Parity-violating electron scattering measures a tiny difference: polarized electrons are fired at a target, their spin direction (helicity) is flipped hundreds of times per second, and the scattering rate for left-handed electrons is compared with the rate for right-handed ones. The difference, the parity-violating asymmetry, is typically between 0.1 and 100 parts per million. It is the only clean handle experiments have on the weak force at these energies, and it is how the neutron skin of lead and the strange-quark content of the proton were measured. The hardest part of the analysis is not the physics. It is subtracting false asymmetries that appear when the beam's position, angle, or energy shifts slightly in step with the helicity flip. Today this subtraction uses a linear regression of detector yield on beam parameters, fitted over whole runs. The regression assumes the detector responds linearly to beam motion, and it says nothing about when its own correction can be trusted.
We propose replacing that run-level linear fit with a nonlinear yield model trained only on helicity-symmetric data, so that it cannot absorb the physics signal. The model is an evidential regressor, meaning a network that outputs in one forward pass both an expected yield and two separate uncertainties: one for irreducible detector noise and one for the model not knowing, for example because the beam has moved into a state it never saw in training. The noise estimate gives per-window statistical weights. The "not knowing" estimate flags windows where the beam correction is extrapolating, such as orbit excursions or recovery after a beam trip. Both are computed only from quantities that do not change sign with helicity, so any cut or weight built from them cannot fake an asymmetry. The central question is whether this beats linear regression, polynomial regression, and deep ensembles on bias and statistical error of the extracted asymmetry. We treat that as open: ensembles are a serious competitor, and the experiments are designed so that a win for them is also a clear answer.
Core Claim
In each helicity pair, the false asymmetry comes from evaluating the detector response at two slightly different beam states. Let $x_R$ and $x_L$ be the beam-parameter vectors (positions, angles, energy) in the right- and left-helicity windows of a pair, and let $f$ be a learned map from beam state to expected detector yield. The corrected asymmetry for that pair is
$$A_{\text{corr}} = A_{\text{raw}} - \frac{f(x_R) - f(x_L)}{f(x_R) + f(x_L)},$$
which reduces to the standard linear correction $\sum_i c_i \Delta x_i$ when $f$ is linear. The safeguard is in how $f$ is trained. It is fitted only to the pair-averaged yield as a function of the pair-averaged beam state. The physics asymmetry flips sign between the two windows, so to first order it cancels in the pair sum, and the network has no signal from which to learn it, even though its capacity is much larger than that of a five-slope linear fit. A flexible network trained on window-level yields could quietly absorb part of the physics signal through small helicity-correlated beam differences. The helicity-symmetric design rules this out by construction.
For $f$ we use Deep Evidential Regression (Deep Evidential Regression). That method places a Normal-Inverse-Gamma prior over the mean and variance of the target and predicts its four parameters $(\gamma, \upsilon, \alpha, \beta)$, where $\gamma$ is the predicted yield and $\upsilon$ acts as a virtual count of how much training data supports that prediction. It also reports a noise variance $\beta/(\alpha-1)$ and a model-ignorance variance $\beta/(\upsilon(\alpha-1))$. We do not use these textbook definitions as given. The Unreasonable Effectiveness of Deep Evidential Regression shows that the parameterization is overdetermined, that its epistemic signal is really an optimizer-speed heuristic, and that under input-dependent noise the original epistemic term mostly reflects data noise. This last failure matters for a beam line whose detector noise changes with beam current. We therefore adopt that paper's corrected quantities: the width of the predicted Student-t distribution measures noise, $1/\sqrt{\upsilon}$ measures ignorance, and the residual loss is normalized by the predicted noise. Because beam parameters are strongly correlated, we also follow Multivariate Deep Evidential Regression in coupling shape parameters to prevent the degenerate optima it identifies. We watch for the evidence-shrinking behavior under long training reported in The Evidence Contraction Issue in Deep Evidential Regression.
Why this gap is real
The motivating evidence comes from the experiments themselves. The electron-deuteron inelastic program (Measurement of parity-violating asymmetry in electron-deuteron inelastic scattering; Measurements of Parity-Violating Asymmetries in Electron-Deuteron Scattering in the Nucleon Resonance Region) corrects beam effects with multivariable linear regression and deliberate beam modulation ("dithering"). Both methods assume purely linear coupling between beam trajectory and yield. That assumption is exactly what can fail during trips and orbit excursions.
The lead measurement (Measurement of the Neutron Radius of Pb208 through Parity Violation in Electron Scattering) sets the constraint. Its asymmetry of 0.656 ppm carries 0.060 ppm statistical error against only 0.014 ppm systematic error, drawn from about $2\times10^7$ helicity quadruplets. Any learned correction must therefore not inflate variance, and it must not discard windows in a way that could correlate with helicity. The proton strange-form-factor measurement (New Precision Limit on the Strange Vector Form Factors of the Proton), with $-23.80 \pm 0.78$ (stat) $\pm 0.36$ (syst) ppm, provides a second, larger-asymmetry regime.
On the machine-learning side, evidential regression has been applied to physical systems such as interatomic potentials (Evidential deep learning for interatomic potentials) and storm extremes (Evidential Deep Learning for Probabilistic Modelling of Extreme Storm Events). Those applications involve static configurations or spatial fields, with no requirement that a learned correction preserve a sign-flipping signal at the part-per-billion level. Single-model uncertainty quantification in neural network potentials does not consistently outperform model ensembles constrains the design: evidential models did not reliably beat ensembles in a physics regression, so an ensemble must be a primary baseline and not an afterthought. The file contains no prior work joining evidential learning with helicity-correlated corrections.
Research Program
The first component is a simulated helicity-window stream, needed because raw window-level data from these experiments is internal to the collaborations. It generates pseudorandom helicity quadruplets with an injected true asymmetry. It includes helicity-correlated beam differences at the scale typical of these experiments, beam jitter, slow drifts, and occasional trips and orbit excursions. Detector noise depends on beam current. The yield response has adjustable nonlinearity: linear, quadratic cross-terms such as position times energy, and a sharp acceptance edge the beam can approach during excursions. Total statistics are matched so that the naive statistical error reproduces the PREX and HAPPEX-III scales. Because the true asymmetry is known, bias can be measured directly.
The second component is the helicity-blind evidential yield model described above. It is trained per run segment on pair-averaged data and applied pair by pair. Its noise estimate supplies inverse-variance weights for each pair. Its ignorance estimate supplies either a soft down-weighting or a threshold cut. Both depend only on pair-averaged inputs, so they are symmetric under helicity reversal. The ignorance term is also aggregated into a proposed "correction-model systematic": the spread of the corrected asymmetry obtained when ignorance-flagged pairs are dropped or kept.
The third component validates the uncertainty machinery before trusting it on beam data. We use the Binary Pulse benchmark from The Unreasonable Effectiveness of Deep Evidential Regression and the variable-density V-shape set from Multivariate Deep Evidential Regression. If noise and ignorance cannot be separated there, they will not be separated on a beam line.
The final stage, contingent on collaboration access, is a blinded reanalysis of archived run segments. There the model is compared against the published linear correction in null tests where helicity labels are randomized.
Baselines
The naive reference is the raw asymmetry with no beam correction. It shows how large the helicity-correlated false asymmetry is in each simulated scenario, and every method must at least remove that.
The standard method, and the one the physics community will judge against, is run-level multivariable linear regression of yield on the five beam parameters, as used in the electron-deuteron analyses (Measurement of parity-violating asymmetry in electron-deuteron inelastic scattering). In simulation, dithering-derived slopes are emulated by fitting on injected modulation periods. For how well this method performs, the file offers only end-to-end totals: PREX reports 13.7 ppb total systematic uncertainty and HAPPEX-III reports 0.36 ppm. For the residual bias or width contributed by the regression step alone, no reported result was found on Lacuna. We will therefore measure it ourselves in the simulator.
The polynomial baseline is linear regression with all quadratic and cross terms, fitted helicity-blind in the same way. It is the cheapest nonlinear competitor.
The strongest uncertainty-aware competitor is a deep ensemble of five copies of the same network with Gaussian-likelihood heads. It gets the identical helicity-symmetric training, and its disagreement serves as the ignorance signal. On depth estimation, Deep Evidential Regression reports a calibration error of 0.0329 for the evidential model against 0.0475 for deep ensembles on NYU Depth v2. The interatomic-potential study shows this ordering does not transfer reliably to physics.
The ablations isolate the proposal's ingredients. The first uses the original evidential uncertainty definitions instead of the corrected ones. The second trains on window-level rather than pair-averaged yields, which tests whether helicity-blindness is actually necessary.
All learned methods get the same inputs, the same run segmentation, the same number of hyperparameter trials, tuning on held-out simulated runs only, and matched training compute. The ensemble gets five times the training compute, which is its natural cost.
Beating a baseline means lower bias in the extracted asymmetry relative to the injected value, together with a statistical error no larger. The pull distribution of bias divided by quoted error must have unit width across 1,000 simulated experiments. For the gating component, beating a baseline also means higher area under the ROC curve at identifying injected excursion windows, without shifting the asymmetry in randomized-helicity null tests.
Why not just…?
Why not just add quadratic terms to the linear regression? With only five beam parameters, a second-order polynomial has about twenty terms and is easy to fit. If the real nonlinearity is smooth and low-order, it will likely match the network on bias. This is a real threat. The simulator settles it by sweeping nonlinearity from quadratic to acceptance-edge behavior. If polynomials tie across the realistic range, the case for a network rests entirely on the ignorance-based gating, which polynomials cannot provide, and we will report it that way.
Why not just use a deep ensemble? The interatomic-potential comparison (Single-model uncertainty quantification in neural network potentials does not consistently outperform model ensembles) says ensembles often win, and the claimed single-pass speed advantage (about 0.85 ms against 3.3 ms in Deep Evidential Regression) matters little for offline analysis. The evidential model's remaining argument is that it separates noise from ignorance within one model, and so could supply both the weights and the gate. If ensemble disagreement flags excursions as well and ensemble noise heads weight as well, we recommend ensembles. The head-to-head is built to decide exactly this.
Why not just keep the existing beam-quality cuts on current and position? Hand-set cuts already remove trips. The open question is whether a learned ignorance score catches subtler out-of-range beam states that pass those cuts. We test this by injecting excursions that stay within standard cut thresholds but cross the nonlinear acceptance edge, then measuring what the gate catches beyond the cuts.
Decisive Checks
Three results would decide the proposal.
First, in simulations with injected cross-term and edge nonlinearities, the helicity-blind evidential correction removes the bias left by linear regression while keeping the statistical error within 2% of the linear method's. The 2% figure matters because PREX-like measurements are limited by statistics.
Second, the window-level-trained ablation shows measurable absorption of the injected asymmetry, while the pair-averaged model shows none within statistical error. This confirms that the helicity-symmetric design is necessary and not merely cautious.
Third, the corrected ignorance score beats both ensemble disagreement and the original evidential definition at flagging within-cuts excursions, with zero asymmetry shift in randomized-helicity null tests.
If the first check fails against polynomials and the third fails against ensembles, the evidential approach offers nothing here, and the note's conclusion becomes a recommendation for polynomial or ensemble corrections with a helicity-symmetric gate.
References
- Deep Evidential Regression
- The Unreasonable Effectiveness of Deep Evidential Regression
- Multivariate Deep Evidential Regression
- The Evidence Contraction Issue in Deep Evidential Regression: Discussion and Solution
- Measurement of parity-violating asymmetry in electron-deuteron inelastic scattering
- Measurements of Parity-Violating Asymmetries in Electron-Deuteron Scattering in the Nucleon Resonance Region
- Measurement of the Neutron Radius of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mmultiscripts><mml:mi>Pb</mml:mi><mml:mprescripts/><mml:none/><mml:mn>208</mml:mn></mml:mmultiscripts></mml:math>through Parity Violation in Electron Scattering
- New Precision Limit on the Strange Vector Form Factors of the Proton
- Evidential deep learning for interatomic potentials
- Evidential Deep Learning for Probabilistic Modelling of Extreme Storm Events
- Single-model uncertainty quantification in neural network potentials does not consistently outperform model ensembles