\sum_{x,\mu>0}\operatorname{Tr}\{U^{\dagger}_{x}U_{x+\mu}+U_{x+\mu}^{\dagger}U_{x}\} \displaystyle\beta=\frac{1}{NT} (3) This model shares properties with QCD such as asymptotic freedom and dynamical mass generation. Still, it only possesses a SU(N)\times SU(N) global symmetry and is not locally gauge invariant. Furthermore, it is possible to obtain analytical predictions using the Bethe ansatz to calculate the exact solution [ 6 ] . By studying the effect of Fourier acceleration in this model, we aim to establish a foundation for investigating more complex theories and, ultimately, QCD. 2.1 HMC & Fourier Acceleration in the Principal Chiral Model The field elements U_{x} belong to the SU(N) group and hence they can be expressed as: U_{x}=\exp\{i\alpha_{x}^{a}\lambda_{a}\} (4) Where \lambda_{a} are a generalization of the Gell-Mann matrices. We define the canonical momentum \pi^{a} conjugate to the \alpha_{x}^{a} and we choose: H=\frac{1}{2}\sum_{x}\pi_{x}^{a}\pi_{x}^{a}+S (5) The equations of motion are: \displaystyle\dot{U}_{x}=i\pi_{x}U_{x} \displaystyle\pi_{x}=\pi_{x}^{a}\lambda_{a} (6) \dot{\pi}_{x}=-2i\beta N\left[\sum_{\mu>0}\left(\left(U_{x+\mu}+U_{x-\mu}\right)U_{x}^{\dagger}-h.c\right)-\frac{1}{N}\operatorname{Tr}\{...\}I\right] (7) The Leapfrog discretisations are: \displaystyle\pi_{x}\left(t+\frac{dt}{2}\right)=\pi_{x}\left(t-\frac{dt}{2}\right)+dt\dot{\pi}_{x}\left(t\right) \displaystyle U_{x}(t+dt)=\exp\left\{i\pi_{x}\left(t+\frac{dt}{2}\right)dt\right\}U_{x}(t) (8) In order to implement FA we introduce the inverse kernel of the action in the canonical momentum distribution, that is: \displaystyle H_{FA}=\frac{1}{2}\sum_{x,y}\pi_{x}K^{-1}_{x,y}(M)\pi_{y}+S \displaystyle S=\beta N\sum_{x,y}\operatorname{Re}\operatorname{Tr}\{U^{\dagger}_{x}K_{x,y}(0)U_{y}\} (9) In Fourier space the kernel is diagonal, simplifying its inversion. The Fourier transform and its inverse on an L^{D} lattice are defined as \displaystyle\tilde{f}(k)=\mathcal{F}[f(x)]_{k}=\sum_{x}f(x)\exp\left\{-2\pi i\frac{k\cdot x}{L}\right\} (10) \displaystyle f(x)=\frac{1}{L^{D}}\mathcal{F}^{-1}[\tilde{f}(k)]_{x}=\sum_{k}f(k)\exp\left\{2\pi i\frac{k\cdot x}{L}\right\} (11) The inverted kernel in Fourier space is given by: \tilde{K}^{-1}_{k,k^{\prime}}(M)=\delta_{k,k^{\prime}}\frac{1}{\sum_{\mu}4\sin^{2}\left(\frac{\pi k_{\mu}}{L}\right)+M^{2}} (12) This modification leaves the momentum’s time evolution unchanged but complicates the equation for the field. In discrete time this becomes: U_{x}(t+dt)=\exp\left\{i\mathcal{F}^{-1}\left[\tilde{K}^{-1}_{k}(M)\tilde{\pi}_{k}\left(t+\frac{dt}{2}\right)\right]_{x}dt\right\}U_{x}(t) (13) For details on the implementation of HMC and FA in this theory see [ 7 ] . 3 Results 3.1 Acceleration To assess the algorithm’s efficiency relative to traditional HMC, we measured the IAT of the model’s susceptibility, a long-wavelength observable prone to critical slowdown. Following [ 8 , 9 ] , the IAT and its error were calculated for N=2,3,4,6 at various correlation lengths using both FA and traditional HMC. The data were fitted to the ansatz \tau_{IAT}=b\xi^{z} , where z is the dynamical critical exponent and \xi=\frac{1}{m_{0}a} is the system’s correlation length (inverse of the lightest mass). This ansatz captures the correct behaviour for the IAT for sufficiently large correlation lengths and it was found to describe the data well. The acceleration mass parameter M was chosen to be M=\xi^{-1} for each point, this was found to yield a near to optimal acceleration rate using Eq. 14 to compare different runs. Details of \xi calculations for each \beta -value are provided in Section 3.2 . To ensure consistency, only runs with acceptance rates near 0.7 were included. Results are shown in Fig. 1 , with red and blue points representing HMC and FA HMC measurements, respectively. (a) SU(2) (b) SU(3) (c) SU(4) (d) SU(6) Figure 1 : The plots show the autocorrelation time \tau_{IAT} as a function of the correlation length for several SU(N) in HMC and FA HMC. The reduction in the critical exponent z demonstrates that FA HMC outperforms traditional HMC, though it does not fully eliminate critical slowing down. For N=3,4,6 , with correlation lengths up to 10, FA achieved a tenfold speedup. For N=2 , the final three data points from [ 7 ] align well with our extrapolated results, and suggest a two-orders-of-magnitude improvement at larger correlation lengths. Notably, our z=0.24\pm 0.03 for the SU(3) model is significantly lower than z=0.45\pm 0.02 reported in [ 10 ] using Multigrid Monte Carlo, highlighting FA HMC’s success when compared to other methods used to reduce autocorrelations. However, FA