Lattice QCD computes properties of the proton and neutron by simulating quarks and gluons on a spacetime grid. Its central bookkeeping problem is a trade-off in "Euclidean time," the imaginary-time direction along which signals are measured. At short times the data are precise but contaminated by excited states, meaning heavier states that the measuring operator also creates. At long times those states have died away, but the noise grows exponentially. The proposal is to stop treating the noisy, expensive part of the measurement as something to fit around. Instead, a bidirectional infilling model, of the kind speech editors use to fill a cut in a recording from the audio on both sides, learns to predict it on each gauge configuration from cheap, precise data. That prediction is used only as a control variate: the model's error is measured on a small random subset and subtracted, so the final estimate stays exactly unbiased. The question is whether this buys statistical precision at fixed computing cost.
The concrete target is the time profile of nucleon three-point functions. A three-point function places a weak-current insertion at time $\tau$ between a nucleon created at time 0 and absorbed at sink time $t$. Its middle region gives the axial coupling $g_A$ that governs neutron beta decay. Its shape near the two ends encodes the excited-state contamination. The CalLat collaboration avoided computing this profile directly by summing the insertion over all times, which gives every sink time from one extra solve. They reached a percent-level $g_A$, but their error budget is still dominated by statistics and the summation discards the profile (A per-cent-level determination of the nucleon axial coupling from quantum chromodynamics; Nucleon axial coupling from Lattice QCD). We propose to recover the profile cheaply. It would be computed exactly on a few sources, and on every other source it would be predicted from that source's summed correlator and two-point function, with context taken from both before and after the sink time.
Core Claim
On each source, the expensive fixed-sink profile $y(\tau; t)$ fluctuates together with cheap quantities computed on the same source. These cheap quantities are the two-point function $C_2(t')$ at all $t'$ and the summed (Feynman-Hellmann) correlator $C_{\rm FH}(t')$ at all $t'$. A predictor $\hat y$ that sees this cheap data on both sides of $t$ can absorb much of the per-source noise. We form the estimator
$$\hat Y(\tau;t) = \frac{1}{|S|}\sum_{s\in S}\big[y_s - \hat y_s\big] + \frac{1}{N}\sum_{s=1}^{N}\hat y_s ,$$
where $N$ is the number of sources with cheap data and $S$ is a random subset of size $|S|\ll N$ on which the profile is also computed exactly. $\hat y$ is trained on held-out sources through cross-fitting, so it is independent of the samples it corrects. The first term removes any model bias in expectation, which satisfies the requirement that machine-learned components in lattice field theory carry exact bias correction (Applications of Machine Learning to Lattice Quantum Field Theory). The variance is approximately $\mathrm{Var}(y-\hat y)/|S| + \mathrm{Var}(\hat y)/N$. When $N \gg |S|$, the gain over using $|S|$ exact sources alone approaches $1/(1-\rho^2)$, where $\rho$ is the per-source correlation between $y$ and $\hat y$; this follows from the standard control-variate variance formula. The proposal stands or falls on whether bidirectional context pushes $\rho$ high enough, roughly above 0.8, to give a gain of about three, after the cost of the exact subset is paid.
"Bidirectional" carries specific content here. On one source, the summed correlator's growth between $t-1$ and $t+1$ approximately sets the integral of the profile at sink $t$. The two-point function at $t'>t$ tracks how much ground state versus excited state that source's sink is absorbing. A predictor that sees only data at $t'\le t$ is the analogue of causal left-to-right generation in speech editing. That literature shows such generation produces seams at the far boundary, while infilling from both sides removes them (VoiceCraft: Zero-Shot Speech Editing and Text-to-Speech in the Wild; EditSpeech: A Text Based Speech Editing System Using Partial Inference and Bidirectional Fusion). We take that observation as a testable hypothesis about correlation structure, not as a borrowed result.
Relation to Prior Work
Three lattice strategies attack excited-state contamination at the operator or analysis level. Variational bases with extra operators, such as adding a scalar sigma operator to two-pion operators, isolated clean $\pi\pi$ states at physical pion mass (S-wave pi-pi I=0 and I=2 scattering at physical pion mass; Studies of I=0 and 2 $\pi\pi$ scattering with physical pion mass). Matrix Prony "calm" operators cancel single-nucleon excited states early (Calm multi-baryon operators). Large multi-fit averages over time windows control fit-range dependence (Decay constants and sigma terms from the lattice.). All three change what is measured or how it is fitted. None of them reduces the variance of a fixed measurement, so our method can be combined with any of them.
CalLat's summation method supplies our cheap context and is also our main competitor. Our proposal keeps their data and adds back the profile they gave up.
Machine learning in this physics area has so far meant configuration generation, spectral inversion (Applications of Machine Learning to Lattice Quantum Field Theory), or collider hadronization with GANs (Integrating particle flavor into deep learning models for hadronization; Fitting a deep generative hadronization model). The hadronization models are phenomenological emulators with no exactness guarantee. We deliberately avoid that role: our model never replaces a measurement, it only reduces variance.
The speech infilling systems supply the masking idea (Voicebox: Text-Guided Multilingual Universal Speech Generation at Scale; CampNet: Context-Aware Mask Prediction for End-to-End Text-Based Speech Editing). They also constrain the design. Voicebox learns a full conditional distribution, but a control variate needs only the conditional mean, so we use masked regression rather than flow-matching sampling.
The obvious idea at this intersection is to generatively infill the noisy middle of the ensemble-averaged correlator from its early and late ends (Gap 1 in our notes). We reject it. That amounts to a multi-exponential fit with a learned prior, and it injects bias that no reweighting step corrects.
Research Program
The data object is one source on one configuration. For that source we have $C_2(t')$ and $C_{\rm FH}(t')$ for all $t'$. For sources in $S$, we also have fixed-sink profiles $y(\tau;t)$ at a ladder of sink times, for example $t\in\{8,10,12,14\}$ in lattice units on the coarser ensembles. The predictor is a small transformer over a grid indexed by $t'$ and $\tau$. Cheap channels are always visible, and the profile at the target sink time is masked, in the style of mask-prediction speech editors. Lattice symmetries are imposed by data augmentation: time reversal combined with the appropriate parity and conjugation map the profile onto itself with $\tau\to t-\tau$. A soft penalty ties the predicted profile's sum to the measured summed-correlator increment. The model is trained in K folds over configurations, never over sources within a configuration, so that gauge-field correlations cannot leak between training and correction.
Training cost is counted honestly in core-hours against the inversions it replaces. We expect it to be negligible, because the inputs are kilobyte-scale per source. This avoids the gigabyte-configuration memory bottleneck described in the review (Applications of Machine Learning to Lattice Quantum Field Theory).
The analysis then fits $C_2$, $C_{\rm FH}$ and the control-variate profiles jointly with the multi-state model CalLat uses. Two outputs follow. The first is $g_A$ itself. The second is the ground-to-excited transition matrix elements, which the summed correlator alone constrains poorly and which the profile shape pins down directly.
Baselines
The naive reference point is "spend the same compute on more of what CalLat did." Every core-hour that our method puts into the exact profile subset and into training is reassigned to additional summed-correlator sources, and the analysis is CalLat's joint two-point plus summed fit. This is the strongest competitor because it is the method that produced the percent-level result. The reported benchmark is the final $g_A = 1.271 \pm 0.013$ (A per-cent-level determination of the nucleon axial coupling from quantum chromodynamics), with a preliminary $g_A = 1.285(17)$ whose 1.29% statistical error dominates chiral (0.21%), finite-volume (0.23%) and discretization (0.10%) errors (Nucleon axial coupling from Lattice QCD). The file gives no per-ensemble statistical errors (no reported result found on Lacuna for per-ensemble $g_A$ uncertainties), so we will regenerate this baseline ourselves on identical ensembles and sources. To beat it, our method must reduce the per-ensemble statistical error on $g_A$ at equal total cost by a factor whose square exceeds 1.5. Equivalently, it must save at least a third of the compute. It must do so while shifting the central value by less than its error relative to a brute-force reference.
The second baseline is the classical fixed-sink approach at equal cost: all budget spent on exact profiles over fewer sources. This tests whether the profile is worth having at all.
The third baseline is a linear control variate. It regresses $y$ on the same cheap features with ordinary least squares, per $\tau$, under the same cross-fitting. This is the analogue of the linear corrections already standard in lattice variance reduction, and it is the most dangerous competitor (see below).
Two ablations isolate the proposal's key ingredient. The first is a prefix-only predictor, identical except that it is blinded to $C_2$ and $C_{\rm FH}$ at $t'>t$. The second is a predictor without the summed correlator. All learned variants share the same training sources, folds, tuning budget (a fixed number of hyperparameter trials, chosen on a held-out ensemble) and inversion budget.
Why not just…?
Why not just use a linear control variate? This is a real threat. If per-source fluctuations are dominated by a few shared modes, for example the overall normalization from the source propagator, a linear regression may capture nearly all of $\rho$. The experiment settles it directly: we report $\rho$ and the cost-equivalent gain for linear and nonlinear predictors with identical context. If the linear version matches, the contribution narrows to the estimator design itself: an exactly corrected, bidirectional-context recovery of the profile from summed data. We would then publish the linear version and drop the network. The bidirectional ablation stays meaningful in either case.
Why not just enlarge the operator basis or use calm operators? These methods remove contamination by changing the operator. They do nothing about statistical noise, which is the dominant error for $g_A$ (Nucleon axial coupling from Lattice QCD). Calm operators also leave residual contamination in some local multi-nucleon correlators and have only been shown at $m_\pi\approx 800$ MeV (Calm multi-baryon operators). Our estimator can be applied to whatever operator is chosen. A secondary experiment repeats the study with a two-operator (point and smeared) source to confirm that the gain survives a better operator.
Why not just keep the summed method and buy more statistics? That is baseline one, and it wins if $\rho$ is low. The cost-equivalent comparison decides this. We will not claim improvement unless the equal-compute error is smaller.
Why not just infill the averaged correlator generatively, as speech models infill audio? An infill of the ensemble mean has no exact correction and is equivalent to a fit with an opaque prior. It would bias exactly the quantity the Exactness Mandate protects (Applications of Machine Learning to Lattice Quantum Field Theory). Our per-source design is unbiased by construction, and we check this empirically below.
Proposed Experiments
The first stage is a ground-truth testbed. On one MILC $N_f=2+1+1$ HISQ ensemble at the coarsest spacing (a15) with heavier pion mass, using the CalLat Möbius domain-wall setup, we compute exact profiles on every source. This makes the full-data estimate available as truth. We then simulate the deployed regime by revealing exact profiles on random subsets of 5–25%. The measurements are $\rho$ as a function of $\tau$ and $t$, the cost-equivalent variance-reduction factor, and the pull distribution of $\hat Y$ against the full-data value over many random subset draws. Pulls must be consistent with zero mean and unit width.
A null test trains the predictor on profiles shuffled across configurations. It must give an unbiased estimate with no gain, which confirms that the correction term, not the model, guarantees correctness.
The decisive ablation compares $\rho$ for bidirectional against prefix-only context at each sink time. If the post-$t$ data add nothing, the speech-editing analogy fails and we will say so.
The second stage is deployment on the near-physical (about 130 MeV) pion-mass ensembles at a15 through a09 used by CalLat (Nucleon axial coupling from Lattice QCD). There we report per-ensemble $g_A$ statistical errors against the equal-cost summed-only baseline. We also report the stability of $g_A$ across fit windows, using a BMW-style histogram over fit choices (Decay constants and sigma terms from the lattice.), and the precision of the leading ground-to-excited axial transition element, the quantity the profile adds.
The kill criterion is applied in stage one: if bidirectional $\rho$ stays below about 0.7 at the sink times that matter (a gain under two), we stop before spending physical-point compute. If the method succeeds, the natural extension is two-nucleon correlators, where noise grows fastest with time (Progress in Multibaryon Spectroscopy). That extension is deferred until the single-nucleon case is settled.
References
- A per-cent-level determination of the nucleon axial coupling from quantum chromodynamics
- Nucleon axial coupling from Lattice QCD
- Applications of Machine Learning to Lattice Quantum Field Theory
- VoiceCraft: Zero-Shot Speech Editing and Text-to-Speech in the Wild
- EditSpeech: A Text Based Speech Editing System Using Partial Inference and Bidirectional Fusion
- S-wave pi-pi I=0 and I=2 scattering at physical pion mass
- Studies of I=0 and 2 $\pi\pi$ scattering with physical pion mass
- Calm multi-baryon operators
- Decay constants and sigma terms from the lattice.
- Integrating particle flavor into deep learning models for hadronization
- Fitting a deep generative hadronization model
- Voicebox: Text-Guided Multilingual Universal Speech Generation at Scale
- CampNet: Context-Aware Mask Prediction for End-to-End Text-Based Speech Editing
- Progress in Multibaryon Spectroscopy