Experiments searching for neutrinoless double-beta decay, a hypothetical nuclear decay that would show neutrinos are their own antiparticles, look for a handful of events per year among a much larger flood of gamma-ray background. In a gas-filled time projection chamber (a detector that records the 3D ionization trail a charged particle leaves as it slows down), the two populations differ in shape. A signal event is two electrons leaving one point, so its track has a dense "blob" of charge at both ends. A background gamma usually produces a single electron, which leaves one blob and a longer, differently scattered trail. The PandaX-III simulation study reduces this difference to a few scalar numbers and applies them as sequential cuts. Its authors state that combining the correlated quantities is the obvious next step.
We propose a specific way to do that combination. We take the active appearance model from face and lip tracking, a statistical model that learns how an object's outline and its pixel texture vary together, and refit it to particle tracks. Shape becomes the track's 3D centerline. Texture becomes the deposited charge sampled along that centerline by arc length. Each track is then summarized by a few coupled coefficients that a discriminant can use. We are not betting on higher accuracy than a deep network on clean simulation; it may not achieve that. The claim is narrower and testable. A low-dimensional generative description, aligned along the track, should transfer better across readout pitch, diffusion, and gain mismatches than black-box classifiers. It should also need less labeled data, and its reconstruction residual gives an anomaly score for backgrounds the simulation never modeled.
Core Claim
The topological variables in Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment are track energy ratio, blob charge, end-track twist, track length, and diffusion width. Each is a hand-chosen projection of one underlying object: a curve carrying a charge profile. A joint shape-plus-charge model of that object, aligned so that position along the track is comparable across events, should capture the correlations these cuts throw away. The decisive question is whether this gives better background rejection at fixed signal efficiency under detector-model shift than three alternatives: a boosted classifier on the same scalars, or a convolutional network on voxels. If it does not, the method still yields interpretable modes of variation, but the headline claim fails. We will report that outcome as such.
Where each direction enters
Signal-to-background discrimination is the application and supplies all the evaluation. Its benchmark is the REST simulation of the PandaX-III high-pressure xenon chamber, which already compares 2D 3 mm strip readouts with 3D pixel readouts from 1 to 10 mm pitch.
Active appearance models supply the method.
- Joint model. Active Appearance Models introduced the idea of learning shape and texture modes and concatenating them.
- Temporal stacking. Improving visual features for lip-reading concatenated shape and appearance parameters and stacked consecutive frames into hyper-vectors. It then projected them with hierarchical linear discriminant analysis (Hi-LDA), a two-stage projection that maximizes class separation. This outperformed pixel-transform features for speaker-independent lip-reading. We reuse this step, with arc length along the track playing the role of time.
- A constraint on the claim. Analysis of correlation between audio and visual speech features for clean audio feature prediction in noise found that appearance-model and plain 2D-DCT features correlated about equally with acoustic targets. A generic frequency transform may therefore do as well, so a DCT-of-track-image baseline is mandatory rather than optional.
The pure spinor worldline formalism gets the smallest role, and we say plainly what we take from it and what we do not. Those papers compute scattering amplitudes by integrating over one-dimensional particle paths with interaction points ("vertex operators") inserted along them. Two examples are The pure spinor superparticle and 10D super-Yang-Mills amplitudes and Tree-level 11D supergravity amplitudes from the pure spinor worldline. Their supersymmetry and constrained-spinor machinery have no role in detector analysis, and we do not pretend otherwise. We borrow only the organizing picture, which acts as a modeling constraint. An event is one curve, its parametrization is arbitrary and must be divided out, and the physics sits in a small number of marked points on it.
Applied to tracks, this gives two hypotheses. A signal event is one curve with an emission vertex somewhere in its interior and two energy-deposition endpoints. A background event is one curve with a single endpoint blob and a start. In practice the model fits each track in two ways: once with a free interior marked point, and once with that point pinned to an end. The fitted marked-point location and the likelihood ratio between the two fits become features. We know of no prior work tying worldline methods to detector reconstruction; the research file finds no such paper. We claim only the parametrization idea.
Research Program
Each simulated event is first reduced to a single connected track, using the REST clustering that PandaX-III already uses. Its centerline is then fit as a smooth 3D curve.
The centerline is resampled at $K$ equally spaced arc-length points. Shape is aligned to a common frame (translation, rotation, scale by total length) before modeling. The curve's orientation is ambiguous, so each event is stored in both directions and the model is fit to the direction with lower residual. This mirrors how appearance models handle symmetric objects.
The model takes the usual linear form:
$$s = \bar s + P_s b_s, \qquad g = \bar g + P_g b_g.$$
The terms are as follows.
- $s$ is the vector of $K$ aligned centerline points, and $g$ is the vector of charge per unit length sampled at the same $K$ arc-length positions, normalized by total energy.
- $\bar s$ and $\bar g$ are the training means.
- $P_s$ and $P_g$ are principal-component bases.
- $b_s$ and $b_g$ are the per-event coefficients.
Following Lan et al., we weight and concatenate $b_s$ and $b_g$, then apply one more PCA to get coupled coefficients $c$. Windows of local curvature and charge are stacked along arc length into hyper-vectors and projected with Hi-LDA trained on the signal/background labels. The final event score combines three things: the Hi-LDA projection, the two-hypothesis marked-point likelihood ratio, and the model's reconstruction residual.
The model is fit separately for each readout configuration. Transfer is tested by fitting on one configuration and applying to another. The 2D strip readout records two projections, so a third question becomes testable. Can a shape-texture prior learned on 3D pixel simulation help reconstruct coefficients from 2D strip data? The readout trade-off flagged in The MAJORANA Project makes this practically important, since denser readouts add cabling and radioactivity.
Baselines
The naive reference point is the published sequential-cut analysis of Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment. It reports gamma background suppression of "approximately seven orders of magnitude" with signal efficiency "above 20%". Diffusion-width cuts separately reduce surface uranium background "by a factor of 0.15 while retaining 75% of the signal". We will re-run these cuts ourselves on the identical simulated sample so that the reference is matched in statistics and energy window. The bar is a statistically significant improvement in background rejection at the same 20% efficiency, or equal rejection at clearly higher efficiency.
The strongest feature-level competitor is a gradient-boosted tree classifier on exactly the five PandaX-III scalars. This is the multivariate analysis the paper itself anticipates, and it captures their correlations without any new representation. No reported result found on Lacuna for a multivariate analysis on PandaX-III topology. It gets the same training events and the same hyperparameter search budget as our discriminant, with tuning on a held-out validation split.
The strongest representation-level competitor is a 3D sparse convolutional network on voxelized charge. This is the image-based route that MicroBooNE's segmentation network exemplifies in Vertex-finding and reconstruction of contained two-track neutrino events in the MicroBooNE detector, which reports about 69% two-track reconstruction efficiency in a different detector and task. No reported result found on Lacuna for a convolutional classifier on PandaX-III tracks. We give it standard augmentation (rotations, diffusion jitter, gain jitter) and a matched GPU-hour budget. We expect it to win on in-distribution simulation. Our claim is beaten if it also matches us under shift.
Ablations isolate each ingredient of our model:
- shape-only coefficients;
- charge-texture-only coefficients;
- coupled coefficients without arc-length alignment, using raw voxel PCA;
- the coupled model without Hi-LDA stacking;
- the full model without the marked-point likelihood ratio;
- a 2D-DCT of projected track images, which tests the Almajai et al. result that frequency transforms match appearance models.
Why not just…?
Why not just feed the five PandaX-III scalars to a boosted classifier? This is the most likely way the proposal loses on raw performance. Those scalars were designed by physicists to carry most of the discriminating information. The case for our model rests on information they compress away, such as the charge profile between the endpoints and the scattering pattern along the whole track rather than only near the ends. If the boosted classifier matches the full model within statistical error at 20% efficiency, in-distribution and under shift, then the scalars are sufficient. We would publish that as a negative result.
Why not just train a 3D convolutional network on voxels? On clean simulation it may well beat us, and we do not claim otherwise. Our expected advantage is under detector-model mismatch and with few labels. Networks trained on simulation can latch onto pitch- or diffusion-specific texture, while an aligned 20–40 coefficient description has little room to do so. The cross-pitch, cross-diffusion and label-scarcity experiments below settle this directly. If an augmented network matches our transfer, our remaining contribution is interpretability and the residual-based anomaly score, and we would say so.
Why not just use a DCT or other fixed transform of the track image? Almajai et al. found such transforms as informative as appearance models for broad acoustic correlates, so this is a real threat. The DCT ablation runs at matched dimensionality. The appearance model must beat it under transfer to justify its fitting cost.
Proposed Experiments
All experiments use the PandaX-III REST simulation:
- signal: double-beta decay of xenon-136;
- background: bismuth-214 and thallium-208 gammas, plus uranium-238 surface events;
- energy window: the region of interest;
- readout: 2D 3 mm strips and 3D pixels at 1, 3 and 10 mm.
The primary metric is background acceptance at fixed signal efficiencies of 20% and 35%. We also report the sensitivity figure of merit, signal efficiency divided by the square root of background acceptance. Confidence intervals come from bootstrap resampling, since accepted-background counts are small.
The decisive check is transfer. Every method is trained at one pitch and diffusion setting and evaluated at all others, with diffusion and gain perturbed by plausible calibration errors. The claim holds only if the appearance model degrades measurably less than both the boosted classifier and the convolutional network.
The second check is label efficiency: performance curves as the training set shrinks by factors of ten.
The third check targets backgrounds absent from training. We withhold the uranium-238 surface class during training and test whether the reconstruction residual flags it. The comparison is the diffusion-width cut, with its published 0.15 reduction at 75% signal retention as the reference.
A final check asks whether a 3D-trained prior lets the 2D strip readout approach 3D-pixel rejection. A positive answer would bear directly on the readout-cost question.
References
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment
- Active Appearance Models
- Improving visual features for lip-reading
- Analysis of correlation between audio and visual speech features for clean audio feature prediction in noise
- The pure spinor superparticle and 10D super-Yang-Mills amplitudes
- Tree-level 11D supergravity amplitudes from the pure spinor worldline
- The MAJORANA Project
- Vertex-finding and reconstruction of contained two-track neutrino events in the MicroBooNE detector