[Paper Review] Improving the dynamical overlap algorithm

Nigel Cundy, Stefan Krieg+1 more|ArXiv.org|Nov 22, 2005

Quantum Chromodynamics and Particle Interactions2 references3 citations

TL;DR

This paper presents algorithmic improvements to the dynamical overlap Hybrid Monte Carlo (HMC) method, including an O(Δt²)-exact correction step for sign changes in the kernel operator eigenvalues and a modified leapfrog integrator for better energy conservation. The key contribution is a 30% performance gain in HMC simulations, with further gains via overlap eigenmode preconditioning in topologically nontrivial configurations.

ABSTRACT

We present algorithmic improvements to the overlap Hybrid Monte Carlo algorithm, including preconditioning techniques and improvements to the correction step, used when one of the eigenvalues of the Kernel operator changes sign, which is now O(Δt^2) exact.

Motivation & Objective

• Address the inefficiency and energy errors in the original overlap HMC algorithm when eigenvalues of the kernel operator cross zero.
• Reduce computational cost in dynamical fermion simulations by improving the molecular dynamics integration and preconditioning strategies.
• Enhance the stability and efficiency of HMC simulations in topologically nontrivial configurations where overlap operator inversions are slow.
• Develop a non-area-conserving correction step with Jacobian correction to improve transmission rates without sacrificing reversibility.
• Optimize the simulation workflow by combining preconditioning, improved integrators, and smearing techniques to reduce overall runtime.

Proposed method

• Introduce a new O(Δt²)-exact correction step that uses multiple orthogonal vectors to cancel energy errors from sign changes in the kernel eigenvalues.
• Implement a modified leapfrog integrator with adjustable parameter λ to improve energy conservation, reducing the need for frequent rejections.
• Apply 'stout smearing' to improve the condition number of the Wilson operator, accelerating overlap operator inversions.
• Use Hasenbusch acceleration with a heavy auxiliary fermion to reduce low-mass inversions, though gains were limited in this study.
• Develop an overlap eigenmode preconditioner using low-precision eigenvectors of the overlap operator to accelerate conjugate gradient solvers.
• Introduce a non-area-conserving correction step with a tunable parameter r₀ to improve transmission rates, where the Jacobian is corrected in the Metropolis step.

Experimental results

Research questions

• RQ1 Can the energy error in the overlap HMC correction step be reduced from O(Δt) to O(Δt²) through improved momentum updates?
• RQ2 Does the modified leapfrog integrator significantly improve energy conservation and trajectory acceptance rates in dynamical overlap simulations?
• RQ3 To what extent does preconditioning with low-precision overlap eigenmodes accelerate the solution of the overlap operator in topologically nontrivial configurations?
• RQ4 Can a non-area-conserving correction step with Jacobian correction improve transmission rates without breaking detailed balance?
• RQ5 How do the combined improvements affect the overall performance and acceptance rate in large-volume dynamical fermion simulations?

Key findings

• The improved correction step achieves O(Δt²) accuracy, eliminating the O(Δt) energy errors present in the original algorithm.
• The modified leapfrog integrator reduces the average trajectory time by approximately 30% on 4⁴, 8⁴, and 12⁴ lattices, with gains more than compensating for two extra overlap inversions per step.
• Overlap eigenmode preconditioning significantly accelerates convergence in topologically nontrivial configurations, with gains increasing with volume and decreasing quark mass.
• The non-area-conserving correction step with r₀ = 1 increases the transmission rate by 50% compared to the area-conserving version.
• On μ = 0.05, the improved HMC (imp) achieves 1479(20) time units per trajectory with 96% acceptance, outperforming the normal HMC (1816(20)) and Hasenbusch (2100(90)) variants.
• The combination of improved integrator and preconditioning (impnap) reduces trajectory time to 1445(50) with 95% acceptance, demonstrating robust performance across different quark masses.