⚛️ lattice

Gauge-Covariant Wavelets for Quark-Propagator Sparsening and Distillation in Lattice QCD

This paper introduces a novel gauge-covariant wavelet framework that utilizes spatial localization and resolution-scale separation to efficiently sparsen quark propagators and construct distillation subspaces without eigenpair calculations, thereby significantly reducing the computational cost of hadronic correlation functions while maintaining accuracy comparable to conventional methods.

Original authors: Debsubhra Chakraborty, Nilmani Mathur, Mrinmoy Basak

Published 2026-10-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Debsubhra Chakraborty, Nilmani Mathur, Mrinmoy Basak

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

To understand the universe at its most fundamental level, physicists often turn to a method called lattice quantum chromodynamics, or lattice QCD. Imagine the fabric of space and time not as a smooth, continuous sheet, but as a vast, three-dimensional grid of tiny points, like the pixels on a screen but in three dimensions. On this grid, scientists simulate the behavior of quarks, the tiny particles that stick together to form protons, neutrons, and other building blocks of matter. The goal is to calculate how these particles interact to create the properties of the atoms that make up our world. However, the universe is incredibly complex, and the grid used for these simulations can be enormous, containing millions of points. Calculating the interactions of particles across such a massive grid requires immense computing power, often pushing the limits of the world's most powerful supercomputers. As researchers try to study more complex systems, like groups of protons and neutrons bound together in atomic nuclei, the calculations become so expensive that they are nearly impossible to complete.

A team of researchers at the Tata Institute of Fundamental Research in India has developed a new way to make these calculations faster and more efficient without losing accuracy. They introduced a mathematical technique based on wavelets, a tool originally designed for compressing digital images and signals. In the context of their work, wavelets allow them to focus on the most important parts of the simulation while ignoring the less critical details, much like how a photographer might zoom in on a subject while blurring the background. By applying this method, they were able to simplify the complex web of particle interactions, reducing the amount of data that needs to be processed. Their approach involves two main strategies: one that smooths out the particle paths to remove unnecessary noise, and another that creates a smaller, more manageable set of "building blocks" to represent the particles. These new blocks are localized, meaning they are concentrated in specific areas of the grid rather than spread out everywhere, which makes the calculations significantly faster.

The researchers tested their new methods by simulating several different types of particles, including single protons and neutrons, pairs of pions, and even a deuteron, which is a nucleus made of one proton and one neutron. They compared their results against traditional, more computationally expensive methods to ensure they were getting the same answers. The findings showed that their wavelet-based approach produced energy levels and particle behaviors that matched the traditional results almost perfectly, within the limits of statistical uncertainty. This means they achieved the same level of precision but with a fraction of the computational effort. In some cases, the new method was found to be five times faster just in setting up the basic calculations, and it offered even greater speedups when calculating how the particles interact with each other.

One of the key challenges the team had to overcome was a problem known as "aliasing," which occurs when a grid is too coarse to capture the true movement of a particle, causing it to appear as if it is moving in a different direction or at a different speed. To fix this, the researchers developed a clever trick where they randomly shifted the starting position of their grid for each calculation. By averaging the results from these many different shifts, they were able to cancel out the errors and recover the true physical behavior of the particles. This technique allowed them to use a coarser, faster grid without introducing false results, a crucial step for studying particles that are moving at high speeds.

The study also explored how to handle the massive amount of data generated when calculating the interactions of multiple particles. In traditional methods, every possible interaction must be calculated exactly, which becomes overwhelming as the number of particles increases. The researchers used a statistical sampling method to estimate these interactions. Instead of calculating every single connection, they focused on the most significant ones, using the localized nature of their wavelet blocks to identify which parts of the grid mattered most. This approach allowed them to speed up the calculations by a factor of three to four while maintaining the same level of accuracy. The results suggest that this new framework could be a game-changer for studying larger and more complex nuclear systems, potentially allowing scientists to explore questions about the structure of matter that were previously out of reach due to computational costs.

The implications of this work extend beyond just faster calculations. By proving that localized, wavelet-based methods can reproduce the results of traditional, global methods, the researchers have opened the door to new ways of thinking about how to simulate the quantum world. Their success with the deuteron and other multi-particle systems indicates that these techniques are robust enough to handle the complexities of real-world nuclear physics. As supercomputers continue to grow in power, these efficient methods will be essential for making the most of that power, allowing researchers to tackle problems involving larger volumes and more intricate operator bases. The work establishes a practical foundation for future studies, suggesting that the combination of spatial localization and smart sampling can unlock new frontiers in our understanding of the subatomic universe.

Technical Summary: Gauge-Covariant Wavelets for Quark-Propagator Sparsening and Distillation in Lattice QCD

Problem Statement
Lattice QCD calculations of hadronic correlation functions face significant computational bottlenecks, particularly for multi-hadron systems and large-volume lattices. Two primary challenges limit the scope of these calculations:

Distillation Scaling: The standard distillation method projects quark fields onto the subspace spanned by the lowest eigenmodes of the gauge-covariant spatial Laplacian. While effective, the number of required eigenmodes ( $n_v$ ) grows linearly with the physical spatial volume to maintain fixed ultraviolet resolution. This leads to expensive eigensolves, massive eigenvector storage requirements, and high input/output costs. Furthermore, Wick contractions within this subspace remain computationally demanding, scaling polynomially with $n_v$ .
Propagator Sparsening Artifacts: An alternative approach, quark-propagator sparsening, evaluates contractions on a coarse spatial sublattice to reduce costs by a factor of $N^3$ . However, direct decimation of the sink introduces two artifacts: enhanced excited-state contamination at early Euclidean times and momentum aliasing (coupling to unwanted momenta). While smearing and random grid shifts can mitigate these, a systematic framework combining multiresolution analysis with gauge covariance has been largely unexplored.

Methodology
The authors introduce a novel gauge-covariant multiresolution framework based on wavelets, realized through two complementary approaches:

Gauge-Covariant Fast Wavelet Transform (FWT) for Sparsening:
- Instead of direct decimation, the authors apply a low-pass filter based on compactly supported Cohen–Daubechies–Feauveau (CDF) biorthogonal wavelets (specifically Bior 4,4 and Bior 5,5) to quark propagators.
- To maintain gauge covariance, parallel transporters (Wilson lines) are inserted to combine fields at different spatial sites before filtering. These transporters are constructed using HYP or APE smeared links and averaged over shortest paths.
- The transform recursively coarse-grains the propagators, reducing the number of sink sites by a factor of $8^s$ (where $s$ is the resolution level).
- To eliminate momentum aliasing induced by the coarse grid, the origin of the coarse grid is randomly shifted on each gauge configuration. This converts systematic bias into statistical fluctuations that average to zero.
Wavelet Distillation:
- This approach constructs a localized approximation space without computing Laplacian eigenpairs.
- Basis functions are generated by applying a resolvent filter, $(I - \tau \nabla^2)^{-m}$ , to color-vector delta sources placed on a coarse sublattice.
- The resulting non-orthogonal basis functions are orthogonalized using symmetric Löwdin orthogonalization to form a distillation projector.
- This method avoids the eigensolve entirely, relying instead on sparse linear solves (conjugate gradient) which are computationally cheaper and allow for on-demand basis regeneration, eliminating persistent storage needs.
- The framework is extended to include momentum smearing for boosted states by modifying the gauge links in the Laplacian.
Stochastic Estimation with Hansen–Hurwitz (HH) Sampling:
- Combining stochastic LapH (using diluted noise vectors in the wavelet basis) with HH sampling for sink contractions.
- Because the wavelet basis is spatially localized, the hadronic elementals (tensor contractions) are sparse. The authors use the magnitude of these elementals to define a probability distribution for importance sampling.
- A "coarse-site" sampling strategy is employed where only the coarse-grid site indices are sampled stochastically, while color sums are evaluated exactly. This significantly reduces the cost of sink contractions.

Key Results
Numerical benchmarks were performed on $N_f=2+1+1$ MILC and $N_f=3$ CLS ensembles for single hadrons (pion, $\rho$ , nucleon), the $I=2$ two-pion system, the deuteron channel, and boosted nucleons.

FWT vs. Direct Decimation: The FWT-based sparsening significantly reduces early-time excited-state contamination compared to direct decimation at equal spatial reduction levels ( $s=1, 2, 3$ ). Random grid shifts successfully remove momentum aliasing, recovering the correct energy levels. At level $s=2$ (64-fold reduction in sink sites), the method achieves sub-percent agreement with full correlators for single-hadron energies while reducing contraction costs by a factor of 64.
Wavelet Distillation vs. Conventional Distillation: The wavelet distillation basis (constructed via resolvent filters) produces smearing profiles comparable to conventional Laplacian distillation (approx. 88% subspace overlap). Extracted energy levels for single nucleons, two-pion systems, and the deuteron channel agree with conventional distillation within statistical uncertainties.
Computational Efficiency:
- Basis Construction: Constructing the wavelet basis is approximately 5 times faster than constructing the Laplacian eigenbasis on the studied lattices, as it avoids eigensolves.
- Contraction Speedup: Using HH sampling with stochastic wavelet distillation, the authors achieved a 3.8-fold speedup for nucleon contractions and a 3.0-fold speedup for deuteron contractions at comparable statistical precision.
- Scaling: The speedup is expected to increase with the dimension of the distillation space ( $n_v$ ). For $n_v=192$ , the speedup increased to roughly 6.9, as the cost of exact contractions scales as $O(n_v^3)$ while the HH cost scales more favorably with the number of sampled sites.

Significance and Claims
The paper claims to establish a practical foundation for wavelet-based propagator compression and operator construction in lattice QCD. Its primary contributions are:

Alternative to Eigensolves: Demonstrating that a distillation subspace can be constructed directly via resolvent filtering of localized sources, bypassing the expensive and storage-intensive Laplacian eigensolve. This enables on-demand basis generation, which is crucial for large-volume calculations.
Systematic Sparsening: Providing a gauge-covariant multiresolution framework that improves upon direct decimation by controlling excited-state contamination and momentum aliasing through filtering and random shifts.
Scalability: Showing that spatial localization enables efficient stochastic contraction methods (HH sampling) that significantly reduce computational costs for multi-hadron systems.
Versatility: Validating the approach across a range of systems, including boosted nucleons and complex multi-quark operators (hexaquarks), suggesting its applicability to broader spectroscopy and scattering calculations on larger lattices.

The authors emphasize that while the methods show substantial promise, the benefits rely on controlling approximations (aliasing, sampling variance) and that further optimization of filters and sampling distributions is required for even larger volumes.

More like this