Variance reduction with probing and Multilevel Monte Carlo in Lattice QCD

arXiv:2607.05157 · hep-lat, cs.NA, math.NA · Submitted 2026-07-06 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: High Energy Physics: Lattice. The day in lattice field theory: what the simulations put on the table.

Kai: We're looking at the paper: "Variance reduction with probing and Multilevel Monte Carlo in Lattice QCD".

Mira: Here's the gist of the paper. Variance reduction with probing and Multilevel Monte Carlo in Lattice QCD investigates two complementary variance reduction strategies to improve trace estimation in lattice QCD computations.

Kai: The results show that multigrid MLMC is most effective on observables which are strongly influenced by low modes, whereas probing is most effective when the dominant fluctuations are short ranged and local.

Mira: Multilevel Monte Carlo Variance Reduction The multigrid multilevel Monte Carlo (multigrid MLMC) method uses a multigrid hierarchy to construct an unbiased multilevel estimator via recursive coarse grid corrections available from the multigrid hierarchy of the solver.

Kai: This approach is based on Multilevel Monte Carlo (MLMC)

6: , which relies on an additive decomposition of a stochastic variable into stochastic variables where the cost to compute a sample from each variable is small when its variance is large.

Title and authors: Kai: We're talking about "Variance reduction with probing and Multilevel Monte Carlo in Lattice QCD" and who wrote it. The title tells you immediately that they are combining two specific techniques to fix the problem of estimating traces in these complex lattice quantum chromodynamics problems.

Mira: The authors are Frommer, Jimenez-Merchan, Lang, Papace, and Ramirez-Hidalgo. They’re coming from a mix of backgrounds: some from the university in Wuppertal and others from Jülich Supercomputing Centre. That suggests they've got expertise in both the theoretical modeling and the high-performance computing side.

Lev: Frommer and Jimenez-Merchan are clearly deep in the lattice QCD machinery, dealing with those Wilson-Dirac operators, while Lang and Papace might be bringing in some of that multigrid structure knowledge. It makes sense they're looking at both low modes and coarse grid corrections together.

Kai: So when you look at their summary, it boils down to this: they show that multigrid multilevel Monte Carlo is really good for observables that are heavily influenced by those very low energy modes of the system.

Mira: And stochastic probing is the other side of the coin; it’s best when the fluctuations you're seeing in your data are short-ranged and local, meaning things happening right next to each other matter most.

Lev: That distinction between long-distance propagation being handled by multigrid MLMC and short-range local fluctuations being handled by probing is crucial for understanding where the computational savings actually land on real hardware.

The paper's summary: Kai: So, to summarize what they found, the paper shows that they can use multigrid multilevel Monte Carlo to build an unbiased multilevel estimator using recursive coarse grid corrections from a multigrid hierarchy of the solver. This is based on Multilevel Monte Carlo principles where you choose samples where the variance is high and cost is low.

Mira: And they do this by creating an oblique projector, which essentially splits the inverse operator into a part that's easy to compute and a part that you only sample with a smaller number of samples. This leads to an estimator called btr(D minus one).

Lev: What I find interesting about the MLMC approach is how it uses that hierarchy of operators, D1 through Dl, along with prolongators and restrictors between levels to create this level splitting. It’s a recursive way to manage the approximation.

Kai: Then they introduce stochastic probing, which is a deterministic method that exploits the sparsity pattern of the operator D to reduce variance by essentially killing off some of those off-diagonal entries in the inverse operator.

Mira: They construct these probing vectors using distance d graph colorings, and they specifically propose a torus based coloring scheme because it requires fewer colors than hierarchical probing at the same distance. This means better scaling for the probing part.

Lev: That torus based coloring is something I need to look into regarding overhead on real hardware; we're always looking for structures that simplify the matrix operations, and reducing the number of colors there sounds like a direct win for complexity.

The paper's improvements: Kai: The paper highlights a big improvement when they test these methods on physical observables, specifically connected pseudoscalar two-point functions. They found that the multilevel decomposition yields a variance reduction of up to O (ten five) at large time separations.

Mira: That massive reduction in variance translates directly into a clear cost reduction when you fix the accuracy level, which is what matters for practical computation on expensive machines. But they note that for disconnected loops, probing combined with dilution gives a substantial cost reduction that improves as you increase the number of probing vectors used.

Lev: So for me, the connected correlator result suggests MLMC is fantastic when dealing with long-distance physics, which aligns perfectly with what we expect from those low modes they were targeting earlier. But I need to see how much that O(ten five) variance reduction actually translates into wall-clock time on a real setup.

Kai: What this means for someone just listening is that if your physics problem has long-distance effects, you can use the MLMC part to get huge savings in computation time without sacrificing accuracy.

Mira: And for those disconnected contributions, it seems like probing with dilution is the way to go when you’re dealing with short-ranged noise dominating those specific loops. It's a tailored approach based on where the error is coming from.

Conclusion: Kai: So wrapping up the "Variance reduction with probing and Multilevel Monte Carlo in Lattice QCD" paper, it seems the main point is that these two variance reduction strategies address different sources of stochastic noise, which is really what makes them so powerful together.

Mira: They've shown that this complementarity means you can target both long-distance and short-distance contributions simultaneously by using the right method for the job. This suggests a path forward where combining MLMC with probing could be really useful in practice.

Lev: I think combining them is the logical next step, trying to target those long-distance and short-distance parts at once, which is where you get the most benefit on real hardware. But we still have to figure out how much that combined strategy actually reduces the overall overhead on our current error correction protocols.

Kai: So, for listeners who just want the big picture, this paper shows that depending on your observable, you pick a strategy—MLMC for low modes and probing for local fluctuations—and you get substantial cost savings.

Mira: Exactly. It's about tailoring the method to the physics problem rather than just applying one universal fix. We’re looking forward to seeing how this idea of combining these approaches plays out in future work.

Bergische University Wuppertal · Jülich Supercomputing Centre Forschungszentrum Jülich GmbH

hep-lat, cs.NA, math.NA

Submitted: 2026-07-06

Updated: 2026-11-09

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

Importance score: 84/100

The gist: Variance reduction with probing and Multilevel Monte Carlo in Lattice QCD investigates two complementary variance reduction strategies to improve trace estimation in lattice QCD computations.

Key concepts

Multigrid MLMC
This method uses a hierarchy of grids to create an unbiased estimator. It recursively uses coarse grid corrections from the solver to estimate the trace, making it very effective for observables strongly affected by low-frequency modes in the system.
Stochastic Probing
This is a deterministic technique that exploits the structure of the matrix (sparsity pattern). It constructs specific 'probing vectors' designed to target and eliminate parts of the off-diagonal entries, significantly reducing variance for local fluctuations.
Connected Correlator
This refers to a specific physical observable in Lattice QCD. The paper found that multigrid MLMC provides a massive variance reduction (up to $O(10^5)$) for this quantity at large time separations, leading to substantial cost savings.
Disconnected Loops
These are contributions from loops that do not connect the initial and final states in the correlator. Probing combined with dilution is shown to substantially reduce the cost for these specific contributions, especially when fluctuations are short-ranged.

Terminology

Summary

Variance reduction with probing and Multilevel Monte Carlo in Lattice QCD investigates two complementary variance reduction strategies to improve trace estimation in lattice QCD computations. p. 0 The results show that multigrid MLMC is most effective on observables which are strongly influenced by low modes, whereas probing is most effective when the dominant fluctuations are short ranged and local. p. 5

Multilevel Monte Carlo Variance Reduction

The multigrid multilevel Monte Carlo (multigrid MLMC) method uses a multigrid hierarchy to construct an unbiased multilevel estimator via recursive coarse grid corrections available from the multigrid hierarchy of the solver. This approach is based on Multilevel Monte Carlo (MLMC) [6], which relies on an additive decomposition of a stochastic variable into stochastic variables where the cost to compute a sample from each variable is small when its variance is large. p. 2 The cost-minimizing choice for the number of samples Nl for Xl to reach a target variance epsilon squared is given by equation (6). p. 2 In a multigrid setting, the operator hierarchy D1 = D, Dl in C n l x n l, for l = 1,. p. 2, L together with prolongators Pl and restrictors Rl between levels l and l + 1. p. 2 Multigrid MLMC realizes a recursive deflation using the oblique projector Pi l = Pl D-1 l+1 P T l D l. p. 2 This induces the level splitting D-1 l = M l + P l D-1 l+1 P T l, where M l:= D-1 l - Pl D-1 l+1 P T l. p. 2 The unbiased multigrid MLMC estimator btr(D-1) is given by equation (12). p. 3

Stochastic Probing Variance Reduction

Stochastic probing is a deterministic method that exploits the sparsity pattern of D to reduce the variance in equation (4) by annihilating parts of the off-diagonal of D-1. This technique involves constructing structurally orthogonal probing vectors to extract specific non-zero entries of the matrix using a distance-d coloring. p. 3 Probing vectors are built in the following way, where v j(i) = (±1, if color(i) = j 0, otherwise. p. 3). This allows probing vectors to replace Rademacher vectors in equation (3), resulting in a new estimator equation (14). p. 3 The trace is then approximated as tr(D-1) ≈ ∑j=1 n c eta j T D-1 v j(i). p. 0

Complementarity in Observable Performance

The results highlight a pronounced complementarity between the two approaches, showing that deflation schemes are most effective for observables dominated by long distance propagation, while probing is most effective for localized quantities. For the connected pseudoscalar two-point function, the multilevel decomposition yields a variance reduction of up to O (10 5) at large time separations and translates into a clear cost reduction at fixed accuracy. p. 0 In contrast, for disconnected loops, probing combined with dilution delivers a substantial cost reduction that improves as the number of probing vectors is increased. p. 0

Application to Physical Observables

The numerical results demonstrate that multigrid MLMC is most effective on observables which are strongly influenced by low modes, while probing is most effective when the dominant fluctuations are short ranged and local. For the connected correlator, the dominant fine-level contribution G1,1 exhibits a pronounced variance reduction relative to plain Hutchinson. p. 6 For disconnected contributions, probing for the disconnected loop contribution, combined with dilution achieves a substantial reduction in cost. p. 8

Conclusions and Future Directions

Taken together, our numerical results show that multigrid MLMC and probing address complementary sources of stochastic variance. The observed complementarity motivates studying combined strategies, in particular the integration of multigrid MLMC with the proposed probing approach, with the goal of simultaneously targeting long-distance and shortdistance contributions to the variance. p. 8 Further work should benchmark the three dimensional torus coloring against hierarchical probing in a controlled comparison at fixed cost. p. 8 The observed complementarity suggests that both approaches can deliver cost reductions of about one order of magnitude in the regimes where they are best matched to the observable. p. 8

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Multigrid MLMC implementation allows AI systems to compute traces of large matrix inverses with a cost reduction of about one order of magnitude for connected correlators at fixed accuracy. This enables more precise results from expensive lattice QCD computations by shifting variance to coarser, cheaper levels where solves are significantly less costly.

  2. Stochastic probing, combined with dilution, allows AI systems to achieve substantial cost reductions for disconnected fermion loops, which is particularly effective when the dominant fluctuations are short ranged and local. This method can be leveraged to compute disconnected correlation functions more efficiently than plain Hutchinson estimators in specific regimes.

  3. The proposed torus based coloring scheme requires substantially fewer colors than hierarchical probing at the same distance, suggesting that AI systems using this coloring can construct more efficient stochastic probing vectors for extracting localized quantities. This leads to a faster and potentially more accurate estimation of traces by replacing purely random vectors with structured probing vectors.

Sources

Related papers