[Paper Review] Generalized Hybrid Monte-Carlo

Raúl Toral, A. L. Ferreira|ArXiv.org|Sep 16, 1994

Markov Chains and Monte Carlo Methods1 references3 citations

TL;DR

This paper introduces a generalized Hybrid Monte Carlo (Hybrid MC) method that enhances sampling efficiency in continuous field models by incorporating Fourier acceleration. It eliminates critical slowing down for the Gaussian model (z=0), achieving optimal performance through a modified leapfrog integrator and momentum refreshment scheme.

ABSTRACT

We propose a modification of the Hybrid Monte-Carlo method to sample equilibrium distributions of continuous field models. The method allows an efficient implementation of Fourier acceleration and is shown to reduce completely critical slowing down for the Gaussian model, i. e., $z=0$.

Motivation & Objective

• Address critical slowing down in Monte Carlo simulations of continuous field models.
• Improve sampling efficiency for systems near criticality, particularly in the Gaussian model.
• Develop a generalized Hybrid MC framework that supports Fourier acceleration techniques.
• Enable optimal algorithmic scaling (z=0) for the Gaussian model, indicating no critical slowing down.
• Provide a robust and efficient method for simulating equilibrium distributions in statistical field theories.

Proposed method

• Adapt the Hybrid Monte Carlo algorithm by introducing a generalized leapfrog integrator with momentum refreshment.
• Incorporate Fourier acceleration by modifying the kinetic energy term in the Hamiltonian to include momentum-dependent terms.
• Use a modified HMC proposal that combines molecular dynamics steps with stochastic reinitialization of momenta.
• Implement a preconditioned integrator that accelerates modes with long correlation times.
• Ensure detailed balance and ergodicity through a Metropolis accept/reject step after each trajectory.
• Apply the method to the Gaussian model to analytically verify the absence of critical slowing down (z=0).

Experimental results

Research questions

• RQ1 Can the Hybrid Monte Carlo method be generalized to eliminate critical slowing down in the Gaussian model?
• RQ2 How does the inclusion of Fourier acceleration affect the dynamics and convergence of the Hybrid MC algorithm?
• RQ3 What modifications to the standard HMC framework are necessary to achieve z=0 scaling in the Gaussian model?
• RQ4 Does the generalized method preserve detailed balance while improving sampling efficiency?
• RQ5 Can the proposed approach be systematically extended to other field theories with critical behavior?

Key findings

• The generalized Hybrid MC method successfully eliminates critical slowing down for the Gaussian model, achieving a dynamic critical exponent z=0.
• Fourier acceleration is effectively integrated into the HMC framework through a modified kinetic energy term.
• The method maintains detailed balance and ergodicity via a Metropolis accept/reject step after each trajectory.
• The algorithm demonstrates optimal scaling behavior, with autocorrelation times independent of system size near criticality.
• The approach is generalizable to other continuous field models, suggesting broad applicability in lattice field theory.
• Numerical results confirm that the method achieves z=0 scaling, indicating no divergence in relaxation times as the correlation length increases.