Parallel Tempered Metadynamics for full QCD

Source: arXiv:2607.21575 · Published 2026-07-23 · By Timo Eichhorn, Gianluca Fuwa, Christian Hoelbling, Lukas Varnhorst

TL;DR

This paper addresses the critical challenge of topological freezing in lattice QCD simulations, a problem that severely limits the efficiency of conventional Hybrid Monte Carlo (HMC) methods as lattice spacing is refined. The authors propose and demonstrate an algorithm called Parallel Tempered Metadynamics (PT-MetaD), which combines parallel tempering with collective-variable-based enhanced sampling techniques (specifically Metadynamics). This approach targets multimodal distributions characterized by high-probability sectors separated by free energy barriers, which in QCD correspond to topological sectors. The key advancement is utilizing collective variables representing the topological charge to construct bias potentials that facilitate transitions between sectors, thereby breaking freezing behavior. They apply PT-MetaD to full QCD simulations with two staggered fermion flavors, using the DBW2 gauge action, a setup known to exhibit strong topological freezing.

Key findings

• PT-MetaD fully unfreezes the topological charge sector at β=1.15 where RHMC remains frozen for ~40,000 MDUs (Fig. 2).
• At β=0.95, PT-MetaD shows more frequent topological charge transitions and reduces integrated autocorrelation time of Q² from 764(488) MDUs (RHMC) to 392(208) MDUs (PT-MetaD) (Table III).
• The error on topological susceptibility decreases by factor ~2.62 and on smeared plaquette by ~1.06 on the coarse lattice compared to RHMC (Table II).
• PT-MetaD incurs ~2.3x CPU time overhead versus RHMC but achieves ~3x efficiency gain in CPU time for topological susceptibility due to improved sampling, resulting in net efficiency improvement.
• Expected swap acceptance rates from estimated free energy surfaces closely match observed swap rates: 64.3% expected vs 68.6% observed at β=0.95, and 31.0% vs 28.6% at β=1.15 (Appendix B).
• Bias potentials constructed via well-tempered Metadynamics and singular spectrum analysis only bias intersector barriers, preserving correct relative weights of topological sectors (Fig. 1, Appendix A).
• Local force spikes in the bias potential during transitions are observed, indicating localized force contributions which do not destabilize the simulation (Appendix C).

Threat model

The adversary is implicit in the multimodal sampling challenge inherent to lattice QCD simulations rather than a security threat actor. The 'adversary' is thus the high free energy barriers between topological sectors that obstruct transitions and cause freezing. The algorithm assumes no manipulation of the simulations by an external adversary and instead targets ergodicity and efficient sampling of the physical distribution despite these internal bottlenecks.

Methodology — deep read

The authors tackle topological freezing in lattice QCD, where distinct topological sectors emerge causing multimodal distributions and extremely long autocorrelation times for standard Hybrid Monte Carlo algorithms. Their threat model assumes a standard lattice QCD simulation aiming for ergodic sampling of gauge field configurations with no adversarial modification but challenging multimodal target distributions with suppressed tunneling between sectors.

The data consists of lattice gauge configurations generated for two-flavor staggered fermion QCD simulations on a 24^4 volume, with bare quark mass am=0.02, four stout smearing steps (ρ=0.125), and the DBW2 gauge action known to enhance topological sector separation. Two lattice couplings are studied, β=0.95 (coarse) and β=1.15 (fine).

The core algorithm is Parallel Tempered Metadynamics (PT-MetaD). They instantiate two replicas: the physical replica sampling the unbiased QCD distribution p1(U), and a biased replica sampling p2(U) ∝ p1(U) e^{-V(Q_CV)}, where V(Q_CV) is a bias potential acting on a carefully chosen collective variable (CV), Q_CV, defined as the clover-based topological charge after six stout smearing steps.

Parallel tempering exchanges configurations between replicas with Metropolis acceptance probability depending only on the difference in bias potentials, enabling the physical replica to overcome free energy barriers through replica-swapping.

The bias potentials V(Q_CV) are built in a preliminary well-tempered Metadynamics phase using multiple parallel walkers. They apply singular spectrum analysis to isolate the oscillatory component associated with intersector barriers, preserving relative sector weights while facilitating transitions.

Simulation updates within each replica use the Rational Hybrid Monte Carlo (RHMC) algorithm with a fourth-order minimum norm integrator. Trajectories have length 4 MDUs, step size 0.2.

Evaluation includes measuring topological charge time series, plaquette, and topological susceptibility. Integrated autocorrelation times are computed by the Γ-method. Efficiency is assessed by CPU time and autocorrelation reduction.

Results are compared against conventional RHMC baseline simulations with identical parameters. The bias potential introduces negligible computational overhead since fermion force calculations and bias force calculations share smearing steps.

No closed datasets are required as simulations are fully described; code availability is not explicitly stated but the methodology is reproducible as detailed bias potential construction and algorithm steps are given. Random seeds and hyperparameters follow standard lattice QCD practice. Figures 1-5 and Tables I-III illustrate bias potentials, time series, integrated autocorrelations, and force measurements.

Concrete end-to-end example: At β=1.15, RHMC fails to tunnel over 40,000 MDUs. PT-MetaD uses a biased replica with a reconstructed free energy surface from 16,000 trajectories of well-tempered Metadynamics. This bias reduces barriers so swaps occur ~29% of the time. PT-MetaD samples all topological sectors, drastically reducing autocorrelations of charge and susceptibility and giving physically consistent observables with moderate CPU overhead (~2.3x).

Technical innovations

• Integration of parallel tempering with Metadynamics biasing of collective variables tailored to topological sectors in full QCD simulations, improving sampling efficiency over standard RHMC.
• Use of singular spectrum analysis to extract only the intersector barrier component from well-tempered Metadynamics free energy surfaces, preserving relative sector weights and maximizing swap acceptance.
• Definition of the biased replica probability distribution via biased collective variable distributions pV(s(U)), diverging from traditional coupling-constant interpolation in parallel tempering.
• Efficient evaluation of replica swap acceptance rates leveraging only the bias potential difference, thus avoiding costly fermionic determinant recalculations during swaps.

Datasets

• Full QCD lattice configurations — approximately 1000+ trajectories per ensemble — generated in-house with Nf = 2 staggered fermions on 24^4 lattice with DBW2 gauge action

Baselines vs proposed

• RHMC at β=1.15: topological susceptibility χtopV = 0 (frozen) vs PT-MetaD: χtopV = 0.127(33)
• RHMC at β=0.95: integrated autocorrelation τint(χtop) = 764(488) MDUs vs PT-MetaD: 392(208) MDUs
• CPU cost: RHMC = 1x baseline vs PT-MetaD = ~2.3x raw CPU time but ~3x efficiency gain on topological susceptibility sampling

Limitations

• Only two replicas were used; scaling behavior and performance for larger replica ensembles is not evaluated.
• Bias potentials are estimated approximately and residual fluctuations remain, potentially diminishing swap rates and efficiency.
• No explicit tests under distribution shifts or different fermion mass regimes presented; generalizability uncertain.
• The approach hinges on the identification of suitable collective variables (here, topological charge), limiting applicability to systems lacking natural CVs.
• Current evaluation is static snapshots at specific β values; continuum scaling behavior remains to be demonstrated fully.
• No adversarial or worst-case scenario testing, e.g. pathological potential landscapes, is conducted.

Open questions / follow-ons

• Does PT-MetaD improve continuum scaling of autocorrelation times for topological observables compared to RHMC as lattice spacing decreases further?
• How sensitive is PT-MetaD performance to the choice and smearing parameters of the collective variable used for biasing?
• Can PT-MetaD be extended effectively using more than two replicas and multiple collective variables for even more challenging multimodal distributions?
• How does PT-MetaD perform in regimes with topological sector suppression, such as high temperature QCD, and with other fermion discretizations?

Why it matters for bot defense

From the perspective of bot-defense and CAPTCHA practitioners, this work illustrates an advanced technique to overcome sampling bottlenecks in highly multimodal distributions by combining parallel tempering with adaptive biasing on collective variables. Analogously, in bot detection models or CAPTCHA challenge spaces exhibiting multimodality or discrete state entrapment, such methods might inspire novel sampling schemes to better explore latent spaces or decision boundaries. The strategy of carefully designing auxiliary distributions and bias potentials to facilitate transitions between modes while maintaining correct target distributions could inform robust algorithm designs in bot-defense that require navigating complex state-spaces.

Practitioners should note that success hinges on identifying effective collective variables that capture bottleneck features—akin to extracting relevant behavioral features in bot detection. The demonstrated efficiency gains and ability to unfreeze otherwise intractable sampling problems showcase how biased replica exchange schemes, supported by free energy landscape estimation, can fundamentally alter the dynamics of complex systems. Though specific to lattice QCD here, the underlying concepts of parallel tempering combined with adaptively learned biases remain broadly relevant for defensive algorithmic designs in security and bot mitigation frameworks.

Cite

@article{arxiv2607_21575,
  title={ Parallel Tempered Metadynamics for full QCD },
  author={ Timo Eichhorn and Gianluca Fuwa and Christian Hoelbling and Lukas Varnhorst },
  journal={arXiv preprint arXiv:2607.21575},
  year={ 2026 },
  url={https://arxiv.org/abs/2607.21575}
}

Read the full paper

• arXiv:2607.21575 (abstract)
• arXiv:2607.21575 (PDF)