Variance reduction for forces and pressure in variational Monte Carlo

arXiv:2603.14521 · cond-mat.str-el, quant-ph · Submitted 2026-03-15 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: Today's paper: "Variance reduction for forces and pressure in variational Monte Carlo".

Mira: Accurate evaluation of atomic forces and pressure is central to first-principles studies of quantum manybody systems,

Kai: First, who's behind it and why it matters.

Title and authors: Kai: So we're diving into "Variance reduction for forces and pressure in variational Monte Carlo" by Linteau, Moroni, Carleo, and Holzmann here. It sounds like the paper is tackling a very specific problem within quantum many-body simulations.

Mira: Exactly what it sounds like is that they are focusing on making the calculations for atomic forces and pressure more reliable when using variational Monte Carlo methods. The title tells us immediately that their main goal is variance reduction in these key quantities.

Lev: From a computational standpoint, I'm wondering how much of this variance reduction actually translates to something useful for running simulations on real quantum hardware. If the noise is still too high, even with these tricks, the results might just be meaningless noise.

Kai: That's a fair point, Lev; it's not just about reducing statistical noise in a simulation; it's about getting results that can actually be used to understand physical systems. I’m curious what kind of systems they are focusing on when they talk about these forces and pressures.

Mira: The paper is showing that these techniques aren't just for forces; the underlying ideas are broader, applying to things like pair correlation functions and angulardistribution functions too. That suggests a much wider applicability beyond just calculating simple derivatives.

Lev: If it applies broadly, then the practical implementation matters a lot. I need to know if these methods require completely rewriting our standard VMC codes or if they are just cheap post-processing improvements that we can slap on top of what we already have.

Kai: The authors seem to be aiming for practical strategies that fit within standard Monte Carlo codes, which is encouraging because it means this isn't some theoretical fantasy for an impossibly complex setup.

Mira: They are presenting simple and practical strategies, suggesting the focus is on making these techniques accessible to researchers who are already working with VMC calculations. It’s about refining existing work rather than inventing something entirely new from scratch.

The paper's summary: Kai: So, what's the core message here? Essentially, this paper argues that standard estimators for Hellmann–Feynman forces and Pulay contributions in VMC calculations have variance problems, specifically a power-law divergence that becomes problematic near nodal surfaces.

Mira: That divergence is the main issue they are addressing; they show how a minor modification based on the Metropolis acceptance ratio can soften that power-law divergence into something more manageable, specifically making it logarithmic for Pulay-type contributions.

Lev: Logarithmic scaling versus polynomial scaling is a big deal in terms of convergence rates. For error correction researchers, that difference between those two scalings could mean the difference between a simulation running for a few seconds and one that requires days to get meaningful statistics.

Kai: It seems they’ve also introduced several other specific methods to tackle these issues, including covariance forms, an acceptance trick, and regularization of the Pulay terms which introduces smooth cutoffs.

Mira: I see how the introduction of the covariance form allows for subtracting a mean term to leave behind a variance-reduced estimator that is straightforward to implement in standard Monte Carlo codes. That's a very practical step for anyone trying to get these results out quickly.

Lev: A covariance form sounds good analytically, but let's talk about the acceptance trick; if it only works by modifying the Metropolis acceptance ratio, does that introduce new assumptions that might limit its use in highly specific physical scenarios?

Kai: The paper suggests this acceptance trick can attenuate variance spikes induced by the nodal surface and improve scaling from polynomial to logarithmic with distance to those nodes, which is a nice feature for stability.

Mira: And they also discuss regularization of Pulay terms using smooth cutoff functions, which allows them to make the variance finite for any fixed regularization parameter by showing specific asymptotic scalings like O(one/epsilon), O(epsilon two), and others depending on epsilon.

The paper's improvements: Kai: So when we look at the specific improvements they propose, it’s a mix of analytical derivations and practical computational fixes, like deriving a covariance form for Pulay estimators.

Mira: That covariance form is key because it provides an analytical guarantee that as the trial wave function gets closer to an exact eigenstate, the variance of the Pulay contribution automatically suppresses itself because Var(EL) to zero provided another term stays finite.

Lev: That analytical guarantee is powerful for theoretical work, but I still need to know how robust this holds up when we introduce realistic noise or imperfections that are common in real quantum hardware setups.

Kai: The paper also introduces the acceptance trick as a cheap post-processing improvement, which it suggests can attenuate those variance spikes caused by the nodal surface and change the scaling behavior from polynomial to logarithmic with distance to those nodes.

Mira: That logarithmic scaling improvement is significant because it shows a systematic way to reduce variance without necessarily needing heavy regularization schemes upfront. It’s a systematic route for reducing noise in parameter-dependent expectation values, as mentioned in page one of this paper.

Lev: If we look at the integration by parts approach for Hellmann–Feynman forces, they introduce IBP1 and IBP2 estimators to cure the short-range divergence of the electron-nucleus interaction term. How does that change the scaling compared to what we see in standard methods?

Kai: The first step, IBP1, partially removes the coalescence singularity at leading order, changing the estimator scaling from O(one/x two) down to O(one/x).

Mira: Then they use a second integration by parts step with a gradient-to-Laplacian identity to cancel that singularity exactly in the IBP2 estimator, which keeps the result finite even at coalescence points.

Lev: So for running these on real hardware, using an estimator that is guaranteed to be finite at singularities is crucial because we can't just throw away data points where those things happen. It’s about ensuring the measurement itself doesn't break down.

Conclusion: Kai: So, to wrap up this discussion on "Variance reduction for forces and pressure in variational Monte Carlo," the paper presents a suite of simple and practical strategies to tame the variance issues in estimating atomic forces and pressure.

Mira: They’ve shown that by using techniques like the covariance form, acceptance tricks, or IBP constructions for forces, we can systematically reduce statistical uncertainty, even when dealing with tricky nodal surfaces.

Lev: For me, what this means is that we might be able to push the limits of what's calculable on current quantum simulators because the underlying variance is lower than previously thought. It’s a necessary step toward running more complex molecular dynamics simulations reliably.

Kai: I think these improvements offer a concrete roadmap for making VMC calculations for forces and pressure more robust and computationally feasible in practice.

Mira: The broader implication is that these ideas aren't just confined to forces; the systematic framework for extending variance reduction to other observables, using things like codimension-zero control variates, opens up ways to handle pair correlations and angular distributions too.

Lev: If we can reliably calculate forces and pressures with this reduced noise, it certainly helps in the long-term goal of accurately predicting structural phase transitions in materials under high pressure.

Kai: It sounds like a solid foundation for improving our quantum simulations, and I think we'll be looking for papers that build on these ideas next.

Mira: Agreed; it’s about refining the estimators so that the statistical uncertainty doesn't overwhelm the physical signal we are trying to extract from those complex many-body systems.

Lev: It’s just a solid piece of methodology, and I think seeing how this translates into error bounds for real hardware will be the most telling test.

Institute of Physics, École Polytechnique Fédérale de Lausanne (EPFL) · Center for Quantum Science and Engineering, École Polytechnique Fédérale de Lausanne (EPFL) · CNR-IOM DEMOCRITOS, Istituto Officina dei Materiali and SISSA Scuola Internazionale Superiore di Studi Avanzati · Univ. Grenoble Alpes, CNRS, LPMMC

cond-mat.str-el, quant-ph

Submitted: 2026-03-15

Updated: 2026-10-01

Journal ref: J. Chem. Phys. 165, 134106 (2026)

DOI: 10.1063/5.0336683

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: Accurate evaluation of atomic forces and pressure is central to first-principles studies of quantum manybody systems, and this work presents simple and practical strategies to reduce the variance of

Key concepts

Pulay contributions
These are specific terms arising when calculating energy derivatives in quantum manybody systems. Standard estimators for these terms suffer from high variance that diverges, meaning the results become unreliable unless special techniques are used to control this instability.
Hellmann–Feynman forces
These represent the forces acting on nuclei derived from the change in energy with respect to nuclear coordinates. The paper introduces integration by parts methods to handle short-range divergences that plague standard force estimators, resulting in more stable calculations.
Virial Pressure ($P_V$)
This component of total pressure is calculated directly from the scaled Hamiltonian. It is noted for having finite variance, making it a reliable and straightforward quantity to calculate precisely in Monte Carlo simulations.
Dilation operator (D)
This mathematical tool is introduced to handle the Pulay wave function pressure ($P_{wf}$). It acts as a generator of isotropic scaling, which helps in expressing the wave function component of pressure in a way that is easier to manage computationally.

Terminology

Summary

Accurate evaluation of atomic forces and pressure is central to first-principles studies of quantum manybody systems, and this work presents simple and practical strategies to reduce the variance of Monte Carlo estimators for these quantities.

The gist: Simple and practical strategies are presented to reduce the variance of Monte Carlo estimators for atomic forces and pressure in variational Monte Carlo calculations, showing that a minor modification based on the Metropolis acceptance ratio softens the power-law divergence of the variance to a logarithmic one for Pulay-type contributions.

Variance Reduction Strategies

The paper introduces several practical methods to address the divergent second moments encountered in standard estimators for Hellmann–Feynman and Pulay contributions. These strategies are focused on improving reliability while maintaining computational feasibility within standard Monte Carlo codes:

  1. Covariance: The general decomposition of energy derivatives into Hellmann–Feynman and Pulay contributions is derived, leading to a covariance form for the Pulay estimators:

∂E/∂p = 2Re Cov(EL, ∂p log Ψ∗) (8)

This formulation allows for the subtraction of the mean term, leaving a variance-reduced estimator that is straightforward to implement in standard Monte Carlo codes.

  1. Acceptance trick: This method revisits an acceptance trick from nearly half a century ago and shows it can attenuate the variance spikes induced by the nodal surface, improving the variance scaling from polynomial to logarithmic with the distance to the nodes. It is applied as a cheap post-processing improvement to reduce numerical instabilities in Pulay terms.

  2. Regularization of Pulay terms: To make variance finite for any fixed regularization parameter, a smooth cutoff function is introduced. The asymptotic scalings for the regularized estimators are summarized as:

estimator variance bias O(r) ξ−1c 0 Oe log(ξ0/ξc) 0 Oe(1)ϵ, Oe(2)ϵ log(ξ0/ϵ) O(ϵ2), Obϵ = χϵOe log(ξ0/ϵ) O(ϵm+2).

Hellmann-Feynman Force Estimators

For Hellmann–Feynman forces, the paper presents two integration by parts (IBP) estimators to cure the variance problems arising from the short-range divergence of the electron-nucleus interaction term.

  1. First IBP step: This step partially removes the coalescence singularity to leading order. The resulting estimator is:

F IBP1I,sr(r) = −2X i v sren(ri − RI)∇ri log Ψ (37)

This estimator scales as O(1/x), compared to the default estimator scaling of O(1/x2).

  1. Second IBP step: This step uses a gradient-to-Laplacian identity to cancel the singularity exactly, yielding the IBP2 estimator:

F IBP2Iα(r) = −X j ∇jQIα · ∇j log Ψ − RIα (41)

This estimator remains finite at coalescence points, unlike the default estimator.

Pressure Estimators

The paper addresses the pressure observable, defined as P = −∂E/∂omega, which splits into a Virial pressure component and a Pulay wave function component: P = P V + P wf.

  1. Virial Pressure (P V): The explicit derivative of the scaled Hamiltonian yields the standard Virial estimator:

P V = 1/3omega2Ke + V (21)

This estimator is noted to have finite variance and is therefore not challenging to calculate precisely.

  1. Pulay Wave Function Pressure (P wf): The wave function contribution is expressed as:

P wf = −2/3omegaRe D ∆E(rR,L)D log Ψ∗(rR,L) (22)

The dilation operator D is introduced as a generator of isotropic scaling.

General Observables and Unification

The framework extends beyond forces and pressure to other observables using divergence theorem identities.

  1. Codimension-0: Zero-mean control variate correction uses the Langevin Stein operator Lπ[A] to show that for any observable O, ⟨O⟩ = ⟨O + Lπ[A]⟩ (C4).

  2. Codimension-1: Surface-to-volume conversion converts an integration over a shell into an integration over a ball, increasing the signal-to-noise ratio by choosing vector fields A such that A(x; s) · ∇ξ(x) = W(x) (C10). This is applied to pair correlation functions, yielding improved estimators for g(s).

Improvements for AI systems

As a fastidious and diligent researcher, I have analyzed this paper, Variance reduction for forces and pressure in variational Monte Carlo, by Linteau et al. The core contribution is the development of several variance-reduction techniques applicable to Variational Monte Carlo (VMC) estimators for atomic forces and pressure.

Here are the specific improvements that can be made to AI systems, followed by what these improved systems can achieve:

) Improvements Applicable to AI Systems:

  1. The integration of practical variance reduction strategies into VMC frameworks (specifically, Covariance forms, Acceptance tricks, and Integration by Parts constructions).

  2. The derivation and implementation of analytical methods for calculating the Hellmann-Feynman forces in both periodic and open boundary conditions using IBP techniques (IBP1 and IBP2 estimators).

  3. The systematic framework for extending variance reduction to higher-order observables, including pair correlation functions, angular distribution functions, and multi-constraint structural observables using codimension-0 to codimension-2 identities.

  4. The development of smooth regularization schemes (polynomial ansatz cutoffs) that systematically reduce bias while maintaining logarithmic variance scaling for Pulay terms.

  5. The application of the Langevin Stein operator framework (Codimension-0 control variates) to create unbiased estimators for structural observables like one-body density and pair correlations by replacing noisy delta function integrals with volume integrals via a specific vector field choice.

) Capabilities of the Improved AI System:

The improved AI system, trained or integrated with VMC methods leveraging these techniques, can perform the following high-precision tasks:

  1. Precision Determination of Quantum Many-Body Properties (Forces and Stress):

  2. Highly Accurate Equations of State Calculations:

  3. Simulation of Complex Molecular Dynamics under High Pressure Conditions:

  4. Reliable Prediction of Structural Phase Transitions in Materials Science:

) Specific Applications Detail:

  1. The AI can calculate the exact atomic forces and pressure (stress tensor) for large, strongly correlated electronic systems (e.g., metallic hydrogen, as demonstrated in the paper). Unlike default VMC estimators that suffer from infinite variance or severe noise near nodal surfaces, this system will provide force estimates with significantly reduced statistical uncertainty.

  2. The system can accurately model the behavior of materials under extreme conditions (e.g., high-pressure hydrogen at 650 GPa), enabling the prediction of structural changes and phase transitions that are currently computationally prohibitive due to noise in standard VMC force calculations.

  3. It can perform long-duration, statistically robust molecular dynamics simulations where the forces are calculated efficiently using the IBP estimators (IBP1/IBP2), allowing for more accurate trajectory predictions and relaxation studies compared to methods relying on noisy default Pulay terms.

  4. The system can calculate complex structural correlation functions (like pair correlations, shown in Fig. 6) with high fidelity, even when probed at short distances, thanks to the codimension-2 volume integration techniques that replace noisy delta-function binning with stable volume averages.

  5. It can serve as a foundation for training highly accurate classical Machine Learning Interatomic Force Fields (ML-IFFs). The variance reduction techniques provide the necessary high-quality, low-noise force data needed to train ML models that are themselves robust and transferable across different geometries, significantly accelerating the design of new materials.

Sources

Related papers