Fast Accurate Quantum Monte Carlo without Metropolis Adjustment

Reuben Cohn-Gordon, Gabriel Pescia, Sumner N. Hearth, Jakob Robnik, Uroš Seljak, Peter LuntsarXiv 2026 · arXiv

TL;DR

Researchers have demonstrated that removing the "adjustment" step in Hamiltonian Monte Carlo (HMC) can speed up complex quantum simulations by an order of magnitude. By trading a tiny, controllable amount of mathematical bias for a massive boost in sampling efficiency, they've unlocked a way to simulate high-dimensional quantum systems that were previously computationally out of reach.

What Problem Does This Paper Address?

In the world of condensed matter physics, understanding how materials behave at the quantum level—like why some become superconductors—requires calculating "expectation values" from incredibly complex probability distributions. Scientists use Quantum Monte Carlo (QMC) methods to do this.

The gold standard for these simulations is Hamiltonian Monte Carlo (HMC). Traditionally, HMC includes a "Metropolis-Hastings" (MH) adjustment step. This step acts as a quality control filter: it checks if a proposed move in the simulation is "good" and rejects it if it isn't. This ensures the simulation is mathematically "unbiased" (exact) in the limit of infinite time.

However, there's a catch: as the system size (dimensionality) grows, the MH step becomes a bottleneck. To keep the filter from rejecting almost every move, you have to take tiny, microscopic steps, making the whole simulation crawl. In modern physics problems where dimensions can exceed $10^5$, this "exactness" comes at a crippling cost in speed.

Methods

The authors propose using unadjusted samplers: uLMC (Unadjusted Langevin Monte Carlo) and uMCLMC (Unadjusted Microcanonical Langevin Monte Carlo).

The key innovation isn't just removing the MH step, but how they manage the resulting bias:

  • Automated Tuning: They developed a procedure to automatically tune the simulation's step size ($\epsilon$). Instead of targeting a specific "acceptance rate" (which doesn't exist without the MH step), they target a specific variance in the energy error.
  • Microcanonical Dynamics: They highlight uMCLMC, which uses "isokinetic" dynamics—essentially keeping the "speed" of the simulation constant on a high-dimensional sphere, which helps maintain stability.
  • Bias Diagnosis: They introduce a way for researchers to estimate the bias of their results without needing a known ground truth, simply by comparing runs with different step sizes.

Datasets

The researchers evaluated their methods on the spin-fermion model, a challenging and physically significant benchmark in condensed matter physics used to study quantum phase transitions and "strange metals." * Clean Systems: Idealized lattices where properties are uniform. * Disordered Systems: More realistic models where interaction strengths vary randomly across the lattice, introducing significant "noise" and complexity.

Key Insights and Ideas

  • Bias is a Bogeyman: We often prioritize "unbiased" algorithms, but for any finite simulation, the error from random sampling (variance) is usually much larger than the systematic bias introduced by these unadjusted methods. If the bias is smaller than the variance, it's effectively invisible.
  • Scaling with Dimension: While adjusted HMC slows down significantly as systems get larger, the efficiency of unadjusted samplers remains remarkably stable. This makes them "future-proof" for increasingly massive simulations.
  • Self-Correction: The authors found that the bias doesn't just grow uncontrollably; it peaks near "critical points" (where the physics gets weird) but can be reliably detected and suppressed by halving the step size.

Key Results

The unadjusted methods (uLMC and uMCLMC) consistently outperformed the standard adjusted HMC (aHMC) across various tests:

  • 10x Efficiency Boost: The unadjusted samplers provided roughly an order of magnitude improvement in Effective Sample Size (ESS) per gradient call. This means they get to the right answer 10 times faster for the same amount of "work."
  • Controlled Accuracy: In tests on the spin-fermion model, the systematic bias remained below the statistical error of 100 independent samples, meaning the results were practically as accurate as the "exact" version but much faster.
Figure 2: An efficiency comparison of aHMC, uLMC and uMCLMC at L = 64, and Nτ = 40 for the spin-fermion model (section 4) at varying λ. The unadjusted samplers (uMCLMC, uLMC) perform an order of magnitude better compared to aHMC both in terms of autocorrelation (ESSa) and samples to low error (ESSn), as defined in section 5.

Limitations and Future Work

  • Hyperparameter Sensitivity: The tuning process depends on a target "energy variance" ($E_V$). While $5 \cdot 10^{-8}$ worked well for this study, different physical systems might require different targets, meaning the process isn't yet 100% "set and forget."
  • Ultra-High Precision: In niche applications where extremely high precision is required (and thus very large sample sizes are used), the systematic bias might eventually become the dominant error source, requiring even smaller step sizes.
  • Future Directions: The authors suggest applying these methods to even larger systems and exploring their use in "sign-problem" systems, where standard QMC usually fails. They also aim to further automate the tuning of the $E_V$ parameter.

Summary updated

Abstract

In quantum systems, physical quantities of interest can be expressed as expectation values of a probability distribution, which allows their estimation by Markov Chain Monte Carlo (MCMC) methods. This approach, an instance of Quantum Monte Carlo (QMC), plays a central role in obtaining predictions of system properties relevant to condensed matter and high energy physics. The use of Hamiltonian dynamics to design a Markov kernel, known as Hybrid or Hamiltonian Monte Carlo (HMC), is widespread in quantum applications. In this context, it is standard to use an HMC kernel that is ``adjusted'' with the Metropolis-Hastings (MH) criterion, so that estimates of expectations converge to their true value in the limit of infinite sample size. However, recent work in computational statistics suggests that an \emph{unadjusted} HMC kernel can be significantly more efficient, while safely maintaining an asymptotic bias which is small relative to the error arising from the variance of the finite sample size. We adapt this approach to the QMC setting, focusing on a challenging and high-dimensional \emph{spin-fermion} model of a quantum phase transition. We find that unadjusted methods outperform adjusted HMC in terms of effective sample size per gradient call, while retaining accurate results.
BibTeX
@misc{cohngordon2026fast,
  author = {Reuben Cohn-Gordon and Gabriel Pescia and Sumner N. Hearth and Jakob Robnik and Uroš Seljak and Peter Lunts},
  title = {{Fast Accurate Quantum Monte Carlo without Metropolis Adjustment}},
  year = {2026},
  eprint = {2611.10187},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.10187},
}

Loading related content...

Figure